Repository navigation
Expand file tree
/
Copy pathmtmd_cpp.py
More file actions
2081 lines (1869 loc) · 70 KB
/
Copy pathmtmd_cpp.py
File metadata and controls
2081 lines (1869 loc) · 70 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
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
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
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
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
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
582
583
584
585
586
587
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
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
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
947
948
949
950
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
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import annotations
import enum
import os
from ctypes import (
c_bool,
c_char_p,
c_int,
c_uint,
c_uint8,
c_int32,
c_uint32,
c_int64,
c_float,
c_void_p,
c_size_t,
POINTER,
_Pointer, # type: ignore
Structure,
CFUNCTYPE
)
import pathlib
from typing import (
Union,
NewType,
Optional,
TYPE_CHECKING,
)
import llama_cpp.llama_cpp as llama_cpp_lib
from llama_cpp._ctypes_extensions import (
load_shared_library,
ctypes_function_for_shared_library,
)
from llama_cpp._ggml import (
ggml_backend_sched_eval_callback,
ggml_log_callback,
)
if TYPE_CHECKING:
from llama_cpp._ctypes_extensions import (
CtypesArray,
)
# --- mtmd library loading ---
_libmtmd_base_name = "mtmd"
_libmtmd_override_path = os.environ.get("MTMD_CPP_LIB")
_libmtmd_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _libmtmd_override_path is None else pathlib.Path(_libmtmd_override_path).parent
# Load the mtmd library
_libmtmd = load_shared_library(_libmtmd_base_name, _libmtmd_base_path)
ctypes_function_mtmd = ctypes_function_for_shared_library(_libmtmd)
################################################
# mtmd.h
# /**
# * libmtmd: A library for multimodal support in llama.cpp.
# *
# * WARNING: This API is experimental and subject to many BREAKING CHANGES.
# * Issues related to API usage may receive lower priority support.
# *
# * For the usage, see an example in mtmd-cli.cpp
# */
################################################
# enum mtmd_input_chunk_type {
# MTMD_INPUT_CHUNK_TYPE_TEXT,
# MTMD_INPUT_CHUNK_TYPE_IMAGE,
# MTMD_INPUT_CHUNK_TYPE_AUDIO,
# MTMD_INPUT_CHUNK_TYPE_COUNT, // for validation
# };
class mtmd_input_chunk_type(enum.IntEnum):
MTMD_INPUT_CHUNK_TYPE_TEXT = 0
MTMD_INPUT_CHUNK_TYPE_IMAGE = 1
MTMD_INPUT_CHUNK_TYPE_AUDIO = 2
MTMD_INPUT_CHUNK_TYPE_COUNT = 3
# // position indexing for decoder model
# enum mtmd_pos_type {
# MTMD_POS_TYPE_NORMAL, // number of positions equals to number of tokens
# MTMD_POS_TYPE_MROPE, // qwen-vl mrope style, each image takes max(t,h,w) position indexes
# MTMD_POS_TYPE_HUNYUANVL, // HunyuanVL mrope + BOI/EOI/newline layout with XD-RoPE dim-3
# MTMD_POS_TYPE_COUNT, // for validation
# };
class mtmd_pos_type(enum.IntEnum):
MTMD_POS_TYPE_NORMAL = 0 # number of positions equals to number of tokens
MTMD_POS_TYPE_MROPE = 1 # qwen-vl mrope style, each image takes max(t,h,w) position indexes
MTMD_POS_TYPE_HUNYUANVL = 2 # HunyuanVL mrope + BOI/EOI/newline layout with XD-RoPE dim-3
MTMD_POS_TYPE_COUNT = 3 # for validation
# // opaque types
# struct mtmd_context;
mtmd_context_p = NewType("mtmd_context_p", int)
mtmd_context_p_ctypes = c_void_p
# // represents raw image data, layout is RGBRGBRGB...
# // length of data must be nx * ny * 3
# struct mtmd_bitmap {
# uint32_t nx;
# uint32_t ny;
# std::vector<unsigned char> data;
# std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
# bool is_audio = false; // true if the bitmap is audio
# };
mtmd_bitmap_p = NewType("mtmd_bitmap_p", int)
mtmd_bitmap_p_ctypes = c_void_p
# struct mtmd_image_tokens {
# uint32_t nx; // number of tokens in x direction
# uint32_t ny; // number of tokens in y direction
# mtmd_pos_type pos = MTMD_POS_TYPE_NORMAL;
# uint32_t n_tokens() const { return nx * ny; }
# clip_image_f32_batch batch_f32; // preprocessed image patches
# std::string id; // optional user-defined ID, useful for KV cache tracking
# mtmd_image_tokens clone() {
# return mtmd_image_tokens{
# nx,
# ny,
# use_mrope_pos,
# batch_f32.clone(),
# id
# };
# }
# };
mtmd_image_tokens_p = NewType("mtmd_image_tokens_p", int)
mtmd_image_tokens_p_ctypes = c_void_p
# struct mtmd_audio_tokens {
# uint32_t n_tokens; // number of tokens
# clip_image_f32_batch batch_f32; // preprocessed image patches
# std::string id; // optional user-defined ID, useful for KV cache tracking
# mtmd_audio_tokens clone() {
# return mtmd_audio_tokens{
# n_tokens,
# batch_f32.clone(),
# id
# };
# }
# };
mtmd_audio_tokens_p = NewType("mtmd_audio_tokens_p", int)
mtmd_audio_tokens_p_ctypes = c_void_p
# struct mtmd_input_chunk {
# mtmd_input_chunk_type type;
# std::vector<llama_token> tokens_text;
# mtmd_image_tokens_ptr tokens_image;
# mtmd_audio_tokens_ptr tokens_audio;
# };
mtmd_input_chunk_p = NewType("mtmd_input_chunk_p", int)
mtmd_input_chunk_p_ctypes = c_void_p
# struct mtmd_input_chunks;
mtmd_input_chunks_p = NewType("mtmd_input_chunks_p", int)
mtmd_input_chunks_p_ctypes = c_void_p
# struct mtmd_batch {
# mtmd_context * ctx;
# std::vector<const mtmd_input_chunk *> entries;
# std::vector<float> output_embd; // aggregated output embedding for the whole batch
# mtmd_batch(mtmd_context * ctx): ctx(ctx) {}
# int32_t n_tokens() const {
# int32_t n = 0;
# for (const auto * chunk : entries) {
# n += mtmd_input_chunk_get_n_tokens(chunk);
# }
# return n;
# }
# };
mtmd_batch_p = NewType("mtmd_batch_p", int)
mtmd_batch_p_ctypes = c_void_p
# typedef bool (*mtmd_progress_callback)(float progress, void * user_data);
mtmd_progress_callback = CFUNCTYPE(
c_bool,
c_float, # progress
c_void_p, # user_data
)
# struct mtmd_input_text {
# const char * text;
# size_t text_len;
# bool add_special;
# bool parse_special;
# };
class mtmd_input_text(Structure):
_fields_ = [
("text", c_char_p),
("text_len", c_size_t),
("add_special", c_bool),
("parse_special", c_bool),
]
mtmd_input_text_p = NewType("mtmd_input_text_p", int)
mtmd_input_text_p_ctypes = POINTER(mtmd_input_text)
# struct mtmd_input_part {
# // only text or bitmap can be set, not both
# const struct mtmd_input_text * text;
# const struct mtmd_bitmap * bitmap;
# };
class mtmd_input_part(Structure):
_fields_ = [
("text", mtmd_input_text_p_ctypes),
("bitmap", mtmd_bitmap_p_ctypes),
]
mtmd_input_part_p = NewType("mtmd_input_part_p", int)
mtmd_input_part_p_ctypes = POINTER(mtmd_input_part)
# enum clip_flash_attn_type {
# CLIP_FLASH_ATTN_TYPE_AUTO = -1,
# CLIP_FLASH_ATTN_TYPE_DISABLED = 0,
# CLIP_FLASH_ATTN_TYPE_ENABLED = 1,
# };
class clip_flash_attn_type (enum.IntEnum):
CLIP_FLASH_ATTN_TYPE_AUTO = -1
CLIP_FLASH_ATTN_TYPE_DISABLED = 0
CLIP_FLASH_ATTN_TYPE_ENABLED = 1
# struct clip_context_params {
# bool use_gpu;
# enum clip_flash_attn_type flash_attn_type;
# int image_min_tokens;
# int image_max_tokens;
# bool warmup;
# };
class clip_context_params(Structure):
_fields_ = [
("use_gpu", c_bool),
("flash_attn_type", c_int),
("image_min_tokens", c_int),
("image_max_tokens", c_int),
("warmup", c_bool),
]
# struct mtmd_context_params {
# bool use_gpu;
# ggml_backend_dev_t device;
# bool print_timings;
# int n_threads;
# const char * image_marker; // deprecated, use media_marker instead
# const char * media_marker;
# enum llama_flash_attn_type flash_attn_type;
# bool warmup; // whether to run a warmup encode pass after initialization
# // limit number of image tokens, only for vision models with dynamic resolution
# int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)
# int image_max_tokens; // maximum number of tokens for image input (default: read from metadata)
# // callback function passed over to mtmd proper
# ggml_backend_sched_eval_callback cb_eval;
# void * cb_eval_user_data;
# // batching params
# int32_t batch_max_tokens; // maximum number of output tokens in a batch
# // (note: this is not a hard-limit, the first image will always be added even if it exceeds this limit)
# // (default: 1024)
# // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
# // If the provided progress_callback returns true, model loading continues.
# // If it returns false, model loading is immediately aborted.
# mtmd_progress_callback progress_callback;
# void * progress_callback_user_data;
# };
class mtmd_context_params(Structure):
_fields_ = [
("use_gpu", c_bool),
("device", c_void_p),
("print_timings", c_bool),
("n_threads", c_int),
("image_marker", c_char_p),
("media_marker", c_char_p),
("flash_attn_type", c_int),
("warmup", c_bool),
("image_min_tokens", c_int),
("image_max_tokens", c_int),
("cb_eval", ggml_backend_sched_eval_callback),
("cb_eval_user_data", c_void_p),
("batch_max_tokens", c_int32),
("progress_callback", mtmd_progress_callback),
("progress_callback_user_data", c_void_p)
]
mtmd_context_params_p_ctypes = POINTER(mtmd_context_params)
# MTMD_API const char * mtmd_default_marker(void);
@ctypes_function_mtmd(
"mtmd_default_marker",
[],
c_char_p,
)
def mtmd_default_marker() -> c_char_p:
...
# MTMD_API struct mtmd_context_params mtmd_context_params_default(void);
@ctypes_function_mtmd(
"mtmd_context_params_default",
[],
mtmd_context_params,
)
def mtmd_context_params_default() -> mtmd_context_params:
...
# // initialize the mtmd context
# // return nullptr on failure
# MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
# const struct llama_model * text_model,
# const struct mtmd_context_params ctx_params);
@ctypes_function_mtmd(
"mtmd_init_from_file", [
c_char_p,
llama_cpp_lib.llama_model_p_ctypes,
mtmd_context_params,
],
mtmd_context_p_ctypes,
)
def mtmd_init_from_file(
mmproj_fname: c_char_p,
text_model: llama_cpp_lib.llama_model_p,
ctx_params: mtmd_context_params,
/,
) -> mtmd_context_p:
"""
initialize the mtmd context
return nullptr on failure
"""
...
# MTMD_API void mtmd_free(mtmd_context * ctx);
@ctypes_function_mtmd("mtmd_free", [mtmd_context_p_ctypes], None)
def mtmd_free(ctx: mtmd_context_p):
...
# // whether we need to set non-causal mask before llama_decode
# // if chunk is nullptr, we assume the default case where chunk is an image chunk
# MTMD_API bool mtmd_decode_use_non_causal(const mtmd_context * ctx, const mtmd_input_chunk * chunk);
@ctypes_function_mtmd(
"mtmd_decode_use_non_causal", [mtmd_context_p_ctypes, mtmd_input_chunk_p_ctypes], c_bool)
def mtmd_decode_use_non_causal(ctx: mtmd_context_p, chunk: mtmd_input_chunk_p) -> c_bool:
...
# // whether the current model use M-RoPE for llama_decode
# MTMD_API bool mtmd_decode_use_mrope(const mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_decode_use_mrope", [mtmd_context_p_ctypes], c_bool)
def mtmd_decode_use_mrope(ctx: mtmd_context_p) -> c_bool:
...
# // whether the current model supports vision input
# MTMD_API bool mtmd_support_vision(const mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_support_vision", [mtmd_context_p_ctypes], c_bool)
def mtmd_support_vision(ctx: mtmd_context_p) -> c_bool:
...
# // whether the current model supports audio input
# MTMD_API bool mtmd_support_audio(const mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_support_audio", [mtmd_context_p_ctypes], c_bool)
def mtmd_support_audio(ctx: mtmd_context_p) -> c_bool:
...
# // get audio sample rate in Hz, for example 16000 for Whisper
# // return -1 if audio is not supported
# MTMD_API int mtmd_get_audio_sample_rate(const mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_get_audio_sample_rate", [mtmd_context_p_ctypes], c_int)
def mtmd_get_audio_sample_rate(ctx: mtmd_context_p) -> c_int:
"""
get audio sample rate in Hz, for example 16000 for Whisper
return -1 if audio is not supported
"""
...
# // get the current marker string
# MTMD_API const char * mtmd_get_marker(const mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_get_marker", [mtmd_context_p_ctypes], c_char_p)
def mtmd_get_marker(ctx: mtmd_context_p) -> c_char_p:
"""
get the current marker string
"""
...
# // mtmd_bitmap
# //
# // if bitmap is image:
# // length of data must be nx * ny * 3
# // the data is in RGBRGBRGB... format
# // note: some video-capable models (i.e. qwen-vl) can merge consecutive bitmaps
# // into one chunk; mtmd_tokenize() handles this, but remember to set
# // mtmd_bitmap_set_mergeable(true) for every frame
# // if bitmap is audio:
# // length of data must be n_samples * sizeof(float)
# // the data is in float format (PCM F32)
# //
# // if data == nullptr:
# // the bitmap is considered "empty", and will be treated as a placeholder for counting tokens
# // you can pass the bitmap via mtmd_tokenize(), then call mtmd_*_get_n_tokens() to count the tokens
# // note: passing a placeholder bitmap to mtmd_encode() will return an error
# MTMD_API mtmd_bitmap * mtmd_bitmap_init (uint32_t nx, uint32_t ny, const unsigned char * data);
@ctypes_function_mtmd(
"mtmd_bitmap_init", [
c_uint32,
c_uint32,
POINTER(c_uint8),
],
mtmd_bitmap_p_ctypes,
)
def mtmd_bitmap_init(
nx: c_uint32,
ny: c_uint32,
data: POINTER(c_uint8), # type: ignore
/,
) -> mtmd_bitmap_p:
...
# MTMD_API mtmd_bitmap * mtmd_bitmap_init_from_audio(size_t n_samples, const float * data);
@ctypes_function_mtmd(
"mtmd_bitmap_init_from_audio", [
c_size_t,
POINTER(c_float)
],
mtmd_bitmap_p_ctypes,
)
def mtmd_bitmap_init_from_audio(
n_samples: c_size_t,
data: POINTER(c_float), # type: ignore
/,
) -> mtmd_bitmap_p:
...
# MTMD_API uint32_t mtmd_bitmap_get_nx (const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_get_nx", [mtmd_bitmap_p_ctypes], c_uint32)
def mtmd_bitmap_get_nx(bitmap: mtmd_bitmap_p) -> c_uint32:
...
# MTMD_API uint32_t mtmd_bitmap_get_ny (const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_get_ny", [mtmd_bitmap_p_ctypes], c_uint32)
def mtmd_bitmap_get_ny(bitmap: mtmd_bitmap_p) -> c_uint32:
...
# MTMD_API const unsigned char * mtmd_bitmap_get_data (const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_get_data", [mtmd_bitmap_p_ctypes], POINTER(c_uint8))
def mtmd_bitmap_get_data(bitmap: mtmd_bitmap_p) -> POINTER(c_uint8): # type: ignore
...
# MTMD_API size_t mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_get_n_bytes", [mtmd_bitmap_p_ctypes], c_size_t)
def mtmd_bitmap_get_n_bytes(bitmap: mtmd_bitmap_p) -> c_size_t:
...
# MTMD_API bool mtmd_bitmap_is_audio (const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_is_audio", [mtmd_bitmap_p_ctypes], c_bool)
def mtmd_bitmap_is_audio(bitmap: mtmd_bitmap_p) -> c_bool:
...
# MTMD_API void mtmd_bitmap_free (mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_free", [mtmd_bitmap_p_ctypes], None)
def mtmd_bitmap_free(bitmap: mtmd_bitmap_p):
...
# // bitmap ID is optional, but useful for KV cache tracking
# // these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
# MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap);
@ctypes_function_mtmd("mtmd_bitmap_get_id", [mtmd_bitmap_p_ctypes], c_char_p)
def mtmd_bitmap_get_id(bitmap: mtmd_bitmap_p) -> c_char_p:
"""
bitmap ID is optional, but useful for KV cache tracking
these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
"""
...
# MTMD_API void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id);
@ctypes_function_mtmd(
"mtmd_bitmap_set_id", [
mtmd_bitmap_p_ctypes,
c_char_p,
], None)
def mtmd_bitmap_set_id(
bitmap: mtmd_bitmap_p,
id: c_char_p,
/,
):
...
# // if true, this bitmap can be merged (temporal merge) with an adjacent mergeable bitmap by certain video input models
# MTMD_API void mtmd_bitmap_set_mergeable(mtmd_bitmap * bitmap, bool mergeable);
@ctypes_function_mtmd(
"mtmd_bitmap_set_mergeable", [
mtmd_bitmap_p_ctypes,
c_bool,
], None)
def mtmd_bitmap_set_mergeable(
bitmap: mtmd_bitmap_p,
mergeable: bool,
/,
):
"""
if true, this bitmap can be merged (temporal merge) with an adjacent mergeable bitmap by certain video input models
"""
...
# // mtmd_bitmap lazy
# //
# // this is a special bitmap that:
# // - does not hold the actual data
# // - can be expanded into one or more chunks (either media to text chunks)
# // user must provide a callback to fill in the data when mtmd_tokenize() is called
# // this is useful for large video inputs:
# // - allow reading video frame by frame, without loading the entire video into memory
# // - allow tracking the whole video with a single ID (for example, the file hash)
# // set (*out_bitmap) to non-nullptr to emit a bitmap chunk; it will be freed automatically
# // set (*out_text) to non-nullptr to emit a text chunk; it must be heap-allocated, null-terminated and will be freed automatically
# // either out_bitmap or out_text can be set, but not both
# // out_bitmap cannot be another lazy bitmap (no nested lazy allowed)
# // return value:
# // 0 on success
# // -1 on EOF (signal to mtmd_tokenize to move on)
# // -2 on error (signal to mtmd_tokenize to abort)
# typedef int(* mtmd_bitmap_lazy_callback)(
# size_t chunk_idx,
# void * user_data,
# mtmd_bitmap ** out_bitmap,
# char ** out_text);
mtmd_bitmap_lazy_callback = CFUNCTYPE(
c_int,
c_size_t, # chunk_idx
c_void_p, # user_data
POINTER(mtmd_bitmap_p_ctypes), # mtmd_bitmap ** out_bitmap
POINTER(c_char_p), # char ** out_text
)
# MTMD_API mtmd_bitmap * mtmd_bitmap_init_lazy(mtmd_context * ctx,
# const char * id, // usually set to file hash
# void * user_data,
# mtmd_bitmap_lazy_callback callback);
@ctypes_function_mtmd(
"mtmd_bitmap_init_lazy", [
mtmd_context_p_ctypes,
c_char_p,
c_void_p,
mtmd_bitmap_lazy_callback,
], mtmd_bitmap_p_ctypes)
def mtmd_bitmap_init_lazy(
ctx: mtmd_context_p,
id: c_char_p,
user_data: c_void_p,
callback: Optional[mtmd_bitmap_lazy_callback], # type: ignore
/,
) -> mtmd_bitmap_p:
...
# // mtmd_input_chunks
# //
# // this is simply a list of mtmd_input_chunk
# // the elements can only be populated via mtmd_tokenize()
# MTMD_API mtmd_input_chunks * mtmd_input_chunks_init(void);
@ctypes_function_mtmd("mtmd_input_chunks_init", [], mtmd_input_chunks_p_ctypes)
def mtmd_input_chunks_init() -> mtmd_input_chunks_p:
"""
this is simply a list of mtmd_input_chunk
the elements can only be populated via mtmd_tokenize()
"""
...
# MTMD_API size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks);
@ctypes_function_mtmd("mtmd_input_chunks_size", [mtmd_input_chunks_p_ctypes], c_size_t)
def mtmd_input_chunks_size(chunks: mtmd_input_chunks_p) -> c_size_t:
...
# MTMD_API const mtmd_input_chunk * mtmd_input_chunks_get (const mtmd_input_chunks * chunks, size_t idx);
@ctypes_function_mtmd(
"mtmd_input_chunks_get", [
mtmd_input_chunks_p_ctypes,
c_size_t,
], mtmd_input_chunk_p_ctypes)
def mtmd_input_chunks_get(
chunks: mtmd_input_chunks_p,
idx: c_size_t,
/,
) -> mtmd_input_chunk_p:
...
# MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
@ctypes_function_mtmd("mtmd_input_chunks_free", [mtmd_input_chunks_p_ctypes], None)
def mtmd_input_chunks_free(chunks: mtmd_input_chunks_p):
...
# // mtmd_input_chunk
# //
# // the instance will be constructed via mtmd_tokenize()
# // it will be freed along with mtmd_input_chunks
# MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type (const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_type", [mtmd_input_chunk_p_ctypes], c_int32)
def mtmd_input_chunk_get_type(chunk: mtmd_input_chunk_p) -> c_int32:
"""
the instance will be constructed via mtmd_tokenize()
it will be freed along with mtmd_input_chunks
"""
...
# MTMD_API const llama_token * mtmd_input_chunk_get_tokens_text(const mtmd_input_chunk * chunk, size_t * n_tokens_output);
@ctypes_function_mtmd(
"mtmd_input_chunk_get_tokens_text",
[mtmd_input_chunk_p_ctypes, POINTER(c_size_t)],
POINTER(llama_cpp_lib.llama_token)
)
def mtmd_input_chunk_get_tokens_text(
chunk: mtmd_input_chunk_p, n_tokens_output: "_Pointer[c_size_t]", /
) -> Optional["_Pointer[llama_cpp_lib.llama_token]"]:
...
# MTMD_API const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_tokens_image", [mtmd_input_chunk_p_ctypes], mtmd_image_tokens_p_ctypes)
def mtmd_input_chunk_get_tokens_image(chunk: mtmd_input_chunk_p) -> mtmd_image_tokens_p:
...
# MTMD_API size_t mtmd_input_chunk_get_n_tokens (const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_n_tokens", [mtmd_input_chunk_p_ctypes], c_size_t)
def mtmd_input_chunk_get_n_tokens(chunk: mtmd_input_chunk_p) -> c_size_t:
...
# // returns nullptr for ID on text chunk
# MTMD_API const char * mtmd_input_chunk_get_id (const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_id", [mtmd_input_chunk_p_ctypes], c_char_p)
def mtmd_input_chunk_get_id(chunk: mtmd_input_chunk_p) -> c_char_p:
"""
returns nullptr for ID on text chunk
"""
...
# // number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
# MTMD_API llama_pos mtmd_input_chunk_get_n_pos (const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_n_pos", [mtmd_input_chunk_p_ctypes], c_int32)
def mtmd_input_chunk_get_n_pos(chunk: mtmd_input_chunk_p) -> c_int32:
"""
number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
"""
...
# // in case you want to use custom logic to handle the chunk (i.e. KV cache management)
# // you can move the chunk ownership to your own code by copying it
# // remember to free the chunk when you are done with it
# MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_copy", [mtmd_input_chunk_p_ctypes], mtmd_input_chunk_p_ctypes)
def mtmd_input_chunk_copy(chunk: mtmd_input_chunk_p) -> mtmd_input_chunk_p:
"""
in case you want to use custom logic to handle the chunk (i.e. KV cache management)
you can move the chunk ownership to your own code by copying it
remember to free the chunk when you are done with it
"""
...
# MTMD_API void mtmd_input_chunk_free(mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_free", [mtmd_input_chunk_p_ctypes], None)
def mtmd_input_chunk_free(chunk: mtmd_input_chunk_p):
"""
remember to free the chunk when you are done with it
"""
...
# // similar to mtmd_input_chunk_copy, but returns a placeholder chunk
# MTMD_API mtmd_input_chunk * mtmd_input_chunk_get_placeholder(const mtmd_input_chunk * chunk);
@ctypes_function_mtmd("mtmd_input_chunk_get_placeholder", [mtmd_input_chunk_p_ctypes], mtmd_input_chunk_p_ctypes)
def mtmd_input_chunk_get_placeholder(chunk: mtmd_input_chunk_p) -> mtmd_input_chunk_p:
"""
similar to mtmd_input_chunk_copy, but returns a placeholder chunk
"""
...
# // save/load an input chunk to/from a buffer (useful for KV save/load)
# // important: only chunk's metadata will be saved, the actual image/audio data will not be saved
# // the loaded chunk will always be a placeholder, cannot be used for mtmd_encode() or mtmd_batch_encode()
# // out_buf can be nullptr (to query expected_out_len)
# // returns 0 on success, non-zero on failure
# MTMD_API int32_t mtmd_input_chunk_save(const mtmd_input_chunk * chunk, char * out_buf, size_t out_len, size_t * expected_out_len);
@ctypes_function_mtmd("mtmd_input_chunk_save",
[
mtmd_input_chunk_p_ctypes,
c_char_p,
c_size_t,
POINTER(c_size_t),
],
c_int32,
)
def mtmd_input_chunk_save(
chunk: mtmd_input_chunk_p,
out_buf: bytes,
out_len: c_size_t,
expected_out_len: POINTER(c_size_t), # type: ignore
) -> int:
"""
save an input chunk to/from a buffer (useful for KV save)
important: only chunk's metadata will be saved, the actual image/audio data will not be saved
the loaded chunk will always be a placeholder, cannot be used for mtmd_encode() or mtmd_batch_encode()
out_buf can be nullptr (to query expected_out_len)
returns 0 on success, non-zero on failure
"""
...
# // returns nullptr on failure
# MTMD_API mtmd_input_chunk * mtmd_input_chunk_load(const char * buf, size_t len);
@ctypes_function_mtmd("mtmd_input_chunk_load",
[
c_char_p,
c_size_t
],
mtmd_input_chunk_p_ctypes,
)
def mtmd_input_chunk_load(
buf: bytes,
len: c_size_t,
) -> mtmd_input_chunk_p:
"""
load an input chunk from a buffer (useful for KV load)
important: only chunk's metadata will be saved, the actual image/audio data will not be saved
the loaded chunk will always be a placeholder, cannot be used for mtmd_encode() or mtmd_batch_encode()
returns nullptr on failure
"""
...
# // mtmd_image_tokens
# //
# // the instance will be constructed via mtmd_tokenize()
# // it will be freed along with mtmd_input_chunk
# MTMD_API size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
@ctypes_function_mtmd(
"mtmd_image_tokens_get_n_tokens", [mtmd_image_tokens_p_ctypes], c_size_t)
def mtmd_image_tokens_get_n_tokens(image_tokens: mtmd_image_tokens_p) -> c_size_t:
...
# MTMD_API const char * mtmd_image_tokens_get_id (const mtmd_image_tokens * image_tokens); // TODO: deprecate
@ctypes_function_mtmd(
"mtmd_image_tokens_get_id", [mtmd_image_tokens_p_ctypes], c_char_p)
def mtmd_image_tokens_get_id(image_tokens: mtmd_image_tokens_p) -> c_char_p:
...
# // number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
# MTMD_API llama_pos mtmd_image_tokens_get_n_pos (const mtmd_image_tokens * image_tokens); // TODO: deprecate
@ctypes_function_mtmd(
"mtmd_image_tokens_get_n_pos", [mtmd_image_tokens_p_ctypes], c_int32)
def mtmd_image_tokens_get_n_pos(image_tokens: mtmd_image_tokens_p) -> c_int32:
"""number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)"""
...
# DEPRECATED(MTMD_API size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens),
# "use mtmd_image_tokens_get_decoder_pos() instead");
@ctypes_function_mtmd(
"mtmd_image_tokens_get_nx", [mtmd_image_tokens_p_ctypes], c_size_t)
def mtmd_image_tokens_get_nx(image_tokens: mtmd_image_tokens_p) -> c_size_t:
"""
use mtmd_image_tokens_get_decoder_pos() instead
"""
...
# DEPRECATED(MTMD_API size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens),
# "use mtmd_image_tokens_get_decoder_pos() instead");
@ctypes_function_mtmd(
"mtmd_image_tokens_get_ny", [mtmd_image_tokens_p_ctypes], c_size_t)
def mtmd_image_tokens_get_ny(image_tokens: mtmd_image_tokens_p) -> c_size_t:
"""
use mtmd_image_tokens_get_decoder_pos() instead
"""
...
# struct mtmd_decoder_pos {
# uint32_t t;
# uint32_t x;
# uint32_t y;
# uint32_t z;
# };
class mtmd_decoder_pos(Structure):
_fields_ = [
("t", c_uint32),
("x", c_uint32),
("y", c_uint32),
("z", c_uint32),
]
if TYPE_CHECKING:
t: c_uint32
x: c_uint32
y: c_uint32
z: c_uint32
mtmd_decoder_pos_p_ctypes = POINTER(mtmd_decoder_pos)
# // get position for decoder attention, to be used by M-RoPE models
# // i is the index of the embedding token, ranging from 0 to mtmd_image_tokens_get_n_tokens() - 1
# // pos_0 is the absolute position of the first token
# // return relative position (for example, embedding 0 will have position (0, 0, 0); remember to adjust it to the current absolute position)
# MTMD_API struct mtmd_decoder_pos mtmd_image_tokens_get_decoder_pos(const mtmd_image_tokens * image_tokens, llama_pos pos_0, size_t i);
@ctypes_function_mtmd(
"mtmd_image_tokens_get_decoder_pos", [mtmd_image_tokens_p_ctypes, c_int32, c_size_t], mtmd_decoder_pos)
def mtmd_image_tokens_get_decoder_pos(image_tokens: mtmd_image_tokens_p, pos_0: c_int32, i: c_size_t) -> mtmd_decoder_pos:
"""
get position for decoder attention, to be used by M-RoPE models
i is the index of the embedding token, ranging from 0 to mtmd_image_tokens_get_n_tokens() - 1
pos_0 is the absolute position of the first token
return relative position (for example, embedding 0 will have position (0, 0, 0); remember to adjust it to the current absolute position)
"""
...
# // tokenize an input text prompt and a list of bitmaps (image/audio)
# // the prompt must have the input media marker (default: "<__media__>") in it
# // the default marker is defined by mtmd_default_marker()
# // the marker will be replaced with the media chunk
# // for example:
# // "here is an image: <__media__>\ndescribe it in detail."
# // this will gives 3 chunks:
# // 1. "here is an image: <start_of_image>"
# // 2. (image/audio tokens)
# // 3. "<end_of_image>\ndescribe it in detail."
# // number of bitmaps must be equal to the number of markers in the prompt
# // this function is thread-safe (shared ctx)
# // return values:
# // 0 on success
# // 1 on number of bitmaps not matching the number of markers
# // 2 on media preprocessing error
# MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
# mtmd_input_chunks * output,
# const mtmd_input_text * text,
# const mtmd_bitmap ** bitmaps,
# size_t n_bitmaps);
@ctypes_function_mtmd(
"mtmd_tokenize", [
mtmd_context_p_ctypes,
mtmd_input_chunks_p_ctypes,
mtmd_input_text_p_ctypes,
POINTER(mtmd_bitmap_p_ctypes),
c_size_t,
],
c_int32,
)
def mtmd_tokenize(
ctx: mtmd_context_p,
output: mtmd_input_chunks_p,
text: mtmd_input_text_p,
bitmaps: POINTER(mtmd_bitmap_p_ctypes), # type: ignore
n_bitmaps: c_size_t,
/,
) -> c_int32:
"""
tokenize an input text prompt and a list of bitmaps (image/audio)
the prompt must have the input media marker (default: "<__media__>") in it
the default marker is defined by mtmd_default_marker()
the marker will be replaced with the media chunk
for example:
"here is an image: <__media__>\ndescribe it in detail."
this will gives 3 chunks:
1. "here is an image: <start_of_image>"
2. (image/audio tokens)
3. "<end_of_image>\\ndescribe it in detail."
number of bitmaps must be equal to the number of markers in the prompt
this function is thread-safe (shared ctx)
return values:
0 on success
1 on number of bitmaps not matching the number of markers
2 on media preprocessing error
"""
...
# // same as mtmd_tokenize(), but takes an array of mtmd_input_part
# // use cases:
# // - when you don't want to use media markers (they will be tokenized as normal text)
# // - when you want to control parse_special for each text part
# // note: per-part add_special will be ignored
# // return 1 if a part has both text and bitmap set (or neither)
# MTMD_API int32_t mtmd_tokenize_from_parts(mtmd_context * ctx,
# mtmd_input_chunks * output,
# const mtmd_input_part ** parts,
# size_t n_parts,
# bool add_special);
@ctypes_function_mtmd(
"mtmd_tokenize_from_parts", [
mtmd_context_p_ctypes,
mtmd_input_chunks_p_ctypes,
POINTER(mtmd_input_part_p_ctypes),
c_size_t,
c_bool,
],
c_int32,
)
def mtmd_tokenize_from_parts(
ctx: mtmd_context_p,
output: mtmd_input_chunks_p,
parts: POINTER(mtmd_input_part_p_ctypes), # type: ignore
n_parts: c_size_t,
add_special: c_bool,
/,
) -> c_int32:
"""
same as mtmd_tokenize(), but takes an array of mtmd_input_part
use cases:
- when you don't want to use media markers (they will be tokenized as normal text)
- when you want to control parse_special for each text part
note: per-part add_special will be ignored.
return 1 if a part has both text and bitmap set (or neither)
"""
...
# // returns 0 on success
# // TODO: deprecate
# DEPRECATED(MTMD_API int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens),
# "use mtmd_encode_chunk() instead");
@ctypes_function_mtmd(
"mtmd_encode", [
mtmd_context_p_ctypes,
mtmd_image_tokens_p_ctypes
],
c_int32,
)
def mtmd_encode(
ctx: mtmd_context_p,
image_tokens: mtmd_image_tokens_p,
/,
) -> c_int32:
"""
DEPRECATED: use mtmd_encode_chunk() instead
"""
...
# // text chunk will be ignored silently, only media chunk will be encoded
# // returns 0 on success
# // returns 1 on generic error
# MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
# const mtmd_input_chunk * chunk);
@ctypes_function_mtmd(
"mtmd_encode_chunk", [
mtmd_context_p_ctypes,
mtmd_input_chunk_p_ctypes
],
c_int32,
)
def mtmd_encode_chunk(
ctx: mtmd_context_p,
chunk: mtmd_input_chunk_p,
/,
) -> c_int32:
"""
text chunk will be ignored silently, only media chunk will be encoded
returns 0 on success
returns 1 on generic error
"""
...
# // get output embeddings from the last encode pass
# // the reading size (in bytes) is equal to:
# // llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
# MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_get_output_embd", [mtmd_context_p_ctypes], POINTER(c_float))
def mtmd_get_output_embd(ctx: mtmd_context_p) -> POINTER(c_float): # type: ignore
"""
get output embeddings from the last encode pass
"""
...
# // batch encoding API
# // chunks are not owned by the batch, they will not be freed by mtmd_batch_free()
# // batch is valid for a given context, cannot be shared across contexts
# MTMD_API mtmd_batch * mtmd_batch_init(mtmd_context * ctx);
@ctypes_function_mtmd(
"mtmd_batch_init",
[mtmd_context_p_ctypes],
mtmd_batch_p_ctypes,
)
def mtmd_batch_init(ctx: mtmd_context_p, /) -> mtmd_batch_p:
...
# MTMD_API void mtmd_batch_free(mtmd_batch * batch);