Function clip_image_batch_encode

examples/llava/clip.cpp:848–885

}

bool clip_image_batch_encode(const clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {

    if (!ctx->has_vision_encoder) {
        printf("This gguf file seems to have no vision encoder\n");
        return false;
    }

    int batch_size = imgs->size;
    if(ctx->has_llava_projector) {
        GGML_ASSERT(batch_size == 1); // TODO: support multiple images
    }

    // reset alloc buffer to clean the memory from previous invocations
    ggml_allocr_reset(ctx->alloc);

    // build the inference graph
    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
    ggml_allocr_alloc_graph(ctx->alloc, gf);

    struct ggml_cplan plan = ggml_graph_plan(gf, n_threads);
    if (plan.work_size > 0) {
        plan.work_data = (uint8_t *)malloc(plan.work_size);
    }

    ggml_graph_compute(gf, &plan);

    // the last node is the embedding tensor
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];

    // copy the embeddings to the location passed by the user
    memcpy(vec, ggml_get_data_f32(embeddings), ggml_nbytes(embeddings));

    if (plan.work_size > 0) {
        free(plan.work_data);
    }

    return true;
}

bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) {