| 1 | #pragma once |
| 2 | |
| 3 | #include "ggml.h" |
| 4 | |
| 5 | #include <stddef.h> |
| 6 | #include <stdint.h> |
| 7 | |
| 8 | // !!! Internal header, to be used by mtmd only !!! |
| 9 | |
| 10 | struct clip_ctx; |
| 11 | |
| 12 | struct clip_image_size { |
| 13 | int width; |
| 14 | int height; |
| 15 | }; |
| 16 | |
| 17 | struct clip_image_f32; |
| 18 | struct clip_image_u8_batch; |
| 19 | struct clip_image_f32_batch; |
| 20 | |
| 21 | enum clip_modality { |
| 22 | CLIP_MODALITY_VISION, |
| 23 | CLIP_MODALITY_AUDIO, |
| 24 | }; |
| 25 | |
| 26 | enum clip_flash_attn_type { |
| 27 | CLIP_FLASH_ATTN_TYPE_AUTO = -1, |
| 28 | CLIP_FLASH_ATTN_TYPE_DISABLED = 0, |
| 29 | CLIP_FLASH_ATTN_TYPE_ENABLED = 1, |
| 30 | }; |
| 31 | |
| 32 | struct clip_context_params { |
| 33 | bool use_gpu; |
| 34 | enum ggml_log_level verbosity; |
| 35 | enum clip_flash_attn_type flash_attn_type; |
| 36 | int image_min_tokens; |
| 37 | int image_max_tokens; |
| 38 | }; |
| 39 | |
| 40 | struct clip_init_result { |
| 41 | struct clip_ctx * ctx_v; // vision context |
| 42 | struct clip_ctx * ctx_a; // audio context |
| 43 | }; |
| 44 | |
| 45 | struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params); |
| 46 | |
| 47 | void clip_free(struct clip_ctx * ctx); |
| 48 | |
| 49 | size_t clip_embd_nbytes(const struct clip_ctx * ctx); |
| 50 | size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h); |
| 51 | |
| 52 | int32_t clip_get_image_size (const struct clip_ctx * ctx); |
| 53 | int32_t clip_get_patch_size (const struct clip_ctx * ctx); |
| 54 | int32_t clip_get_hidden_size(const struct clip_ctx * ctx); |
| 55 | |
| 56 | // TODO: should be enum, not string |
| 57 | const char * clip_patch_merge_type(const struct clip_ctx * ctx); |
| 58 | |
| 59 | int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img); |
| 60 | |
| 61 | // for M-RoPE, this will be the number of token positions in X and Y directions |
| 62 | // for other models, X will be the total number of tokens and Y will be 1 |
| 63 | int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img); |
| 64 | int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img); |
| 65 | |
| 66 | // this should be equal to the embedding dimension of the text model |
| 67 | int clip_n_mmproj_embd(const struct clip_ctx * ctx); |
| 68 | |
| 69 | struct clip_image_size * clip_image_size_init(void); |
| 70 | struct clip_image_u8 * clip_image_u8_init (void); |
| 71 | struct clip_image_f32 * clip_image_f32_init(void); |
| 72 | struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava |
| 73 | |
| 74 | // nx, ny are the output image dimensions |
| 75 | unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny); |
| 76 | |
| 77 | void clip_image_size_free (struct clip_image_size * img_size); |
| 78 | void clip_image_u8_free (struct clip_image_u8 * img); |
| 79 | void clip_image_f32_free(struct clip_image_f32 * img); |
| 80 | void clip_image_u8_batch_free (struct clip_image_u8_batch * batch); |
| 81 | void clip_image_f32_batch_free(struct clip_image_f32_batch * batch); |
| 82 | |
| 83 | // use for accessing underlay data of clip_image_f32_batch |
| 84 | size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size() |
| 85 | size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx |
| 86 | size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny |
| 87 | struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data |
| 88 | |
| 89 | /** |
| 90 | * Build image from pixels decoded by other libraries instead of stb_image.h for better performance. |
| 91 | * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes |
| 92 | */ |
| 93 | void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img); |
| 94 | |
| 95 | /** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */ |
| 96 | bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs ); |
| 97 | |
| 98 | struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx); |
| 99 | |
| 100 | bool clip_image_encode (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec); |
| 101 | bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec); |
| 102 | |
| 103 | int clip_is_minicpmv(const struct clip_ctx * ctx); |
| 104 | bool clip_is_glm(const struct clip_ctx * ctx); |
| 105 | bool clip_is_qwen2vl(const struct clip_ctx * ctx); |
| 106 | bool clip_is_llava(const struct clip_ctx * ctx); |
| 107 | bool clip_is_gemma3(const struct clip_ctx * ctx); |
| 108 | |
| 109 | bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec); |
| 110 | |
| 111 | // use by audio input |
| 112 | void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel); |
| 113 | |
| 114 | bool clip_has_vision_encoder(const struct clip_ctx * ctx); |
| 115 | bool clip_has_audio_encoder(const struct clip_ctx * ctx); |
| 116 | bool clip_has_whisper_encoder(const struct clip_ctx * ctx); |
| 117 | |