ggml.c 244 KB
Newer Older
1
/**
2
 * llama.cpp - commit 40c6d79fb52f995f47507fedfeaae2ac05d9b35c - do not edit this file
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
 *
 * MIT License
 *
 * Copyright (c) 2023-2024 The ggml authors
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

27
#define _CRT_SECURE_NO_DEPRECATE // Disables "unsafe" warnings on Windows
28
29
#define _USE_MATH_DEFINES // For M_PI on MSVC

30
#include "ggml-backend.h"
31
#include "ggml-impl.h"
32
#include "ggml-threading.h"
33
#include "ggml.h"
34
35
36

// FIXME: required here for quantization functions
#include "ggml-quants.h"
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
#include "ggml-aarch64.h"

#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
#include <alloca.h>
#endif

#include <assert.h>
#include <errno.h>
#include <time.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <inttypes.h>
#include <stdio.h>
#include <float.h>
#include <limits.h>
#include <stdarg.h>
#include <signal.h>
#if defined(__gnu_linux__)
#include <syscall.h>
#endif

62
#if defined(__APPLE__)
63
#include <unistd.h>
64
65
#include <mach/mach.h>
#include <TargetConditionals.h>
66
67
68
69
70
71
72
73
74
75
#endif

#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
    #define NOMINMAX
#endif
#include <windows.h>
#endif

76
#define UNUSED GGML_UNUSED
77

78
79
80
#if defined(_MSC_VER)
#define m512bh(p) p
#define m512i(p) p
81
#else
82
83
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
84
85
#endif

86
87
// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
float ggml_table_f32_f16[1 << 16];
88
89
90

#if (defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)) && \
    (!defined(TARGET_OS_TV) && !defined(TARGET_OS_WATCH))
91
92
93
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
#include <sys/wait.h>

#if defined(__ANDROID__)
#include <unwind.h>
#include <dlfcn.h>
#include <stdio.h>

struct backtrace_state {
    void ** current;
    void ** end;
};

static _Unwind_Reason_Code unwind_callback(struct _Unwind_Context* context, void* arg) {
    struct backtrace_state * state = (struct backtrace_state *)arg;
    uintptr_t pc = _Unwind_GetIP(context);
    if (pc) {
        if (state->current == state->end) {
            return _URC_END_OF_STACK;
        } else {
            *state->current++ = (void*)pc;
        }
    }
    return _URC_NO_REASON;
}

static void ggml_print_backtrace_symbols(void) {
    const int max = 100;
    void* buffer[max];

    struct backtrace_state state = {buffer, buffer + max};
    _Unwind_Backtrace(unwind_callback, &state);

    int count = state.current - buffer;

    for (int idx = 0; idx < count; ++idx) {
        const void * addr = buffer[idx];
        const char * symbol = "";

        Dl_info info;
        if (dladdr(addr, &info) && info.dli_sname) {
            symbol = info.dli_sname;
        }

        fprintf(stderr, "%d: %p %s\n", idx, addr, symbol);
    }
}
#elif defined(__linux__) && defined(__GLIBC__)
#include <execinfo.h>
static void ggml_print_backtrace_symbols(void) {
    void * trace[100];
    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
    backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
}
#else
static void ggml_print_backtrace_symbols(void) {
    // platform not supported
}
#endif

static void ggml_print_backtrace(void) {
    char attach[32];
    snprintf(attach, sizeof(attach), "attach %d", getpid());
    int pid = fork();
    if (pid == 0) {
        // try gdb
        execlp("gdb", "gdb", "--batch",
            "-ex", "set style enabled on",
            "-ex", attach,
            "-ex", "bt -frame-info source-and-location",
            "-ex", "detach",
            "-ex", "quit",
            (char *) NULL);
        // try lldb
        execlp("lldb", "lldb", "--batch",
            "-o", "bt",
            "-o", "quit",
            "-p", attach,
            (char *) NULL);
        exit(EXIT_FAILURE);
    } else {
        int wstatus;
        waitpid(pid, &wstatus, 0);
        if (WIFEXITED(wstatus)) {
            if (WEXITSTATUS(wstatus) == EXIT_FAILURE) {
                // gdb failed, fallback to backtrace_symbols
                ggml_print_backtrace_symbols();
            }
        }
    }
}
#else
static void ggml_print_backtrace(void) {
    // platform not supported
}
#endif

void ggml_abort(const char * file, int line, const char * fmt, ...) {
    fflush(stdout);

    fprintf(stderr, "%s:%d: ", file, line);

    va_list args;
    va_start(args, fmt);
    vfprintf(stderr, fmt, args);
    va_end(args);

    fprintf(stderr, "\n");

    ggml_print_backtrace();
    abort();
}

//
// logging
//

210
211
212
213
214
struct ggml_logger_state {
    ggml_log_callback log_callback;
    void * log_callback_user_data;
};
static struct ggml_logger_state g_logger_state = {ggml_log_callback_default, NULL};
215

216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
static void ggml_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
    if (format == NULL) {
        return;
    }
    va_list args_copy;
    va_copy(args_copy, args);
    char buffer[128];
    int len = vsnprintf(buffer, 128, format, args);
    if (len < 128) {
        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
    } else {
        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
        vsnprintf(buffer2, len + 1, format, args_copy);
        buffer2[len] = 0;
        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
        free(buffer2);
    }
    va_end(args_copy);
}
235

236
237
238
239
240
241
void ggml_log_internal(enum ggml_log_level level, const char * format, ...) {
    va_list args;
    va_start(args, format);
    ggml_log_internal_v(level, format, args);
    va_end(args);
}
242

243
244
245
246
247
248
void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
    (void) level;
    (void) user_data;
    fputs(text, stderr);
    fflush(stderr);
}
249
250
251
252
253
254
255
256
257
258
259

//
// end of logging block
//

#ifdef GGML_USE_ACCELERATE
// uncomment to use vDSP for soft max computation
// note: not sure if it is actually faster
//#define GGML_SOFT_MAX_ACCELERATE
#endif

260
261
262
263

void * ggml_aligned_malloc(size_t size) {
    const int alignment = 64;

264
#if defined(_MSC_VER) || defined(__MINGW32__)
265
    return _aligned_malloc(size, alignment);
266
267
#else
    if (size == 0) {
268
        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
269
270
271
        return NULL;
    }
    void * aligned_memory = NULL;
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
  #ifdef GGML_USE_CPU_HBM
    int result = hbw_posix_memalign(&aligned_memory, alignment, size);
  #elif TARGET_OS_OSX
    GGML_UNUSED(alignment);
    kern_return_t alloc_status = vm_allocate((vm_map_t) mach_task_self(), (vm_address_t *) &aligned_memory, size, VM_FLAGS_ANYWHERE);
    int result = EFAULT;
    switch (alloc_status) {
        case KERN_SUCCESS:
            result = 0;
            break;
        case KERN_INVALID_ADDRESS:
            result = EINVAL;
            break;
        case KERN_NO_SPACE:
            result = ENOMEM;
            break;
        default:
            result = EFAULT;
            break;
    }
  #else
    int result = posix_memalign(&aligned_memory, alignment, size);
  #endif
295
296
297
298
299
300
301
302
303
304
305
    if (result != 0) {
        // Handle allocation failure
        const char *error_desc = "unknown allocation error";
        switch (result) {
            case EINVAL:
                error_desc = "invalid alignment value";
                break;
            case ENOMEM:
                error_desc = "insufficient memory";
                break;
        }
306
        GGML_LOG_ERROR("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0));
307
308
309
        return NULL;
    }
    return aligned_memory;
310
#endif
311
}
312
313
314
315
316
317
318
319
320
321
322
323
324

void ggml_aligned_free(void * ptr, size_t size) {
    GGML_UNUSED(size);
#if defined(_MSC_VER) || defined(__MINGW32__)
    _aligned_free(ptr);
#elif GGML_USE_CPU_HBM
    if (ptr != NULL) {
        hbw_free(ptr);
    }
#elif TARGET_OS_OSX
    if (ptr != NULL) {
        vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ptr, size);
    }
325
#else
326
    free(ptr);
327
#endif
328
329
}

330
331
332

inline static void * ggml_malloc(size_t size) {
    if (size == 0) {
333
        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_malloc!\n");
334
335
336
337
        return NULL;
    }
    void * result = malloc(size);
    if (result == NULL) {
338
        GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
339
340
341
342
343
344
345
346
        GGML_ABORT("fatal error");
    }
    return result;
}

// calloc
inline static void * ggml_calloc(size_t num, size_t size) {
    if (num == 0 || size == 0) {
347
        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_calloc!\n");
348
349
350
351
        return NULL;
    }
    void * result = calloc(num, size);
    if (result == NULL) {
352
        GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
353
354
355
356
357
358
359
360
361
362
        GGML_ABORT("fatal error");
    }
    return result;
}

#define GGML_MALLOC(size)      ggml_malloc(size)
#define GGML_CALLOC(num, size) ggml_calloc(num, size)

#define GGML_FREE(ptr) free(ptr)

363
const char * ggml_status_to_string(enum ggml_status status) {
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
    switch (status) {
        case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
        case GGML_STATUS_FAILED:       return "GGML status: error (operation failed)";
        case GGML_STATUS_SUCCESS:      return "GGML status: success";
        case GGML_STATUS_ABORTED:      return "GGML status: warning (operation aborted)";
    }

    return "GGML status: unknown";
}

float ggml_fp16_to_fp32(ggml_fp16_t x) {
#define ggml_fp16_to_fp32 do_not_use__ggml_fp16_to_fp32__in_ggml
    return GGML_FP16_TO_FP32(x);
}

ggml_fp16_t ggml_fp32_to_fp16(float x) {
#define ggml_fp32_to_fp16 do_not_use__ggml_fp32_to_fp16__in_ggml
    return GGML_FP32_TO_FP16(x);
}

float ggml_bf16_to_fp32(ggml_bf16_t x) {
#define ggml_bf16_to_fp32 do_not_use__ggml_bf16_to_fp32__in_ggml
    return GGML_BF16_TO_FP32(x);  // it just left shifts
}

ggml_bf16_t ggml_fp32_to_bf16(float x) {
#define ggml_fp32_to_bf16 do_not_use__ggml_fp32_to_bf16__in_ggml
    return GGML_FP32_TO_BF16(x);
}

void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
    for (int64_t i = 0; i < n; i++) {
        y[i] = GGML_FP16_TO_FP32(x[i]);
    }
}

400
401
// FIXME: these functions must detect the instruction set at runtime, since they are part of the core ggml library
//        currently, the ggml_cpu_has_* functions are entirely compile-time
402
403
404
void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
    int64_t i = 0;
#if defined(__F16C__)
405
406
407
408
409
410
411
412
413
414
415
416
    //if (ggml_cpu_has_f16c()) {
        for (; i + 7 < n; i += 8) {
            __m256 x_vec = _mm256_loadu_ps(x + i);
            __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
            _mm_storeu_si128((__m128i *)(y + i), y_vec);
        }
        for(; i + 3 < n; i += 4) {
            __m128 x_vec = _mm_loadu_ps(x + i);
            __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
            _mm_storel_epi64((__m128i *)(y + i), y_vec);
        }
    //}
417
418
419
420
421
422
423
424
425
#endif
    for (; i < n; i++) {
        y[i] = GGML_FP32_TO_FP16(x[i]);
    }
}

void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
    int64_t i = 0;
#if defined(__AVX512F__)
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
    //if (ggml_cpu_has_avx512()) {
        for (; i + 16 <= n; i += 16) {
            _mm512_storeu_ps(y + i,
                            _mm512_castsi512_ps(
                                _mm512_slli_epi32(
                                    _mm512_cvtepu16_epi32(
                                        _mm256_loadu_si256(
                                            (const __m256i *)(x + i))),
                                    16)));
        }
    //}
#endif
#if defined(__AVX2__)
    //if (ggml_cpu_has_avx2()) {
        for (; i + 8 <= n; i += 8) {
            _mm256_storeu_ps(y + i,
                            _mm256_castsi256_ps(
                                _mm256_slli_epi32(
                                    _mm256_cvtepu16_epi32(
                                        _mm_loadu_si128(
                                            (const __m128i *)(x + i))),
                                    16)));
        }
    //}
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
#endif
    for (; i < n; i++) {
        y[i] = GGML_BF16_TO_FP32(x[i]);
    }
}

void ggml_fp32_to_bf16_row_ref(const float * x, ggml_bf16_t * y, int64_t n) {
    for (int i = 0; i < n; i++) {
        y[i] = ggml_compute_fp32_to_bf16(x[i]);
    }
}

void ggml_fp32_to_bf16_row(const float * x, ggml_bf16_t * y, int64_t n) {
  int i = 0;
#if defined(__AVX512BF16__)
  // subnormals are flushed to zero on this platform
  for (; i + 32 <= n; i += 32) {
        _mm512_storeu_si512(
            (__m512i *)(y + i),
            m512i(_mm512_cvtne2ps_pbh(_mm512_loadu_ps(x + i + 16),
                                _mm512_loadu_ps(x + i))));
  }
#endif
    for (; i < n; i++) {
        y[i] = GGML_FP32_TO_BF16(x[i]);
    }
}

bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b) {
    return memcmp(guid_a, guid_b, sizeof(ggml_guid)) == 0;
}

//
// timing
//

#if defined(_MSC_VER) || defined(__MINGW32__)
static int64_t timer_freq, timer_start;
void ggml_time_init(void) {
    LARGE_INTEGER t;
    QueryPerformanceFrequency(&t);
    timer_freq = t.QuadPart;

    // The multiplication by 1000 or 1000000 below can cause an overflow if timer_freq
    // and the uptime is high enough.
    // We subtract the program start time to reduce the likelihood of that happening.
    QueryPerformanceCounter(&t);
    timer_start = t.QuadPart;
}
int64_t ggml_time_ms(void) {
    LARGE_INTEGER t;
    QueryPerformanceCounter(&t);
    return ((t.QuadPart-timer_start) * 1000) / timer_freq;
}
int64_t ggml_time_us(void) {
    LARGE_INTEGER t;
    QueryPerformanceCounter(&t);
    return ((t.QuadPart-timer_start) * 1000000) / timer_freq;
}
#else
void ggml_time_init(void) {}
int64_t ggml_time_ms(void) {
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC, &ts);
    return (int64_t)ts.tv_sec*1000 + (int64_t)ts.tv_nsec/1000000;
}

int64_t ggml_time_us(void) {
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC, &ts);
    return (int64_t)ts.tv_sec*1000000 + (int64_t)ts.tv_nsec/1000;
}
#endif

int64_t ggml_cycles(void) {
    return clock();
}

int64_t ggml_cycles_per_ms(void) {
    return CLOCKS_PER_SEC/1000;
}

//
// cross-platform UTF-8 file paths
//

#ifdef _WIN32
static wchar_t * ggml_mbstowcs(const char * mbs) {
    int wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, NULL, 0);
    if (!wlen) {
        errno = EINVAL;
        return NULL;
    }

    wchar_t * wbuf = GGML_MALLOC(wlen * sizeof(wchar_t));
    wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, wbuf, wlen);
    if (!wlen) {
        GGML_FREE(wbuf);
        errno = EINVAL;
        return NULL;
    }

    return wbuf;
}
#endif

FILE * ggml_fopen(const char * fname, const char * mode) {
#ifdef _WIN32
    FILE * file = NULL;

    // convert fname (UTF-8)
    wchar_t * wfname = ggml_mbstowcs(fname);
    if (wfname) {
        // convert mode (ANSI)
        wchar_t * wmode = GGML_MALLOC((strlen(mode) + 1) * sizeof(wchar_t));
        wchar_t * wmode_p = wmode;
        do {
            *wmode_p++ = (wchar_t)*mode;
        } while (*mode++);

        // open file
        file = _wfopen(wfname, wmode);

        GGML_FREE(wfname);
        GGML_FREE(wmode);
    }

    return file;
#else
    return fopen(fname, mode);
#endif

582
}
583
584
585
586
static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);

587
static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
    [GGML_TYPE_I8] = {
        .type_name                = "i8",
        .blck_size                = 1,
        .type_size                = sizeof(int8_t),
        .is_quantized             = false,
    },
    [GGML_TYPE_I16] = {
        .type_name                = "i16",
        .blck_size                = 1,
        .type_size                = sizeof(int16_t),
        .is_quantized             = false,
    },
    [GGML_TYPE_I32] = {
        .type_name                = "i32",
        .blck_size                = 1,
        .type_size                = sizeof(int32_t),
        .is_quantized             = false,
    },
    [GGML_TYPE_I64] = {
        .type_name                = "i64",
        .blck_size                = 1,
        .type_size                = sizeof(int64_t),
        .is_quantized             = false,
    },
    [GGML_TYPE_F64] = {
        .type_name                = "f64",
        .blck_size                = 1,
        .type_size                = sizeof(double),
        .is_quantized             = false,
    },
    [GGML_TYPE_F32] = {
        .type_name                = "f32",
        .blck_size                = 1,
        .type_size                = sizeof(float),
        .is_quantized             = false,
    },
    [GGML_TYPE_F16] = {
        .type_name                = "f16",
        .blck_size                = 1,
        .type_size                = sizeof(ggml_fp16_t),
        .is_quantized             = false,
        .to_float                 = (ggml_to_float_t) ggml_fp16_to_fp32_row,
        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_fp16_row,
    },
    [GGML_TYPE_Q4_0] = {
        .type_name                = "q4_0",
        .blck_size                = QK4_0,
        .type_size                = sizeof(block_q4_0),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q4_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_0_ref,
    },
    [GGML_TYPE_Q4_1] = {
        .type_name                = "q4_1",
        .blck_size                = QK4_1,
        .type_size                = sizeof(block_q4_1),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q4_1,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_1_ref,
    },
    [4] = { // GGML_TYPE_Q4_2
        .type_name                = "DEPRECATED",
        .blck_size                = 0,
        .type_size                = 0,
        .is_quantized             = false,
    },
    [5] = { // GGML_TYPE_Q4_3
        .type_name                = "DEPRECATED",
        .blck_size                = 0,
        .type_size                = 0,
        .is_quantized             = false,
    },
    [GGML_TYPE_Q5_0] = {
        .type_name                = "q5_0",
        .blck_size                = QK5_0,
        .type_size                = sizeof(block_q5_0),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q5_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_0_ref,
    },
    [GGML_TYPE_Q5_1] = {
        .type_name                = "q5_1",
        .blck_size                = QK5_1,
        .type_size                = sizeof(block_q5_1),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q5_1,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_1_ref,
    },
    [GGML_TYPE_Q8_0] = {
        .type_name                = "q8_0",
        .blck_size                = QK8_0,
        .type_size                = sizeof(block_q8_0),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q8_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q8_0_ref,
    },
    [GGML_TYPE_Q8_1] = {
        .type_name                = "q8_1",
        .blck_size                = QK8_1,
        .type_size                = sizeof(block_q8_1),
        .is_quantized             = true,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q8_1_ref,
    },
    [GGML_TYPE_Q2_K] = {
        .type_name                = "q2_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q2_K),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q2_K,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q2_K_ref,
    },
    [GGML_TYPE_Q3_K] = {
        .type_name                = "q3_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q3_K),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q3_K,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q3_K_ref,
    },
    [GGML_TYPE_Q4_K] = {
        .type_name                = "q4_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q4_K),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q4_K,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_K_ref,
    },
    [GGML_TYPE_Q5_K] = {
        .type_name                = "q5_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q5_K),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q5_K,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_K_ref,
    },
    [GGML_TYPE_Q6_K] = {
        .type_name                = "q6_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q6_K),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_q6_K,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q6_K_ref,
    },
    [GGML_TYPE_IQ2_XXS] = {
        .type_name                = "iq2_xxs",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq2_xxs),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_IQ2_XS] = {
        .type_name                = "iq2_xs",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq2_xs),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_IQ3_XXS] = {
        .type_name                = "iq3_xxs",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq3_xxs),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_xxs,
        .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_xxs_ref,
    },
    [GGML_TYPE_IQ3_S] = {
        .type_name                = "iq3_s",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq3_s),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_s,
        .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_s_ref,
    },
    [GGML_TYPE_IQ2_S] = {
        .type_name                = "iq2_s",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq2_s),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_s,
        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_s_ref,
    },
    [GGML_TYPE_IQ1_S] = {
        .type_name                = "iq1_s",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq1_s),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_s,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_IQ1_M] = {
        .type_name                = "iq1_m",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq1_m),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_m,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_IQ4_NL] = {
        .type_name                = "iq4_nl",
        .blck_size                = QK4_NL,
        .type_size                = sizeof(block_iq4_nl),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_nl,
        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_nl_ref,
    },
    [GGML_TYPE_IQ4_XS] = {
        .type_name                = "iq4_xs",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_iq4_xs),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_xs,
        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_xs_ref,
    },
    [GGML_TYPE_Q8_K] = {
        .type_name                = "q8_K",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_q8_K),
        .is_quantized             = true,
    },
    [GGML_TYPE_BF16] = {
        .type_name                = "bf16",
        .blck_size                = 1,
        .type_size                = sizeof(ggml_bf16_t),
        .is_quantized             = false,
        .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
    },
    [GGML_TYPE_Q4_0_4_4] = {
        .type_name                = "q4_0_4x4",
        .blck_size                = QK4_0,
        .blck_size_interleave     = 4,
        .type_size                = sizeof(block_q4_0),
        .is_quantized             = true,
        .to_float                 = NULL,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_Q4_0_4_8] = {
        .type_name                = "q4_0_4x8",
        .blck_size                = QK4_0,
        .blck_size_interleave     = 8,
        .type_size                = sizeof(block_q4_0),
        .is_quantized             = true,
        .to_float                 = NULL,
        .from_float_ref           = NULL,
    },
    [GGML_TYPE_Q4_0_8_8] = {
        .type_name                = "q4_0_8x8",
        .blck_size                = QK4_0,
        .blck_size_interleave     = 8,
        .type_size                = sizeof(block_q4_0),
        .is_quantized             = true,
        .to_float                 = NULL,
        .from_float_ref           = NULL,
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
    },
    [GGML_TYPE_TQ1_0] = {
        .type_name                = "tq1_0",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_tq1_0),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_tq1_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_tq1_0_ref,
    },
    [GGML_TYPE_TQ2_0] = {
        .type_name                = "tq2_0",
        .blck_size                = QK_K,
        .type_size                = sizeof(block_tq2_0),
        .is_quantized             = true,
        .to_float                 = (ggml_to_float_t) dequantize_row_tq2_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_tq2_0_ref,
859
860
861
862
863
864
865
866
867
    },
    [GGML_TYPE_IQ4_NL_4_4] = {
        .type_name                = "iq4_nl_4x4",
        .blck_size                = QK4_NL,
        .blck_size_interleave     = 4,
        .type_size                = sizeof(block_iq4_nl),
        .is_quantized             = true,
        .to_float                 = NULL,
        .from_float_ref           = NULL,
868
    },
869
870
};

871
const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
872
    GGML_ASSERT(type < GGML_TYPE_COUNT);
873
    return &type_traits[type];
874
875
}

876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
//
// ggml object
//

struct ggml_object {
    size_t offs;
    size_t size;

    struct ggml_object * next;

    enum ggml_object_type type;

    char padding[4];
};

static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);

893
894
895
896
897
898
//
// ggml context
//

struct ggml_context {
    size_t mem_size;
899
    void * mem_buffer;
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
    bool   mem_buffer_owned;
    bool   no_alloc;

    int    n_objects;

    struct ggml_object * objects_begin;
    struct ggml_object * objects_end;
};

struct ggml_context_container {
    bool used;

    struct ggml_context context;
};

//
916
// data types
917
918
//

919
920
static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
    "NONE",
921

922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
    "DUP",
    "ADD",
    "ADD1",
    "ACC",
    "SUB",
    "MUL",
    "DIV",
    "SQR",
    "SQRT",
    "LOG",
    "SIN",
    "COS",
    "SUM",
    "SUM_ROWS",
    "MEAN",
    "ARGMAX",
    "COUNT_EQUAL",
    "REPEAT",
    "REPEAT_BACK",
    "CONCAT",
    "SILU_BACK",
    "NORM",
    "RMS_NORM",
    "RMS_NORM_BACK",
    "GROUP_NORM",
947

948
949
950
    "MUL_MAT",
    "MUL_MAT_ID",
    "OUT_PROD",
951

952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
    "SCALE",
    "SET",
    "CPY",
    "CONT",
    "RESHAPE",
    "VIEW",
    "PERMUTE",
    "TRANSPOSE",
    "GET_ROWS",
    "GET_ROWS_BACK",
    "DIAG",
    "DIAG_MASK_INF",
    "DIAG_MASK_ZERO",
    "SOFT_MAX",
    "SOFT_MAX_BACK",
    "ROPE",
    "ROPE_BACK",
    "CLAMP",
    "CONV_TRANSPOSE_1D",
    "IM2COL",
    "IM2COL_BACK",
    "CONV_TRANSPOSE_2D",
    "POOL_1D",
    "POOL_2D",
    "POOL_2D_BACK",
    "UPSCALE",
    "PAD",
    "UNPAD",
    "ARANGE",
    "TIMESTEP_EMBEDDING",
    "ARGSORT",
    "LEAKY_RELU",
984

985
986
987
988
989
990
991
992
993
    "FLASH_ATTN_EXT",
    "FLASH_ATTN_BACK",
    "SSM_CONV",
    "SSM_SCAN",
    "WIN_PART",
    "WIN_UNPART",
    "GET_REL_POS",
    "ADD_REL_POS",
    "RWKV_WKV6",
994

995
    "UNARY",
996

997
998
    "MAP_UNARY",
    "MAP_BINARY",
999

1000
1001
1002
    "MAP_CUSTOM1_F32",
    "MAP_CUSTOM2_F32",
    "MAP_CUSTOM3_F32",
1003

1004
1005
1006
    "MAP_CUSTOM1",
    "MAP_CUSTOM2",
    "MAP_CUSTOM3",
1007

1008
1009
1010
1011
    "CROSS_ENTROPY_LOSS",
    "CROSS_ENTROPY_LOSS_BACK",
    "OPT_STEP_ADAMW",
};
1012

1013
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
1014

1015
1016
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
    "none",
1017

1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
    "x",
    "x+y",
    "x+y",
    "view(x,nb,offset)+=y->x",
    "x-y",
    "x*y",
    "x/y",
    "x^2",
    "√x",
    "log(x)",
    "sin(x)",
    "cos(x)",
    "Σx",
    "Σx_k",
    "Σx/n",
    "argmax(x)",
    "count_equal(x)",
    "repeat(x)",
    "repeat_back(x)",
    "concat(x, y)",
    "silu_back(x)",
    "norm(x)",
    "rms_norm(x)",
    "rms_norm_back(x)",
    "group_norm(x)",
1043

1044
1045
1046
    "X*Y",
    "X[i]*Y",
    "X*Y",
1047

1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
    "x*v",
    "y-\\>view(x)",
    "x-\\>y",
    "cont(x)",
    "reshape(x)",
    "view(x)",
    "permute(x)",
    "transpose(x)",
    "get_rows(x)",
    "get_rows_back(x)",
    "diag(x)",
    "diag_mask_inf(x)",
    "diag_mask_zero(x)",
    "soft_max(x)",
    "soft_max_back(x)",
    "rope(x)",
    "rope_back(x)",
    "clamp(x)",
    "conv_transpose_1d(x)",
    "im2col(x)",
    "im2col_back(x)",
    "conv_transpose_2d(x)",
    "pool_1d(x)",
    "pool_2d(x)",
    "pool_2d_back(x)",
    "upscale(x)",
    "pad(x)",
    "unpad(x)",
    "arange(start, stop, step)",
    "timestep_embedding(timesteps, dim, max_period)",
    "argsort(x)",
    "leaky_relu(x)",
1080

1081
1082
1083
1084
1085
1086
1087
1088
1089
    "flash_attn_ext(x)",
    "flash_attn_back(x)",
    "ssm_conv(x)",
    "ssm_scan(x)",
    "win_part(x)",
    "win_unpart(x)",
    "get_rel_pos(x)",
    "add_rel_pos(x)",
    "rwkv_wkv6(k, v, r, tf, td, s)",
1090

1091
    "unary(x)",
1092

1093
1094
    "f(x)",
    "f(x,y)",
1095

1096
1097
1098
    "custom_f32(x)",
    "custom_f32(x,y)",
    "custom_f32(x,y,z)",
1099

1100
1101
1102
    "custom(x)",
    "custom(x,y)",
    "custom(x,y,z)",
1103

1104
1105
1106
1107
    "cross_entropy_loss(x,y)",
    "cross_entropy_loss_back(x,y)",
    "adamw(x)",
};
1108

1109
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
1110

1111
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
1112
1113


1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
    "ABS",
    "SGN",
    "NEG",
    "STEP",
    "TANH",
    "ELU",
    "RELU",
    "SIGMOID",
    "GELU",
    "GELU_QUICK",
    "SILU",
    "HARDSWISH",
    "HARDSIGMOID",
    "EXP",
};
1130

1131
static_assert(GGML_UNARY_OP_COUNT == 14, "GGML_UNARY_OP_COUNT != 14");
1132
1133


1134
1135
static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
1136
1137


1138
////////////////////////////////////////////////////////////////////////////////
1139

1140
1141
1142
void ggml_print_object(const struct ggml_object * obj) {
    GGML_LOG_INFO(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n",
            obj->type, obj->offs, obj->size, (const void *) obj->next);
1143
1144
}

1145
1146
void ggml_print_objects(const struct ggml_context * ctx) {
    struct ggml_object * obj = ctx->objects_begin;
1147

1148
    GGML_LOG_INFO("%s: objects in context %p:\n", __func__, (const void *) ctx);
1149

1150
1151
1152
    while (obj != NULL) {
        ggml_print_object(obj);
        obj = obj->next;
1153
1154
    }

1155
1156
    GGML_LOG_INFO("%s: --- end ---\n", __func__);
}
1157

1158
1159
int64_t ggml_nelements(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1160

1161
1162
    return tensor->ne[0]*tensor->ne[1]*tensor->ne[2]*tensor->ne[3];
}
1163

1164
1165
int64_t ggml_nrows(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1166

1167
1168
    return tensor->ne[1]*tensor->ne[2]*tensor->ne[3];
}
1169

1170
1171
1172
1173
1174
1175
1176
size_t ggml_nbytes(const struct ggml_tensor * tensor) {
    size_t nbytes;
    const size_t blck_size = ggml_blck_size(tensor->type);
    if (blck_size == 1) {
        nbytes = ggml_type_size(tensor->type);
        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
1177
1178
        }
    }
1179
1180
1181
1182
    else {
        nbytes = tensor->ne[0]*tensor->nb[0]/blck_size;
        for (int i = 1; i < GGML_MAX_DIMS; ++i) {
            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
1183
1184
1185
        }
    }

1186
    return nbytes;
1187
1188
}

1189
1190
1191
size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
    return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
}
1192

1193
1194
1195
int64_t ggml_blck_size(enum ggml_type type) {
    return type_traits[type].blck_size;
}
1196

1197
1198
1199
size_t ggml_type_size(enum ggml_type type) {
    return type_traits[type].type_size;
}
1200

1201
1202
1203
1204
size_t ggml_row_size(enum ggml_type type, int64_t ne) {
    assert(ne % ggml_blck_size(type) == 0);
    return ggml_type_size(type)*ne/ggml_blck_size(type);
}
1205

1206
1207
1208
double ggml_type_sizef(enum ggml_type type) {
    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
}
1209

1210
1211
const char * ggml_type_name(enum ggml_type type) {
    return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
1212
1213
}

1214
1215
1216
bool ggml_is_quantized(enum ggml_type type) {
    return type_traits[type].is_quantized;
}
1217

1218
1219
1220
const char * ggml_op_name(enum ggml_op op) {
    return GGML_OP_NAME[op];
}
1221

1222
1223
1224
const char * ggml_op_symbol(enum ggml_op op) {
    return GGML_OP_SYMBOL[op];
}
1225

1226
1227
1228
const char * ggml_unary_op_name(enum ggml_unary_op op) {
    return GGML_UNARY_OP_NAME[op];
}
1229

1230
1231
1232
1233
const char * ggml_op_desc(const struct ggml_tensor * t) {
    if (t->op == GGML_OP_UNARY) {
        enum ggml_unary_op uop = ggml_get_unary_op(t);
        return ggml_unary_op_name(uop);
1234
    }
1235
1236
    return ggml_op_name(t->op);
}
1237

1238
1239
size_t ggml_element_size(const struct ggml_tensor * tensor) {
    return ggml_type_size(tensor->type);
1240
1241
}

1242
1243
bool ggml_is_scalar(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1244

1245
1246
    return tensor->ne[0] == 1 && tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
1247

1248
1249
bool ggml_is_vector(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1250

1251
1252
    return tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
1253

1254
1255
bool ggml_is_matrix(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1256

1257
1258
    return tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
1259

1260
1261
1262
bool ggml_is_3d(const struct ggml_tensor * tensor) {
    return tensor->ne[3] == 1;
}
1263

1264
1265
1266
1267
int ggml_n_dims(const struct ggml_tensor * tensor) {
    for (int i = GGML_MAX_DIMS - 1; i >= 1; --i) {
        if (tensor->ne[i] > 1) {
            return i + 1;
1268
1269
        }
    }
1270
    return 1;
1271
1272
}

1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
    enum ggml_type wtype = GGML_TYPE_COUNT;

    switch (ftype) {
        case GGML_FTYPE_ALL_F32:              wtype = GGML_TYPE_F32;   break;
        case GGML_FTYPE_MOSTLY_F16:           wtype = GGML_TYPE_F16;   break;
        case GGML_FTYPE_MOSTLY_BF16:          wtype = GGML_TYPE_BF16;  break;
        case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
        case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
        case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
        case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
        case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
        case GGML_FTYPE_MOSTLY_Q2_K:          wtype = GGML_TYPE_Q2_K;  break;
        case GGML_FTYPE_MOSTLY_Q3_K:          wtype = GGML_TYPE_Q3_K;  break;
        case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
        case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;  break;
        case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
        case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
        case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
        case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
        case GGML_FTYPE_MOSTLY_IQ1_S:         wtype = GGML_TYPE_IQ1_S;    break;
        case GGML_FTYPE_MOSTLY_IQ1_M:         wtype = GGML_TYPE_IQ1_M;    break;
        case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
        case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
        case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
        case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
        case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
        case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
        case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
        case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
    }
1305

1306
    GGML_ASSERT(wtype != GGML_TYPE_COUNT);
1307

1308
1309
    return wtype;
}
1310

1311
1312
1313
size_t ggml_tensor_overhead(void) {
    return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE;
}
1314

1315
1316
1317
bool ggml_is_transposed(const struct ggml_tensor * tensor) {
    return tensor->nb[0] > tensor->nb[1];
}
1318

1319
1320
1321
1322
static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) {
    size_t next_nb = ggml_type_size(tensor->type);
    if (tensor->ne[0] != ggml_blck_size(tensor->type) && tensor->nb[0] != next_nb) {
        return false;
1323
    }
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
    next_nb *= tensor->ne[0]/ggml_blck_size(tensor->type);
    for (int i = 1; i < GGML_MAX_DIMS; i++) {
        if (tensor->ne[i] != 1) {
            if (i > n) {
                if (tensor->nb[i] != next_nb) {
                    return false;
                }
                next_nb *= tensor->ne[i];
            } else {
                // this dimension does not need to be contiguous
                next_nb = tensor->ne[i]*tensor->nb[i];
            }
        }
1337
    }
1338
    return true;
1339
1340
}

1341
1342
1343
bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
    return ggml_is_contiguous_0(tensor);
}
1344

1345
1346
1347
bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
    return ggml_is_contiguous_n(tensor, 0);
}
1348

1349
1350
1351
bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
    return ggml_is_contiguous_n(tensor, 1);
}
1352

1353
1354
1355
bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
    return ggml_is_contiguous_n(tensor, 2);
}
1356

1357
1358
bool ggml_is_permuted(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1359

1360
    return tensor->nb[0] > tensor->nb[1] || tensor->nb[1] > tensor->nb[2] || tensor->nb[2] > tensor->nb[3];
1361
1362
}

1363
1364
1365
1366
1367
1368
1369
static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");

    return
        tensor->nb[0] == ggml_type_size(tensor->type) &&
        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
1370
1371
}

1372
1373
1374
1375
1376
bool ggml_is_empty(const struct ggml_tensor * tensor) {
    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
        if (tensor->ne[i] == 0) {
            // empty if any dimension has no elements
            return true;
1377
1378
        }
    }
1379
    return false;
1380
1381
}

1382
1383
1384
1385
1386
1387
1388
1389
bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");

    return
        (t0->ne[0] == t1->ne[0]) &&
        (t0->ne[1] == t1->ne[1]) &&
        (t0->ne[2] == t1->ne[2]) &&
        (t0->ne[3] == t1->ne[3]);
1390
1391
}

1392
1393
bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
1394

1395
1396
1397
1398
1399
    return
        (t0->nb[0] == t1->nb[0]) &&
        (t0->nb[1] == t1->nb[1]) &&
        (t0->nb[2] == t1->nb[2]) &&
        (t0->nb[3] == t1->nb[3]);
1400
}
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410

// check if t1 can be represented as a repeatition of t0
bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");

    return ggml_is_empty(t0) ? ggml_is_empty(t1) :
        (t1->ne[0]%t0->ne[0] == 0) &&
        (t1->ne[1]%t0->ne[1] == 0) &&
        (t1->ne[2]%t0->ne[2] == 0) &&
        (t1->ne[3]%t0->ne[3] == 0);
1411
1412
}

1413
1414
1415
1416
static inline bool ggml_can_repeat_rows(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");

    return (t0->ne[0] == t1->ne[0]) && ggml_can_repeat(t0, t1);
1417
1418
}

1419
1420
1421
// assert that pointer is aligned to GGML_MEM_ALIGN
#define GGML_ASSERT_ALIGNED(ptr) \
    GGML_ASSERT(((uintptr_t) (ptr))%GGML_MEM_ALIGN == 0)
1422

1423
////////////////////////////////////////////////////////////////////////////////
1424

1425
1426
struct ggml_context * ggml_init(struct ggml_init_params params) {
    static bool is_first_call = true;
1427

1428
1429
1430
1431
1432
    ggml_critical_section_start();

    if (is_first_call) {
        // initialize time system (required on Windows)
        ggml_time_init();
1433

1434
1435
1436
1437
1438
1439
1440
        for (int i = 0; i < (1 << 16); ++i) {
            union {
                uint16_t u16;
                ggml_fp16_t fp16;
            } u = {i};
            ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
        }
1441

1442
        is_first_call = false;
1443
1444
    }

1445
1446
1447
    ggml_critical_section_end();

    struct ggml_context * ctx = GGML_MALLOC(sizeof(struct ggml_context));
1448

1449
1450
1451
    // allow to call ggml_init with 0 size
    if (params.mem_size == 0) {
        params.mem_size = GGML_MEM_ALIGN;
1452
    }
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472

    const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);

    *ctx = (struct ggml_context) {
        /*.mem_size           =*/ mem_size,
        /*.mem_buffer         =*/ params.mem_buffer ? params.mem_buffer : ggml_aligned_malloc(mem_size),
        /*.mem_buffer_owned   =*/ params.mem_buffer ? false : true,
        /*.no_alloc           =*/ params.no_alloc,
        /*.n_objects          =*/ 0,
        /*.objects_begin      =*/ NULL,
        /*.objects_end        =*/ NULL,
    };

    GGML_ASSERT(ctx->mem_buffer != NULL);

    GGML_ASSERT_ALIGNED(ctx->mem_buffer);

    GGML_PRINT_DEBUG("%s: context initialized\n", __func__);

    return ctx;
1473
1474
}

1475
1476
1477
void ggml_reset(struct ggml_context * ctx) {
    if (ctx == NULL) {
        return;
1478
    }
1479
1480
1481
1482

    ctx->n_objects     = 0;
    ctx->objects_begin = NULL;
    ctx->objects_end   = NULL;
1483
1484
}

1485
1486
1487
void ggml_free(struct ggml_context * ctx) {
    if (ctx == NULL) {
        return;
1488
1489
    }

1490
1491
    if (ctx->mem_buffer_owned) {
        ggml_aligned_free(ctx->mem_buffer, ctx->mem_size);
1492
    }
1493
1494

    GGML_FREE(ctx);
1495
1496
}

1497
1498
size_t ggml_used_mem(const struct ggml_context * ctx) {
    return ctx->objects_end == NULL ? 0 : ctx->objects_end->offs + ctx->objects_end->size;
1499
1500
}

1501
1502
bool ggml_get_no_alloc(struct ggml_context * ctx) {
    return ctx->no_alloc;
1503
1504
}

1505
1506
void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc) {
    ctx->no_alloc = no_alloc;
1507
1508
}

1509
1510
void * ggml_get_mem_buffer(const struct ggml_context * ctx) {
    return ctx->mem_buffer;
1511
1512
}

1513
1514
1515
size_t ggml_get_mem_size(const struct ggml_context * ctx) {
    return ctx->mem_size;
}
1516

1517
1518
size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
    size_t max_size = 0;
1519

1520
1521
1522
1523
    for (struct ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor != NULL; tensor = ggml_get_next_tensor(ctx, tensor)) {
        size_t bytes = ggml_nbytes(tensor);
        max_size = MAX(max_size, bytes);
    }
1524

1525
1526
    return max_size;
}
1527

1528
////////////////////////////////////////////////////////////////////////////////
1529

1530
1531
1532
static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml_object_type type, size_t size) {
    // always insert objects at the end of the context's memory pool
    struct ggml_object * obj_cur = ctx->objects_end;
1533

1534
1535
1536
    const size_t cur_offs = obj_cur == NULL ? 0 : obj_cur->offs;
    const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size;
    const size_t cur_end  = cur_offs + cur_size;
1537

1538
1539
    // align to GGML_MEM_ALIGN
    size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);
1540

1541
1542
    char * const mem_buffer = ctx->mem_buffer;
    struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
1543

1544
1545
1546
1547
1548
1549
1550
1551
    if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
        GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
                __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
#ifndef NDEBUG
        GGML_ABORT("not enough space in the context's memory pool");
#endif
        return NULL;
    }
1552

1553
1554
1555
1556
1557
1558
    *obj_new = (struct ggml_object) {
        .offs = cur_end + GGML_OBJECT_SIZE,
        .size = size_needed,
        .next = NULL,
        .type = type,
    };
1559

1560
    GGML_ASSERT_ALIGNED(mem_buffer + obj_new->offs);
1561

1562
1563
1564
1565
1566
1567
    if (obj_cur != NULL) {
        obj_cur->next = obj_new;
    } else {
        // this is the first object in this context
        ctx->objects_begin = obj_new;
    }
1568

1569
    ctx->objects_end = obj_new;
1570

1571
    //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size);
1572

1573
1574
    return obj_new;
}
1575

1576
1577
1578
1579
1580
1581
1582
static struct ggml_tensor * ggml_new_tensor_impl(
        struct ggml_context * ctx,
        enum   ggml_type      type,
        int                   n_dims,
        const int64_t       * ne,
        struct ggml_tensor  * view_src,
        size_t                view_offs) {
1583

1584
1585
    GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
    GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
1586

1587
1588
1589
1590
1591
    // find the base tensor and absolute offset
    if (view_src != NULL && view_src->view_src != NULL) {
        view_offs += view_src->view_offs;
        view_src   = view_src->view_src;
    }
1592

1593
1594
1595
1596
    size_t data_size = ggml_row_size(type, ne[0]);
    for (int i = 1; i < n_dims; i++) {
        data_size *= ne[i];
    }
1597

1598
    GGML_ASSERT(view_src == NULL || data_size == 0 || data_size + view_offs <= ggml_nbytes(view_src));
1599

1600
1601
1602
1603
    void * data = view_src != NULL ? view_src->data : NULL;
    if (data != NULL) {
        data = (char *) data + view_offs;
    }
1604

1605
    size_t obj_alloc_size = 0;
1606

1607
1608
1609
1610
    if (view_src == NULL && !ctx->no_alloc) {
        // allocate tensor data in the context's memory pool
        obj_alloc_size = data_size;
    }
1611

1612
1613
    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
    GGML_ASSERT(obj_new);
1614

1615
    struct ggml_tensor * const result = (struct ggml_tensor *)((char *)ctx->mem_buffer + obj_new->offs);
1616

1617
1618
1619
1620
1621
#ifdef __clang__
    // temporary until ggml_tensor::backend is removed
    #pragma clang diagnostic push
    #pragma clang diagnostic ignored "-Wdeprecated-declarations"
#endif
1622

1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
    *result = (struct ggml_tensor) {
        /*.type         =*/ type,
        /*.backend      =*/ GGML_BACKEND_TYPE_CPU,
        /*.buffer       =*/ NULL,
        /*.ne           =*/ { 1, 1, 1, 1 },
        /*.nb           =*/ { 0, 0, 0, 0 },
        /*.op           =*/ GGML_OP_NONE,
        /*.op_params    =*/ { 0 },
        /*.flags        =*/ 0,
        /*.src          =*/ { NULL },
        /*.view_src     =*/ view_src,
        /*.view_offs    =*/ view_offs,
        /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
        /*.name         =*/ { 0 },
        /*.extra        =*/ NULL,
        /*.padding      =*/ { 0 },
    };
1640

1641
1642
#ifdef __clang__
    #pragma clang diagnostic pop
1643
1644
#endif

1645
1646
    // TODO: this should not be needed as long as we don't rely on aligned SIMD loads
    //GGML_ASSERT_ALIGNED(result->data);
1647

1648
1649
    for (int i = 0; i < n_dims; i++) {
        result->ne[i] = ne[i];
1650
1651
    }

1652
1653
1654
1655
    result->nb[0] = ggml_type_size(type);
    result->nb[1] = result->nb[0]*(result->ne[0]/ggml_blck_size(type));
    for (int i = 2; i < GGML_MAX_DIMS; i++) {
        result->nb[i] = result->nb[i - 1]*result->ne[i - 1];
1656
1657
    }

1658
    ctx->n_objects++;
1659

1660
1661
    return result;
}
1662

1663
1664
1665
1666
1667
1668
struct ggml_tensor * ggml_new_tensor(
        struct ggml_context * ctx,
        enum   ggml_type      type,
        int                   n_dims,
        const int64_t       * ne) {
    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL, 0);
1669
}
1670

1671
1672
1673
1674
1675
struct ggml_tensor * ggml_new_tensor_1d(
        struct ggml_context * ctx,
        enum   ggml_type      type,
        int64_t ne0) {
    return ggml_new_tensor(ctx, type, 1, &ne0);
1676
1677
}

1678
1679
1680
1681
1682
1683
1684
struct ggml_tensor * ggml_new_tensor_2d(
        struct ggml_context * ctx,
        enum   ggml_type      type,
        int64_t ne0,
        int64_t ne1) {
    const int64_t ne[2] = { ne0, ne1 };
    return ggml_new_tensor(ctx, type, 2, ne);
1685
}
1686
1687
1688
1689
1690
1691
1692
1693
1694

struct ggml_tensor * ggml_new_tensor_3d(
        struct ggml_context * ctx,
        enum   ggml_type      type,
        int64_t ne0,
        int64_t ne1,
        int64_t ne2) {
    const int64_t ne[3] = { ne0, ne1, ne2 };
    return ggml_new_tensor(ctx, type, 3, ne);
1695
1696
}

1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
struct ggml_tensor * ggml_new_tensor_4d(
        struct ggml_context * ctx,
        enum   ggml_type type,
        int64_t ne0,
        int64_t ne1,
        int64_t ne2,
        int64_t ne3) {
    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
    return ggml_new_tensor(ctx, type, 4, ne);
}
1707

1708
1709
void * ggml_new_buffer(struct ggml_context * ctx, size_t nbytes) {
    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, nbytes);
1710

1711
1712
    return (uint8_t *)ctx->mem_buffer + obj->offs;
}
1713

1714
1715
1716
struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
    return ggml_new_tensor(ctx, src->type, GGML_MAX_DIMS, src->ne);
}
1717

1718
1719
1720
1721
void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3) {
    const int64_t ne2 = tensor->ne[2];
    const int64_t ne1 = tensor->ne[1];
    const int64_t ne0 = tensor->ne[0];
1722

1723
1724
1725
1726
    const int64_t i3_ = (i/(ne2*ne1*ne0));
    const int64_t i2_ = (i - i3_*ne2*ne1*ne0)/(ne1*ne0);
    const int64_t i1_ = (i - i3_*ne2*ne1*ne0 - i2_*ne1*ne0)/ne0;
    const int64_t i0_ = (i - i3_*ne2*ne1*ne0 - i2_*ne1*ne0 - i1_*ne0);
1727

1728
1729
    if (i0) {
        * i0 = i0_;
1730
    }
1731
1732
1733
1734
1735
1736
1737
1738
    if (i1) {
        * i1 = i1_;
    }
    if (i2) {
        * i2 = i2_;
    }
    if (i3) {
        * i3 = i3_;
1739
    }
1740
}
1741

1742
1743
1744
void * ggml_get_data(const struct ggml_tensor * tensor) {
    return tensor->data;
}
1745

1746
1747
1748
1749
float * ggml_get_data_f32(const struct ggml_tensor * tensor) {
    assert(tensor->type == GGML_TYPE_F32);
    return (float *)(tensor->data);
}
1750

1751
1752
1753
1754
enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
    GGML_ASSERT(tensor->op == GGML_OP_UNARY);
    return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0);
}
1755

1756
1757
1758
const char * ggml_get_name(const struct ggml_tensor * tensor) {
    return tensor->name;
}
1759

1760
1761
1762
1763
struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name) {
    size_t i;
    for (i = 0; i < sizeof(tensor->name) - 1 && name[i] != '\0'; i++) {
        tensor->name[i] = name[i];
1764
    }
1765
1766
    tensor->name[i] = '\0';
    return tensor;
1767
1768
}

1769
1770
1771
1772
1773
1774
struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...) {
    va_list args;
    va_start(args, fmt);
    vsnprintf(tensor->name, sizeof(tensor->name), fmt, args);
    va_end(args);
    return tensor;
1775
1776
}

1777
1778
1779
1780
1781
struct ggml_tensor * ggml_view_tensor(
        struct ggml_context * ctx,
        struct ggml_tensor  * src) {
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, GGML_MAX_DIMS, src->ne, src, 0);
    ggml_format_name(result, "%s (view)", src->name);
1782

1783
1784
1785
1786
1787
    for (int i = 0; i < GGML_MAX_DIMS; i++) {
        result->nb[i] = src->nb[i];
    }

    return result;
1788
1789
}

1790
struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx) {
1791
1792
    struct ggml_object * obj = ctx->objects_begin;

1793
    char * const mem_buffer = ctx->mem_buffer;
1794
1795

    while (obj != NULL) {
1796
1797
1798
1799
        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
            return (struct ggml_tensor *)(mem_buffer + obj->offs);
        }

1800
1801
1802
        obj = obj->next;
    }

1803
    return NULL;
1804
1805
}

1806
1807
1808
struct ggml_tensor * ggml_get_next_tensor(const struct ggml_context * ctx, struct ggml_tensor * tensor) {
    struct ggml_object * obj = (struct ggml_object *) ((char *)tensor - GGML_OBJECT_SIZE);
    obj = obj->next;
1809

1810
    char * const mem_buffer = ctx->mem_buffer;
1811

1812
1813
1814
1815
    while (obj != NULL) {
        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
            return (struct ggml_tensor *)(mem_buffer + obj->offs);
        }
1816

1817
1818
1819
1820
        obj = obj->next;
    }

    return NULL;
1821
1822
}

1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name) {
    struct ggml_object * obj = ctx->objects_begin;

    char * const mem_buffer = ctx->mem_buffer;

    while (obj != NULL) {
        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
            struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs);
            if (strcmp(cur->name, name) == 0) {
                return cur;
            }
1834
        }
1835
1836

        obj = obj->next;
1837
1838
    }

1839
    return NULL;
1840
1841
}

1842
////////////////////////////////////////////////////////////////////////////////
1843

1844
// ggml_dup
1845

1846
1847
1848
1849
1850
static struct ggml_tensor * ggml_dup_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
1851

1852
1853
    result->op     = GGML_OP_DUP;
    result->src[0] = a;
1854

1855
    return result;
1856
1857
}

1858
1859
1860
1861
struct ggml_tensor * ggml_dup(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_dup_impl(ctx, a, false);
1862
1863
}

1864
1865
1866
1867
struct ggml_tensor * ggml_dup_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_dup_impl(ctx, a, true);
1868
1869
}

1870
// ggml_add
1871

1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
static struct ggml_tensor * ggml_add_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    GGML_ASSERT(ggml_can_repeat(b, a));

    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    result->op     = GGML_OP_ADD;
    result->src[0] = a;
    result->src[1] = b;

    return result;
1886
1887
}

1888
1889
1890
1891
1892
struct ggml_tensor * ggml_add(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_add_impl(ctx, a, b, false);
1893
1894
}

1895
1896
1897
1898
1899
struct ggml_tensor * ggml_add_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_add_impl(ctx, a, b, true);
1900
1901
}

1902
// ggml_add_cast
1903

1904
1905
1906
1907
1908
1909
1910
1911
static struct ggml_tensor * ggml_add_cast_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        enum   ggml_type      type) {
    // TODO: support less-strict constraint
    //       GGML_ASSERT(ggml_can_repeat(b, a));
    GGML_ASSERT(ggml_can_repeat_rows(b, a));
1912

1913
1914
1915
1916
    // currently only supported for quantized input and f16
    GGML_ASSERT(ggml_is_quantized(a->type) ||
                a->type == GGML_TYPE_F16 ||
                a->type == GGML_TYPE_BF16);
1917

1918
    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
1919

1920
1921
1922
    result->op     = GGML_OP_ADD;
    result->src[0] = a;
    result->src[1] = b;
1923

1924
    return result;
1925
1926
}

1927
1928
1929
1930
1931
1932
struct ggml_tensor * ggml_add_cast(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        enum   ggml_type      type) {
    return ggml_add_cast_impl(ctx, a, b, type);
1933
1934
}

1935
// ggml_add1
1936

1937
1938
1939
1940
1941
1942
1943
static struct ggml_tensor * ggml_add1_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    GGML_ASSERT(ggml_is_scalar(b));
    GGML_ASSERT(ggml_is_padded_1d(a));
1944

1945
1946
1947
1948
1949
1950
1951
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    result->op     = GGML_OP_ADD1;
    result->src[0] = a;
    result->src[1] = b;

    return result;
1952
1953
}

1954
1955
1956
1957
1958
1959
struct ggml_tensor * ggml_add1(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_add1_impl(ctx, a, b, false);
}
1960

1961
1962
1963
1964
1965
struct ggml_tensor * ggml_add1_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_add1_impl(ctx, a, b, true);
1966
1967
}

1968
// ggml_acc
1969

1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
static struct ggml_tensor * ggml_acc_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset,
        bool                  inplace) {
    GGML_ASSERT(ggml_nelements(b) <= ggml_nelements(a));
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(a->type == GGML_TYPE_F32);
    GGML_ASSERT(b->type == GGML_TYPE_F32);
1983

1984
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
1985

1986
1987
    int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
    ggml_set_op_params(result, params, sizeof(params));
1988

1989
1990
1991
    result->op     = GGML_OP_ACC;
    result->src[0] = a;
    result->src[1] = b;
1992

1993
    return result;
1994
1995
}

1996
1997
1998
1999
2000
2001
2002
2003
2004
struct ggml_tensor * ggml_acc(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset) {
    return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
2005
2006
}

2007
2008
2009
2010
2011
2012
2013
2014
2015
struct ggml_tensor * ggml_acc_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset) {
    return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
2016
2017
}

2018
// ggml_sub
2019

2020
2021
2022
2023
2024
2025
static struct ggml_tensor * ggml_sub_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    GGML_ASSERT(ggml_can_repeat(b, a));
2026

2027
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2028

2029
2030
2031
    result->op     = GGML_OP_SUB;
    result->src[0] = a;
    result->src[1] = b;
2032

2033
    return result;
2034
2035
}

2036
2037
2038
2039
2040
struct ggml_tensor * ggml_sub(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_sub_impl(ctx, a, b, false);
2041
2042
}

2043
2044
2045
2046
2047
struct ggml_tensor * ggml_sub_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_sub_impl(ctx, a, b, true);
2048
2049
}

2050
// ggml_mul
2051

2052
2053
2054
2055
2056
2057
static struct ggml_tensor * ggml_mul_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    GGML_ASSERT(ggml_can_repeat(b, a));
2058

2059
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2060

2061
2062
2063
    result->op     = GGML_OP_MUL;
    result->src[0] = a;
    result->src[1] = b;
2064

2065
    return result;
2066
2067
}

2068
2069
2070
2071
2072
struct ggml_tensor * ggml_mul(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_mul_impl(ctx, a, b, false);
2073
2074
}

2075
2076
2077
2078
2079
struct ggml_tensor * ggml_mul_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_mul_impl(ctx, a, b, true);
2080
2081
}

2082
// ggml_div
2083

2084
2085
2086
2087
2088
2089
static struct ggml_tensor * ggml_div_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    GGML_ASSERT(ggml_can_repeat(b, a));
2090

2091
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2092

2093
2094
2095
    result->op     = GGML_OP_DIV;
    result->src[0] = a;
    result->src[1] = b;
2096

2097
2098
    return result;
}
2099

2100
2101
2102
2103
2104
2105
struct ggml_tensor * ggml_div(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_div_impl(ctx, a, b, false);
}
2106

2107
2108
2109
2110
2111
2112
struct ggml_tensor * ggml_div_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_div_impl(ctx, a, b, true);
}
2113

2114
// ggml_sqr
2115

2116
2117
2118
2119
2120
static struct ggml_tensor * ggml_sqr_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2121

2122
2123
    result->op     = GGML_OP_SQR;
    result->src[0] = a;
2124

2125
2126
    return result;
}
2127

2128
2129
2130
2131
2132
struct ggml_tensor * ggml_sqr(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sqr_impl(ctx, a, false);
}
2133

2134
2135
2136
2137
struct ggml_tensor * ggml_sqr_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sqr_impl(ctx, a, true);
2138
2139
}

2140
// ggml_sqrt
2141

2142
2143
2144
2145
2146
static struct ggml_tensor * ggml_sqrt_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2147

2148
2149
    result->op     = GGML_OP_SQRT;
    result->src[0] = a;
2150

2151
2152
    return result;
}
2153

2154
2155
2156
2157
2158
struct ggml_tensor * ggml_sqrt(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sqrt_impl(ctx, a, false);
}
2159

2160
2161
2162
2163
2164
struct ggml_tensor * ggml_sqrt_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sqrt_impl(ctx, a, true);
}
2165

2166
// ggml_log
2167

2168
2169
2170
2171
2172
static struct ggml_tensor * ggml_log_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2173

2174
2175
    result->op     = GGML_OP_LOG;
    result->src[0] = a;
2176

2177
2178
    return result;
}
2179

2180
2181
2182
2183
2184
struct ggml_tensor * ggml_log(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_log_impl(ctx, a, false);
}
2185

2186
2187
2188
2189
2190
struct ggml_tensor * ggml_log_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_log_impl(ctx, a, true);
}
2191

2192
// ggml_sin
2193

2194
2195
2196
2197
2198
static struct ggml_tensor * ggml_sin_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2199

2200
2201
    result->op     = GGML_OP_SIN;
    result->src[0] = a;
2202

2203
2204
    return result;
}
2205

2206
2207
2208
2209
2210
struct ggml_tensor * ggml_sin(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sin_impl(ctx, a, false);
}
2211

2212
2213
2214
2215
2216
struct ggml_tensor * ggml_sin_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_sin_impl(ctx, a, true);
}
2217

2218
// ggml_cos
2219

2220
2221
2222
2223
2224
static struct ggml_tensor * ggml_cos_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2225

2226
2227
    result->op     = GGML_OP_COS;
    result->src[0] = a;
2228

2229
2230
    return result;
}
2231

2232
2233
2234
2235
2236
struct ggml_tensor * ggml_cos(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_cos_impl(ctx, a, false);
}
2237

2238
2239
2240
2241
2242
struct ggml_tensor * ggml_cos_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_cos_impl(ctx, a, true);
}
2243

2244
// ggml_sum
2245

2246
2247
2248
2249
struct ggml_tensor * ggml_sum(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1);
2250

2251
2252
2253
2254
    result->op     = GGML_OP_SUM;
    result->src[0] = a;

    return result;
2255
2256
}

2257
// ggml_sum_rows
2258

2259
2260
2261
2262
2263
2264
2265
struct ggml_tensor * ggml_sum_rows(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    int64_t ne[GGML_MAX_DIMS] = { 1 };
    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
        ne[i] = a->ne[i];
    }
2266

2267
    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
2268

2269
2270
    result->op     = GGML_OP_SUM_ROWS;
    result->src[0] = a;
2271

2272
2273
    return result;
}
2274

2275
// ggml_mean
2276

2277
2278
2279
2280
2281
struct ggml_tensor * ggml_mean(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
2282

2283
2284
    result->op     = GGML_OP_MEAN;
    result->src[0] = a;
2285

2286
    return result;
2287
2288
}

2289
2290
2291
2292
2293
2294
2295
// ggml_argmax

struct ggml_tensor * ggml_argmax(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    GGML_ASSERT(ggml_is_matrix(a));
    GGML_ASSERT(a->ne[0] <= INT32_MAX);
2296

2297
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
2298

2299
2300
    result->op     = GGML_OP_ARGMAX;
    result->src[0] = a;
2301
2302
2303
2304

    return result;
}

2305
// ggml_count_equal
2306

2307
2308
2309
2310
2311
struct ggml_tensor * ggml_count_equal(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_are_same_shape(a, b));
2312

2313
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, 1);
2314

2315
2316
2317
2318
2319
    result->op     = GGML_OP_COUNT_EQUAL;
    result->src[0] = a;
    result->src[1] = b;

    return result;
2320
2321
}

2322
// ggml_repeat
2323

2324
2325
2326
2327
2328
struct ggml_tensor * ggml_repeat(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_can_repeat(a, b));
2329

2330
    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
2331

2332
2333
    result->op     = GGML_OP_REPEAT;
    result->src[0] = a;
2334

2335
    return result;
2336
2337
}

2338
// ggml_repeat_back
2339

2340
2341
2342
2343
2344
struct ggml_tensor * ggml_repeat_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_can_repeat(b, a));
2345

2346
    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
2347

2348
2349
    result->op     = GGML_OP_REPEAT_BACK;
    result->src[0] = a;
2350

2351
2352
    return result;
}
2353

2354
// ggml_concat
2355

2356
2357
2358
2359
2360
2361
struct ggml_tensor * ggml_concat(
    struct ggml_context * ctx,
    struct ggml_tensor  * a,
    struct ggml_tensor  * b,
    int                   dim) {
    GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
2362

2363
2364
2365
2366
2367
2368
2369
2370
    int64_t ne[GGML_MAX_DIMS];
    for (int d = 0; d < GGML_MAX_DIMS; ++d) {
        if (d == dim) {
            ne[d] = a->ne[d] + b->ne[d];
            continue;
        }
        GGML_ASSERT(a->ne[d] == b->ne[d]);
        ne[d] = a->ne[d];
2371
2372
    }

2373
    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
2374

2375
    ggml_set_op_params_i32(result, 0, dim);
2376

2377
2378
2379
    result->op     = GGML_OP_CONCAT;
    result->src[0] = a;
    result->src[1] = b;
2380

2381
2382
    return result;
}
2383

2384
// ggml_abs
2385

2386
2387
2388
2389
struct ggml_tensor * ggml_abs(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_ABS);
2390
2391
}

2392
struct ggml_tensor * ggml_abs_inplace(
2393
        struct ggml_context * ctx,
2394
2395
2396
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_ABS);
}
2397

2398
// ggml_sgn
2399

2400
2401
2402
2403
2404
struct ggml_tensor * ggml_sgn(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_SGN);
}
2405

2406
2407
2408
2409
2410
struct ggml_tensor * ggml_sgn_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SGN);
}
2411

2412
// ggml_neg
2413

2414
2415
2416
2417
2418
struct ggml_tensor * ggml_neg(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_NEG);
}
2419

2420
2421
2422
2423
2424
struct ggml_tensor * ggml_neg_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_NEG);
}
2425

2426
// ggml_step
2427

2428
2429
2430
2431
2432
struct ggml_tensor * ggml_step(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_STEP);
}
2433

2434
2435
2436
2437
2438
struct ggml_tensor * ggml_step_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_STEP);
}
2439

2440
// ggml_tanh
2441

2442
2443
2444
2445
2446
struct ggml_tensor * ggml_tanh(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_TANH);
}
2447

2448
2449
2450
2451
2452
struct ggml_tensor * ggml_tanh_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_TANH);
}
2453

2454
// ggml_elu
2455

2456
2457
2458
2459
2460
struct ggml_tensor * ggml_elu(
    struct ggml_context * ctx,
    struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_ELU);
}
2461

2462
2463
2464
2465
2466
struct ggml_tensor * ggml_elu_inplace(
    struct ggml_context * ctx,
    struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_ELU);
}
2467

2468
// ggml_relu
2469

2470
2471
2472
2473
2474
struct ggml_tensor * ggml_relu(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_RELU);
}
2475

2476
2477
2478
2479
2480
struct ggml_tensor * ggml_relu_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_RELU);
}
2481

2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
// ggml_leaky_relu

struct ggml_tensor * ggml_leaky_relu(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 negative_slope,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    ggml_set_op_params(result, &negative_slope, sizeof(negative_slope));

    result->op     = GGML_OP_LEAKY_RELU;
    result->src[0] = a;
2495
2496
2497
2498

    return result;
}

2499
2500
2501
// ggml_sigmoid

struct ggml_tensor * ggml_sigmoid(
2502
        struct ggml_context * ctx,
2503
2504
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_SIGMOID);
2505
2506
}

2507
struct ggml_tensor * ggml_sigmoid_inplace(
2508
        struct ggml_context * ctx,
2509
2510
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SIGMOID);
2511
2512
}

2513
2514
2515
// ggml_gelu

struct ggml_tensor * ggml_gelu(
2516
        struct ggml_context * ctx,
2517
2518
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_GELU);
2519
2520
}

2521
struct ggml_tensor * ggml_gelu_inplace(
2522
        struct ggml_context * ctx,
2523
2524
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU);
2525
2526
}

2527
2528
2529
// ggml_gelu_quick

struct ggml_tensor * ggml_gelu_quick(
2530
        struct ggml_context * ctx,
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_GELU_QUICK);
}

struct ggml_tensor * ggml_gelu_quick_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU_QUICK);
}

// ggml_silu

struct ggml_tensor * ggml_silu(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_SILU);
2547
2548
}

2549
2550
2551
2552
2553
struct ggml_tensor * ggml_silu_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SILU);
}
2554

2555
// ggml_silu_back
2556

2557
2558
2559
2560
2561
struct ggml_tensor * ggml_silu_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
2562

2563
2564
2565
    result->op     = GGML_OP_SILU_BACK;
    result->src[0] = a;
    result->src[1] = b;
2566
2567
2568
2569

    return result;
}

2570
2571
2572
2573
2574
2575
2576
// ggml hardswish

struct ggml_tensor * ggml_hardswish(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSWISH);
}
2577

2578
// ggml hardsigmoid
2579

2580
2581
2582
2583
2584
struct ggml_tensor * ggml_hardsigmoid(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID);
}
2585

2586
// ggml exp
2587

2588
2589
2590
2591
struct ggml_tensor * ggml_exp(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary(ctx, a, GGML_UNARY_OP_EXP);
2592
2593
}

2594
2595
2596
2597
struct ggml_tensor * ggml_exp_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXP);
2598
2599
}

2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
// ggml_norm

static struct ggml_tensor * ggml_norm_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    ggml_set_op_params(result, &eps, sizeof(eps));

    result->op     = GGML_OP_NORM;
    result->src[0] = a;

    return result;
2615
2616
}

2617
2618
2619
2620
2621
struct ggml_tensor * ggml_norm(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps) {
    return ggml_norm_impl(ctx, a, eps, false);
2622
2623
}

2624
2625
2626
2627
2628
struct ggml_tensor * ggml_norm_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps) {
    return ggml_norm_impl(ctx, a, eps, true);
2629
2630
}

2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
// ggml_rms_norm

static struct ggml_tensor * ggml_rms_norm_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    ggml_set_op_params(result, &eps, sizeof(eps));

    result->op     = GGML_OP_RMS_NORM;
    result->src[0] = a;

    return result;
2646
2647
}

2648
2649
2650
2651
2652
struct ggml_tensor * ggml_rms_norm(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps) {
    return ggml_rms_norm_impl(ctx, a, eps, false);
2653
2654
}

2655
2656
2657
2658
2659
struct ggml_tensor * ggml_rms_norm_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 eps) {
    return ggml_rms_norm_impl(ctx, a, eps, true);
2660
2661
}

2662
// ggml_rms_norm_back
2663

2664
2665
2666
2667
2668
2669
struct ggml_tensor * ggml_rms_norm_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        float                 eps) {
    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
2670

2671
    ggml_set_op_params(result, &eps, sizeof(eps));
2672

2673
2674
2675
2676
2677
    result->op     = GGML_OP_RMS_NORM_BACK;
    result->src[0] = a;
    result->src[1] = b;

    return result;
2678
2679
}

2680
// ggml_group_norm
2681

2682
2683
2684
2685
2686
2687
2688
static struct ggml_tensor * ggml_group_norm_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_groups,
        float                 eps,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
2689

2690
2691
    ggml_set_op_params_i32(result, 0, n_groups);
    ggml_set_op_params_f32(result, 1, eps);
2692

2693
2694
    result->op     = GGML_OP_GROUP_NORM;
    result->src[0] = a;
2695

2696
    return result;
2697
2698
}

2699
2700
2701
2702
2703
2704
struct ggml_tensor * ggml_group_norm(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_groups,
        float                 eps) {
    return ggml_group_norm_impl(ctx, a, n_groups, eps, false);
2705
2706
}

2707
2708
2709
2710
2711
2712
struct ggml_tensor * ggml_group_norm_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_groups,
        float                 eps) {
    return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
2713
2714
}

2715
// ggml_mul_mat
2716

2717
2718
static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
2719

2720
2721
2722
    return (t0->ne[0]           == t1->ne[0])  &&
           (t1->ne[2]%t0->ne[2] == 0)          && // verify t0 is broadcastable
           (t1->ne[3]%t0->ne[3] == 0);
2723
2724
}

2725
2726
2727
2728
2729
2730
struct ggml_tensor * ggml_mul_mat(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_can_mul_mat(a, b));
    GGML_ASSERT(!ggml_is_transposed(a));
2731

2732
2733
    const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
2734

2735
2736
2737
    result->op     = GGML_OP_MUL_MAT;
    result->src[0] = a;
    result->src[1] = b;
2738

2739
    return result;
2740
2741
}

2742
2743
2744
2745
void ggml_mul_mat_set_prec(
        struct ggml_tensor * a,
        enum ggml_prec       prec) {
    GGML_ASSERT(a->op == GGML_OP_MUL_MAT);
2746

2747
    const int32_t prec_i32 = (int32_t) prec;
2748

2749
    ggml_set_op_params_i32(a, 0, prec_i32);
2750
2751
}

2752
// ggml_mul_mat_id
2753

2754
2755
/*
    c = ggml_mul_mat_id(ctx, as, b, ids);
2756

2757
2758
2759
2760
    as  -> [cols, rows, n_expert]
    ids -> [n_experts_used, n_tokens] (i32)
    b   -> [cols, n_expert_used, n_tokens]
    c   -> [rows, n_expert_used, n_tokens]
2761

2762
    in b, n_experts_used can be broadcasted to match the n_expert_used of ids
2763

2764
2765
2766
    c ~= as[:,:,i] @ b[:,i%r,t], i = ids[e,t] for all e,t in ids
*/
struct ggml_tensor * ggml_mul_mat_id(
2767
        struct ggml_context * ctx,
2768
        struct ggml_tensor  * as,
2769
        struct ggml_tensor  * b,
2770
2771
2772
        struct ggml_tensor  * ids) {
    GGML_ASSERT(!ggml_is_transposed(as));
    GGML_ASSERT(ids->type == GGML_TYPE_I32);
2773

2774
2775
2776
2777
2778
2779
    GGML_ASSERT(as->ne[3] == 1); // as is 3d (one matrix per expert)
    GGML_ASSERT(b->ne[3] == 1); // b is 3d
    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1); // ids is 2d
    GGML_ASSERT(ids->ne[1] == b->ne[2]); // must have an expert list per b row
    GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat
    GGML_ASSERT(ids->ne[0] % b->ne[1] == 0); // can broadcast
2780

2781
2782
2783
2784
2785
    const int64_t ne[4] = { as->ne[1], ids->ne[0], b->ne[2], 1 };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);

    result->op     = GGML_OP_MUL_MAT_ID;
    result->src[0] = as;
2786
    result->src[1] = b;
2787
    result->src[2] = ids;
2788
2789
2790
2791

    return result;
}

2792
// ggml_out_prod
2793

2794
2795
static inline bool ggml_can_out_prod(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
2796

2797
2798
2799
2800
    return (t0->ne[1] == t1->ne[1])   &&
           (t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
           (t1->ne[3]%t0->ne[3] == 0);
}
2801

2802
struct ggml_tensor * ggml_out_prod(
2803
        struct ggml_context * ctx,
2804
        struct ggml_tensor  * a,
2805
2806
2807
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_can_out_prod(a, b));
    GGML_ASSERT(!ggml_is_transposed(a));
2808

2809
2810
2811
    // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
    const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
2812

2813
    result->op     = GGML_OP_OUT_PROD;
2814
2815
2816
2817
2818
2819
    result->src[0] = a;
    result->src[1] = b;

    return result;
}

2820
// ggml_scale
2821

2822
static struct ggml_tensor * ggml_scale_impl(
2823
        struct ggml_context * ctx,
2824
        struct ggml_tensor  * a,
2825
        float                 s,
2826
        bool                  inplace) {
2827
2828
2829
2830
    GGML_ASSERT(ggml_is_padded_1d(a));

    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

2831
2832
2833
    ggml_set_op_params(result, &s, sizeof(s));

    result->op     = GGML_OP_SCALE;
2834
2835
2836
2837
2838
    result->src[0] = a;

    return result;
}

2839
struct ggml_tensor * ggml_scale(
2840
        struct ggml_context * ctx,
2841
        struct ggml_tensor  * a,
2842
2843
        float                 s) {
    return ggml_scale_impl(ctx, a, s, false);
2844
2845
}

2846
struct ggml_tensor * ggml_scale_inplace(
2847
        struct ggml_context * ctx,
2848
        struct ggml_tensor  * a,
2849
2850
        float                 s) {
    return ggml_scale_impl(ctx, a, s, true);
2851
2852
}

2853
// ggml_set
2854

2855
static struct ggml_tensor * ggml_set_impl(
2856
        struct ggml_context * ctx,
2857
2858
2859
2860
2861
2862
2863
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset,
        bool                  inplace) {
2864
    GGML_ASSERT(ggml_nelements(a) >= ggml_nelements(b));
2865

2866
    // make a view of the destination
2867
2868
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

2869
    GGML_ASSERT(offset < (size_t)(1 << 30));
2870
2871
2872
    int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
    ggml_set_op_params(result, params, sizeof(params));

2873
    result->op     = GGML_OP_SET;
2874
2875
2876
2877
2878
2879
    result->src[0] = a;
    result->src[1] = b;

    return result;
}

2880
struct ggml_tensor * ggml_set(
2881
        struct ggml_context * ctx,
2882
2883
2884
2885
2886
2887
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset) {
2888
    return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
2889
2890
}

2891
struct ggml_tensor * ggml_set_inplace(
2892
        struct ggml_context * ctx,
2893
2894
2895
2896
2897
2898
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset) {
2899
    return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
2900
2901
}

2902
struct ggml_tensor * ggml_set_1d(
2903
        struct ggml_context * ctx,
2904
2905
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
2906
2907
        size_t                offset) {
    return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, false);
2908
2909
}

2910
struct ggml_tensor * ggml_set_1d_inplace(
2911
        struct ggml_context * ctx,
2912
        struct ggml_tensor  * a,
2913
2914
2915
        struct ggml_tensor  * b,
        size_t                offset) {
    return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, true);
2916
2917
}

2918
struct ggml_tensor * ggml_set_2d(
2919
        struct ggml_context * ctx,
2920
        struct ggml_tensor  * a,
2921
2922
2923
2924
        struct ggml_tensor  * b,
        size_t                nb1,
        size_t                offset) {
    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
2925
2926
}

2927
struct ggml_tensor * ggml_set_2d_inplace(
2928
        struct ggml_context * ctx,
2929
2930
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
2931
2932
2933
2934
        size_t                nb1,
        size_t                offset) {
    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true);
}
2935

2936
// ggml_cpy
2937

2938
static struct ggml_tensor * ggml_cpy_impl(
2939
2940
2941
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
2942
    GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
2943

2944
2945
2946
2947
2948
2949
2950
    // make a view of the destination
    struct ggml_tensor * result = ggml_view_tensor(ctx, b);
    if (strlen(b->name) > 0) {
        ggml_format_name(result, "%s (copy of %s)", b->name, a->name);
    } else {
        ggml_format_name(result, "%s (copy)", a->name);
    }
2951

2952
    result->op     = GGML_OP_CPY;
2953
2954
2955
2956
2957
2958
    result->src[0] = a;
    result->src[1] = b;

    return result;
}

2959
struct ggml_tensor * ggml_cpy(
2960
        struct ggml_context * ctx,
2961
2962
2963
        struct ggml_tensor * a,
        struct ggml_tensor * b) {
    return ggml_cpy_impl(ctx, a, b);
2964
2965
}

2966
struct ggml_tensor * ggml_cast(
2967
        struct ggml_context * ctx,
2968
        struct ggml_tensor  * a,
2969
2970
2971
        enum   ggml_type      type) {
    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
    ggml_format_name(result, "%s (copy)", a->name);
2972

2973
    result->op     = GGML_OP_CPY;
2974
    result->src[0] = a;
2975
    result->src[1] = result;
2976
2977
2978
2979

    return result;
}

2980
// ggml_cont
2981

2982
static struct ggml_tensor * ggml_cont_impl(
2983
2984
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
2985
2986
    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
    ggml_format_name(result, "%s (cont)", a->name);
2987

2988
    result->op     = GGML_OP_CONT;
2989
2990
2991
2992
2993
    result->src[0] = a;

    return result;
}

2994
struct ggml_tensor * ggml_cont(
2995
        struct ggml_context * ctx,
2996
2997
        struct ggml_tensor * a) {
    return ggml_cont_impl(ctx, a);
2998
2999
}

3000
3001
// make contiguous, with new shape
GGML_API struct ggml_tensor * ggml_cont_1d(
3002
3003
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
3004
3005
        int64_t               ne0) {
    return ggml_cont_4d(ctx, a, ne0, 1, 1, 1);
3006
3007
}

3008
GGML_API struct ggml_tensor * ggml_cont_2d(
3009
        struct ggml_context * ctx,
3010
3011
3012
3013
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1) {
    return ggml_cont_4d(ctx, a, ne0, ne1, 1, 1);
3014
3015
}

3016
GGML_API struct ggml_tensor * ggml_cont_3d(
3017
        struct ggml_context * ctx,
3018
3019
3020
3021
3022
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2) {
    return ggml_cont_4d(ctx, a, ne0, ne1, ne2, 1);
3023
3024
}

3025
struct ggml_tensor * ggml_cont_4d(
3026
3027
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
3028
3029
3030
3031
3032
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2,
        int64_t               ne3) {
    GGML_ASSERT(ggml_nelements(a) == (ne0*ne1*ne2*ne3));
3033

3034
3035
3036
3037
    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
    ggml_format_name(result, "%s (cont)", a->name);

    result->op     = GGML_OP_CONT;
3038
3039
3040
3041
3042
    result->src[0] = a;

    return result;
}

3043
// ggml_reshape
3044

3045
struct ggml_tensor * ggml_reshape(
3046
        struct ggml_context * ctx,
3047
3048
3049
3050
3051
        struct ggml_tensor * a,
        struct ggml_tensor * b) {
    GGML_ASSERT(ggml_is_contiguous(a));
    // as only the shape of b is relevant, and not its memory layout, b is allowed to be non contiguous.
    GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
3052

3053
3054
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
    ggml_format_name(result, "%s (reshaped)", a->name);
3055

3056
    result->op     = GGML_OP_RESHAPE;
3057
3058
3059
3060
3061
    result->src[0] = a;

    return result;
}

3062
struct ggml_tensor * ggml_reshape_1d(
3063
        struct ggml_context * ctx,
3064
3065
3066
3067
        struct ggml_tensor  * a,
        int64_t               ne0) {
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(ggml_nelements(a) == ne0);
3068

3069
3070
3071
    const int64_t ne[1] = { ne0 };
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0);
    ggml_format_name(result, "%s (reshaped)", a->name);
3072

3073
    result->op     = GGML_OP_RESHAPE;
3074
3075
3076
3077
3078
    result->src[0] = a;

    return result;
}

3079
struct ggml_tensor * ggml_reshape_2d(
3080
        struct ggml_context * ctx,
3081
3082
3083
3084
3085
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1) {
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(ggml_nelements(a) == ne0*ne1);
3086

3087
3088
3089
    const int64_t ne[2] = { ne0, ne1 };
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0);
    ggml_format_name(result, "%s (reshaped)", a->name);
3090

3091
    result->op     = GGML_OP_RESHAPE;
3092
3093
3094
3095
3096
    result->src[0] = a;

    return result;
}

3097
struct ggml_tensor * ggml_reshape_3d(
3098
        struct ggml_context * ctx,
3099
3100
3101
3102
3103
3104
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2) {
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2);
3105

3106
3107
3108
3109
3110
    const int64_t ne[3] = { ne0, ne1, ne2 };
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0);
    ggml_format_name(result, "%s (reshaped)", a->name);

    result->op     = GGML_OP_RESHAPE;
3111
3112
3113
3114
3115
    result->src[0] = a;

    return result;
}

3116
struct ggml_tensor * ggml_reshape_4d(
3117
        struct ggml_context * ctx,
3118
3119
3120
3121
3122
3123
3124
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2,
        int64_t               ne3) {
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3);
3125

3126
3127
3128
    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0);
    ggml_format_name(result, "%s (reshaped)", a->name);
3129

3130
    result->op     = GGML_OP_RESHAPE;
3131
3132
3133
3134
3135
    result->src[0] = a;

    return result;
}

3136
static struct ggml_tensor * ggml_view_impl(
3137
        struct ggml_context * ctx,
3138
        struct ggml_tensor  * a,
3139
3140
3141
3142
3143
        int                   n_dims,
        const int64_t       * ne,
        size_t                offset) {
    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset);
    ggml_format_name(result, "%s (view)", a->name);
3144

3145
    ggml_set_op_params(result, &offset, sizeof(offset));
3146

3147
    result->op     = GGML_OP_VIEW;
3148
3149
3150
3151
3152
    result->src[0] = a;

    return result;
}

3153
// ggml_view_1d
3154

3155
struct ggml_tensor * ggml_view_1d(
3156
        struct ggml_context * ctx,
3157
        struct ggml_tensor  * a,
3158
3159
3160
        int64_t               ne0,
        size_t                offset) {
    struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset);
3161
3162
3163
3164

    return result;
}

3165
// ggml_view_2d
3166

3167
3168
3169
3170
3171
3172
3173
3174
struct ggml_tensor * ggml_view_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        size_t                nb1,
        size_t                offset) {
    const int64_t ne[2] = { ne0, ne1 };
3175

3176
    struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset);
3177

3178
3179
3180
    result->nb[1] = nb1;
    result->nb[2] = result->nb[1]*ne1;
    result->nb[3] = result->nb[2];
3181
3182
3183
3184

    return result;
}

3185
// ggml_view_3d
3186

3187
struct ggml_tensor * ggml_view_3d(
3188
        struct ggml_context * ctx,
3189
3190
3191
3192
3193
3194
3195
3196
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2,
        size_t                nb1,
        size_t                nb2,
        size_t                offset) {
    const int64_t ne[3] = { ne0, ne1, ne2 };
3197

3198
    struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset);
3199

3200
3201
3202
    result->nb[1] = nb1;
    result->nb[2] = nb2;
    result->nb[3] = result->nb[2]*ne2;
3203

3204
    return result;
3205
3206
}

3207
// ggml_view_4d
3208

3209
struct ggml_tensor * ggml_view_4d(
3210
        struct ggml_context * ctx,
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2,
        int64_t               ne3,
        size_t                nb1,
        size_t                nb2,
        size_t                nb3,
        size_t                offset) {
    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
3221

3222
    struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset);
3223

3224
3225
3226
    result->nb[1] = nb1;
    result->nb[2] = nb2;
    result->nb[3] = nb3;
3227

3228
    return result;
3229
3230
}

3231
// ggml_permute
3232

3233
struct ggml_tensor * ggml_permute(
3234
        struct ggml_context * ctx,
3235
3236
3237
3238
3239
3240
3241
3242
3243
        struct ggml_tensor  * a,
        int                   axis0,
        int                   axis1,
        int                   axis2,
        int                   axis3) {
    GGML_ASSERT(axis0 >= 0 && axis0 < GGML_MAX_DIMS);
    GGML_ASSERT(axis1 >= 0 && axis1 < GGML_MAX_DIMS);
    GGML_ASSERT(axis2 >= 0 && axis2 < GGML_MAX_DIMS);
    GGML_ASSERT(axis3 >= 0 && axis3 < GGML_MAX_DIMS);
3244

3245
3246
3247
3248
3249
3250
    GGML_ASSERT(axis0 != axis1);
    GGML_ASSERT(axis0 != axis2);
    GGML_ASSERT(axis0 != axis3);
    GGML_ASSERT(axis1 != axis2);
    GGML_ASSERT(axis1 != axis3);
    GGML_ASSERT(axis2 != axis3);
3251

3252
3253
    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
    ggml_format_name(result, "%s (permuted)", a->name);
3254

3255
3256
    int ne[GGML_MAX_DIMS];
    int nb[GGML_MAX_DIMS];
3257

3258
3259
3260
3261
    ne[axis0] = a->ne[0];
    ne[axis1] = a->ne[1];
    ne[axis2] = a->ne[2];
    ne[axis3] = a->ne[3];
3262

3263
3264
3265
3266
    nb[axis0] = a->nb[0];
    nb[axis1] = a->nb[1];
    nb[axis2] = a->nb[2];
    nb[axis3] = a->nb[3];
3267

3268
3269
3270
3271
    result->ne[0] = ne[0];
    result->ne[1] = ne[1];
    result->ne[2] = ne[2];
    result->ne[3] = ne[3];
3272

3273
3274
3275
3276
    result->nb[0] = nb[0];
    result->nb[1] = nb[1];
    result->nb[2] = nb[2];
    result->nb[3] = nb[3];
3277

3278
    result->op     = GGML_OP_PERMUTE;
3279
3280
    result->src[0] = a;

3281
3282
    int32_t params[] = { axis0, axis1, axis2, axis3 };
    ggml_set_op_params(result, params, sizeof(params));
3283

3284
    return result;
3285
3286
}

3287
// ggml_transpose
3288

3289
struct ggml_tensor * ggml_transpose(
3290
3291
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
3292
3293
    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
    ggml_format_name(result, "%s (transposed)", a->name);
3294

3295
3296
    result->ne[0] = a->ne[1];
    result->ne[1] = a->ne[0];
3297

3298
3299
    result->nb[0] = a->nb[1];
    result->nb[1] = a->nb[0];
3300

3301
3302
    result->op     = GGML_OP_TRANSPOSE;
    result->src[0] = a;
3303

3304
    return result;
3305
3306
}

3307
// ggml_get_rows
3308

3309
3310
3311
3312
3313
3314
3315
struct ggml_tensor * ggml_get_rows(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(a->ne[2] == b->ne[1]);
    GGML_ASSERT(b->ne[3] == 1);
    GGML_ASSERT(b->type == GGML_TYPE_I32);
3316

3317
3318
3319
3320
    // TODO: implement non F32 return
    enum ggml_type type = GGML_TYPE_F32;
    if (a->type == GGML_TYPE_I32) {
        type = a->type;
3321
    }
3322
    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
3323

3324
3325
3326
    result->op     = GGML_OP_GET_ROWS;
    result->src[0] = a;
    result->src[1] = b;
3327

3328
3329
    return result;
}
3330

3331
// ggml_get_rows_back
3332

3333
3334
3335
3336
3337
3338
3339
struct ggml_tensor * ggml_get_rows_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c) {
    GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32);
    GGML_ASSERT(ggml_is_matrix(c) && (a->ne[0] == c->ne[0]));
3340

3341
3342
3343
3344
3345
3346
3347
3348
3349
    // TODO: implement non F32 return
    //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
    struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, c->ne[0], c->ne[1]);

    result->op     = GGML_OP_GET_ROWS_BACK;
    result->src[0] = a;
    result->src[1] = b;

    return result;
3350
3351
}

3352
// ggml_diag
3353

3354
3355
3356
3357
struct ggml_tensor * ggml_diag(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    GGML_ASSERT(a->ne[1] == 1);
3358

3359
3360
    const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne);
3361

3362
3363
    result->op     = GGML_OP_DIAG;
    result->src[0] = a;
3364

3365
    return result;
3366
3367
}

3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
// ggml_diag_mask_inf

static struct ggml_tensor * ggml_diag_mask_inf_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    int32_t params[] = { n_past };
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_DIAG_MASK_INF;
    result->src[0] = a;

    return result;
3384
3385
}

3386
3387
3388
3389
3390
3391
struct ggml_tensor * ggml_diag_mask_inf(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past) {
    return ggml_diag_mask_inf_impl(ctx, a, n_past, false);
}
3392

3393
3394
3395
3396
3397
3398
struct ggml_tensor * ggml_diag_mask_inf_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past) {
    return ggml_diag_mask_inf_impl(ctx, a, n_past, true);
}
3399

3400
// ggml_diag_mask_zero
3401

3402
3403
3404
3405
3406
3407
static struct ggml_tensor * ggml_diag_mask_zero_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
3408

3409
3410
    int32_t params[] = { n_past };
    ggml_set_op_params(result, params, sizeof(params));
3411

3412
3413
    result->op     = GGML_OP_DIAG_MASK_ZERO;
    result->src[0] = a;
3414

3415
3416
    return result;
}
3417

3418
3419
3420
3421
3422
3423
struct ggml_tensor * ggml_diag_mask_zero(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past) {
    return ggml_diag_mask_zero_impl(ctx, a, n_past, false);
}
3424

3425
3426
3427
3428
3429
3430
struct ggml_tensor * ggml_diag_mask_zero_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   n_past) {
    return ggml_diag_mask_zero_impl(ctx, a, n_past, true);
}
3431

3432
// ggml_soft_max
3433

3434
3435
3436
3437
3438
3439
3440
3441
static struct ggml_tensor * ggml_soft_max_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * mask,
        float                 scale,
        float                 max_bias,
        bool                  inplace) {
    GGML_ASSERT(ggml_is_contiguous(a));
3442

3443
3444
3445
3446
3447
3448
    if (mask) {
        GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
        GGML_ASSERT(ggml_is_contiguous(mask));
        GGML_ASSERT(ggml_is_matrix(mask));
        GGML_ASSERT(mask->ne[0] == a->ne[0]);
        GGML_ASSERT(mask->ne[1] >= a->ne[1]);
3449
3450
    }

3451
3452
    if (max_bias > 0.0f) {
        GGML_ASSERT(mask);
3453
3454
    }

3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

    float params[] = { scale, max_bias };
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_SOFT_MAX;
    result->src[0] = a;
    result->src[1] = mask;

    return result;
}

struct ggml_tensor * ggml_soft_max(
        struct ggml_context * ctx,
        struct ggml_tensor  * a) {
    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, false);
3471
3472
}

3473
struct ggml_tensor * ggml_soft_max_inplace(
3474
        struct ggml_context * ctx,
3475
3476
3477
        struct ggml_tensor  * a) {
    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, true);
}
3478

3479
3480
3481
3482
3483
3484
3485
struct ggml_tensor * ggml_soft_max_ext(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * mask,
        float                 scale,
        float                 max_bias) {
    return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
3486
3487
}

3488
// ggml_soft_max_back
3489

3490
3491
3492
3493
3494
3495
static struct ggml_tensor * ggml_soft_max_back_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        bool                  inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
3496

3497
3498
3499
    result->op     = GGML_OP_SOFT_MAX_BACK;
    result->src[0] = a;
    result->src[1] = b;
3500

3501
    return result;
3502
3503
}

3504
3505
3506
3507
3508
struct ggml_tensor * ggml_soft_max_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_soft_max_back_impl(ctx, a, b, false);
3509
3510
}

3511
3512
3513
3514
3515
struct ggml_tensor * ggml_soft_max_back_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_soft_max_back_impl(ctx, a, b, true);
3516
3517
}

3518
// ggml_rope
3519

3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
static struct ggml_tensor * ggml_rope_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow,
        bool                  inplace) {
    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
3536

3537
3538
3539
    GGML_ASSERT(ggml_is_vector(b));
    GGML_ASSERT(b->type == GGML_TYPE_I32);
    GGML_ASSERT(a->ne[2] == b->ne[0]);
3540

3541
3542
3543
3544
    if (c) {
        GGML_ASSERT(c->type == GGML_TYPE_F32);
        GGML_ASSERT(c->ne[0] >= n_dims / 2);
    }
3545

3546
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
3547

3548
3549
3550
3551
3552
3553
3554
3555
    int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
    memcpy(params +  5, &freq_base,    sizeof(float));
    memcpy(params +  6, &freq_scale,   sizeof(float));
    memcpy(params +  7, &ext_factor,   sizeof(float));
    memcpy(params +  8, &attn_factor,  sizeof(float));
    memcpy(params +  9, &beta_fast,    sizeof(float));
    memcpy(params + 10, &beta_slow,    sizeof(float));
    ggml_set_op_params(result, params, sizeof(params));
3556

3557
3558
3559
3560
    result->op     = GGML_OP_ROPE;
    result->src[0] = a;
    result->src[1] = b;
    result->src[2] = c;
3561

3562
    return result;
3563
3564
}

3565
3566
3567
3568
3569
3570
3571
3572
3573
struct ggml_tensor * ggml_rope(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   n_dims,
        int                   mode) {
    return ggml_rope_impl(
        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
    );
3574
3575
}

3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
struct ggml_tensor * ggml_rope_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   n_dims,
        int                   mode) {
    return ggml_rope_impl(
        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
    );
}
3586

3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
struct ggml_tensor * ggml_rope_ext(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow) {
    return ggml_rope_impl(
        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
        ext_factor, attn_factor, beta_fast, beta_slow, false
    );
3605
3606
}

3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
struct ggml_tensor * ggml_rope_ext_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow) {
    return ggml_rope_impl(
        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
        ext_factor, attn_factor, beta_fast, beta_slow, true
    );
}
3626

3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
struct ggml_tensor * ggml_rope_custom(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow) {
    return ggml_rope_impl(
        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
        ext_factor, attn_factor, beta_fast, beta_slow, false
    );
}
3645

3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
struct ggml_tensor * ggml_rope_custom_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow) {
    return ggml_rope_impl(
        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
        ext_factor, attn_factor, beta_fast, beta_slow, true
    );
}

// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
}

void ggml_rope_yarn_corr_dims(
    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
) {
    // start and end correction dims
    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
    float end   =  ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
    dims[0] = MAX(0, start);
    dims[1] = MIN(n_dims - 1, end);
}

// ggml_rope_back

struct ggml_tensor * ggml_rope_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c,
        int                   n_dims,
        int                   mode,
        int                   n_ctx_orig,
        float                 freq_base,
        float                 freq_scale,
        float                 ext_factor,
        float                 attn_factor,
        float                 beta_fast,
        float                 beta_slow) {
    GGML_ASSERT(ggml_is_vector(b));
    GGML_ASSERT(b->type == GGML_TYPE_I32);
    GGML_ASSERT(a->ne[2] == b->ne[0]);
3700

3701
    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
3702

3703
3704
3705
3706
3707
3708
3709
3710
    int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
    memcpy(params +  5, &freq_base,    sizeof(float));
    memcpy(params +  6, &freq_scale,   sizeof(float));
    memcpy(params +  7, &ext_factor,   sizeof(float));
    memcpy(params +  8, &attn_factor,  sizeof(float));
    memcpy(params +  9, &beta_fast,    sizeof(float));
    memcpy(params + 10, &beta_slow,    sizeof(float));
    ggml_set_op_params(result, params, sizeof(params));
3711

3712
3713
3714
3715
    result->op     = GGML_OP_ROPE_BACK;
    result->src[0] = a;
    result->src[1] = b;
    result->src[2] = c;
3716
3717
3718
3719

    return result;
}

3720
// ggml_clamp
3721

3722
3723
3724
3725
3726
3727
3728
struct ggml_tensor * ggml_clamp(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        float                 min,
        float                 max) {
    // TODO: when implement backward, fix this:
    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
3729

3730
3731
    float params[] = { min, max };
    ggml_set_op_params(result, params, sizeof(params));
3732

3733
3734
    result->op     = GGML_OP_CLAMP;
    result->src[0] = a;
3735

3736
    return result;
3737
3738
}

3739
// ggml_conv_1d
3740

3741
3742
static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
    return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
3743
3744
}

3745
3746
3747
3748
3749
3750
3751
3752
GGML_API struct ggml_tensor * ggml_conv_1d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s0,
        int                   p0,
        int                   d0) {
    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16); // [N, OL, IC * K]
3753

3754
3755
3756
3757
    struct ggml_tensor * result =
        ggml_mul_mat(ctx,
                ggml_reshape_2d(ctx, im2col, im2col->ne[0], (im2col->ne[2] * im2col->ne[1])), // [N, OL, IC * K] => [N*OL, IC * K]
                ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1]), a->ne[2]));                    // [OC,IC, K] => [OC, IC * K]
3758

3759
    result = ggml_reshape_3d(ctx, result, im2col->ne[1], a->ne[2], im2col->ne[2]); // [N, OC, OL]
3760

3761
    return result;
3762
3763
}

3764
3765
3766
3767
3768
3769
3770
3771
3772
// ggml_conv_1d_ph

struct ggml_tensor* ggml_conv_1d_ph(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s,
        int                   d) {
    return ggml_conv_1d(ctx, a, b, s, a->ne[0] / 2, d);
3773
3774
}

3775
// ggml_conv_transpose_1d
3776

3777
3778
3779
static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
}
3780

3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s0,
        int                   p0,
        int                   d0) {
    GGML_ASSERT(ggml_is_matrix(b));
    GGML_ASSERT(a->ne[2] == b->ne[1]);
    GGML_ASSERT(a->ne[3] == 1);
3791

3792
3793
    GGML_ASSERT(p0 == 0);
    GGML_ASSERT(d0 == 1);
3794

3795
3796
3797
3798
3799
    const int64_t ne[4] = {
        ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
        a->ne[1], b->ne[2], 1,
    };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
3800

3801
3802
    int32_t params[] = { s0, p0, d0 };
    ggml_set_op_params(result, params, sizeof(params));
3803

3804
3805
3806
    result->op     = GGML_OP_CONV_TRANSPOSE_1D;
    result->src[0] = a;
    result->src[1] = b;
3807

3808
3809
    return result;
}
3810

3811
// ggml_conv_depthwise
3812

3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
struct ggml_tensor * ggml_conv_depthwise_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s0,
        int                   s1,
        int                   p0,
        int                   p1,
        int                   d0,
        int                   d1) {
    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a,
                                        ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
                                        s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
    struct ggml_tensor * new_b = ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
3828

3829
3830
3831
    new_a = ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC,1, KH, KW] => [1, OC, 1, KH * KW]
    struct ggml_tensor * result = ggml_mul_mat(ctx, new_a, new_b);
    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
3832

3833
    return result;
3834
}
3835
// ggml_conv_2d
3836

3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
// a: [OC,IC, KH, KW]
// b: [N, IC, IH, IW]
// result: [N, OH, OW, IC*KH*KW]
struct ggml_tensor * ggml_im2col(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s0,
        int                   s1,
        int                   p0,
        int                   p1,
        int                   d0,
        int                   d1,
        bool                  is_2D,
        enum ggml_type        dst_type) {
    if(is_2D) {
        GGML_ASSERT(a->ne[2] == b->ne[2]);
    } else {
        GGML_ASSERT(a->ne[1] == b->ne[1]);
        GGML_ASSERT(b->ne[3] == 1);
3858
3859
    }

3860
3861
    const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
    const int64_t OW =         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
3862

3863
3864
    GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
    GGML_ASSERT((OW > 0)           && "b too small compared to a");
3865

3866
3867
3868
3869
3870
3871
    const int64_t ne[4] = {
        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
        OW,
        is_2D ? OH : b->ne[2],
        is_2D ?      b->ne[3] : 1,
    };
3872

3873
3874
3875
    struct ggml_tensor * result = ggml_new_tensor(ctx, dst_type, 4, ne);
    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
    ggml_set_op_params(result, params, sizeof(params));
3876

3877
3878
3879
3880
3881
3882
    result->op     = GGML_OP_IM2COL;
    result->src[0] = a;
    result->src[1] = b;

    return result;
}
3883

3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
struct ggml_tensor * ggml_im2col_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int64_t             * ne,
        int                   s0,
        int                   s1,
        int                   p0,
        int                   p1,
        int                   d0,
        int                   d1,
        bool                  is_2D) {
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
    ggml_set_op_params(result, params, sizeof(params));
3899

3900
3901
3902
    result->op     = GGML_OP_IM2COL_BACK;
    result->src[0] = a;
    result->src[1] = b;
3903

3904
    return result;
3905
3906
}

3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
// a: [OC,IC, KH, KW]
// b: [N, IC, IH, IW]
// result: [N, OC, OH, OW]
struct ggml_tensor * ggml_conv_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   s0,
        int                   s1,
        int                   p0,
        int                   p1,
        int                   d0,
        int                   d1) {
    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW]
3921

3922
3923
3924
3925
    struct ggml_tensor * result =
        ggml_mul_mat(ctx,
                ggml_reshape_2d(ctx, im2col, im2col->ne[0],  im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW]
                ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]),  a->ne[3]));                       // [OC,IC, KH, KW] => [OC, IC * KH * KW]
3926

3927
3928
    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], im2col->ne[3], a->ne[3]); // [OC, N, OH, OW]
    result = ggml_cont(ctx, ggml_permute(ctx, result, 0, 1, 3, 2)); // [N, OC, OH, OW]
3929

3930
3931

    return result;
3932
3933
}

3934
// ggml_conv_2d_sk_p0
3935

3936
3937
3938
3939
3940
struct ggml_tensor * ggml_conv_2d_sk_p0(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_conv_2d(ctx, a, b, a->ne[0], a->ne[1], 0, 0, 1, 1);
3941
3942
}

3943
// ggml_conv_2d_s1_ph
3944

3945
3946
3947
3948
3949
3950
struct ggml_tensor * ggml_conv_2d_s1_ph(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    return ggml_conv_2d(ctx, a, b, 1, 1, a->ne[0] / 2, a->ne[1] / 2, 1, 1);
}
3951

3952
// ggml_conv_transpose_2d_p0
3953

3954
3955
static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
    return (ins - 1) * s - 2 * p + ks;
3956
3957
}

3958
3959
3960
3961
3962
3963
struct ggml_tensor * ggml_conv_transpose_2d_p0(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        int                   stride) {
    GGML_ASSERT(a->ne[3] == b->ne[2]);
3964

3965
3966
3967
3968
3969
    const int64_t ne[4] = {
        ggml_calc_conv_transpose_output_size(b->ne[0], a->ne[0], stride, 0 /*p0*/),
        ggml_calc_conv_transpose_output_size(b->ne[1], a->ne[1], stride, 0 /*p1*/),
        a->ne[2], b->ne[3],
    };
3970

3971
    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
3972

3973
    ggml_set_op_params_i32(result, 0, stride);
3974

3975
3976
3977
    result->op     = GGML_OP_CONV_TRANSPOSE_2D;
    result->src[0] = a;
    result->src[1] = b;
3978

3979
    return result;
3980
3981
}

3982
// ggml_pool_*
3983

3984
3985
3986
static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, float p) {
    return (ins + 2 * p - ks) / s + 1;
}
3987

3988
// ggml_pool_1d
3989

3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
struct ggml_tensor * ggml_pool_1d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        enum ggml_op_pool     op,
        int                   k0,
        int                   s0,
        int                   p0) {
    const int64_t ne[4] = {
        ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
        a->ne[1],
        a->ne[2],
        a->ne[3],
    };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
4004

4005
4006
    int32_t params[] = { op, k0, s0, p0 };
    ggml_set_op_params(result, params, sizeof(params));
4007

4008
4009
    result->op     = GGML_OP_POOL_1D;
    result->src[0] = a;
4010

4011
    return result;
4012
4013
}

4014
// ggml_pool_2d
4015

4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
struct ggml_tensor * ggml_pool_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        enum ggml_op_pool     op,
        int                   k0,
        int                   k1,
        int                   s0,
        int                   s1,
        float                 p0,
        float                 p1) {
    struct ggml_tensor * result;
    const int64_t ne[4] = {
        ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
        ggml_calc_pool_output_size(a->ne[1], k1, s1, p1),
        a->ne[2],
        a->ne[3],
    };
    result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
4034

4035
4036
4037
4038
4039
4040
4041
    int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_POOL_2D;
    result->src[0] = a;

    return result;
4042
4043
}

4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
struct ggml_tensor * ggml_pool_2d_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * af,
        enum ggml_op_pool     op,
        int                   k0,
        int                   k1,
        int                   s0,
        int                   s1,
        float                 p0,
        float                 p1) {
    struct ggml_tensor * result;
    result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, af->ne);

    int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_POOL_2D_BACK;
    result->src[0] = a;
    result->src[1] = af;

    return result;
}
4067

4068
// ggml_upscale
4069

4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
static struct ggml_tensor * ggml_upscale_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   ne0,
        int                   ne1,
        int                   ne2,
        int                   ne3) {
    GGML_ASSERT(a->ne[0] <= ne0);
    GGML_ASSERT(a->ne[1] <= ne1);
    GGML_ASSERT(a->ne[2] <= ne2);
    GGML_ASSERT(a->ne[3] <= ne3);
4081

4082
    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
4083

4084
4085
    result->op     = GGML_OP_UPSCALE;
    result->src[0] = a;
4086

4087
4088
    return result;
}
4089

4090
4091
4092
4093
4094
4095
struct ggml_tensor * ggml_upscale(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   scale_factor) {
    return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3]);
}
4096

4097
4098
4099
4100
4101
4102
4103
4104
struct ggml_tensor * ggml_upscale_ext(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   ne0,
        int                   ne1,
        int                   ne2,
        int                   ne3) {
    return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3);
4105
4106
}

4107
// ggml_pad
4108

4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
struct ggml_tensor * ggml_pad(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   p0,
        int                   p1,
        int                   p2,
        int                   p3) {
    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
            a->ne[0] + p0,
            a->ne[1] + p1,
            a->ne[2] + p2,
            a->ne[3] + p3);
4121

4122
4123
    result->op     = GGML_OP_PAD;
    result->src[0] = a;
4124

4125
    return result;
4126
4127
}

4128
// ggml_unpad
4129

4130
4131
4132
4133
struct ggml_tensor * ggml_unpad(
    struct ggml_context * ctx,
    struct ggml_tensor  * a,
    int p0, int p1, int p2, int p3) {
4134

4135
4136
4137
4138
4139
    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
            a->ne[0] - p0,
            a->ne[1] - p1,
            a->ne[2] - p2,
            a->ne[3] - p3);
4140

4141
4142
    result->op = GGML_OP_UNPAD;
    result->src[0] = a;
4143

4144
4145
    return result;
}
4146

4147
// ggml_arange
4148

4149
4150
4151
4152
4153
4154
struct ggml_tensor * ggml_arange(
        struct ggml_context * ctx,
        float                 start,
        float                 stop,
        float                 step) {
    GGML_ASSERT(stop > start);
4155

4156
    const int64_t steps = (int64_t) ceilf((stop - start) / step);
4157

4158
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
4159

4160
4161
4162
    ggml_set_op_params_f32(result, 0, start);
    ggml_set_op_params_f32(result, 1, stop);
    ggml_set_op_params_f32(result, 2, step);
4163

4164
    result->op = GGML_OP_ARANGE;
4165

4166
4167
    return result;
}
4168

4169
// ggml_timestep_embedding
4170

4171
4172
4173
4174
4175
4176
4177
4178
struct ggml_tensor * ggml_timestep_embedding(
        struct ggml_context * ctx,
        struct ggml_tensor  * timesteps,
        int                   dim,
        int                   max_period) {
    int actual_dim = dim;
    if (dim % 2 != 0) {
        actual_dim = dim + 1;
4179
4180
    }

4181
    struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, actual_dim, timesteps->ne[0]);
4182

4183
4184
    ggml_set_op_params_i32(result, 0, dim);
    ggml_set_op_params_i32(result, 1, max_period);
4185

4186
4187
    result->op     = GGML_OP_TIMESTEP_EMBEDDING;
    result->src[0] = timesteps;
4188

4189
4190
    return result;
}
4191

4192
// ggml_argsort
4193

4194
4195
4196
4197
4198
4199
struct ggml_tensor * ggml_argsort(
        struct ggml_context  * ctx,
        struct ggml_tensor   * a,
        enum ggml_sort_order   order) {
    GGML_ASSERT(a->ne[0] <= INT32_MAX);
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
4200

4201
    ggml_set_op_params_i32(result, 0, (int32_t) order);
4202

4203
4204
    result->op     = GGML_OP_ARGSORT;
    result->src[0] = a;
4205

4206
4207
    return result;
}
4208

4209
// ggml_top_k
4210

4211
4212
4213
4214
4215
struct ggml_tensor * ggml_top_k(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   k) {
    GGML_ASSERT(a->ne[0] >= k);
4216

4217
    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
4218

4219
4220
4221
4222
    result = ggml_view_4d(ctx, result,
                k, result->ne[1], result->ne[2], result->ne[3],
                   result->nb[1], result->nb[2], result->nb[3],
                0);
4223

4224
    return result;
4225
4226
}

4227
// ggml_flash_attn_ext
4228

4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
struct ggml_tensor * ggml_flash_attn_ext(
        struct ggml_context * ctx,
        struct ggml_tensor  * q,
        struct ggml_tensor  * k,
        struct ggml_tensor  * v,
        struct ggml_tensor  * mask,
        float                 scale,
        float                 max_bias,
        float                 logit_softcap) {
    GGML_ASSERT(ggml_can_mul_mat(k, q));
    // TODO: check if vT can be multiplied by (k*qT)
4240

4241
4242
4243
4244
4245
4246
4247
    if (mask) {
        GGML_ASSERT(ggml_is_contiguous(mask));
        GGML_ASSERT(mask->ne[2] == 1);
        GGML_ASSERT(mask->ne[3] == 1);
        GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
        //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
4248
4249
    }

4250
4251
    if (max_bias > 0.0f) {
        GGML_ASSERT(mask);
4252
4253
    }

4254
4255
4256
4257
4258
4259
    // permute(0, 2, 1, 3)
    int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);

    float params[] = { scale, max_bias, logit_softcap };
    ggml_set_op_params(result, params, sizeof(params));
4260

4261
4262
4263
4264
4265
    result->op     = GGML_OP_FLASH_ATTN_EXT;
    result->src[0] = q;
    result->src[1] = k;
    result->src[2] = v;
    result->src[3] = mask;
4266

4267
    return result;
4268
4269
}

4270
4271
4272
4273
void ggml_flash_attn_ext_set_prec(
        struct ggml_tensor * a,
        enum ggml_prec       prec) {
    GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT);
4274

4275
    const int32_t prec_i32 = (int32_t) prec;
4276

4277
    ggml_set_op_params_i32(a, 3, prec_i32); // scale is on first pos, max_bias on second
4278
4279
}

4280
4281
4282
enum ggml_prec ggml_flash_attn_ext_get_prec(
        const struct ggml_tensor * a) {
    GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT);
4283

4284
    const int32_t prec_i32 = ggml_get_op_params_i32(a, 3);
4285

4286
4287
    return (enum ggml_prec) prec_i32;
}
4288

4289
// ggml_flash_attn_back
4290

4291
4292
4293
4294
4295
4296
4297
4298
struct ggml_tensor * ggml_flash_attn_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * q,
        struct ggml_tensor  * k,
        struct ggml_tensor  * v,
        struct ggml_tensor  * d,
        bool                  masked) {
    GGML_ABORT("TODO: adapt to ggml_flash_attn_ext() changes");
4299

4300
4301
    GGML_ASSERT(ggml_can_mul_mat(k, q));
    // TODO: check if vT can be multiplied by (k*qT)
4302

4303
4304
4305
4306
    // d shape [D,N,ne2,ne3]
    // q shape [D,N,ne2,ne3]
    // k shape [D,M,kvne2,ne3]
    // v shape [M,D,kvne2,ne3]
4307

4308
4309
4310
4311
4312
4313
    const int64_t     D = q->ne[0];
    const int64_t     N = q->ne[1];
    const int64_t     M = k->ne[1];
    const int64_t   ne2 = q->ne[2];
    const int64_t   ne3 = q->ne[3];
    const int64_t kvne2 = k->ne[2];
4314

4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
    GGML_ASSERT(k->ne[0] == D);
    GGML_ASSERT(v->ne[0] == M);
    GGML_ASSERT(v->ne[1] == D);
    GGML_ASSERT(d->ne[0] == D);
    GGML_ASSERT(d->ne[1] == N);
    GGML_ASSERT(k->ne[2] == kvne2);
    GGML_ASSERT(k->ne[3] == ne3);
    GGML_ASSERT(v->ne[2] == kvne2);
    GGML_ASSERT(v->ne[3] == ne3);
    GGML_ASSERT(d->ne[2] == ne2);
    GGML_ASSERT(d->ne[3] == ne3);
4326

4327
    GGML_ASSERT(ne2 % kvne2 == 0);
4328

4329
4330
4331
4332
4333
    // store gradients of q, k and v as continuous tensors concatenated in result.
    // note: v and gradv are actually transposed, i.e. v->ne[0] != D.
    const int64_t elem_q = ggml_nelements(q);
    const int64_t elem_k = ggml_nelements(k);
    const int64_t elem_v = ggml_nelements(v);
4334

4335
4336
4337
    enum ggml_type result_type = GGML_TYPE_F32;
    GGML_ASSERT(ggml_blck_size(result_type) == 1);
    const size_t tsize = ggml_type_size(result_type);
4338

4339
4340
4341
4342
    const size_t offs_q = 0;
    const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
    const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
    const size_t end    = offs_v + GGML_PAD(elem_v * tsize, GGML_MEM_ALIGN);
4343

4344
    const size_t nelements = (end + tsize - 1)/tsize;
4345

4346
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nelements);
4347

4348
4349
    int32_t masked_i = masked ? 1 : 0;
    ggml_set_op_params(result, &masked_i, sizeof(masked_i));
4350

4351
4352
4353
4354
4355
    result->op     = GGML_OP_FLASH_ATTN_BACK;
    result->src[0] = q;
    result->src[1] = k;
    result->src[2] = v;
    result->src[3] = d;
4356

4357
    return result;
4358
4359
}

4360
// ggml_ssm_conv
4361

4362
4363
4364
4365
4366
4367
struct ggml_tensor * ggml_ssm_conv(
        struct ggml_context * ctx,
        struct ggml_tensor  * sx,
        struct ggml_tensor  * c) {
    GGML_ASSERT(ggml_is_3d(sx));
    GGML_ASSERT(ggml_is_matrix(c));
4368

4369
4370
4371
4372
    const int64_t d_conv  = c->ne[0];
    const int64_t d_inner = c->ne[1];
    const int64_t n_t     = sx->ne[0] - d_conv + 1; // tokens per sequence
    const int64_t n_s     = sx->ne[2];
4373

4374
4375
4376
4377
4378
    // TODO: maybe support other strides than 1?
    // FIXME: this is always true?
    GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
    GGML_ASSERT(sx->ne[1] == d_inner);
    GGML_ASSERT(n_t >= 0);
4379

4380
    struct ggml_tensor * result = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_t, n_s);
4381

4382
4383
4384
    result->op     = GGML_OP_SSM_CONV;
    result->src[0] = sx;
    result->src[1] = c;
4385

4386
4387
    return result;
}
4388

4389
// ggml_ssm_scan
4390

4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
struct ggml_tensor * ggml_ssm_scan(
        struct ggml_context * ctx,
        struct ggml_tensor  * s,
        struct ggml_tensor  * x,
        struct ggml_tensor  * dt,
        struct ggml_tensor  * A,
        struct ggml_tensor  * B,
        struct ggml_tensor  * C) {
    GGML_ASSERT(ggml_is_contiguous(s));
    GGML_ASSERT(ggml_is_contiguous(x));
    GGML_ASSERT(ggml_is_contiguous(dt));
    GGML_ASSERT(ggml_is_contiguous(A));
    GGML_ASSERT(ggml_is_matrix(A));
    GGML_ASSERT(ggml_is_3d(B));
    GGML_ASSERT(ggml_is_3d(s));
    GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
    GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
    GGML_ASSERT(ggml_are_same_shape(x, dt));
    GGML_ASSERT(ggml_are_same_shape(B, C));
4410

4411
4412
4413
4414
4415
    {
        const int64_t d_state      = s->ne[0];
        const int64_t d_inner      = s->ne[1];
        const int64_t n_seq_tokens = x->ne[1];
        const int64_t n_seqs       = x->ne[2];
4416

4417
4418
4419
4420
4421
4422
4423
        GGML_ASSERT(s->ne[2] == n_seqs);
        GGML_ASSERT(x->ne[0] == d_inner);
        GGML_ASSERT(A->ne[0] == d_state);
        GGML_ASSERT(A->ne[1] == d_inner);
        GGML_ASSERT(B->ne[0] == d_state);
        GGML_ASSERT(B->ne[1] == n_seq_tokens);
        GGML_ASSERT(B->ne[2] == n_seqs);
4424
4425
    }

4426
4427
    // concatenated y + ssm_states
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
4428

4429
4430
4431
4432
4433
4434
4435
    result->op   = GGML_OP_SSM_SCAN;
    result->src[0] = s;
    result->src[1] = x;
    result->src[2] = dt;
    result->src[3] = A;
    result->src[4] = B;
    result->src[5] = C;
4436

4437
    return result;
4438
4439
}

4440
// ggml_win_part
4441

4442
4443
4444
4445
4446
4447
struct ggml_tensor * ggml_win_part(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   w) {
    GGML_ASSERT(a->ne[3] == 1);
    GGML_ASSERT(a->type  == GGML_TYPE_F32);
4448

4449
4450
4451
    // padding
    const int px = (w - a->ne[1]%w)%w;
    const int py = (w - a->ne[2]%w)%w;
4452

4453
4454
4455
    const int npx = (px + a->ne[1])/w;
    const int npy = (py + a->ne[2])/w;
    const int np  = npx*npy;
4456

4457
4458
    const int64_t ne[4] = { a->ne[0], w, w, np, };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
4459

4460
4461
    int32_t params[] = { npx, npy, w };
    ggml_set_op_params(result, params, sizeof(params));
4462

4463
4464
    result->op     = GGML_OP_WIN_PART;
    result->src[0] = a;
4465

4466
4467
    return result;
}
4468

4469
// ggml_win_unpart
4470

4471
4472
4473
4474
4475
4476
4477
struct ggml_tensor * ggml_win_unpart(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   w0,
        int                   h0,
        int                   w) {
    GGML_ASSERT(a->type == GGML_TYPE_F32);
4478

4479
4480
    const int64_t ne[4] = { a->ne[0], w0, h0, 1, };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
4481

4482
4483
    int32_t params[] = { w };
    ggml_set_op_params(result, params, sizeof(params));
4484

4485
4486
    result->op     = GGML_OP_WIN_UNPART;
    result->src[0] = a;
4487

4488
    return result;
4489
4490
}

4491
4492
4493
4494
4495
4496
4497
4498
4499
// ggml_get_rel_pos

struct ggml_tensor * ggml_get_rel_pos(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int                   qh,
        int                   kh) {
    GGML_ASSERT(qh == kh);
    GGML_ASSERT(2*MAX(qh, kh) - 1 == a->ne[1]);
4500

4501
4502
    const int64_t ne[4] = { a->ne[0], kh, qh, 1, };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 3, ne);
4503

4504
4505
    result->op     = GGML_OP_GET_REL_POS;
    result->src[0] = a;
4506

4507
    return result;
4508
4509
}

4510
// ggml_add_rel_pos
4511

4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
static struct ggml_tensor * ggml_add_rel_pos_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * pw,
        struct ggml_tensor  * ph,
        bool                  inplace) {
    GGML_ASSERT(ggml_are_same_shape(pw, ph));
    GGML_ASSERT(ggml_is_contiguous(a));
    GGML_ASSERT(ggml_is_contiguous(pw));
    GGML_ASSERT(ggml_is_contiguous(ph));
    GGML_ASSERT(ph->type == GGML_TYPE_F32);
    GGML_ASSERT(pw->type == GGML_TYPE_F32);
    GGML_ASSERT(pw->ne[3] == a->ne[2]);
    GGML_ASSERT(pw->ne[0]*pw->ne[0] == a->ne[0]);
    GGML_ASSERT(pw->ne[1]*pw->ne[2] == a->ne[1]);
4527

4528
4529
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
    ggml_set_op_params_i32(result, 0, inplace ? 1 : 0);
4530

4531
4532
4533
4534
    result->op     = GGML_OP_ADD_REL_POS;
    result->src[0] = a;
    result->src[1] = pw;
    result->src[2] = ph;
4535

4536
    return result;
4537
4538
}

4539
4540
4541
4542
4543
4544
struct ggml_tensor * ggml_add_rel_pos(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * pw,
        struct ggml_tensor  * ph) {
    return ggml_add_rel_pos_impl(ctx, a, pw, ph, false);
4545
4546
}

4547
4548
4549
4550
4551
4552
struct ggml_tensor * ggml_add_rel_pos_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * pw,
        struct ggml_tensor  * ph) {
    return ggml_add_rel_pos_impl(ctx, a, pw, ph, true);
4553
4554
}

4555
// ggml_rwkv_wkv6
4556

4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
struct ggml_tensor * ggml_rwkv_wkv6(
        struct ggml_context * ctx,
        struct ggml_tensor  * k,
        struct ggml_tensor  * v,
        struct ggml_tensor  * r,
        struct ggml_tensor  * tf,
        struct ggml_tensor  * td,
        struct ggml_tensor  * state) {
    GGML_ASSERT(ggml_is_contiguous(k));
    GGML_ASSERT(ggml_is_contiguous(v));
    GGML_ASSERT(ggml_is_contiguous(r));
    GGML_ASSERT(ggml_is_contiguous(tf));
    GGML_ASSERT(ggml_is_contiguous(td));
    GGML_ASSERT(ggml_is_contiguous(state));
4571

4572
4573
4574
4575
    const int64_t S = k->ne[0];
    const int64_t H = k->ne[2];
    const int64_t n_tokens = k->ne[3];
    const int64_t n_seqs = state->ne[1];
4576
    {
4577
4578
4579
4580
4581
4582
4583
        GGML_ASSERT(k->ne[1] == 1);
        GGML_ASSERT(v->ne[0] == 1 && v->ne[1] == S && v->ne[2] == H && v->ne[3] == n_tokens);
        GGML_ASSERT(r->ne[0] == 1 && r->ne[1] == S && r->ne[2] == H && r->ne[3] == n_tokens);
        // TODO: RWKV v4 and v5
        GGML_ASSERT(td->ne[0] == 1 && td->ne[1] == S && td->ne[2] == H && td->ne[3] == n_tokens);
        GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs);
    }
4584

4585
4586
4587
    // concat output and new_state
    const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
4588

4589
4590
4591
4592
4593
4594
4595
    result->op     = GGML_OP_RWKV_WKV6;
    result->src[0] = k;
    result->src[1] = v;
    result->src[2] = r;
    result->src[3] = tf;
    result->src[4] = td;
    result->src[5] = state;
4596

4597
4598
    return result;
}
4599

4600
// ggml_unary
4601

4602
4603
4604
4605
4606
4607
static struct ggml_tensor * ggml_unary_impl(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        enum ggml_unary_op    op,
        bool                  inplace) {
    GGML_ASSERT(ggml_is_contiguous_1(a));
4608

4609
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4610

4611
    ggml_set_op_params_i32(result, 0, (int32_t) op);
4612

4613
4614
    result->op     = GGML_OP_UNARY;
    result->src[0] = a;
4615

4616
4617
    return result;
}
4618

4619
4620
4621
4622
4623
4624
struct ggml_tensor * ggml_unary(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        enum ggml_unary_op    op) {
    return ggml_unary_impl(ctx, a, op, false);
}
4625

4626
4627
4628
4629
4630
4631
struct ggml_tensor * ggml_unary_inplace(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        enum ggml_unary_op    op) {
    return ggml_unary_impl(ctx, a, op, true);
}
4632

4633
// ggml_map_unary
4634

4635
4636
4637
4638
4639
4640
static struct ggml_tensor * ggml_map_unary_impl_f32(
        struct ggml_context        * ctx,
        struct ggml_tensor         * a,
        const  ggml_unary_op_f32_t   fun,
        bool                         inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4641

4642
    ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
4643

4644
4645
    result->op     = GGML_OP_MAP_UNARY;
    result->src[0] = a;
4646

4647
4648
    return result;
}
4649

4650
4651
4652
4653
4654
4655
struct ggml_tensor * ggml_map_unary_f32(
        struct ggml_context        * ctx,
        struct ggml_tensor         * a,
        const  ggml_unary_op_f32_t   fun) {
    return ggml_map_unary_impl_f32(ctx, a, fun, false);
}
4656

4657
4658
4659
4660
4661
4662
struct ggml_tensor * ggml_map_unary_inplace_f32(
        struct ggml_context        * ctx,
        struct ggml_tensor         * a,
        const  ggml_unary_op_f32_t   fun) {
    return ggml_map_unary_impl_f32(ctx, a, fun, true);
}
4663

4664
// ggml_map_binary
4665

4666
4667
4668
4669
4670
4671
4672
static struct ggml_tensor * ggml_map_binary_impl_f32(
        struct ggml_context         * ctx,
        struct ggml_tensor          * a,
        struct ggml_tensor          * b,
        const  ggml_binary_op_f32_t   fun,
        bool                          inplace) {
    GGML_ASSERT(ggml_are_same_shape(a, b));
4673

4674
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4675

4676
    ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
4677

4678
4679
4680
    result->op     = GGML_OP_MAP_BINARY;
    result->src[0] = a;
    result->src[1] = b;
4681

4682
4683
    return result;
}
4684

4685
4686
4687
4688
4689
4690
4691
struct ggml_tensor * ggml_map_binary_f32(
        struct ggml_context         * ctx,
        struct ggml_tensor          * a,
        struct ggml_tensor          * b,
        const  ggml_binary_op_f32_t   fun) {
    return ggml_map_binary_impl_f32(ctx, a, b, fun, false);
}
4692

4693
4694
4695
4696
4697
4698
4699
struct ggml_tensor * ggml_map_binary_inplace_f32(
        struct ggml_context         * ctx,
        struct ggml_tensor          * a,
        struct ggml_tensor          * b,
        const  ggml_binary_op_f32_t   fun) {
    return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
}
4700

4701
// ggml_map_custom1_f32
4702

4703
4704
4705
4706
4707
4708
static struct ggml_tensor * ggml_map_custom1_impl_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        const  ggml_custom1_op_f32_t   fun,
        bool                           inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4709

4710
    ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
4711

4712
4713
    result->op     = GGML_OP_MAP_CUSTOM1_F32;
    result->src[0] = a;
4714

4715
4716
    return result;
}
4717

4718
4719
4720
4721
4722
struct ggml_tensor * ggml_map_custom1_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        const  ggml_custom1_op_f32_t   fun) {
    return ggml_map_custom1_impl_f32(ctx, a, fun, false);
4723
4724
}

4725
4726
4727
4728
4729
4730
struct ggml_tensor * ggml_map_custom1_inplace_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        const  ggml_custom1_op_f32_t   fun) {
    return ggml_map_custom1_impl_f32(ctx, a, fun, true);
}
4731

4732
// ggml_map_custom2_f32
4733

4734
4735
4736
4737
4738
4739
4740
static struct ggml_tensor * ggml_map_custom2_impl_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        const  ggml_custom2_op_f32_t   fun,
        bool                           inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4741

4742
    ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
4743

4744
4745
4746
    result->op     = GGML_OP_MAP_CUSTOM2_F32;
    result->src[0] = a;
    result->src[1] = b;
4747

4748
4749
    return result;
}
4750

4751
4752
4753
4754
4755
4756
4757
struct ggml_tensor * ggml_map_custom2_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        const  ggml_custom2_op_f32_t   fun) {
    return ggml_map_custom2_impl_f32(ctx, a, b, fun, false);
}
4758

4759
4760
4761
4762
4763
4764
4765
struct ggml_tensor * ggml_map_custom2_inplace_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        const  ggml_custom2_op_f32_t   fun) {
    return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
}
4766

4767
// ggml_map_custom3_f32
4768

4769
4770
4771
4772
4773
4774
4775
4776
static struct ggml_tensor * ggml_map_custom3_impl_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        struct ggml_tensor           * c,
        const  ggml_custom3_op_f32_t   fun,
        bool                           inplace) {
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4777

4778
    ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
4779

4780
4781
4782
4783
    result->op     = GGML_OP_MAP_CUSTOM3_F32;
    result->src[0] = a;
    result->src[1] = b;
    result->src[2] = c;
4784

4785
4786
    return result;
}
4787

4788
4789
4790
4791
4792
4793
4794
4795
struct ggml_tensor * ggml_map_custom3_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        struct ggml_tensor           * c,
        const  ggml_custom3_op_f32_t   fun) {
    return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, false);
}
4796

4797
4798
4799
4800
4801
4802
4803
4804
struct ggml_tensor * ggml_map_custom3_inplace_f32(
        struct ggml_context          * ctx,
        struct ggml_tensor           * a,
        struct ggml_tensor           * b,
        struct ggml_tensor           * c,
        const  ggml_custom3_op_f32_t   fun) {
    return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
}
4805

4806
// ggml_map_custom1
4807

4808
4809
4810
4811
4812
4813
4814
4815
static struct ggml_tensor * ggml_map_custom1_impl(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        const  ggml_custom1_op_t   fun,
        int                        n_tasks,
        void                     * userdata,
        bool                       inplace) {
    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
4816

4817
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4818

4819
4820
4821
4822
4823
4824
    struct ggml_map_custom1_op_params params = {
        /*.fun      =*/ fun,
        /*.n_tasks  =*/ n_tasks,
        /*.userdata =*/ userdata
    };
    ggml_set_op_params(result, (const void *) &params, sizeof(params));
4825

4826
4827
    result->op     = GGML_OP_MAP_CUSTOM1;
    result->src[0] = a;
4828

4829
4830
    return result;
}
4831

4832
4833
4834
4835
4836
4837
4838
4839
struct ggml_tensor * ggml_map_custom1(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        const  ggml_custom1_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false);
}
4840

4841
4842
4843
4844
4845
4846
4847
4848
struct ggml_tensor * ggml_map_custom1_inplace(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        const  ggml_custom1_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true);
}
4849

4850
// ggml_map_custom2
4851

4852
4853
4854
4855
4856
4857
4858
4859
4860
static struct ggml_tensor * ggml_map_custom2_impl(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        const  ggml_custom2_op_t   fun,
        int                        n_tasks,
        void                     * userdata,
        bool                       inplace) {
    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
4861

4862
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4863

4864
4865
4866
4867
4868
4869
    struct ggml_map_custom2_op_params params = {
        /*.fun      =*/ fun,
        /*.n_tasks  =*/ n_tasks,
        /*.userdata =*/ userdata
    };
    ggml_set_op_params(result, (const void *) &params, sizeof(params));
4870

4871
4872
4873
    result->op     = GGML_OP_MAP_CUSTOM2;
    result->src[0] = a;
    result->src[1] = b;
4874

4875
4876
    return result;
}
4877

4878
4879
4880
4881
4882
4883
4884
4885
4886
struct ggml_tensor * ggml_map_custom2(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        const  ggml_custom2_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false);
}
4887

4888
4889
4890
4891
4892
4893
4894
4895
4896
struct ggml_tensor * ggml_map_custom2_inplace(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        const  ggml_custom2_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true);
}
4897

4898
// ggml_map_custom3
4899

4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
static struct ggml_tensor * ggml_map_custom3_impl(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        struct ggml_tensor       * c,
        const  ggml_custom3_op_t   fun,
        int                        n_tasks,
        void                     * userdata,
        bool                       inplace) {
    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
4910

4911
    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
4912

4913
4914
4915
4916
4917
4918
    struct ggml_map_custom3_op_params params = {
        /*.fun      =*/ fun,
        /*.n_tasks  =*/ n_tasks,
        /*.userdata =*/ userdata
    };
    ggml_set_op_params(result, (const void *) &params, sizeof(params));
4919

4920
4921
4922
4923
    result->op     = GGML_OP_MAP_CUSTOM3;
    result->src[0] = a;
    result->src[1] = b;
    result->src[2] = c;
4924

4925
4926
    return result;
}
4927

4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
struct ggml_tensor * ggml_map_custom3(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        struct ggml_tensor       * c,
        const  ggml_custom3_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false);
}
4938

4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
struct ggml_tensor * ggml_map_custom3_inplace(
        struct ggml_context      * ctx,
        struct ggml_tensor       * a,
        struct ggml_tensor       * b,
        struct ggml_tensor       * c,
        const  ggml_custom3_op_t   fun,
        int                        n_tasks,
        void                     * userdata) {
    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
}
4949

4950
// ggml_cross_entropy_loss
4951

4952
4953
4954
4955
4956
struct ggml_tensor * ggml_cross_entropy_loss(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b) {
    GGML_ASSERT(ggml_are_same_shape(a, b));
4957

4958
    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1);
4959

4960
4961
4962
    result->op     = GGML_OP_CROSS_ENTROPY_LOSS;
    result->src[0] = a;
    result->src[1] = b;
4963

4964
4965
    return result;
}
4966

4967
// ggml_cross_entropy_loss_back
4968

4969
4970
4971
4972
4973
4974
4975
struct ggml_tensor * ggml_cross_entropy_loss_back(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * b,
        struct ggml_tensor  * c) {
    GGML_ASSERT(ggml_are_same_shape(a, b));
    GGML_ASSERT(ggml_is_scalar(c));
4976

4977
    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
4978

4979
4980
4981
4982
    result->op     = GGML_OP_CROSS_ENTROPY_LOSS_BACK;
    result->src[0] = a;
    result->src[1] = b;
    result->src[2] = c;
4983

4984
4985
    return result;
}
4986

4987
// opt_step_adamw
4988

4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
struct ggml_tensor * ggml_opt_step_adamw(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        struct ggml_tensor  * grad,
        struct ggml_tensor  * m,
        struct ggml_tensor  * v,
        struct ggml_tensor  * adamw_params) {
    GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
    GGML_ASSERT(ggml_are_same_shape(a, grad));
    GGML_ASSERT(ggml_are_same_shape(a, m));
    GGML_ASSERT(ggml_are_same_shape(a, v));
    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
5002

5003
    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
5004

5005
5006
5007
5008
5009
5010
    result->op     = GGML_OP_OPT_STEP_ADAMW;
    result->src[0] = a;
    result->src[1] = grad;
    result->src[2] = m;
    result->src[3] = v;
    result->src[4] = adamw_params;
5011

5012
5013
    return result;
}
5014

5015
////////////////////////////////////////////////////////////////////////////////
5016

5017
5018
5019
5020
5021
5022
5023
5024
struct ggml_hash_set ggml_hash_set_new(size_t size) {
    size = ggml_hash_size(size);
    struct ggml_hash_set result;
    result.size = size;
    result.keys = GGML_MALLOC(sizeof(struct ggml_tensor *) * size);
    result.used = GGML_CALLOC(ggml_bitset_size(size), sizeof(ggml_bitset_t));
    return result;
}
5025

5026
5027
5028
void ggml_hash_set_reset(struct ggml_hash_set * hash_set) {
    memset(hash_set->used, 0, sizeof(ggml_bitset_t) * ggml_bitset_size(hash_set->size));
}
5029

5030
5031
5032
5033
void ggml_hash_set_free(struct ggml_hash_set * hash_set) {
    GGML_FREE(hash_set->used);
    GGML_FREE(hash_set->keys);
}
5034

5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
size_t ggml_hash_size(size_t min_sz) {
    // next primes after powers of two
    static const size_t primes[] = {
        2, 3, 5, 11, 17, 37, 67, 131, 257, 521, 1031,
        2053, 4099, 8209, 16411, 32771, 65537, 131101,
        262147, 524309, 1048583, 2097169, 4194319, 8388617,
        16777259, 33554467, 67108879, 134217757, 268435459,
        536870923, 1073741827, 2147483659
    };
    static const size_t n_primes = sizeof(primes)/sizeof(primes[0]);
5045

5046
5047
5048
5049
5050
5051
5052
5053
5054
    // find the smallest prime that is larger or equal than min_sz
    size_t l = 0;
    size_t r = n_primes;
    while (l < r) {
        size_t m = (l + r)/2;
        if (primes[m] < min_sz) {
            l = m + 1;
        } else {
            r = m;
5055
5056
        }
    }
5057
5058
5059
5060
5061
5062
5063
5064
    size_t sz = l < n_primes ? primes[l] : min_sz | 1;
    return sz;
}

struct hash_map {
    struct ggml_hash_set set;
    struct ggml_tensor ** vals;
};
5065

5066
5067
5068
5069
static struct hash_map * ggml_new_hash_map(size_t size) {
    struct hash_map * result = GGML_MALLOC(sizeof(struct hash_map));
    result->set = ggml_hash_set_new(size);
    result->vals = GGML_CALLOC(result->set.size, sizeof(struct ggml_tensor *));
5070
5071
5072
    return result;
}

5073
5074
5075
5076
5077
static void ggml_hash_map_free(struct hash_map * map) {
    ggml_hash_set_free(&map->set);
    GGML_FREE(map->vals);
    GGML_FREE(map);
}
5078

5079
5080
5081
5082
5083
5084
// utility functions to change gradients
// isrc is the index of tensor in cgraph->visited_has_set.keys
// the corresponding gradient (accumulators) are also at position isrc
// if tensor has a gradient accumulator, modify that accumulator in-place
// else if there is no gradient for tensor, set the corresponding value
// else, just add/subtract/etc. the gradients
5085

5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
static void ggml_add_or_set(
        struct ggml_context * ctx,
        struct ggml_cgraph  * cgraph,
        size_t                isrc,
        struct ggml_tensor  * tensor) {
    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
    GGML_ASSERT(src);
    if (cgraph->grads[isrc]) {
        cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, /*inplace =*/ cgraph->grad_accs[isrc]);
    } else {
        cgraph->grads[isrc] = tensor;
5097
    }
5098
5099
5100
    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
}
5101

5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
static void ggml_acc_or_set(
        struct ggml_context * ctx,
        struct ggml_cgraph  * cgraph,
        size_t                isrc,
        struct ggml_tensor  * tensor,
        const  size_t         nb1,
        const  size_t         nb2,
        const  size_t         nb3,
        const  size_t         offset) {
    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
    GGML_ASSERT(src);
    if (cgraph->grads[isrc]) {
        cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]);
    } else {
        struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
        cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false);
5118
    }
5119
5120
    ggml_format_name(cgraph->grads[isrc], "grad for %s", cgraph->visited_hash_set.keys[isrc]->name);
    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
5121
5122
}

5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
static void ggml_add1_or_set(
        struct ggml_context * ctx,
        struct ggml_cgraph  * cgraph,
        size_t                isrc,
        struct ggml_tensor  * tensor) {
    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
    GGML_ASSERT(src);
    if (cgraph->grads[isrc]) {
        cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
    } else {
        cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src);
5134
    }
5135
5136
5137
    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
}
5138

5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
static void ggml_sub_or_set(
        struct ggml_context * ctx,
        struct ggml_cgraph  * cgraph,
        size_t                isrc,
        struct ggml_tensor  * tensor) {
    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
    GGML_ASSERT(src);
    if (cgraph->grads[isrc]) {
        cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
    } else {
        cgraph->grads[isrc] = ggml_neg(ctx, tensor);
5150
    }
5151
5152
    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
5153
5154
}

5155
5156
5157
5158
static void ggml_compute_backward(
        struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, bool * grads_needed) {
    struct ggml_tensor * tensor = cgraph->nodes[i];
    struct ggml_tensor * grad   = ggml_graph_get_grad(cgraph, tensor);
5159

5160
5161
    if (!grad) {
        return;
5162
5163
    }

5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
    struct ggml_tensor * src0 = tensor->src[0];
    struct ggml_tensor * src1 = tensor->src[1];
    struct ggml_tensor * src2 = tensor->src[2];
    struct ggml_hash_set * hash_set = &cgraph->visited_hash_set;
    const size_t isrc0 = src0 ? ggml_hash_find(hash_set, src0) : (size_t) -1;
    const size_t isrc1 = src1 ? ggml_hash_find(hash_set, src1) : (size_t) -1;
    const size_t isrc2 = src2 ? ggml_hash_find(hash_set, src2) : (size_t) -1;
    const bool src0_needs_grads = src0 && isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
    const bool src1_needs_grads = src1 && isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
    const bool src2_needs_grads = src2 && isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
5174

5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
    switch (tensor->op) {
        case GGML_OP_DUP: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
        } break;
        case GGML_OP_ADD: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
            if (src1_needs_grads) {
                struct ggml_tensor * tmp = grad;
                if (!ggml_are_same_shape(src0, src1)) {
                    tmp = ggml_repeat_back(ctx, tmp, src1);
                }
                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
            }
        } break;
        case GGML_OP_ADD1: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
            if (src1_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc1, ggml_mean(ctx, grad)); // TODO: should probably be sum instead of mean
            }
        } break;
        case GGML_OP_ACC: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
            if (src1_needs_grads) {
                const size_t nb1    = ((int32_t *) tensor->op_params)[0];
                const size_t nb2    = ((int32_t *) tensor->op_params)[1];
                const size_t nb3    = ((int32_t *) tensor->op_params)[2];
                const size_t offset = ((int32_t *) tensor->op_params)[3];

                struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx,
                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                    nb1, nb2, nb3, offset);

                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
            }
        } break;
        case GGML_OP_SUB: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
            if (src1_needs_grads) {
                ggml_sub_or_set(ctx, cgraph, isrc1, grad);
            }
        } break;
        case GGML_OP_MUL: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, src1, grad));
            }
            if (src1_needs_grads) {
                struct ggml_tensor * tmp = ggml_mul(ctx, src0, grad);
                if (!ggml_are_same_shape(src0, src1)) {
                    tmp = ggml_repeat_back(ctx, tmp, src1);
                }
                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
            }
        } break;
        case GGML_OP_DIV: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src1));
            }
            if (src1_needs_grads) {
                ggml_sub_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, grad, ggml_div(ctx, tensor, src1)));
            }
        } break;
        case GGML_OP_SQR: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_mul(ctx, src0, grad), 2.0f));
            }
        } break;
        case GGML_OP_SQRT: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_div(ctx, grad, tensor), 0.5f));
            }
        } break;
        case GGML_OP_LOG: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src0));
            }
        } break;
        case GGML_OP_SIN: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_cos(ctx, src0)));
            }
        } break;
        case GGML_OP_COS: {
            if (src0_needs_grads) {
                ggml_sub_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sin(ctx, src0)));
            }
        } break;
        case GGML_OP_SUM: {
            if (src0_needs_grads) {
                ggml_add1_or_set(ctx, cgraph, isrc0, grad);
            }
        } break;
        case GGML_OP_SUM_ROWS: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
            }
        } break;
        case GGML_OP_MEAN: {
            if (src0_needs_grads) {
                ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
            }
        } break;
        case GGML_OP_REPEAT: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat_back(ctx, grad, src0));
            }
        } break;
        case GGML_OP_REPEAT_BACK: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
            }
        } break;
        case GGML_OP_RMS_NORM: {
            if (src0_needs_grads) {
                float eps;
                memcpy(&eps, tensor->op_params, sizeof(float));
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, src0, grad, eps));
            }
        } break;
        case GGML_OP_MUL_MAT: {
            // https://cs231n.github.io/optimization-2/#staged
            // # forward pass
            // s0 = np.random.randn(5, 10)
            // s1 = np.random.randn(10, 3)
            // t = s0.dot(s1)

            // # now suppose we had the gradient on t from above in the circuit
            // dt = np.random.randn(*t.shape) # same shape as t
            // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
            // ds1 = t.T.dot(dt)

            // tensor.shape [m,p,qq,rr]
            // src0.shape   [n,m,q1,r1]
            // src1.shape   [n,p,qq,rr]

            if (src0_needs_grads) {
                struct ggml_tensor * s1_tg =
                    ggml_out_prod(ctx, // [n,m,qq,rr]
                        src1,          // [n,p,qq,rr]
                        grad);         // [m,p,qq,rr]
                const int64_t qq = s1_tg->ne[2];
                const int64_t rr = s1_tg->ne[3];
                const int64_t q1 = src0->ne[2];
                const int64_t r1 = src0->ne[3];
                const bool ne2_broadcasted = qq > q1;
                const bool ne3_broadcasted = rr > r1;
                if (ne2_broadcasted || ne3_broadcasted) {
                    // sum broadcast repetitions of s1_tg into shape of src0
                    s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
                }
                ggml_add_or_set(ctx, cgraph, isrc0, s1_tg /*= [n,m,q1,r1]*/);
            }
            if (src1_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc1,
                        // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
                        //     ggml_cont(ctx,                  // [m,n,q1,r1]
                        //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
                        //     grad),                          // [m,p,qq,rr]

                        // when src0 is bigger than tensor->grad (this is mostly the case in llama),
                        // avoid transpose of src0, rather transpose smaller tensor->grad
                        // and then use ggml_out_prod
                        ggml_out_prod(ctx,      // [n,p,qq,rr]
                            src0,               // [n,m,q1,r1]
                            ggml_transpose(ctx, // [p,m,qq,rr]
                                grad)));        // [m,p,qq,rr]
            }
        } break;
        case GGML_OP_SCALE: {
            if (src0_needs_grads) {
                float s;
                memcpy(&s, tensor->op_params, sizeof(float));
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, s, false));
            }
        } break;
        case GGML_OP_SET: {
            const size_t nb1    = ((const int32_t *) tensor->op_params)[0];
            const size_t nb2    = ((const int32_t *) tensor->op_params)[1];
            const size_t nb3    = ((const int32_t *) tensor->op_params)[2];
            const size_t offset = ((const int32_t *) tensor->op_params)[3];

            struct ggml_tensor * tensor_grad_view = NULL;

            if (src0_needs_grads || src1_needs_grads) {
                GGML_ASSERT(src0->type == tensor->type);
                GGML_ASSERT(!cgraph->grads[isrc0] ||                      cgraph->grads[isrc0]->type == grad->type);
                GGML_ASSERT(!cgraph->grads[isrc1] || !src1_needs_grads || cgraph->grads[isrc1]->type == grad->type);

                tensor_grad_view = ggml_view_4d(ctx,
                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                    nb1, nb2, nb3, offset);
            }

            if (src0_needs_grads) {
                struct ggml_tensor * tmp = ggml_neg(ctx, tensor_grad_view);
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_acc_impl(ctx, grad, tmp, nb1, nb2, nb3, offset, false));
            }

            if (src1_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
            }
        } break;
        case GGML_OP_CPY: {
            // cpy overwrites value of src1 by src0 and returns view(src1)
            // the overwriting is mathematically equivalent to:
            // tensor = src0 * 1 + src1 * 0
            if (src0_needs_grads) {
                // dsrc0 = dtensor * 1
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
            if (src1_needs_grads) {
                // dsrc1 = dtensor * 0 -> noop
            }
        } break;
        case GGML_OP_CONT: {
            // same as cpy
            if (src0_needs_grads) {
                GGML_ASSERT(!cgraph->grads[isrc0] || ggml_is_contiguous(cgraph->grads[isrc0]));
                GGML_ASSERT(ggml_is_contiguous(grad));
                ggml_add_or_set(ctx, cgraph, isrc0, grad);
            }
        } break;
        case GGML_OP_RESHAPE: {
            if (src0_needs_grads) {
                struct ggml_tensor * grad_cont = ggml_is_contiguous(grad) ? grad : ggml_cont(ctx, grad);
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad_cont, src0));
            }
        } break;
        case GGML_OP_VIEW: {
            if (src0_needs_grads) {
                size_t offset;

                memcpy(&offset, tensor->op_params, sizeof(offset));

                size_t nb1 = tensor->nb[1];
                size_t nb2 = tensor->nb[2];
                size_t nb3 = tensor->nb[3];

                if (cgraph->grads[isrc0] && src0->type != cgraph->grads[isrc0]->type) {
                    // gradient is typically F32, but src0 could be other type
                    size_t ng = ggml_element_size(cgraph->grads[isrc0]);
                    size_t n0 = ggml_element_size(src0);
                    GGML_ASSERT(offset % n0 == 0);
                    GGML_ASSERT(nb1 % n0 == 0);
                    GGML_ASSERT(nb2 % n0 == 0);
                    GGML_ASSERT(nb3 % n0 == 0);
                    offset = (offset / n0) * ng;
                    nb1 = (nb1 / n0) * ng;
                    nb2 = (nb2 / n0) * ng;
                    nb3 = (nb3 / n0) * ng;
                }

                ggml_acc_or_set(ctx, cgraph, isrc0, grad, nb1, nb2, nb3, offset);
            }
        } break;
        case GGML_OP_PERMUTE: {
            if (src0_needs_grads) {
                const int32_t * axes = (const int32_t *) tensor->op_params;
                const int axis0 = axes[0] & 0x3;
                const int axis1 = axes[1] & 0x3;
                const int axis2 = axes[2] & 0x3;
                const int axis3 = axes[3] & 0x3;
                int axb[4] = {0,0,0,0}; // axes backward
                axb[axis0] = 0;
                axb[axis1] = 1;
                axb[axis2] = 2;
                axb[axis3] = 3;
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_permute(ctx, grad, axb[0], axb[1], axb[2], axb[3]));
            }
        } break;
        case GGML_OP_TRANSPOSE: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_transpose(ctx, grad));
            }
        } break;
        case GGML_OP_GET_ROWS: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_get_rows_back(ctx, grad, src1, src0));
            }
            if (src1_needs_grads) {
                // noop
            }
        } break;
        case GGML_OP_DIAG_MASK_INF: {
            if (src0_needs_grads) {
                /* ggml_diag_mask_inf_impl() shouldn't be here */
                /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
                const int n_past = ((const int32_t *) tensor->op_params)[0];
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
            }
        } break;
        case GGML_OP_DIAG_MASK_ZERO: {
            if (src0_needs_grads) {
                const int n_past = ((const int32_t *) tensor->op_params)[0];
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
            }
        } break;
        case GGML_OP_SOFT_MAX: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_back(ctx, grad, tensor));
            }
            GGML_ASSERT((!src1 || !src1_needs_grads) && "backward pass for softmax mask not implemented");
        } break;
        case GGML_OP_ROPE: {
            if (src0_needs_grads) {
                //const int n_past = ((int32_t *) tensor->op_params)[0];
                const int n_dims     = ((const int32_t *) tensor->op_params)[1];
                const int mode       = ((const int32_t *) tensor->op_params)[2];
                //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
                const int n_ctx_orig = ((const int32_t *) tensor->op_params)[4];
                float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;

                memcpy(&freq_base,   (const float *) tensor->op_params +  5, sizeof(float));
                memcpy(&freq_scale,  (const float *) tensor->op_params +  6, sizeof(float));
                memcpy(&ext_factor,  (const float *) tensor->op_params +  7, sizeof(float));
                memcpy(&attn_factor, (const float *) tensor->op_params +  8, sizeof(float));
                memcpy(&beta_fast,   (const float *) tensor->op_params +  9, sizeof(float));
                memcpy(&beta_slow,   (const float *) tensor->op_params + 10, sizeof(float));

                ggml_add_or_set(ctx, cgraph, isrc0,
                    ggml_rope_back(ctx, grad, src1, src2, n_dims, mode, n_ctx_orig, freq_base,
                        freq_scale, ext_factor, attn_factor, beta_fast, beta_slow));
            }
            GGML_ASSERT((!src2 || !src2_needs_grads) && "gradients for freq factors not implemented");
        } break;
        case GGML_OP_IM2COL: {
            if (src1_needs_grads) {
                const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
                const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
                const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
                const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
                const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
                const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
                const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;

                ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, src0, grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
            }
        } break;
        case GGML_OP_POOL_2D: {
            if (src0_needs_grads) {
                const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
                const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
                const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
                const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
                const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
                const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
                const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);

                ggml_add_or_set(ctx, cgraph, isrc0, ggml_pool_2d_back(ctx, grad, src0, op, k0, k1, s0, s1, p0, p1));
            }
        } break;
        case GGML_OP_WIN_PART:
        case GGML_OP_WIN_UNPART:
        case GGML_OP_UNARY: {
            switch (ggml_get_unary_op(tensor)) {
                case GGML_UNARY_OP_ABS: {
                    if (src0_needs_grads) {
                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_sgn(ctx, src0), grad));
                    }
                } break;
                case GGML_UNARY_OP_SGN: {
                    // noop
                } break;
                case GGML_UNARY_OP_NEG: {
                    if (src0_needs_grads) {
                        ggml_sub_or_set(ctx, cgraph, isrc0, grad);
                    }
                } break;
                case GGML_UNARY_OP_STEP: {
                    // noop
                } break;
                case GGML_UNARY_OP_RELU: {
                    if (src0_needs_grads) {
                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_step(ctx, src0), grad));
                    }
                } break;
                case GGML_UNARY_OP_SILU: {
                    if (src0_needs_grads) {
                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, src0, grad));
                    }
                } break;
                case GGML_UNARY_OP_EXP: {
                    if (src0_needs_grads) {
                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
                    }
                } break;
                default: {
                    fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
                        __func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));
                    GGML_ABORT("fatal error");
                } //break;
            }
        } break;
        case GGML_OP_CROSS_ENTROPY_LOSS: {
            if (src0_needs_grads) {
                ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, src0, src1, grad));
            }
            GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
        } break;
        case GGML_OP_NONE: {
            // noop
        } break;
        case GGML_OP_COUNT:
        default: {
            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
            GGML_ABORT("fatal error");
        } //break;
    }
5592

5593
5594
5595
    GGML_ASSERT(!src0_needs_grads || ggml_are_same_shape(src0, cgraph->grads[isrc0]));
    GGML_ASSERT(!src1_needs_grads || ggml_are_same_shape(src1, cgraph->grads[isrc1]));
    GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
5596
5597
}

5598
5599
5600
5601
5602
static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
    // check if already visited
    if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
        return;
    }
5603

5604
5605
5606
5607
5608
5609
5610
    for (int i = 0; i < GGML_MAX_SRC; ++i) {
        const int k =
            (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
            (cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) :
            /* unknown order, just fall back to using i*/ i;
        if (node->src[k]) {
            ggml_visit_parents(cgraph, node->src[k]);
5611
        }
5612
    }
5613

5614
5615
5616
    if (node->op == GGML_OP_NONE && !(node->flags & GGML_TENSOR_FLAG_PARAM)) {
        // reached a leaf node, not part of the gradient graph (e.g. a constant)
        GGML_ASSERT(cgraph->n_leafs < cgraph->size);
5617

5618
5619
        if (strlen(node->name) == 0) {
            ggml_format_name(node, "leaf_%d", cgraph->n_leafs);
5620
5621
        }

5622
5623
5624
5625
        cgraph->leafs[cgraph->n_leafs] = node;
        cgraph->n_leafs++;
    } else {
        GGML_ASSERT(cgraph->n_nodes < cgraph->size);
5626

5627
5628
        if (strlen(node->name) == 0) {
            ggml_format_name(node, "node_%d", cgraph->n_nodes);
5629
        }
5630
5631
5632

        cgraph->nodes[cgraph->n_nodes] = node;
        cgraph->n_nodes++;
5633
    }
5634
}
5635

5636
5637
5638
5639
5640
static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
    if (!expand) {
        // TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand
        ggml_graph_clear(cgraph);
    }
5641

5642
    const int n0 = cgraph->n_nodes;
5643

5644
    ggml_visit_parents(cgraph, tensor);
5645

5646
5647
    const int n_new = cgraph->n_nodes - n0;
    GGML_PRINT_DEBUG("%s: visited %d new nodes\n", __func__, n_new);
5648

5649
5650
5651
    if (n_new > 0) {
        // the last added node should always be starting point
        GGML_ASSERT(cgraph->nodes[cgraph->n_nodes - 1] == tensor);
5652
    }
5653
}
5654

5655
5656
5657
void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
    ggml_build_forward_impl(cgraph, tensor, true);
}
5658

5659
5660
5661
5662
5663
5664
5665
5666
void ggml_build_backward_expand(
        struct ggml_context * ctx_static,
        struct ggml_context * ctx_compute,
        struct ggml_cgraph  * cgraph,
        bool                  accumulate) {
    GGML_ASSERT(cgraph->n_nodes > 0);
    GGML_ASSERT(cgraph->grads);
    GGML_ASSERT(cgraph->grad_accs);
5667

5668
    const int n_nodes_f = cgraph->n_nodes;
5669

5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
    memset(cgraph->grads,     0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
    memset(cgraph->grad_accs, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
    bool * grads_needed = calloc(cgraph->visited_hash_set.size, sizeof(bool));

    {
        bool any_params = false;
        bool any_loss   = false;
        for (int i = 0; i < n_nodes_f; ++i) {
            struct ggml_tensor * node = cgraph->nodes[i];
            any_params = any_params || (node->flags & GGML_TENSOR_FLAG_PARAM);
            any_loss   = any_loss   || (node->flags & GGML_TENSOR_FLAG_LOSS);
5681
        }
5682
5683
        GGML_ASSERT(any_params && "no trainable parameters found, did you forget to call ggml_set_param?");
        GGML_ASSERT(any_loss && "no training loss found, did you forget to call ggml_set_loss?");
5684
5685
    }

5686
5687
    for (int i = 0; i < n_nodes_f; ++i) {
        struct ggml_tensor * node = cgraph->nodes[i];
5688

5689
5690
5691
        if (node->type == GGML_TYPE_I32) {
            continue;
        }
5692

5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
        bool node_needs_grad = (node->flags & GGML_TENSOR_FLAG_PARAM) || (node->flags & GGML_TENSOR_FLAG_LOSS);
        bool ignore_src[GGML_MAX_SRC] = {false};
        switch (node->op) {
            // gradients in node->src[0] for one reason or another have no effect on output gradients
            case GGML_OP_IM2COL:      // only used for its shape
            case GGML_OP_IM2COL_BACK: // same as IM2COL
                ignore_src[0] = true;
                break;
            case GGML_OP_UNARY: {
                const enum ggml_unary_op uop = ggml_get_unary_op(node);
                // SGN and STEP unary ops are piecewise constant
                if (uop == GGML_UNARY_OP_SGN || uop == GGML_UNARY_OP_STEP) {
                    ignore_src[0] = true;
                }
            } break;
5708

5709
5710
5711
5712
5713
5714
5715
            // gradients in node->src[1] for one reason or another have no effect on output gradients
            case GGML_OP_CPY:           // gradients in CPY target are irrelevant
            case GGML_OP_GET_ROWS:      // row indices not differentiable
            case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS
            case GGML_OP_ROPE:          // positions not differentiable
                ignore_src[1] = true;
                break;
5716

5717
5718
            default:
                break;
5719
        }
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
        for (int j = 0; j < GGML_MAX_SRC; ++j) {
            if (!node->src[j] || ignore_src[j] || !grads_needed[ggml_hash_find(&cgraph->visited_hash_set, node->src[j])]) {
                continue;
            }
            GGML_ASSERT(node->src[j]->type == GGML_TYPE_F32 || node->src[j]->type == GGML_TYPE_F16);
            node_needs_grad = true;
            break;
        }
        if (!node_needs_grad) {
            continue;
5730
5731
        }

5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
        // inplace operations are currently not supported
        GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
            node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);

        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
5743
        }
5744
        grads_needed[igrad] = true;
5745
5746
    }

5747
5748
5749
5750
    for (int i = n_nodes_f - 1; i >= 0; --i) {
        // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
        // use allocator to automatically make inplace operations
        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
5751
    }
5752
5753

    free(grads_needed);
5754
5755
}

5756
5757
5758
5759
5760
5761
static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
    void * ptr = *p;
    ptr = (void *) GGML_PAD((uintptr_t) ptr, align);
    *p = (void *) ((char *) ptr + size);
    return ptr;
}
5762

5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
static size_t ggml_graph_nbytes(size_t size, bool grads) {
    size_t hash_size = ggml_hash_size(size * 2);
    void * p = 0;
    incr_ptr_aligned(&p, sizeof(struct ggml_cgraph), 1);
    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // nodes
    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
    incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
    if (grads) {
        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grad_accs
5773
    }
5774
    incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
5775

5776
5777
5778
    size_t nbytes = (size_t) p;
    return nbytes;
}
5779

5780
5781
5782
size_t ggml_graph_overhead_custom(size_t size, bool grads) {
    return GGML_OBJECT_SIZE + GGML_PAD(ggml_graph_nbytes(size, grads), GGML_MEM_ALIGN);
}
5783

5784
5785
5786
size_t ggml_graph_overhead(void) {
    return ggml_graph_overhead_custom(GGML_DEFAULT_GRAPH_SIZE, false);
}
5787

5788
5789
5790
5791
struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads) {
    const size_t obj_size = ggml_graph_nbytes(size, grads);
    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_GRAPH, obj_size);
    struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
5792

5793
5794
    // the size of the hash table is doubled since it needs to hold both nodes and leafs
    size_t hash_size = ggml_hash_size(size * 2);
5795

5796
    void * p = cgraph + 1;
5797

5798
5799
5800
5801
5802
    struct ggml_tensor ** nodes_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
    struct ggml_tensor ** leafs_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
    struct ggml_tensor ** hash_keys_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
    struct ggml_tensor ** grads_ptr     = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
    struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
5803

5804
    ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
5805

5806
5807
    // check that we allocated the correct amount of memory
    assert(obj_size == (size_t)((char *)p - (char *)cgraph));
5808

5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
    *cgraph = (struct ggml_cgraph) {
        /*.size         =*/ size,
        /*.n_nodes      =*/ 0,
        /*.n_leafs      =*/ 0,
        /*.nodes        =*/ nodes_ptr,
        /*.grads        =*/ grads_ptr,
        /*.grad_accs    =*/ grad_accs_ptr,
        /*.leafs        =*/ leafs_ptr,
        /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
        /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
    };
5820

5821
5822
5823
5824
5825
    ggml_hash_set_reset(&cgraph->visited_hash_set);
    if (grads) {
        memset(cgraph->grads,     0, hash_size*sizeof(struct ggml_tensor *));
        memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
    }
5826

5827
5828
    return cgraph;
}
5829

5830
5831
5832
struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
    return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false);
}
5833

5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
    struct ggml_cgraph cgraph = {
        /*.size             =*/ 0,
        /*.n_nodes          =*/ i1 - i0,
        /*.n_leafs          =*/ 0,
        /*.nodes            =*/ cgraph0->nodes + i0,
        /*.grads            =*/ NULL, // gradients would need visited_hash_set
        /*.grad_accs        =*/ NULL,
        /*.leafs            =*/ NULL,
        /*.visited_hash_set =*/ { 0, NULL, NULL },
        /*.order            =*/ cgraph0->order,
    };
5846

5847
5848
    return cgraph;
}
5849

5850
5851
5852
5853
void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
    GGML_ASSERT(dst->size >= src->n_leafs);
    GGML_ASSERT(dst->size >= src->n_nodes);
    GGML_ASSERT(dst->visited_hash_set.size >= src->visited_hash_set.size);
5854

5855
5856
5857
    dst->n_leafs = src->n_leafs;
    dst->n_nodes = src->n_nodes;
    dst->order   = src->order;
5858

5859
5860
5861
    for (int i = 0; i < src->n_leafs; ++i) {
        dst->leafs[i] = src->leafs[i];
    }
5862

5863
5864
5865
    for (int i = 0; i < src->n_nodes; ++i) {
        dst->nodes[i] = src->nodes[i];
    }
5866

5867
5868
5869
5870
    for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
        // copy all hashset keys (tensors) that are in use
        if (ggml_bitset_get(src->visited_hash_set.used, i)) {
            ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
5871
        }
5872
    }
5873

5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
    if (dst->grads) {
        memset(dst->grads,     0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
        memset(dst->grad_accs, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
    }
    if (src->grads) {
        GGML_ASSERT(dst->grads     != NULL);
        GGML_ASSERT(dst->grad_accs != NULL);
        for (int i = 0; i < src->n_nodes; ++i) {
            const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
            const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
5884

5885
5886
5887
5888
            GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
            GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
            GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
            GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
5889

5890
5891
            dst->grads[igrad_dst]     = src->grads[igrad_src];
            dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
5892
5893
5894
5895
        }
    }
}

5896
5897
5898
5899
5900
struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL);
    ggml_graph_cpy(cgraph, result);
    return result;
}
5901

5902
5903
5904
struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
    if (ggml_is_empty(tensor)) {
        return tensor;
5905
    }
5906
5907
5908
5909
5910
    if (tensor->buffer) {
        ggml_backend_tensor_memset(tensor, 0, 0, ggml_nbytes(tensor));
    } else {
        GGML_ASSERT(tensor->data);
        memset(tensor->data, 0, ggml_nbytes(tensor));
5911
    }
5912
5913
    return tensor;
}
5914

5915
5916
void ggml_graph_reset(struct ggml_cgraph * cgraph) {
    GGML_ASSERT(cgraph->grads != NULL);
5917

5918
5919
5920
    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * node     = cgraph->nodes[i];
        struct ggml_tensor * grad_acc = ggml_graph_get_grad_acc(cgraph, node);
5921

5922
5923
5924
5925
5926
        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
            // clear momenta
            ggml_set_zero(node->src[2]);
            ggml_set_zero(node->src[3]);
        }
5927

5928
5929
5930
5931
5932
        // initial gradients of loss should be 1, 0 otherwise
        if (grad_acc) {
            if (node->flags & GGML_TENSOR_FLAG_LOSS) {
                GGML_ASSERT(grad_acc->type == GGML_TYPE_F32);
                GGML_ASSERT(ggml_is_scalar(grad_acc));
5933

5934
5935
5936
                const float onef = 1.0f;
                if (grad_acc->buffer) {
                    ggml_backend_tensor_set(grad_acc, &onef, 0, sizeof(float));
5937
                } else {
5938
5939
                    GGML_ASSERT(grad_acc->data);
                    *((float *) grad_acc->data) = onef;
5940
                }
5941
5942
            } else {
                ggml_set_zero(grad_acc);
5943
5944
5945
            }
        }
    }
5946
}
5947

5948
5949
5950
5951
5952
void ggml_graph_clear(struct ggml_cgraph * cgraph) {
    cgraph->n_leafs = 0;
    cgraph->n_nodes = 0;
    ggml_hash_set_reset(&cgraph->visited_hash_set);
}
5953

5954
5955
int ggml_graph_size(struct ggml_cgraph * cgraph) {
    return cgraph->size;
5956
5957
}

5958
5959
5960
5961
struct ggml_tensor * ggml_graph_node(struct ggml_cgraph * cgraph, int i) {
    if (i < 0) {
        GGML_ASSERT(cgraph->n_nodes + i >= 0);
        return cgraph->nodes[cgraph->n_nodes + i];
5962
5963
    }

5964
5965
5966
    GGML_ASSERT(i < cgraph->n_nodes);
    return cgraph->nodes[i];
}
5967

5968
5969
5970
struct ggml_tensor ** ggml_graph_nodes(struct ggml_cgraph * cgraph) {
    return cgraph->nodes;
}
5971

5972
5973
5974
int ggml_graph_n_nodes(struct ggml_cgraph * cgraph) {
    return cgraph->n_nodes;
}
5975

5976
5977
5978
5979
5980
void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
    GGML_ASSERT(cgraph->size > cgraph->n_nodes);
    cgraph->nodes[cgraph->n_nodes] = tensor;
    cgraph->n_nodes++;
}
5981

5982
5983
5984
struct ggml_tensor * ggml_graph_get_tensor(const struct ggml_cgraph * cgraph, const char * name) {
    for (int i = 0; i < cgraph->n_leafs; i++) {
        struct ggml_tensor * leaf = cgraph->leafs[i];
5985

5986
5987
5988
        if (strcmp(leaf->name, name) == 0) {
            return leaf;
        }
5989
5990
    }

5991
5992
    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * node = cgraph->nodes[i];
5993

5994
5995
5996
5997
        if (strcmp(node->name, name) == 0) {
            return node;
        }
    }
5998

5999
6000
    return NULL;
}
6001

6002
6003
6004
6005
struct ggml_tensor * ggml_graph_get_grad(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grads[igrad] : NULL;
}
6006

6007
6008
6009
6010
struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grad_accs[igrad] : NULL;
}
6011

6012
6013
void ggml_graph_print(const struct ggml_cgraph * cgraph) {
    GGML_LOG_INFO("=== GRAPH ===\n");
6014

6015
6016
6017
    GGML_LOG_INFO("n_nodes = %d\n", cgraph->n_nodes);
    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * node = cgraph->nodes[i];
6018

6019
6020
6021
6022
6023
        GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n",
                i,
                node->ne[0], node->ne[1], node->ne[2],
                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" :
                      ggml_graph_get_grad(cgraph, node) ? "g" : " ");
6024
6025
    }

6026
6027
6028
    GGML_LOG_INFO("n_leafs = %d\n", cgraph->n_leafs);
    for (int i = 0; i < cgraph->n_leafs; i++) {
        struct ggml_tensor * node = cgraph->leafs[i];
6029

6030
6031
6032
6033
6034
        GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n",
                i,
                node->ne[0], node->ne[1],
                ggml_op_name(node->op),
                ggml_get_name(node));
6035
6036
    }

6037
6038
6039
6040
6041
6042
6043
    GGML_LOG_INFO("========================================\n");
}

// check if node is part of the graph
static bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
    if (cgraph == NULL) {
        return true;
6044
6045
    }

6046
6047
6048
    for (int i = 0; i < cgraph->n_nodes; i++) {
        if (cgraph->nodes[i] == node) {
            return true;
6049
6050
6051
        }
    }

6052
6053
    return false;
}
6054

6055
6056
6057
6058
static struct ggml_tensor * ggml_graph_get_parent(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * parent = cgraph->nodes[i];
        struct ggml_tensor * grad = ggml_graph_get_grad(cgraph, parent);
6059

6060
6061
        if (grad == node) {
            return parent;
6062
        }
6063
6064
6065
6066
    }

    return NULL;
}
6067

6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
static void ggml_graph_dump_dot_node_edge(FILE * fp, const struct ggml_cgraph * gb, struct ggml_tensor * node, struct ggml_tensor * parent, const char * label)  {
    struct ggml_tensor * gparent = ggml_graph_get_parent(gb, node);
    struct ggml_tensor * gparent0 = ggml_graph_get_parent(gb, parent);
    fprintf(fp, "  \"%p\":%s -> \"%p\":%s [ arrowhead = %s; style = %s; label = \"%s\"; ]\n",
            gparent0 ? (void *) gparent0 : (void *) parent,
            gparent0 ? "g" : "x",
            gparent ? (void *) gparent : (void *) node,
            gparent ? "g" : "x",
            gparent ? "empty" : "vee",
            gparent ? "dashed" : "solid",
            label);
}
6080

6081
6082
6083
6084
6085
6086
static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node, struct ggml_tensor * parent, const char * label)  {
    fprintf(fp, "  \"%p\":%s -> \"%p\":%s [ label = \"%s\"; ]\n",
            (void *) parent, "x",
            (void *) node, "x",
            label);
}
6087

6088
6089
void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
    char color[16];
6090

6091
6092
    FILE * fp = ggml_fopen(filename, "w");
    GGML_ASSERT(fp);
6093

6094
6095
6096
    fprintf(fp, "digraph G {\n");
    fprintf(fp, "  newrank = true;\n");
    fprintf(fp, "  rankdir = TB;\n");
6097

6098
6099
6100
    for (int i = 0; i < gb->n_nodes; i++) {
        struct ggml_tensor * node = gb->nodes[i];
        struct ggml_tensor * grad = ggml_graph_get_grad(gb, node);
6101

6102
6103
        if (ggml_graph_get_parent(gb, node) != NULL) {
            continue;
6104
6105
        }

6106
6107
6108
6109
6110
        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
            snprintf(color, sizeof(color), "yellow");
        } else if (grad) {
            if (ggml_graph_find(gf, node)) {
                snprintf(color, sizeof(color), "green");
6111
            } else {
6112
                snprintf(color, sizeof(color), "lightblue");
6113
            }
6114
6115
        } else {
            snprintf(color, sizeof(color), "white");
6116
6117
        }

6118
6119
6120
6121
        fprintf(fp, "  \"%p\" [ "
                    "style = filled; fillcolor = %s; shape = record; "
                    "label=\"",
                (void *) node, color);
6122

6123
6124
6125
6126
        if (strlen(node->name) > 0) {
            fprintf(fp, "%s (%s)|", node->name, ggml_type_name(node->type));
        } else {
            fprintf(fp, "(%s)|", ggml_type_name(node->type));
6127
6128
        }

6129
6130
6131
6132
        if (ggml_is_matrix(node)) {
            fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], ggml_op_symbol(node->op));
        } else {
            fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
6133
6134
        }

6135
6136
6137
6138
6139
        if (grad) {
            fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(grad->op));
        } else {
            fprintf(fp, "\"; ]\n");
        }
6140
6141
    }

6142
6143
    for (int i = 0; i < gb->n_leafs; i++) {
        struct ggml_tensor * node = gb->leafs[i];
6144

6145
        snprintf(color, sizeof(color), "pink");
6146

6147
6148
6149
6150
        fprintf(fp, "  \"%p\" [ "
                    "style = filled; fillcolor = %s; shape = record; "
                    "label=\"<x>",
                (void *) node, color);
6151

6152
6153
6154
6155
        if (strlen(node->name) > 0) {
            fprintf(fp, "%s (%s)|", node->name, ggml_type_name(node->type));
        } else {
            fprintf(fp, "(%s)|", ggml_type_name(node->type));
6156
6157
        }

6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
        fprintf(fp, "CONST %d [%" PRId64 ", %" PRId64 "]", i, node->ne[0], node->ne[1]);
        if (ggml_nelements(node) < 5 && node->data != NULL) {
            fprintf(fp, " | (");
            for (int j = 0; j < ggml_nelements(node); j++) {
                // FIXME: use ggml-backend to obtain the tensor data
                //if (node->type == GGML_TYPE_I8 || node->type == GGML_TYPE_I16 || node->type == GGML_TYPE_I32) {
                //    fprintf(fp, "%d", ggml_get_i32_1d(node, j));
                //}
                //else if (node->type == GGML_TYPE_F32 ||
                //         node->type == GGML_TYPE_F16 ||
                //         node->type == GGML_TYPE_BF16) {
                //    fprintf(fp, "%.1e", (double)ggml_get_f32_1d(node, j));
                //}
                //else
                {
                    fprintf(fp, "#");
                }
                if (j < ggml_nelements(node) - 1) {
                    fprintf(fp, ", ");
                }
            }
            fprintf(fp, ")");
6180
        }
6181
        fprintf(fp, "\"; ]\n");
6182
6183
    }

6184
6185
    for (int i = 0; i < gb->n_nodes; i++) {
        struct ggml_tensor * node = gb->nodes[i];
6186

6187
6188
6189
6190
6191
6192
6193
        for (int j = 0; j < GGML_MAX_SRC; j++) {
            if (node->src[j]) {
                char label[16];
                snprintf(label, sizeof(label), "src %d", j);
                ggml_graph_dump_dot_node_edge(fp, gb, node, node->src[j], label);
            }
        }
6194
6195
    }

6196
6197
    for (int i = 0; i < gb->n_leafs; i++) {
        struct ggml_tensor * node = gb->leafs[i];
6198

6199
6200
6201
6202
6203
6204
6205
        for (int j = 0; j < GGML_MAX_SRC; j++) {
            if (node->src[j]) {
                char label[16];
                snprintf(label, sizeof(label), "src %d", j);
                ggml_graph_dump_dot_leaf_edge(fp, node, node->src[j], label);
            }
        }
6206
6207
    }

6208
    fprintf(fp, "}\n");
6209

6210
    fclose(fp);
6211

6212
    GGML_LOG_INFO("%s: dot -Tpng %s -o %s.png && open %s.png\n", __func__, filename, filename, filename);
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
}

////////////////////////////////////////////////////////////////////////////////

void ggml_set_input(struct ggml_tensor * tensor) {
    tensor->flags |= GGML_TENSOR_FLAG_INPUT;
}

void ggml_set_output(struct ggml_tensor * tensor) {
    tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
}

6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) {
    GGML_UNUSED(ctx); // TODO: remove this parameter
    tensor->flags |= GGML_TENSOR_FLAG_PARAM;
}

void ggml_set_loss(struct ggml_tensor * tensor) {
    GGML_ASSERT(ggml_is_scalar(tensor));
    GGML_ASSERT(tensor->type == GGML_TYPE_F32);
    tensor->flags |= GGML_TENSOR_FLAG_LOSS;
}

6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
////////////////////////////////////////////////////////////////////////////////

void ggml_quantize_init(enum ggml_type type) {
    ggml_critical_section_start();

    switch (type) {
        case GGML_TYPE_IQ2_XXS:
        case GGML_TYPE_IQ2_XS:
        case GGML_TYPE_IQ2_S:
        case GGML_TYPE_IQ1_S:
        case GGML_TYPE_IQ1_M:   iq2xs_init_impl(type); break;
        case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
        case GGML_TYPE_IQ3_S:   iq3xs_init_impl(512); break;
        default: // nothing
            break;
    }

    ggml_critical_section_end();
}

void ggml_quantize_free(void) {
    ggml_critical_section_start();

    iq2xs_free_impl(GGML_TYPE_IQ2_XXS);
    iq2xs_free_impl(GGML_TYPE_IQ2_XS);
    iq2xs_free_impl(GGML_TYPE_IQ1_S);
    iq3xs_free_impl(256);

    ggml_critical_section_end();
}

bool ggml_quantize_requires_imatrix(enum ggml_type type) {
    return
        type == GGML_TYPE_IQ2_XXS ||
        type == GGML_TYPE_IQ2_XS  ||
        type == GGML_TYPE_IQ1_S;//   ||
        //type == GGML_TYPE_IQ1_M;
}

size_t ggml_quantize_chunk(
        enum ggml_type   type,
           const float * src,
                  void * dst,
               int64_t   start,
               int64_t   nrows,
               int64_t   n_per_row,
           const float * imatrix) {
    const int64_t n = (int64_t) nrows * n_per_row;

    if (ggml_quantize_requires_imatrix(type)) {
        GGML_ASSERT(imatrix != NULL);
    }

    GGML_ASSERT(start % type_traits[type].blck_size == 0);
    GGML_ASSERT(start % n_per_row == 0);

    ggml_quantize_init(type); // this is noop if already initialized

    const size_t start_row = start / n_per_row;
    const size_t row_size  = ggml_row_size(type, n_per_row);

    size_t result = 0;

    switch (type) {
        case GGML_TYPE_Q4_0:    result = quantize_q4_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_1:    result = quantize_q4_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q5_0:    result = quantize_q5_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q5_1:    result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q8_0:    result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q5_K:    result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q6_K:    result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
6310
6311
        case GGML_TYPE_TQ1_0:   result = quantize_tq1_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_TQ2_0:   result = quantize_tq2_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
        case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ2_XS:  result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ3_S:   result = quantize_iq3_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_F16:
            {
                size_t elemsize = sizeof(ggml_fp16_t);
                ggml_fp32_to_fp16_row(src + start, (ggml_fp16_t *)dst + start, n);
                result = n * elemsize;
            } break;
        case GGML_TYPE_BF16:
            {
                size_t elemsize = sizeof(ggml_bf16_t);
                ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n);
                result = n * elemsize;
            } break;
        case GGML_TYPE_F32:
            {
                size_t elemsize = sizeof(float);
                result = n * elemsize;
                memcpy((uint8_t *)dst + start * elemsize, src + start, result);
            } break;
        default:
            assert(false);
    }

    GGML_ASSERT(result == nrows * row_size);

    return result;
}

////////////////////////////////////////////////////////////////////////////////

struct gguf_str {
    uint64_t n;  // GGUFv2
    char * data;
};

static const size_t GGUF_TYPE_SIZE[GGUF_TYPE_COUNT] = {
    [GGUF_TYPE_UINT8]   = sizeof(uint8_t),
    [GGUF_TYPE_INT8]    = sizeof(int8_t),
    [GGUF_TYPE_UINT16]  = sizeof(uint16_t),
    [GGUF_TYPE_INT16]   = sizeof(int16_t),
    [GGUF_TYPE_UINT32]  = sizeof(uint32_t),
    [GGUF_TYPE_INT32]   = sizeof(int32_t),
    [GGUF_TYPE_FLOAT32] = sizeof(float),
    [GGUF_TYPE_BOOL]    = sizeof(bool),
    [GGUF_TYPE_STRING]  = sizeof(struct gguf_str),
    [GGUF_TYPE_UINT64]  = sizeof(uint64_t),
    [GGUF_TYPE_INT64]   = sizeof(int64_t),
    [GGUF_TYPE_FLOAT64] = sizeof(double),
    [GGUF_TYPE_ARRAY]   = 0, // undefined
};
static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");

static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = {
    [GGUF_TYPE_UINT8]   = "u8",
    [GGUF_TYPE_INT8]    = "i8",
    [GGUF_TYPE_UINT16]  = "u16",
    [GGUF_TYPE_INT16]   = "i16",
    [GGUF_TYPE_UINT32]  = "u32",
    [GGUF_TYPE_INT32]   = "i32",
    [GGUF_TYPE_FLOAT32] = "f32",
    [GGUF_TYPE_BOOL]    = "bool",
    [GGUF_TYPE_STRING]  = "str",
    [GGUF_TYPE_ARRAY]   = "arr",
    [GGUF_TYPE_UINT64]  = "u64",
    [GGUF_TYPE_INT64]   = "i64",
    [GGUF_TYPE_FLOAT64] = "f64",
};
static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");

union gguf_value {
    uint8_t  uint8;
    int8_t   int8;
    uint16_t uint16;
    int16_t  int16;
    uint32_t uint32;
    int32_t  int32;
    float    float32;
    uint64_t uint64;
    int64_t  int64;
    double   float64;
    bool     bool_;

    struct gguf_str str;

    struct {
        enum gguf_type type;

        uint64_t n;  // GGUFv2
        void * data;
    } arr;
};

struct gguf_kv {
    struct gguf_str key;

    enum  gguf_type  type;
    union gguf_value value;
};

struct gguf_header {
    char magic[4];

    uint32_t version;
    uint64_t n_tensors; // GGUFv2
    uint64_t n_kv;      // GGUFv2
};

struct gguf_tensor_info {
    struct gguf_str name;

    uint32_t n_dims;
    uint64_t ne[GGML_MAX_DIMS];

    enum ggml_type type;

    uint64_t offset; // offset from start of `data`, must be a multiple of `ALIGNMENT`

    // for writing API
    const void * data;
    size_t size;
};

struct gguf_context {
    struct gguf_header header;

    struct gguf_kv          * kv;
    struct gguf_tensor_info * infos;

    size_t alignment;
    size_t offset;    // offset of `data` from beginning of file
    size_t size;      // size of `data` in bytes

    //uint8_t * padding;
    void * data;
};

static size_t gguf_type_size(enum gguf_type type) {
    GGML_ASSERT(0 <= type && type < GGUF_TYPE_COUNT);
    return GGUF_TYPE_SIZE[type];
}

6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
static bool gguf_tensor_info_sanitize(struct gguf_tensor_info * info) {
    if (info->n_dims > GGML_MAX_DIMS) {
        fprintf(stderr, "%s: invalid number of dimensions (%" PRIu32 ")\n", __func__, info->n_dims);
        return false;
    }

    if (info->type < 0 || info->type >= GGML_TYPE_COUNT) {
        fprintf(stderr, "%s: invalid type (%d)\n", __func__, info->type);
        return false;
    }

    if (strlen(info->name.data) >= GGML_MAX_NAME) {
        fprintf(stderr, "%s: tensor '%s' name is too long\n", __func__, info->name.data);
        return false;
    }
6479
6480

    for (uint32_t i = 0; i < info->n_dims; ++i) {
6481
6482
6483
6484
        if (info->ne[i] <= 0) {
            fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[i]);
            return false;
        }
6485
6486
6487
    }

    // prevent overflow for total number of elements
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
    if (INT64_MAX/info->ne[1] <= info->ne[0]) {
        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[1]);
        return false;
    }

    if (INT64_MAX/info->ne[2] <= info->ne[0]*info->ne[1]) {
        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[2]);
        return false;
    }

    if (INT64_MAX/info->ne[3] <= info->ne[0]*info->ne[1]*info->ne[2]) {
        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[3]);
        return false;
    }

    return true;
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
}

static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) {
    const size_t n = fread(dst, 1, size, file);
    *offset += n;
    return n == size;
}

static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
    p->n    = 0;
    p->data = NULL;

    bool ok = true;

    ok = ok && gguf_fread_el(file, &p->n, sizeof(p->n), offset);

    // early exit if string length is invalid, prevents from integer overflow
    if (p->n == SIZE_MAX) {
        fprintf(stderr, "%s: invalid string length (%" PRIu64 ")\n", __func__, p->n);
        return false;
    }

6526
6527
6528
6529
6530
    p->data = calloc(p->n + 1, 1);
    if (!p->data) {
        fprintf(stderr, "%s: failed to allocate memory for string of length %" PRIu64 "\n", __func__, p->n);
        return false;
    }
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563

    ok = ok && gguf_fread_el(file,  p->data, p->n, offset);

    return ok;
}

static void gguf_free_kv(struct gguf_kv * kv) {
    if (kv->key.data) {
        GGML_FREE(kv->key.data);
    }

    if (kv->type == GGUF_TYPE_STRING) {
        if (kv->value.str.data) {
            GGML_FREE(kv->value.str.data);
        }
    }

    if (kv->type == GGUF_TYPE_ARRAY) {
        if (kv->value.arr.data) {
            if (kv->value.arr.type == GGUF_TYPE_STRING) {
                for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                    struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
                    if (str->data) {
                        GGML_FREE(str->data);
                    }
                }
            }
            GGML_FREE(kv->value.arr.data);
        }
    }
}

struct gguf_context * gguf_init_empty(void) {
6564
6565
6566
6567
6568
    struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
    if (!ctx) {
        fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
        return NULL;
    }
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613

    memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
    ctx->header.version   = GGUF_VERSION;
    ctx->header.n_tensors = 0;
    ctx->header.n_kv      = 0;

    ctx->kv    = NULL;
    ctx->infos = NULL;

    ctx->alignment = GGUF_DEFAULT_ALIGNMENT;
    ctx->offset    = 0;
    ctx->size      = 0;

    ctx->data = NULL;

    return ctx;
}

struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) {
    FILE * file = ggml_fopen(fname, "rb");
    if (!file) {
        fprintf(stderr, "%s: failed to open '%s': '%s'\n", __func__, fname, strerror(errno));
        return NULL;
    }

    // offset from start of file
    size_t offset = 0;

    char magic[4];

    // check the magic before making allocations
    {
        gguf_fread_el(file, &magic, sizeof(magic), &offset);

        for (uint32_t i = 0; i < sizeof(magic); i++) {
            if (magic[i] != GGUF_MAGIC[i]) {
                fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                fclose(file);
                return NULL;
            }
        }
    }

    bool ok = true;

6614
6615
6616
6617
6618
6619
    struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
    if (!ctx) {
        fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
        fclose(file);
        return NULL;
    }
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657

    // read the header
    {
        strncpy(ctx->header.magic, magic, 4);

        ctx->kv    = NULL;
        ctx->infos = NULL;
        ctx->data  = NULL;

        ok = ok && gguf_fread_el(file, &ctx->header.version,   sizeof(ctx->header.version),   &offset);
        ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset);
        ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);

        if (ctx->header.version == 1) {
            fprintf(stderr, "%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }

        // sanity-checks to prevent from integer/buffer overflows

        ok = ok && (ctx->header.n_tensors < (SIZE_MAX/2)/sizeof(struct gguf_tensor_info));
        ok = ok && (ctx->header.n_tensors < (SIZE_MAX/2)/ggml_tensor_overhead());
        ok = ok && (ctx->header.n_kv      < (SIZE_MAX/2)/sizeof(struct gguf_kv));

        if (!ok) {
            fprintf(stderr, "%s: failed to read header\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }
    }

    // read the kv pairs
    {
        const uint64_t n_kv = ctx->header.n_kv;

6658
6659
6660
6661
6662
6663
6664
        ctx->kv = calloc(n_kv, sizeof(struct gguf_kv));
        if (!ctx->kv) {
            fprintf(stderr, "%s: failed to allocate memory for kv pairs\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714

        for (uint64_t i = 0; i < n_kv; ++i) {
            struct gguf_kv * kv = &ctx->kv[i];

            //fprintf(stderr, "%s: reading kv %d\n", __func__, i);

            ok = ok && gguf_fread_str(file, &kv->key,                    &offset);
            ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset);

            //fprintf(stderr, "%s: reading kv with key %s\n", __func__, kv->key.data);

            switch (kv->type) {
                case GGUF_TYPE_UINT8:   ok = ok && gguf_fread_el (file, &kv->value.uint8,   sizeof(kv->value.uint8),   &offset); break;
                case GGUF_TYPE_INT8:    ok = ok && gguf_fread_el (file, &kv->value.int8,    sizeof(kv->value.int8),    &offset); break;
                case GGUF_TYPE_UINT16:  ok = ok && gguf_fread_el (file, &kv->value.uint16,  sizeof(kv->value.uint16),  &offset); break;
                case GGUF_TYPE_INT16:   ok = ok && gguf_fread_el (file, &kv->value.int16,   sizeof(kv->value.int16),   &offset); break;
                case GGUF_TYPE_UINT32:  ok = ok && gguf_fread_el (file, &kv->value.uint32,  sizeof(kv->value.uint32),  &offset); break;
                case GGUF_TYPE_INT32:   ok = ok && gguf_fread_el (file, &kv->value.int32,   sizeof(kv->value.int32),   &offset); break;
                case GGUF_TYPE_FLOAT32: ok = ok && gguf_fread_el (file, &kv->value.float32, sizeof(kv->value.float32), &offset); break;
                case GGUF_TYPE_UINT64:  ok = ok && gguf_fread_el (file, &kv->value.uint64,  sizeof(kv->value.uint64),  &offset); break;
                case GGUF_TYPE_INT64:   ok = ok && gguf_fread_el (file, &kv->value.int64,   sizeof(kv->value.int64),   &offset); break;
                case GGUF_TYPE_FLOAT64: ok = ok && gguf_fread_el (file, &kv->value.float64, sizeof(kv->value.float64), &offset); break;
                case GGUF_TYPE_BOOL:    ok = ok && gguf_fread_el (file, &kv->value.bool_,   sizeof(kv->value.bool_),   &offset); break;
                case GGUF_TYPE_STRING:  ok = ok && gguf_fread_str(file, &kv->value.str,                                &offset); break;
                case GGUF_TYPE_ARRAY:
                    {
                        ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset);
                        ok = ok && gguf_fread_el(file, &kv->value.arr.n,    sizeof(kv->value.arr.n),    &offset);

                        switch (kv->value.arr.type) {
                            case GGUF_TYPE_UINT8:
                            case GGUF_TYPE_INT8:
                            case GGUF_TYPE_UINT16:
                            case GGUF_TYPE_INT16:
                            case GGUF_TYPE_UINT32:
                            case GGUF_TYPE_INT32:
                            case GGUF_TYPE_FLOAT32:
                            case GGUF_TYPE_UINT64:
                            case GGUF_TYPE_INT64:
                            case GGUF_TYPE_FLOAT64:
                            case GGUF_TYPE_BOOL:
                                {
                                    // prevent from integer overflow in the malloc below
                                    if (kv->value.arr.n >= SIZE_MAX/gguf_type_size(kv->value.arr.type)) {
                                        fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
                                        fclose(file);
                                        gguf_free(ctx);
                                        return NULL;
                                    }

6715
6716
6717
6718
6719
6720
6721
                                    kv->value.arr.data = calloc(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
                                    if (!kv->value.arr.data) {
                                        fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
                                        fclose(file);
                                        gguf_free(ctx);
                                        return NULL;
                                    }
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734

                                    ok = ok && gguf_fread_el(file, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type), &offset);
                                } break;
                            case GGUF_TYPE_STRING:
                                {
                                    // prevent from integer overflow in the malloc below
                                    if (kv->value.arr.n >= SIZE_MAX/sizeof(struct gguf_str)) {
                                        fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
                                        fclose(file);
                                        gguf_free(ctx);
                                        return NULL;
                                    }

6735
6736
6737
6738
6739
6740
6741
                                    kv->value.arr.data = calloc(kv->value.arr.n, sizeof(struct gguf_str));
                                    if (!kv->value.arr.data) {
                                        fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
                                        fclose(file);
                                        gguf_free(ctx);
                                        return NULL;
                                    }
6742
6743
6744
6745
6746
6747

                                    for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                                        ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset);
                                    }
                                } break;
                            case GGUF_TYPE_ARRAY:
6748
6749
6750
6751
6752
                            default:
                                {
                                    fprintf(stderr, "%s: invalid array type %d\n", __func__, kv->value.arr.type);
                                    ok = false;
                                } break;
6753
6754
                        }
                    } break;
6755
6756
6757
6758
6759
                default:
                    {
                        fprintf(stderr, "%s: invalid type %d\n", __func__, kv->type);
                        ok = false;
                    } break;
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
            }

            if (!ok) {
                break;
            }
        }

        if (!ok) {
            fprintf(stderr, "%s: failed to read key-value pairs\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }
    }

    // read the tensor infos
    if (ctx->header.n_tensors > 0) {
6777
6778
6779
6780
6781
6782
6783
        ctx->infos = calloc(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
        if (!ctx->infos) {
            fprintf(stderr, "%s: failed to allocate memory for tensor infos\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803

        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
            struct gguf_tensor_info * info = &ctx->infos[i];

            for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                info->ne[j] = 1;
            }

            ok = ok && gguf_fread_str(file, &info->name,                          &offset);
            ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims),  &offset);

            ok = ok && (info->n_dims <= GGML_MAX_DIMS);

            for (uint32_t j = 0; j < info->n_dims; ++j) {
                ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
            }

            ok = ok && gguf_fread_el (file, &info->type,   sizeof(info->type),    &offset);
            ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset),  &offset);

6804
            ok = ok && gguf_tensor_info_sanitize(info);
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625

            // make sure there is no duplicated tensor names
            for (uint64_t j = 0; j < i && ok; ++j) {
                if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) {
                    fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data);
                    ok = false;
                }
            }

            if (!ok) {
                fprintf(stderr, "%s: failed to read tensor info\n", __func__);
                fclose(file);
                gguf_free(ctx);
                return NULL;
            }
        }
    }

    ctx->alignment = GGUF_DEFAULT_ALIGNMENT;

    int alignment_idx = gguf_find_key(ctx, "general.alignment");
    if (alignment_idx != -1) {
        ctx->alignment = gguf_get_val_u32(ctx, alignment_idx);
    }

    // we require the data section to be aligned, so take into account any padding
    {
        const size_t offset_pad = offset % ctx->alignment;

        if (offset_pad != 0) {
            offset += ctx->alignment - offset_pad;
            fseek(file, offset, SEEK_SET);
        }
    }

    // store the current file offset - this is where the data section starts
    ctx->offset = offset;

    // compute the total size of the data section, taking into account the alignment
    {
        ctx->size = 0;
        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
            struct gguf_tensor_info * info = &ctx->infos[i];

            const int64_t ne =
                (int64_t) info->ne[0] *
                (int64_t) info->ne[1] *
                (int64_t) info->ne[2] *
                (int64_t) info->ne[3];

            if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
                fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                        __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                fclose(file);
                gguf_free(ctx);
                return NULL;
            }

            const size_t size_cur = ggml_row_size(info->type, ne);

            ctx->size += GGML_PAD(size_cur, ctx->alignment);
        }
    }

    // load the tensor data only if requested
    if (params.ctx != NULL) {
        // if the provided gguf_context is no_alloc, then we create "empty" tensors and do not read the binary blob
        // otherwise, we load the binary blob into the created ggml_context as well, and point the "data" members of
        // the ggml_tensor structs to the appropriate locations in the binary blob

        // compute the exact size needed for the new ggml_context
        const size_t mem_size =
            params.no_alloc ?
            (ctx->header.n_tensors    )*ggml_tensor_overhead() :
            (ctx->header.n_tensors + 1)*ggml_tensor_overhead() + ctx->size;

        struct ggml_init_params pdata = {
            .mem_size   = mem_size,
            .mem_buffer = NULL,
            .no_alloc   = params.no_alloc,
        };

        *params.ctx = ggml_init(pdata);
        if (*params.ctx == NULL) {
            fprintf(stderr, "%s: failed to initialize context\n", __func__);
            fclose(file);
            gguf_free(ctx);
            return NULL;
        }

        struct ggml_context * ctx_data = *params.ctx;

        struct ggml_tensor * data = NULL;

        if (!params.no_alloc) {
            data = ggml_new_tensor_1d(ctx_data, GGML_TYPE_I8, ctx->size);

            ok = ok && data != NULL;

            // read the binary blob with the tensor data
            ok = ok && gguf_fread_el(file, data->data, ctx->size, &offset);

            if (!ok) {
                fprintf(stderr, "%s: failed to read tensor data\n", __func__);
                fclose(file);
                ggml_free(ctx_data);
                gguf_free(ctx);
                return NULL;
            }

            ctx->data = data->data;
        }

        ggml_set_no_alloc(ctx_data, true);

        // create the tensors
        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
            const int64_t ne[GGML_MAX_DIMS] = {
                ctx->infos[i].ne[0],
                ctx->infos[i].ne[1],
                ctx->infos[i].ne[2],
                ctx->infos[i].ne[3],
            };

            struct ggml_tensor * cur = ggml_new_tensor(ctx_data, ctx->infos[i].type, ctx->infos[i].n_dims, ne);

            ok = ok && cur != NULL;

            if (!ok) {
                break;
            }

            ggml_set_name(cur, ctx->infos[i].name.data);

            // point the data member to the appropriate location in the binary blob using the tensor infos
            if (!params.no_alloc) {
              //cur->data = (char *) data->data + ctx->infos[i].offset - ctx->offset; // offset from start of file
                cur->data = (char *) data->data + ctx->infos[i].offset;               // offset from data
            }
        }

        if (!ok) {
            fprintf(stderr, "%s: failed to read the tensor data\n", __func__);
            fclose(file);
            ggml_free(ctx_data);
            gguf_free(ctx);
            return NULL;
        }

        ggml_set_no_alloc(ctx_data, params.no_alloc);
    }

    fclose(file);

    return ctx;
}

void gguf_free(struct gguf_context * ctx) {
    if (ctx == NULL) {
        return;
    }

    if (ctx->kv) {
        // free string memory - not great..
        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
            gguf_free_kv(&ctx->kv[i]);
        }

        GGML_FREE(ctx->kv);
    }

    if (ctx->infos) {
        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
            struct gguf_tensor_info * info = &ctx->infos[i];

            if (info->name.data) {
                GGML_FREE(info->name.data);
            }
        }

        GGML_FREE(ctx->infos);
    }

    GGML_FREE(ctx);
}

const char * gguf_type_name(enum gguf_type type) {
    return GGUF_TYPE_NAME[type];
}

int gguf_get_version(const struct gguf_context * ctx) {
    return ctx->header.version;
}

size_t gguf_get_alignment(const struct gguf_context * ctx) {
    return ctx->alignment;
}

size_t gguf_get_data_offset(const struct gguf_context * ctx) {
    return ctx->offset;
}

void * gguf_get_data(const struct gguf_context * ctx) {
    return ctx->data;
}

int gguf_get_n_kv(const struct gguf_context * ctx) {
    return ctx->header.n_kv;
}

int gguf_find_key(const struct gguf_context * ctx, const char * key) {
    // return -1 if key not found
    int keyfound = -1;

    const int n_kv = gguf_get_n_kv(ctx);

    for (int i = 0; i < n_kv; ++i) {
        if (strcmp(key, gguf_get_key(ctx, i)) == 0) {
            keyfound = i;
            break;
        }
    }

    return keyfound;
}

const char * gguf_get_key(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    return ctx->kv[key_id].key.data;
}

enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    return ctx->kv[key_id].type;
}

enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
    return ctx->kv[key_id].value.arr.type;
}

const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
    return ctx->kv[key_id].value.arr.data;
}

const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
    struct gguf_kv * kv = &ctx->kv[key_id];
    struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i];
    return str->data;
}

int gguf_get_arr_n(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
    return ctx->kv[key_id].value.arr.n;
}

uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT8);
    return ctx->kv[key_id].value.uint8;
}

int8_t gguf_get_val_i8(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT8);
    return ctx->kv[key_id].value.int8;
}

uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT16);
    return ctx->kv[key_id].value.uint16;
}

int16_t gguf_get_val_i16(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT16);
    return ctx->kv[key_id].value.int16;
}

uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT32);
    return ctx->kv[key_id].value.uint32;
}

int32_t gguf_get_val_i32(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT32);
    return ctx->kv[key_id].value.int32;
}

float gguf_get_val_f32(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT32);
    return ctx->kv[key_id].value.float32;
}

uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT64);
    return ctx->kv[key_id].value.uint64;
}

int64_t gguf_get_val_i64(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT64);
    return ctx->kv[key_id].value.int64;
}

double gguf_get_val_f64(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT64);
    return ctx->kv[key_id].value.float64;
}

bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_BOOL);
    return ctx->kv[key_id].value.bool_;
}

const char * gguf_get_val_str(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_STRING);
    return ctx->kv[key_id].value.str.data;
}

const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id) {
    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
    GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_ARRAY);
    GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_STRING);
    return &ctx->kv[key_id].value;
}

int gguf_get_n_tensors(const struct gguf_context * ctx) {
    return ctx->header.n_tensors;
}

int gguf_find_tensor(const struct gguf_context * ctx, const char * name) {
    // return -1 if tensor not found
    int tensorfound = -1;

    const int n_tensors = gguf_get_n_tensors(ctx);

    for (int i = 0; i < n_tensors; ++i) {
        if (strcmp(name, gguf_get_tensor_name(ctx, i)) == 0) {
            tensorfound = i;
            break;
        }
    }

    return tensorfound;
}

size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i) {
    return ctx->infos[i].offset;
}

char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) {
    return ctx->infos[i].name.data;
}

enum ggml_type gguf_get_tensor_type(const struct gguf_context * ctx, int i) {
    return ctx->infos[i].type;
}

// returns the index
static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
    const int idx = gguf_find_key(ctx, key);
    if (idx >= 0) {
        return idx;
    }

    const int n_kv = gguf_get_n_kv(ctx);

    ctx->kv = realloc(ctx->kv, (n_kv + 1) * sizeof(struct gguf_kv));
    ctx->kv[n_kv].key.n    = strlen(key);
    ctx->kv[n_kv].key.data = strdup(key);
    ctx->header.n_kv++;

    return n_kv;
}

void gguf_remove_key(struct gguf_context * ctx, const char * key) {
    const int idx = gguf_find_key(ctx, key);
    if (idx >= 0) {
        const int n_kv = gguf_get_n_kv(ctx);
        gguf_free_kv(&ctx->kv[idx]);
        for (int i = idx; i < n_kv-1; ++i) {
            ctx->kv[i] = ctx->kv[i+1];
        }
        ctx->kv = realloc(ctx->kv, (n_kv - 1) * sizeof(struct gguf_kv));
        ctx->header.n_kv--;
    }
}

void gguf_set_val_u8(struct gguf_context * ctx, const char * key, uint8_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type        = GGUF_TYPE_UINT8;
    ctx->kv[idx].value.uint8 = val;
}

void gguf_set_val_i8(struct gguf_context * ctx, const char * key, int8_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type       = GGUF_TYPE_INT8;
    ctx->kv[idx].value.int8 = val;
}

void gguf_set_val_u16(struct gguf_context * ctx, const char * key, uint16_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type         = GGUF_TYPE_UINT16;
    ctx->kv[idx].value.uint16 = val;
}

void gguf_set_val_i16(struct gguf_context * ctx, const char * key, int16_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type        = GGUF_TYPE_INT16;
    ctx->kv[idx].value.int16 = val;
}

void gguf_set_val_u32(struct gguf_context * ctx, const char * key, uint32_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type         = GGUF_TYPE_UINT32;
    ctx->kv[idx].value.uint32 = val;
}

void gguf_set_val_i32(struct gguf_context * ctx, const char * key, int32_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type        = GGUF_TYPE_INT32;
    ctx->kv[idx].value.int32 = val;
}

void gguf_set_val_f32(struct gguf_context * ctx, const char * key, float val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type          = GGUF_TYPE_FLOAT32;
    ctx->kv[idx].value.float32 = val;
}

void gguf_set_val_u64(struct gguf_context * ctx, const char * key, uint64_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type         = GGUF_TYPE_UINT64;
    ctx->kv[idx].value.uint64 = val;
}

void gguf_set_val_i64(struct gguf_context * ctx, const char * key, int64_t val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type        = GGUF_TYPE_INT64;
    ctx->kv[idx].value.int64 = val;
}

void gguf_set_val_f64(struct gguf_context * ctx, const char * key, double val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type          = GGUF_TYPE_FLOAT64;
    ctx->kv[idx].value.float64 = val;
}

void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type        = GGUF_TYPE_BOOL;
    ctx->kv[idx].value.bool_ = val;
}

void gguf_set_val_str(struct gguf_context * ctx, const char * key, const char * val) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type           = GGUF_TYPE_STRING;
    ctx->kv[idx].value.str.n    = strlen(val);
    ctx->kv[idx].value.str.data = strdup(val);
}

void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type           = GGUF_TYPE_ARRAY;
    ctx->kv[idx].value.arr.type = type;
    ctx->kv[idx].value.arr.n    = n;
    ctx->kv[idx].value.arr.data = GGML_CALLOC(n, gguf_type_size(type));
    memcpy(ctx->kv[idx].value.arr.data, data, n*gguf_type_size(type));
}

void gguf_set_arr_str(struct gguf_context * ctx, const char * key, const char ** data, int n) {
    const int idx = gguf_get_or_add_key(ctx, key);

    ctx->kv[idx].type           = GGUF_TYPE_ARRAY;
    ctx->kv[idx].value.arr.type = GGUF_TYPE_STRING;
    ctx->kv[idx].value.arr.n    = n;
    ctx->kv[idx].value.arr.data = GGML_CALLOC(n, sizeof(struct gguf_str));
    for (int i = 0; i < n; i++) {
        struct gguf_str * str = &((struct gguf_str *)ctx->kv[idx].value.arr.data)[i];
        str->n    = strlen(data[i]);
        str->data = strdup(data[i]);
    }
}

// set or add KV pairs from another context
void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
    for (uint32_t i = 0; i < src->header.n_kv; i++) {
        switch (src->kv[i].type) {
            case GGUF_TYPE_UINT8:   gguf_set_val_u8  (ctx, src->kv[i].key.data, src->kv[i].value.uint8);    break;
            case GGUF_TYPE_INT8:    gguf_set_val_i8  (ctx, src->kv[i].key.data, src->kv[i].value.int8);     break;
            case GGUF_TYPE_UINT16:  gguf_set_val_u16 (ctx, src->kv[i].key.data, src->kv[i].value.uint16);   break;
            case GGUF_TYPE_INT16:   gguf_set_val_i16 (ctx, src->kv[i].key.data, src->kv[i].value.int16);    break;
            case GGUF_TYPE_UINT32:  gguf_set_val_u32 (ctx, src->kv[i].key.data, src->kv[i].value.uint32);   break;
            case GGUF_TYPE_INT32:   gguf_set_val_i32 (ctx, src->kv[i].key.data, src->kv[i].value.int32);    break;
            case GGUF_TYPE_FLOAT32: gguf_set_val_f32 (ctx, src->kv[i].key.data, src->kv[i].value.float32);  break;
            case GGUF_TYPE_UINT64:  gguf_set_val_u64 (ctx, src->kv[i].key.data, src->kv[i].value.uint64);   break;
            case GGUF_TYPE_INT64:   gguf_set_val_i64 (ctx, src->kv[i].key.data, src->kv[i].value.int64);    break;
            case GGUF_TYPE_FLOAT64: gguf_set_val_f64 (ctx, src->kv[i].key.data, src->kv[i].value.float64);  break;
            case GGUF_TYPE_BOOL:    gguf_set_val_bool(ctx, src->kv[i].key.data, src->kv[i].value.bool_);    break;
            case GGUF_TYPE_STRING:  gguf_set_val_str (ctx, src->kv[i].key.data, src->kv[i].value.str.data); break;
            case GGUF_TYPE_ARRAY:
                {
                    if (src->kv[i].value.arr.type == GGUF_TYPE_STRING) {
                        const char ** data = GGML_CALLOC(src->kv[i].value.arr.n, sizeof(char *));
                        for (uint32_t j = 0; j < src->kv[i].value.arr.n; j++) {
                            data[j] = ((struct gguf_str *)src->kv[i].value.arr.data)[j].data;
                        }
                        gguf_set_arr_str(ctx, src->kv[i].key.data, data, src->kv[i].value.arr.n);
                        GGML_FREE((void *)data);
                    } else if (src->kv[i].value.arr.type == GGUF_TYPE_ARRAY) {
                        GGML_ABORT("nested arrays not supported");
                    } else {
                        gguf_set_arr_data(ctx, src->kv[i].key.data, src->kv[i].value.arr.type, src->kv[i].value.arr.data, src->kv[i].value.arr.n);
                    }
                } break;
            default: GGML_ABORT("invalid type");
        }
    }
}

void gguf_add_tensor(
             struct gguf_context * ctx,
        const struct ggml_tensor * tensor) {
    GGML_ASSERT(tensor);
    if (gguf_find_tensor(ctx, tensor->name) != -1) {
        GGML_ABORT("duplicated tensor name");
    }

    const int idx = ctx->header.n_tensors;
    ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info));

    ctx->infos[idx].name.n    = strlen(tensor->name);
    ctx->infos[idx].name.data = strdup(tensor->name);

    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
        ctx->infos[idx].ne[i] = 1;
    }

    ctx->infos[idx].n_dims = ggml_n_dims(tensor);
    for (uint32_t i = 0; i < ctx->infos[idx].n_dims; i++) {
        ctx->infos[idx].ne[i] = tensor->ne[i];
    }

    ctx->infos[idx].type   = tensor->type;
    ctx->infos[idx].offset = 0;
    ctx->infos[idx].data   = tensor->data;
    ctx->infos[idx].size   = ggml_nbytes(tensor);

    if (ctx->header.n_tensors > 0) {
        ctx->infos[idx].offset = ctx->infos[idx - 1].offset + GGML_PAD(ctx->infos[idx - 1].size, ctx->alignment);
    }

    ctx->header.n_tensors++;
}

void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type) {
    const int idx = gguf_find_tensor(ctx, name);
    if (idx < 0) {
        GGML_ABORT("tensor not found");
    }

    ctx->infos[idx].type = type;
}

void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size) {
    const int idx = gguf_find_tensor(ctx, name);
    if (idx < 0) {
        GGML_ABORT("tensor not found");
    }

    ctx->infos[idx].data = data;
    ctx->infos[idx].size = size;

    // update offsets
    for (uint32_t i = idx + 1; i < ctx->header.n_tensors; ++i) {
        ctx->infos[i].offset = ctx->infos[i - 1].offset + GGML_PAD(ctx->infos[i - 1].size, ctx->alignment);
    }
}

//static void gguf_fwrite_str(FILE * file, const struct gguf_str * val) {
//    fwrite(&val->n,   sizeof(val->n),    1, file);
//    fwrite(val->data, sizeof(char), val->n, file);
//}
//
//static void gguf_fwrite_el(FILE * file, const void * val, size_t size) {
//    fwrite(val, sizeof(char), size, file);
//}

struct gguf_buf {
    void * data;
    size_t size;
    size_t offset;
};

static struct gguf_buf gguf_buf_init(size_t size) {
    struct gguf_buf buf = {
        /*buf.data   =*/ size == 0 ? NULL : GGML_CALLOC(1, size),
        /*buf.size   =*/ size,
        /*buf.offset =*/ 0,
    };

    return buf;
}

static void gguf_buf_free(struct gguf_buf buf) {
    if (buf.data) {
        GGML_FREE(buf.data);
    }
}

static void gguf_buf_grow(struct gguf_buf * buf, size_t size) {
    if (buf->offset + size > buf->size) {
        buf->size = 1.5*(buf->offset + size);
        if (buf->data) {
            buf->data = realloc(buf->data, buf->size);
        }
    }
}

static void gguf_bwrite_str(struct gguf_buf * buf, const struct gguf_str * val) {
    gguf_buf_grow(buf, sizeof(val->n) + val->n);

    if (buf->data) {
        memcpy((char *) buf->data + buf->offset, &val->n, sizeof(val->n));
    }
    buf->offset += sizeof(val->n);

    if (buf->data) {
        memcpy((char *) buf->data + buf->offset, val->data, val->n);
    }
    buf->offset += val->n;
}

static void gguf_bwrite_el(struct gguf_buf * buf, const void * val, size_t el_size) {
    gguf_buf_grow(buf, el_size);

    if (buf->data) {
        memcpy((char *) buf->data + buf->offset, val, el_size);
    }
    buf->offset += el_size;
}

static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta) {
    // write header
    gguf_bwrite_el(buf, &ctx->header.magic,     sizeof(ctx->header.magic));
    gguf_bwrite_el(buf, &ctx->header.version,   sizeof(ctx->header.version));
    gguf_bwrite_el(buf, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors));
    gguf_bwrite_el(buf, &ctx->header.n_kv,      sizeof(ctx->header.n_kv));

    // write key-value pairs
    for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
        struct gguf_kv * kv = &ctx->kv[i];

        gguf_bwrite_str(buf, &kv->key);
        gguf_bwrite_el (buf, &kv->type, sizeof(kv->type));

        switch (kv->type) {
            case GGUF_TYPE_UINT8:   gguf_bwrite_el( buf, &kv->value.uint8,   sizeof(kv->value.uint8)  ); break;
            case GGUF_TYPE_INT8:    gguf_bwrite_el (buf, &kv->value.int8,    sizeof(kv->value.int8)   ); break;
            case GGUF_TYPE_UINT16:  gguf_bwrite_el (buf, &kv->value.uint16,  sizeof(kv->value.uint16) ); break;
            case GGUF_TYPE_INT16:   gguf_bwrite_el (buf, &kv->value.int16,   sizeof(kv->value.int16)  ); break;
            case GGUF_TYPE_UINT32:  gguf_bwrite_el (buf, &kv->value.uint32,  sizeof(kv->value.uint32) ); break;
            case GGUF_TYPE_INT32:   gguf_bwrite_el (buf, &kv->value.int32,   sizeof(kv->value.int32)  ); break;
            case GGUF_TYPE_FLOAT32: gguf_bwrite_el (buf, &kv->value.float32, sizeof(kv->value.float32)); break;
            case GGUF_TYPE_UINT64:  gguf_bwrite_el (buf, &kv->value.uint64,  sizeof(kv->value.uint64) ); break;
            case GGUF_TYPE_INT64:   gguf_bwrite_el (buf, &kv->value.int64,   sizeof(kv->value.int64)  ); break;
            case GGUF_TYPE_FLOAT64: gguf_bwrite_el (buf, &kv->value.float64, sizeof(kv->value.float64)); break;
            case GGUF_TYPE_BOOL:    gguf_bwrite_el (buf, &kv->value.bool_,   sizeof(kv->value.bool_)  ); break;
            case GGUF_TYPE_STRING:  gguf_bwrite_str(buf, &kv->value.str                               ); break;
            case GGUF_TYPE_ARRAY:
                {
                    gguf_bwrite_el(buf, &kv->value.arr.type, sizeof(kv->value.arr.type));
                    gguf_bwrite_el(buf, &kv->value.arr.n,    sizeof(kv->value.arr.n)   );

                    switch (kv->value.arr.type) {
                        case GGUF_TYPE_UINT8:
                        case GGUF_TYPE_INT8:
                        case GGUF_TYPE_UINT16:
                        case GGUF_TYPE_INT16:
                        case GGUF_TYPE_UINT32:
                        case GGUF_TYPE_INT32:
                        case GGUF_TYPE_FLOAT32:
                        case GGUF_TYPE_UINT64:
                        case GGUF_TYPE_INT64:
                        case GGUF_TYPE_FLOAT64:
                        case GGUF_TYPE_BOOL:
                            {
                                gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type));
                            } break;
                        case GGUF_TYPE_STRING:
                            {
                                for (uint32_t j = 0; j < kv->value.arr.n; ++j) {
                                    gguf_bwrite_str(buf, &((struct gguf_str *) kv->value.arr.data)[j]);
                                }
                            } break;
                        case GGUF_TYPE_ARRAY:
                        default: GGML_ABORT("invalid type");
                    }
                } break;
            default: GGML_ABORT("invalid type");
        }
    }

    // write tensor infos
    for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
        struct gguf_tensor_info * info = &ctx->infos[i];

        gguf_bwrite_str(buf, &info->name);
        gguf_bwrite_el (buf, &info->n_dims, sizeof(info->n_dims));
        for (uint32_t j = 0; j < info->n_dims; ++j) {
            gguf_bwrite_el(buf, &info->ne[j], sizeof(info->ne[j]));
        }
        gguf_bwrite_el(buf, &info->type,   sizeof(info->type));
        gguf_bwrite_el(buf, &info->offset, sizeof(info->offset));
    }

    // we require the data section to be aligned, so take into account any padding
    {
        const size_t offset     = buf->offset;
        const size_t offset_pad = GGML_PAD(offset, ctx->alignment);

        if (offset_pad != offset) {
            uint8_t pad = 0;
            for (size_t i = 0; i < offset_pad - offset; ++i) {
                gguf_bwrite_el(buf, &pad, sizeof(pad));
            }
        }
    }

    if (only_meta) {
        return;
    }

    size_t offset = 0;

    // write tensor data
    for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
        struct gguf_tensor_info * info = &ctx->infos[i];

        const size_t size     = info->size;
        const size_t size_pad = GGML_PAD(size, ctx->alignment);

        gguf_bwrite_el(buf, info->data, size);

        if (size_pad != size) {
            uint8_t pad = 0;
            for (size_t j = 0; j < size_pad - size; ++j) {
                gguf_bwrite_el(buf, &pad, sizeof(pad));
            }
        }

        GGML_ASSERT(offset == info->offset);

        offset += size_pad;
    }
}

void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
    FILE * file = ggml_fopen(fname, "wb");
    if (!file) {
        GGML_ABORT("failed to open file for writing");
    }

    struct gguf_buf buf = gguf_buf_init(16*1024);

    gguf_write_to_buf(ctx, &buf, only_meta);

    fwrite(buf.data, 1, buf.offset, file);

    gguf_buf_free(buf);

    fclose(file);
}

size_t gguf_get_meta_size(const struct gguf_context * ctx) {
    // no allocs - only compute size
    struct gguf_buf buf = gguf_buf_init(0);

    gguf_write_to_buf(ctx, &buf, true);

    return buf.offset;
}

void gguf_get_meta_data(const struct gguf_context * ctx, void * data) {
    struct gguf_buf buf = gguf_buf_init(16*1024);

    gguf_write_to_buf(ctx, &buf, true);

    memcpy(data, buf.data, buf.offset);

    gguf_buf_free(buf);
}

7626
7627
7628
void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
    g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
    g_logger_state.log_callback_user_data = user_data;
7629
7630
}

7631
7632
7633
7634
7635
7636
7637
void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
    p->n_threads  = n_threads;
    p->prio       = 0;     // default priority (usually means normal or inherited)
    p->poll       = 50;    // hybrid-polling enabled
    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
    p->paused     = false; // threads are ready to go
    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
7638
7639
}

7640
7641
7642
7643
struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
    struct ggml_threadpool_params p;
    ggml_threadpool_params_init(&p, n_threads);
    return p;
7644
7645
}

7646
7647
7648
7649
7650
7651
bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
    if (p0->n_threads      != p1->n_threads  )    return false;
    if (p0->prio           != p1->prio       )    return false;
    if (p0->poll           != p1->poll       )    return false;
    if (p0->strict_cpu     != p1->strict_cpu )    return false;
    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
7652
}