1#pragma once
2
3#include "ggml.h"
4#include "mtmd.h"
5
6#include <stddef.h>
7#include <stdint.h>
8
9// !!! Internal header, to be used by mtmd only !!!
10
11#define MTMD_INTERNAL_HEADER
12
13struct clip_ctx;
14
15struct clip_image_size {
16 int width;
17 int height;
18};
19
20struct clip_image_f32;
21struct clip_image_u8_batch;
22struct clip_image_f32_batch;
23
24enum clip_modality {
25 CLIP_MODALITY_VISION,
26 CLIP_MODALITY_AUDIO,
27};
28
29enum clip_flash_attn_type {
30 CLIP_FLASH_ATTN_TYPE_AUTO = -1,
31 CLIP_FLASH_ATTN_TYPE_DISABLED = 0,
32 CLIP_FLASH_ATTN_TYPE_ENABLED = 1,
33};
34
35struct clip_context_params {
36 bool use_gpu;
37 enum clip_flash_attn_type flash_attn_type;
38 int image_min_tokens;
39 int image_max_tokens;
40 bool warmup;
41 ggml_backend_sched_eval_callback cb_eval;
42 void * cb_eval_user_data;
43};
44
45struct clip_init_result {
46 struct clip_ctx * ctx_v; // vision context
47 struct clip_ctx * ctx_a; // audio context
48};
49
50struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params);
51
52void clip_free(struct clip_ctx * ctx);
53
54size_t clip_embd_nbytes(const struct clip_ctx * ctx);
55size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
56
57int32_t clip_get_image_size (const struct clip_ctx * ctx);
58int32_t clip_get_patch_size (const struct clip_ctx * ctx);
59int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
60
61// TODO: should be enum, not string
62const char * clip_patch_merge_type(const struct clip_ctx * ctx);
63
64int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
65
66// for M-RoPE, this will be the number of token positions in X and Y directions
67// for other models, X will be the total number of tokens and Y will be 1
68int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
69int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
70
71// this should be equal to the embedding dimension of the text model
72int clip_n_mmproj_embd(const struct clip_ctx * ctx);
73
74struct clip_image_size * clip_image_size_init(void);
75struct clip_image_u8 * clip_image_u8_init (void);
76struct clip_image_f32 * clip_image_f32_init(void);
77struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava
78
79// nx, ny are the output image dimensions
80unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
81
82void clip_image_size_free (struct clip_image_size * img_size);
83void clip_image_u8_free (struct clip_image_u8 * img);
84void clip_image_f32_free(struct clip_image_f32 * img);
85void clip_image_u8_batch_free (struct clip_image_u8_batch * batch);
86void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
87
88// use for accessing underlay data of clip_image_f32_batch
89size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
90size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
91size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
92struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
93
94/**
95 * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
96 * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
97 */
98void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
99
100/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
101bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
102
103struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
104
105bool clip_image_encode (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
106bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
107
108int clip_is_minicpmv(const struct clip_ctx * ctx);
109bool clip_is_glm(const struct clip_ctx * ctx);
110bool clip_is_llava(const struct clip_ctx * ctx);
111// note for contributor: this clip_is_(model) pattern is deprecated
112// do NOT add new functions like this
113
114bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
115
116// use by audio input
117void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel);
118
119bool clip_has_vision_encoder(const struct clip_ctx * ctx);
120bool clip_has_audio_encoder(const struct clip_ctx * ctx);
121bool clip_has_whisper_encoder(const struct clip_ctx * ctx);