diff options
Diffstat (limited to 'examples/llava/llava.cpp')
| -rw-r--r-- | examples/llava/llava.cpp | 75 |
1 files changed, 72 insertions, 3 deletions
diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index 63878d17..916d9dc4 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> return true; } +static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) { + int width = image->nx; + int height = image->ny; + int num_patches = (height / patch_size) * (width / patch_size); + clip_image_f32 * patch = clip_image_f32_init(); + patch->nx = patch_size * num_patches; + patch->ny = patch_size; + patch->buf.resize(3 * patch->nx * patch->ny); + + int patch_index = 0; + + for (int i = 0; i < height; i += patch_size) { + for (int j = 0; j < width; j += patch_size) { + for (int pi = 0; pi < patch_size; ++pi) { + for (int pj = 0; pj < patch_size; ++pj) { + int input_index = ((i + pi) * width + (j + pj)) * 3; + int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3; + patch->buf[output_index] = image->buf[input_index]; + patch->buf[output_index+1] = image->buf[input_index+1]; + patch->buf[output_index+2] = image->buf[input_index+2]; + } + } + patch_index++; + } + } + return patch; +} static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336 @@ -218,7 +245,44 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); - if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { + if (clip_is_minicpmv(ctx_clip)) { + std::vector<float *> image_embd_v; + image_embd_v.resize(img_res_v.size); + struct clip_image_size * load_image_size = clip_image_size_init(); + for (size_t i = 0; i < img_res_v.size; i++) { + const int64_t t_img_enc_step_start_us = ggml_time_us(); + image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); + int patch_size=14; + load_image_size->width = img_res_v.data[i].nx; + load_image_size->height = img_res_v.data[i].ny; + clip_add_load_image_size(ctx_clip, load_image_size); + const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); + if (!encoded) { + LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + return false; + } + const int64_t t_img_enc_steop_batch_us = ggml_time_us(); + LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0); + } + const int64_t t_img_enc_batch_us = ggml_time_us(); + LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + + int n_img_pos_out = 0; + for (size_t i = 0; i < image_embd_v.size(); i++) { + std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip)); + n_img_pos_out += clip_n_patches(ctx_clip); + } + *n_img_pos = n_img_pos_out; + for (size_t i = 0; i < image_embd_v.size(); i++) { + free(image_embd_v[i]); + } + image_embd_v.clear(); + load_image_size->width = img->nx; + load_image_size->height = img->ny; + clip_add_load_image_size(ctx_clip, load_image_size); + LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height); + } + else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { // flat / default llava-1.5 type embedding *n_img_pos = clip_n_patches(ctx_clip); bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096 @@ -228,7 +292,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli return false; } - } else { + } + else { // spatial_unpad llava-1.6 type embedding // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working std::vector<float *> image_embd_v; @@ -297,7 +362,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * } bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { - float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model + int num_max_patches = 6; + if (clip_is_minicpmv(ctx_clip)) { + num_max_patches = 10; + } + float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model if (!image_embd) { LOG_TEE("Unable to allocate memory for image embeddings\n"); return false; |