llama.cpp/tools/mtmd/clip.cpp

// NOTE: This is modified from clip.cpp only for LLaVA,
// so there might be still unnecessary artifacts hanging around
// I'll gradually clean and extend it
// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
#include "clip.h"
#include "clip-impl.h"
#include "ggml.h"
#include "ggml-cpp.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include "gguf.h"

#include <cassert>
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <fstream>
#include <map>
#include <regex>
#include <stdexcept>
#include <unordered_set>
#include <vector>
#include <sstream>
#include <cinttypes>
#include <limits>
#include <array>
#include <numeric>
#include <functional>

struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};

enum ffn_op_type {
    FFN_GELU,
    FFN_GELU_ERF,
    FFN_SILU,
    FFN_GELU_QUICK,
};

enum norm_type {
    NORM_TYPE_NORMAL,
    NORM_TYPE_RMS,
};

//#define CLIP_DEBUG_FUNCTIONS

#ifdef CLIP_DEBUG_FUNCTIONS
static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
    std::ofstream file(filename, std::ios::binary);
    if (!file.is_open()) {
        LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
        return;
    }

    // PPM header: P6 format, width, height, and max color value
    file << "P6\n" << img.nx << " " << img.ny << "\n255\n";

    // Write pixel data
    for (size_t i = 0; i < img.buf.size(); i += 3) {
        // PPM expects binary data in RGB format, which matches our image buffer
        file.write(reinterpret_cast<const char*>(&img.buf[i]), 3);
    }

    file.close();
}

static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
    std::ofstream file(filename, std::ios::binary);
    if (!file.is_open()) {
        LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
        return;
    }

    int fileSize = 54 + 3 * img.nx * img.ny; // File header + info header + pixel data
    int bytesPerPixel = 3;
    int widthInBytes = img.nx * bytesPerPixel;
    int paddingAmount = (4 - (widthInBytes % 4)) % 4;
    int stride = widthInBytes + paddingAmount;

    // Bitmap file header
    unsigned char fileHeader[14] = {
        'B','M',     // Signature
        0,0,0,0,    // Image file size in bytes
        0,0,0,0,    // Reserved
        54,0,0,0    // Start of pixel array
    };

    // Total file size
    fileSize = 54 + (stride * img.ny);
    fileHeader[2] = (unsigned char)(fileSize);
    fileHeader[3] = (unsigned char)(fileSize >> 8);
    fileHeader[4] = (unsigned char)(fileSize >> 16);
    fileHeader[5] = (unsigned char)(fileSize >> 24);

    // Bitmap information header (BITMAPINFOHEADER)
    unsigned char infoHeader[40] = {
        40,0,0,0,   // Size of this header (40 bytes)
        0,0,0,0,    // Image width
        0,0,0,0,    // Image height
        1,0,        // Number of color planes
        24,0,       // Bits per pixel
        0,0,0,0,    // No compression
        0,0,0,0,    // Image size (can be 0 for no compression)
        0,0,0,0,    // X pixels per meter (not specified)
        0,0,0,0,    // Y pixels per meter (not specified)
        0,0,0,0,    // Total colors (color table not used)
        0,0,0,0     // Important colors (all are important)
    };

    // Width and height in the information header
    infoHeader[4] = (unsigned char)(img.nx);
    infoHeader[5] = (unsigned char)(img.nx >> 8);
    infoHeader[6] = (unsigned char)(img.nx >> 16);
    infoHeader[7] = (unsigned char)(img.nx >> 24);
    infoHeader[8] = (unsigned char)(img.ny);
    infoHeader[9] = (unsigned char)(img.ny >> 8);
    infoHeader[10] = (unsigned char)(img.ny >> 16);
    infoHeader[11] = (unsigned char)(img.ny >> 24);

    // Write file headers
    file.write(reinterpret_cast<char*>(fileHeader), sizeof(fileHeader));
    file.write(reinterpret_cast<char*>(infoHeader), sizeof(infoHeader));

    // Pixel data
    std::vector<unsigned char> padding(3, 0); // Max padding size to be added to each row
    for (int y = img.ny - 1; y >= 0; --y) { // BMP files are stored bottom-to-top
        for (int x = 0; x < img.nx; ++x) {
            // Each pixel
            size_t pixelIndex = (y * img.nx + x) * 3;
            unsigned char pixel[3] = {
                img.buf[pixelIndex + 2], // BMP stores pixels in BGR format
                img.buf[pixelIndex + 1],
                img.buf[pixelIndex]
            };
            file.write(reinterpret_cast<char*>(pixel), 3);
        }
        // Write padding for the row
        file.write(reinterpret_cast<char*>(padding.data()), paddingAmount);
    }

    file.close();
}

// debug function to convert f32 to u8
static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u8& dst) {
    dst.nx = src.nx;
    dst.ny = src.ny;
    dst.buf.resize(3 * src.nx * src.ny);
    for (size_t i = 0; i < src.buf.size(); ++i) {
        dst.buf[i] = static_cast<uint8_t>(std::min(std::max(int(src.buf[i] * 255.0f), 0), 255));
    }
}
#endif


//
// clip layers
//

enum patch_merge_type {
    PATCH_MERGE_FLAT,
    PATCH_MERGE_SPATIAL_UNPAD,
};

struct clip_hparams {
    int32_t image_size;
    int32_t patch_size;
    int32_t n_embd;
    int32_t n_ff;
    int32_t projection_dim;
    int32_t n_head;
    int32_t n_layer;
    int32_t proj_scale_factor = 0; // idefics3

    float image_mean[3];
    float image_std[3];

    // for models using dynamic image size, we need to have a smaller image size to warmup
    // otherwise, user will get OOM everytime they load the model
    int32_t warmup_image_size = 0;
    int32_t warmup_audio_size = 3000;

    ffn_op_type ffn_op = FFN_GELU;

    patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT;

    float eps = 1e-6;
    float rope_theta = 0.0;

    std::vector<clip_image_size> image_res_candidates; // for llava-uhd style models
    int32_t image_crop_resolution;
    std::unordered_set<int32_t> vision_feature_layer;
    int32_t attn_window_size = 0;
    int32_t n_wa_pattern = 0;
    int32_t spatial_merge_size = 0;

    // audio
    int32_t n_mel_bins = 0; // whisper preprocessor
    int32_t proj_stack_factor = 0; // ultravox

    // legacy
    bool has_llava_projector = false;
    int minicpmv_version = 0;
};

struct clip_layer {
    // attention
    ggml_tensor * k_w = nullptr;
    ggml_tensor * k_b = nullptr;
    ggml_tensor * q_w = nullptr;
    ggml_tensor * q_b = nullptr;
    ggml_tensor * v_w = nullptr;
    ggml_tensor * v_b = nullptr;

    ggml_tensor * o_w = nullptr;
    ggml_tensor * o_b = nullptr;

    ggml_tensor * k_norm = nullptr;
    ggml_tensor * q_norm = nullptr;

    // layernorm 1
    ggml_tensor * ln_1_w = nullptr;
    ggml_tensor * ln_1_b = nullptr;

    ggml_tensor * ff_up_w = nullptr;
    ggml_tensor * ff_up_b = nullptr;
    ggml_tensor * ff_gate_w = nullptr;
    ggml_tensor * ff_gate_b = nullptr;
    ggml_tensor * ff_down_w = nullptr;
    ggml_tensor * ff_down_b = nullptr;

    // layernorm 2
    ggml_tensor * ln_2_w = nullptr;
    ggml_tensor * ln_2_b = nullptr;

    // layer scale (no bias)
    ggml_tensor * ls_1_w = nullptr;
    ggml_tensor * ls_2_w = nullptr;
};

struct clip_model {
    clip_modality modality = CLIP_MODALITY_VISION;
    projector_type proj_type = PROJECTOR_TYPE_MLP;
    clip_hparams hparams;

    // embeddings
    ggml_tensor * class_embedding = nullptr;
    ggml_tensor * patch_embeddings_0 = nullptr;
    ggml_tensor * patch_embeddings_1 = nullptr;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
    ggml_tensor * patch_bias = nullptr;
    ggml_tensor * position_embeddings = nullptr;

    ggml_tensor * pre_ln_w = nullptr;
    ggml_tensor * pre_ln_b = nullptr;

    std::vector<clip_layer> layers;

    ggml_tensor * post_ln_w;
    ggml_tensor * post_ln_b;

    ggml_tensor * projection; // TODO: rename it to fc (fully connected layer)
    ggml_tensor * mm_fc_w;
    ggml_tensor * mm_fc_b;

    // LLaVA projection
    ggml_tensor * mm_input_norm_w = nullptr;
    ggml_tensor * mm_0_w = nullptr;
    ggml_tensor * mm_0_b = nullptr;
    ggml_tensor * mm_2_w = nullptr;
    ggml_tensor * mm_2_b = nullptr;

    ggml_tensor * image_newline = nullptr;

    // Yi type models with mlp+normalization projection
    ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4
    ggml_tensor * mm_1_b = nullptr;
    ggml_tensor * mm_3_w = nullptr;
    ggml_tensor * mm_3_b = nullptr;
    ggml_tensor * mm_4_w = nullptr;
    ggml_tensor * mm_4_b = nullptr;

    // GLMV-Edge projection
    ggml_tensor * mm_model_adapter_conv_w = nullptr;
    ggml_tensor * mm_model_adapter_conv_b = nullptr;
    ggml_tensor * mm_glm_tok_boi = nullptr;
    ggml_tensor * mm_glm_tok_eoi = nullptr;

    // MobileVLM projection
    ggml_tensor * mm_model_mlp_1_w = nullptr;
    ggml_tensor * mm_model_mlp_1_b = nullptr;
    ggml_tensor * mm_model_mlp_3_w = nullptr;
    ggml_tensor * mm_model_mlp_3_b = nullptr;
    ggml_tensor * mm_model_block_1_block_0_0_w = nullptr;
    ggml_tensor * mm_model_block_1_block_0_1_w = nullptr;
    ggml_tensor * mm_model_block_1_block_0_1_b = nullptr;
    ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr;
    ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr;
    ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr;
    ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr;
    ggml_tensor * mm_model_block_1_block_2_0_w = nullptr;
    ggml_tensor * mm_model_block_1_block_2_1_w = nullptr;
    ggml_tensor * mm_model_block_1_block_2_1_b = nullptr;
    ggml_tensor * mm_model_block_2_block_0_0_w = nullptr;
    ggml_tensor * mm_model_block_2_block_0_1_w = nullptr;
    ggml_tensor * mm_model_block_2_block_0_1_b = nullptr;
    ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr;
    ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr;
    ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr;
    ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr;
    ggml_tensor * mm_model_block_2_block_2_0_w = nullptr;
    ggml_tensor * mm_model_block_2_block_2_1_w = nullptr;
    ggml_tensor * mm_model_block_2_block_2_1_b = nullptr;

    // MobileVLM_V2 projection
    ggml_tensor * mm_model_mlp_0_w = nullptr;
    ggml_tensor * mm_model_mlp_0_b = nullptr;
    ggml_tensor * mm_model_mlp_2_w = nullptr;
    ggml_tensor * mm_model_mlp_2_b = nullptr;
    ggml_tensor * mm_model_peg_0_w = nullptr;
    ggml_tensor * mm_model_peg_0_b = nullptr;

    // MINICPMV projection
    ggml_tensor * mm_model_pos_embed_k = nullptr;
    ggml_tensor * mm_model_query = nullptr;
    ggml_tensor * mm_model_proj = nullptr;
    ggml_tensor * mm_model_kv_proj = nullptr;
    ggml_tensor * mm_model_attn_q_w = nullptr;
    ggml_tensor * mm_model_attn_q_b = nullptr;
    ggml_tensor * mm_model_attn_k_w = nullptr;
    ggml_tensor * mm_model_attn_k_b = nullptr;
    ggml_tensor * mm_model_attn_v_w = nullptr;
    ggml_tensor * mm_model_attn_v_b = nullptr;
    ggml_tensor * mm_model_attn_o_w = nullptr;
    ggml_tensor * mm_model_attn_o_b = nullptr;
    ggml_tensor * mm_model_ln_q_w = nullptr;
    ggml_tensor * mm_model_ln_q_b = nullptr;
    ggml_tensor * mm_model_ln_kv_w = nullptr;
    ggml_tensor * mm_model_ln_kv_b = nullptr;
    ggml_tensor * mm_model_ln_post_w = nullptr;
    ggml_tensor * mm_model_ln_post_b = nullptr;

    // gemma3
    ggml_tensor * mm_input_proj_w = nullptr;
    ggml_tensor * mm_soft_emb_norm_w = nullptr;

    // pixtral
    ggml_tensor * token_embd_img_break = nullptr;
    ggml_tensor * mm_patch_merger_w = nullptr;

    // ultravox / whisper encoder
    ggml_tensor * conv1d_1_w = nullptr;
    ggml_tensor * conv1d_1_b = nullptr;
    ggml_tensor * conv1d_2_w = nullptr;
    ggml_tensor * conv1d_2_b = nullptr;
    ggml_tensor * mm_norm_pre_w = nullptr;
    ggml_tensor * mm_norm_mid_w = nullptr;

    bool audio_has_avgpool() const {
        return proj_type == PROJECTOR_TYPE_QWEN2A
            || proj_type == PROJECTOR_TYPE_VOXTRAL;
    }

    bool audio_has_stack_frames() const {
        return proj_type == PROJECTOR_TYPE_ULTRAVOX
            || proj_type == PROJECTOR_TYPE_VOXTRAL;
    }
};

struct clip_ctx {
    clip_model model;

    gguf_context_ptr ctx_gguf;
    ggml_context_ptr ctx_data;

    std::vector<uint8_t> buf_compute_meta;

    std::vector<ggml_backend_t> backend_ptrs;
    std::vector<ggml_backend_buffer_type_t> backend_buft;

    ggml_backend_t backend = nullptr;
    ggml_backend_t backend_cpu = nullptr;
    ggml_backend_buffer_ptr buf;

    int max_nodes = 8192;
    ggml_backend_sched_ptr sched;

    // for debugging
    bool debug_graph = false;
    std::vector<ggml_tensor *> debug_print_tensors;

    clip_ctx(clip_context_params & ctx_params) {
        debug_graph = std::getenv("MTMD_DEBUG_GRAPH") != nullptr;
        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
        if (!backend_cpu) {
            throw std::runtime_error("failed to initialize CPU backend");
        }
        if (ctx_params.use_gpu) {
            auto backend_name = std::getenv("MTMD_BACKEND_DEVICE");
            if (backend_name != nullptr) {
                backend = ggml_backend_init_by_name(backend_name, nullptr);
                if (!backend) {
                    LOG_WRN("%s: Warning: Failed to initialize \"%s\" backend, falling back to default GPU backend\n", __func__, backend_name);
                }
            }
            if (!backend) {
                backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
            }
        }

        if (backend) {
            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));
            backend_ptrs.push_back(backend);
            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend));
        } else {
            backend = backend_cpu;
            LOG_INF("%s: CLIP using CPU backend\n", __func__);
        }

        backend_ptrs.push_back(backend_cpu);
        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));

        sched.reset(
            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false, true)
        );
    }

    ~clip_ctx() {
        ggml_backend_free(backend);
        if (backend != backend_cpu) {
            ggml_backend_free(backend_cpu);
        }
    }

    // this function is added so that we don't change too much of the existing code
    projector_type proj_type() const {
        return model.proj_type;
    }
};

struct clip_graph {
    clip_ctx * ctx;
    const clip_model & model;
    const clip_hparams & hparams;

    // we only support single image per batch
    const clip_image_f32 & img;

    const int patch_size;
    const int n_patches_x;
    const int n_patches_y;
    const int n_patches;
    const int n_embd;
    const int n_head;
    const int d_head;
    const int n_layer;
    const float eps;
    const float kq_scale;

    ggml_context_ptr ctx0_ptr;
    ggml_context * ctx0;
    ggml_cgraph * gf;

    clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
            ctx(ctx),
            model(ctx->model),
            hparams(model.hparams),
            img(img),
            patch_size(hparams.patch_size),
            n_patches_x(img.nx / patch_size),
            n_patches_y(img.ny / patch_size),
            n_patches(n_patches_x * n_patches_y),
            n_embd(hparams.n_embd),
            n_head(hparams.n_head),
            d_head(n_embd / n_head),
            n_layer(hparams.n_layer),
            eps(hparams.eps),
            kq_scale(1.0f / sqrtf((float)d_head)) {
        struct ggml_init_params params = {
            /*.mem_size   =*/ ctx->buf_compute_meta.size(),
            /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
            /*.no_alloc   =*/ true,
        };
        ctx0_ptr.reset(ggml_init(params));
        ctx0 = ctx0_ptr.get();
        gf = ggml_new_graph_custom(ctx0, ctx->max_nodes, false);
    }

    ggml_cgraph * build_siglip() {
        ggml_tensor * inp = build_inp();
        ggml_tensor * cur = build_vit(
                                inp, n_patches,
                                NORM_TYPE_NORMAL,
                                hparams.ffn_op,
                                model.position_embeddings,
                                nullptr);

        if (ctx->proj_type() == PROJECTOR_TYPE_GEMMA3) {
            const int batch_size = 1;
            GGML_ASSERT(n_patches_x == n_patches_y);
            const int patches_per_image = n_patches_x;
            const int kernel_size = hparams.proj_scale_factor;

            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
            cur = ggml_reshape_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);

            // doing a pool2d to reduce the number of output tokens
            cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
            cur = ggml_reshape_3d(ctx0, cur, cur->ne[0] * cur->ne[0], n_embd, batch_size);
            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));

            // apply norm before projection
            cur = ggml_rms_norm(ctx0, cur, eps);
            cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w);

            // apply projection
            cur = ggml_mul_mat(ctx0,
                ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)),
                cur);

        } else if (ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3) {
            // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578

            const int scale_factor = model.hparams.proj_scale_factor;
            const int n_embd = cur->ne[0];
            const int seq    = cur->ne[1];
            const int bsz    = 1; // batch size, always 1 for now since we don't support batching
            const int height = std::sqrt(seq);
            const int width  = std::sqrt(seq);
            GGML_ASSERT(scale_factor != 0);
            cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                height / scale_factor,
                width / scale_factor,
                bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                seq / (scale_factor * scale_factor),
                bsz);

            cur = ggml_mul_mat(ctx0, model.projection, cur);
        } else {
            GGML_ABORT("SigLIP: Unsupported projector type");
        }

        // build the graph
        ggml_build_forward_expand(gf, cur);

        return gf;
    }

    ggml_cgraph * build_pixtral() {
        const int n_merge = hparams.spatial_merge_size;

        // 2D input positions
        ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
        ggml_set_name(pos_h, "pos_h");
        ggml_set_input(pos_h);

        ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
        ggml_set_name(pos_w, "pos_w");
        ggml_set_input(pos_w);

        auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
            return build_rope_2d(ctx0, cur, pos_h, pos_w, hparams.rope_theta, true);
        };

        ggml_tensor * inp = build_inp();
        ggml_tensor * cur = build_vit(
                                inp, n_patches,
                                NORM_TYPE_RMS,
                                hparams.ffn_op,
                                nullptr, // no learned pos embd
                                add_pos);

        // mistral small 3.1 patch merger
        // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
        if (model.mm_patch_merger_w) {
            GGML_ASSERT(hparams.spatial_merge_size > 0);

            cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);

            // reshape image tokens to 2D grid
            cur = ggml_reshape_3d(ctx0, cur, n_embd, n_patches_x, n_patches_y);
            cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, n_embd]
            cur = ggml_cont(ctx0, cur);

            // torch.nn.functional.unfold is just an im2col under the hood
            // we just need a dummy kernel to make it work
            ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
            cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);

            // project to n_embd
            cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
            cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
        }

        // LlavaMultiModalProjector (always using GELU activation)
        {
            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
            if (model.mm_1_b) {
                cur = ggml_add(ctx0, cur, model.mm_1_b);
            }

            cur = ggml_gelu(ctx0, cur);
            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
            if (model.mm_2_b) {
                cur = ggml_add(ctx0, cur, model.mm_2_b);
            }
        }

        // arrangement of the [IMG_BREAK] token
        {
            // not efficient, but works
            // the trick is to view the embeddings as a 3D tensor with shape [n_embd, n_patches_per_row, n_rows]
            // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
            // after the concatenation, we have a tensor with shape [n_embd, n_patches_per_row + 1, n_rows]

            const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
            const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
            const int p_total         = p_x * p_y;
            const int n_embd_text     = cur->ne[0];
            const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row

            ggml_tensor * tmp = ggml_reshape_3d(ctx0, cur, n_embd_text, p_x, p_y);
            ggml_tensor * tok = ggml_new_tensor_3d(ctx0, tmp->type, n_embd_text, 1, p_y);
            tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
            tok = ggml_add(ctx0, tok, model.token_embd_img_break);
            tmp = ggml_concat(ctx0, tmp, tok, 1);
            cur = ggml_view_2d(ctx0, tmp,
                n_embd_text, n_tokens_output,
                ggml_row_size(tmp->type, n_embd_text), 0);
        }

        // build the graph
        ggml_build_forward_expand(gf, cur);

        return gf;
    }

    // Qwen2VL and Qwen2.5VL use M-RoPE
    ggml_cgraph * build_qwen2vl() {
        GGML_ASSERT(model.patch_bias == nullptr);
        GGML_ASSERT(model.class_embedding == nullptr);

        const int batch_size       = 1;
        const bool use_window_attn = hparams.n_wa_pattern > 0;
        const int n_wa_pattern     = hparams.n_wa_pattern;
        const int n_pos            = n_patches;
        const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position

        norm_type norm_t = ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL
            ? NORM_TYPE_RMS // qwen 2.5 vl
            : NORM_TYPE_NORMAL; // qwen 2 vl

        int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};

        ggml_tensor * inp_raw = build_inp_raw();
        ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);

        GGML_ASSERT(img.nx % (patch_size * 2) == 0);
        GGML_ASSERT(img.ny % (patch_size * 2) == 0);

        // second conv dimension
        {
            auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
            inp = ggml_add(ctx0, inp, inp_1);

            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));  // [w, h, c, b] -> [c, w, h, b]
            inp = ggml_reshape_4d(
                ctx0, inp,
                n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
            inp = ggml_reshape_4d(
                ctx0, inp,
                n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
            inp = ggml_reshape_3d(
                ctx0, inp,
                n_embd, n_patches_x * n_patches_y, batch_size);
        }

        ggml_tensor * inpL           = inp;
        ggml_tensor * window_mask    = nullptr;
        ggml_tensor * window_idx     = nullptr;
        ggml_tensor * inv_window_idx = nullptr;

        ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
        ggml_set_name(positions, "positions");
        ggml_set_input(positions);

        // pre-layernorm
        if (model.pre_ln_w) {
            inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
        }

        if (use_window_attn) {
            // handle window attention inputs
            inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
            ggml_set_name(inv_window_idx, "inv_window_idx");
            ggml_set_input(inv_window_idx);
            // mask for window attention
            window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
            ggml_set_name(window_mask, "window_mask");
            ggml_set_input(window_mask);

            // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
            GGML_ASSERT(batch_size == 1);
            inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
            inpL = ggml_get_rows(ctx0, inpL, inv_window_idx);
            inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size);
        }

        // loop over layers
        for (int il = 0; il < n_layer; il++) {
            auto & layer = model.layers[il];
            const bool full_attn = use_window_attn ? (il + 1) % n_wa_pattern == 0 : true;

            ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states

            // layernorm1
            cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
            cb(cur, "ln1", il);

            // self-attention
            {
                ggml_tensor * Qcur = ggml_add(ctx0,
                    ggml_mul_mat(ctx0, layer.q_w, cur), layer.q_b);
                ggml_tensor * Kcur = ggml_add(ctx0,
                    ggml_mul_mat(ctx0, layer.k_w, cur), layer.k_b);
                ggml_tensor * Vcur = ggml_add(ctx0,
                    ggml_mul_mat(ctx0, layer.v_w, cur), layer.v_b);

                Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_patches);
                Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_patches);
                Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_patches);

                cb(Qcur, "Qcur", il);
                cb(Kcur, "Kcur", il);
                cb(Vcur, "Vcur", il);

                // apply M-RoPE
                Qcur = ggml_rope_multi(
                    ctx0, Qcur, positions, nullptr,
                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
                Kcur = ggml_rope_multi(
                    ctx0, Kcur, positions, nullptr,
                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);

                cb(Qcur, "Qcur_rope", il);
                cb(Kcur, "Kcur_rope", il);

                ggml_tensor * attn_mask = full_attn ? nullptr : window_mask;

                cur = build_attn(layer.o_w, layer.o_b,
                    Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
                cb(cur, "attn_out", il);
            }

            // re-add the layer input, e.g., residual
            cur = ggml_add(ctx0, cur, inpL);

            inpL = cur; // inpL = residual, cur = hidden_states

            cb(cur, "ffn_inp", il);

            // layernorm2
            cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
            cb(cur, "ffn_inp_normed", il);

            // ffn
            cur = build_ffn(cur,
                layer.ff_up_w, layer.ff_up_b,
                layer.ff_gate_w, layer.ff_gate_b,
                layer.ff_down_w, layer.ff_down_b,
                hparams.ffn_op, il);

            cb(cur, "ffn_out", il);

            // residual 2
            cur = ggml_add(ctx0, inpL, cur);
            cb(cur, "layer_out", il);

            inpL = cur;
        }

        // post-layernorm
        if (model.post_ln_w) {
            inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
        }

        // multimodal projection
        ggml_tensor * embeddings = inpL;
        embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);

        embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
        embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);

        // GELU activation
        embeddings = ggml_gelu(ctx0, embeddings);

        // Second linear layer
        embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);

        if (use_window_attn) {
            window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
            ggml_set_name(window_idx, "window_idx");
            ggml_set_input(window_idx);

            // embeddings shape: [n_embd, n_patches_x * n_patches_y, batch_size]
            GGML_ASSERT(batch_size == 1);
            embeddings = ggml_reshape_2d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4);
            embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
            embeddings = ggml_reshape_3d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4, batch_size);
        }

        // build the graph
        ggml_build_forward_expand(gf, embeddings);

        return gf;
    }

    ggml_cgraph * build_minicpmv() {
        const int batch_size = 1;

        GGML_ASSERT(model.class_embedding == nullptr);
        const int n_pos = n_patches;

        // position embeddings for the projector (not for ViT)
        int n_output_dim = clip_n_mmproj_embd(ctx);
        ggml_tensor * pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_output_dim, n_pos, batch_size);
        ggml_set_name(pos_embed, "pos_embed");
        ggml_set_input(pos_embed);

        // for selecting learned pos embd, used by ViT
        struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
        ggml_set_name(positions, "positions");
        ggml_set_input(positions);

        ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);

        ggml_tensor * inp = build_inp();
        ggml_tensor * embeddings = build_vit(
                                inp, n_patches,
                                NORM_TYPE_NORMAL,
                                hparams.ffn_op,
                                learned_pos_embd,
                                nullptr);

        // resampler projector (it is just another transformer)

        ggml_tensor * q = model.mm_model_query;
        ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);

        // norm
        q = build_norm(q, model.mm_model_ln_q_w, model.mm_model_ln_q_b, NORM_TYPE_NORMAL, eps, -1);
        v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1);

        // k = v + pos_embed
        ggml_tensor * k = ggml_add(ctx0, v, pos_embed);

        // attention
        {
            int n_embd = clip_n_mmproj_embd(ctx);
            const int d_head = 128;
            int n_head = n_embd/d_head;
            int num_query = 96;
            if (ctx->model.hparams.minicpmv_version == 2) {
                // MiniCPM-V 2.5
                num_query = 96;
            } else if (ctx->model.hparams.minicpmv_version == 3) {
                // MiniCPM-V 2.6
                num_query = 64;
            } else if (ctx->model.hparams.minicpmv_version == 4) {
                // MiniCPM-o 2.6
                num_query = 64;
            } else if (ctx->model.hparams.minicpmv_version == 5) {
                // MiniCPM-V 4.0
                num_query = 64;
            }

            ggml_tensor * Q = ggml_add(ctx0,
                ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
                model.mm_model_attn_q_b);
            ggml_tensor * K = ggml_add(ctx0,
                ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k),
                model.mm_model_attn_k_b);
            ggml_tensor * V = ggml_add(ctx0,
                ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v),
                model.mm_model_attn_v_b);

            Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_query);
            K = ggml_reshape_3d(ctx0, K, d_head, n_head, n_pos);
            V = ggml_reshape_3d(ctx0, V, d_head, n_head, n_pos);

            cb(Q, "resampler_Q", -1);
            cb(K, "resampler_K", -1);
            cb(V, "resampler_V", -1);

            embeddings = build_attn(
                model.mm_model_attn_o_w,
                model.mm_model_attn_o_b,
                Q, K, V, nullptr, kq_scale, -1);
            cb(embeddings, "resampler_attn_out", -1);
        }
        // layernorm
        embeddings = build_norm(embeddings, model.mm_model_ln_post_w, model.mm_model_ln_post_b, NORM_TYPE_NORMAL, eps, -1);

        // projection
        embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);

        // build the graph
        ggml_build_forward_expand(gf, embeddings);

        return gf;
    }

    ggml_cgraph * build_internvl() {
        GGML_ASSERT(model.class_embedding != nullptr);
        GGML_ASSERT(model.position_embeddings != nullptr);

        const int n_pos = n_patches + 1;
        ggml_tensor * inp = build_inp();

        // add CLS token
        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);

        // The larger models use a different ViT, which uses RMS norm instead of layer norm
        // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
        norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45)
            ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B)
            : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models)

        ggml_tensor * cur = build_vit(
                                inp, n_pos,
                                norm_t,
                                hparams.ffn_op,
                                model.position_embeddings,
                                nullptr);

        // remove CLS token
        cur = ggml_view_2d(ctx0, cur,
            n_embd, n_patches,
            ggml_row_size(cur->type, n_embd), 0);

        // pixel shuffle
        {
            const int scale_factor = model.hparams.proj_scale_factor;
            const int bsz    = 1; // batch size, always 1 for now since we don't support batching
            const int height = n_patches_y;
            const int width  = n_patches_x;
            GGML_ASSERT(scale_factor > 0);
            cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                height / scale_factor,
                width / scale_factor,
                bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            // flatten to 2D
            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                cur->ne[1] * cur->ne[2]);
        }

        // projector (always using GELU activation)
        {
            // projector LayerNorm uses pytorch's default eps = 1e-5
            // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79
            cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
            cur = ggml_add(ctx0, cur, model.mm_1_b);
            cur = ggml_gelu(ctx0, cur);
            cur = ggml_mul_mat(ctx0, model.mm_3_w, cur);
            cur = ggml_add(ctx0, cur, model.mm_3_b);
        }

        // build the graph
        ggml_build_forward_expand(gf, cur);

        return gf;
    }

    ggml_cgraph * build_llama4() {
        GGML_ASSERT(model.class_embedding != nullptr);
        GGML_ASSERT(model.position_embeddings != nullptr);

        const int n_pos = n_patches + 1; // +1 for [CLS]

        // 2D input positions
        ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
        ggml_set_name(pos_h, "pos_h");
        ggml_set_input(pos_h);

        ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
        ggml_set_name(pos_w, "pos_w");
        ggml_set_input(pos_w);

        ggml_tensor * inp = build_inp_raw();

        // Llama4UnfoldConvolution
        {
            ggml_tensor * kernel = ggml_reshape_4d(ctx0, model.patch_embeddings_0,
                                                    patch_size, patch_size, 3, n_embd);
            inp = ggml_im2col(ctx0, kernel, inp, patch_size, patch_size, 0, 0, 1, 1, true, inp->type);
            inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
            inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
            cb(inp, "patch_conv", -1);
        }

        // add CLS token
        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);

        // build ViT with 2D position embeddings
        auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
            // first half is X axis and second half is Y axis
            // ref: https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/llama4/modeling_llama4.py#L1312
            // ref: https://github.com/Blaizzy/mlx-vlm/blob/a57156aa87b33cca6e5ee6cfc14dd4ef8f611be6/mlx_vlm/models/llama4/vision.py#L441
            return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
        };
        ggml_tensor * cur = build_vit(
                                inp, n_pos,
                                NORM_TYPE_NORMAL,
                                hparams.ffn_op,
                                model.position_embeddings,
                                add_pos);

        // remove CLS token
        cur = ggml_view_2d(ctx0, cur,
            n_embd, n_patches,
            ggml_row_size(cur->type, n_embd), 0);

        // pixel shuffle
        // based on Llama4VisionPixelShuffleMLP
        // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
        {
            const int scale_factor = model.hparams.proj_scale_factor;
            const int bsz = 1; // batch size, always 1 for now since we don't support batching
            GGML_ASSERT(scale_factor > 0);
            GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
            cur = ggml_reshape_4d(ctx0, cur,
                n_embd * scale_factor,
                n_patches_x / scale_factor,
                n_patches_y,
                bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                n_patches_x / scale_factor,
                n_patches_y / scale_factor,
                bsz);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            // flatten to 2D
            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
                n_embd * scale_factor * scale_factor,
                n_patches / scale_factor / scale_factor);
            cb(cur, "pixel_shuffle", -1);
        }

        // based on Llama4VisionMLP2 (always uses GELU activation, no bias)
        {
            cur = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, cur);
            cur = ggml_gelu(ctx0, cur);
            cur = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, cur);
            cur = ggml_gelu(ctx0, cur);
            cb(cur, "adapter_mlp", -1);
        }

        // Llama4MultiModalProjector
        cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
        cb(cur, "projected", -1);

        // build the graph
        ggml_build_forward_expand(gf, cur);

        return gf;
    }

    // this graph is used by llava, granite and glm
    // due to having embedding_stack (used by granite), we cannot reuse build_vit
    ggml_cgraph * build_llava() {
        const int batch_size = 1;
        const int n_pos = n_patches + (model.class_embedding ? 1 : 0);

        GGML_ASSERT(n_patches_x == n_patches_y && "only square images supported");

        // Calculate the deepest feature layer based on hparams and projector type
        int max_feature_layer = n_layer;
        {
            // Get the index of the second to last layer; this is the default for models that have a llava projector
            int il_last = hparams.n_layer - 1;
            int deepest_feature_layer = -1;

            if (ctx->proj_type() == PROJECTOR_TYPE_MINICPMV || ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE) {
                il_last += 1;
            }

            // If we set explicit vision feature layers, only go up to the deepest one
            // NOTE: only used by granite-vision models for now
            for (const auto & feature_layer : hparams.vision_feature_layer) {
                if (feature_layer > deepest_feature_layer) {
                    deepest_feature_layer = feature_layer;
                }
            }
            max_feature_layer = deepest_feature_layer < 0 ? il_last : deepest_feature_layer;
        }

        ggml_tensor * inp = build_inp();

        // concat class_embeddings and patch_embeddings
        if (model.class_embedding) {
            inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
        }

        ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
        ggml_set_name(positions, "positions");
        ggml_set_input(positions);

        inp = ggml_add(ctx0, inp, ggml_get_rows(ctx0, model.position_embeddings, positions));

        ggml_tensor * inpL = inp;

        // pre-layernorm
        if (model.pre_ln_w) {
            inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, NORM_TYPE_NORMAL, eps, -1);
            cb(inpL, "pre_ln", -1);
        }

        std::vector<ggml_tensor *> embedding_stack;
        const auto & vision_feature_layer = hparams.vision_feature_layer;

        // loop over layers
        for (int il = 0; il < max_feature_layer; il++) {
            auto & layer = model.layers[il];
            ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states

            // If this is an embedding feature layer, save the output.
            // NOTE: 0 index here refers to the input to the encoder.
            if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
                embedding_stack.push_back(cur);
            }

            // layernorm1
            cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
            cb(cur, "layer_inp_normed", il);

            // self-attention
            {
                ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur);
                if (layer.q_b) {
                    Qcur = ggml_add(ctx0, Qcur, layer.q_b);
                }

                ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
                if (layer.k_b) {
                    Kcur = ggml_add(ctx0, Kcur, layer.k_b);
                }

                ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
                if (layer.v_b) {
                    Vcur = ggml_add(ctx0, Vcur, layer.v_b);
                }

                Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
                Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
                Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);

                cb(Qcur, "Qcur", il);
                cb(Kcur, "Kcur", il);
                cb(Vcur, "Vcur", il);

                cur = build_attn(layer.o_w, layer.o_b,
                    Qcur, Kcur, Vcur, nullptr, kq_scale, il);
                cb(cur, "attn_out", il);
            }

            // re-add the layer input, e.g., residual
            cur = ggml_add(ctx0, cur, inpL);

            inpL = cur; // inpL = residual, cur = hidden_states

            cb(cur, "ffn_inp", il);

            // layernorm2
            cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
            cb(cur, "ffn_inp_normed", il);

            // ffn
            cur = build_ffn(cur,
                layer.ff_up_w, layer.ff_up_b,
                layer.ff_gate_w, layer.ff_gate_b,
                layer.ff_down_w, layer.ff_down_b,
                hparams.ffn_op, il);

            cb(cur, "ffn_out", il);

            // residual 2
            cur = ggml_add(ctx0, inpL, cur);
            cb(cur, "layer_out", il);

            inpL = cur;
        }

        // post-layernorm
        if (model.post_ln_w) {
            inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
        }

        ggml_tensor * embeddings = inpL;

        // process vision feature layers (used by granite)
        {
            // final layer is a vision feature layer
            if (vision_feature_layer.find(max_feature_layer) != vision_feature_layer.end()) {
                embedding_stack.push_back(inpL);
            }

            // If feature layers are explicitly set, stack them (if we have multiple)
            if (!embedding_stack.empty()) {
                embeddings = embedding_stack[0];
                for (size_t i = 1; i < embedding_stack.size(); i++) {
                    embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
                }
            }
        }

        // llava projector (also used by granite)
        if (ctx->model.hparams.has_llava_projector) {
            embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);

            ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
            ggml_set_name(patches, "patches");
            ggml_set_input(patches);

            // shape [1, 576, 1024]
            // ne is whcn, ne = [1024, 576, 1, 1]
            embeddings = ggml_get_rows(ctx0, embeddings, patches);

            // print_tensor_info(embeddings, "embeddings");

            // llava projector
            if (ctx->proj_type() == PROJECTOR_TYPE_MLP) {
                embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
                embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);

                embeddings = ggml_gelu(ctx0, embeddings);
                if (model.mm_2_w) {
                    embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
                    embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
                }
            }
            else if (ctx->proj_type() == PROJECTOR_TYPE_MLP_NORM) {
                embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
                embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
                // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
                // First LayerNorm
                embeddings = ggml_norm(ctx0, embeddings, eps);
                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
                                    model.mm_1_b);

                // GELU activation
                embeddings = ggml_gelu(ctx0, embeddings);

                // Second linear layer
                embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
                embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);

                // Second LayerNorm
                embeddings = ggml_norm(ctx0, embeddings, eps);
                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
                                    model.mm_4_b);
            }
            else if (ctx->proj_type() == PROJECTOR_TYPE_LDP) {
                // MobileVLM projector
                int n_patch = 24;
                ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
                mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
                mlp_1 = ggml_gelu(ctx0, mlp_1);
                ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
                mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
                // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]

                // block 1
                ggml_tensor * block_1 = nullptr;
                {
                    // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
                    mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
                    mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
                    // stride = 1, padding = 1, bias is nullptr
                    block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);

                    // layer norm
                    // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
                    // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
                    block_1 = ggml_norm(ctx0, block_1, eps);
                    block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));

                    // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
                    // hardswish
                    ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);

                    block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
                    // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
                    // pointwise conv
                    block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
                    block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
                    block_1 = ggml_relu(ctx0, block_1);
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
                    block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
                    block_1 = ggml_hardsigmoid(ctx0, block_1);
                    // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
                    block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
                    block_1 = ggml_mul(ctx0, block_1_hw, block_1);

                    int w = block_1->ne[0], h = block_1->ne[1];
                    block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));

                    // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
                    block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);

                    // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
                    block_1 = ggml_norm(ctx0, block_1, eps);
                    block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
                    // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
                    // residual
                    block_1 = ggml_add(ctx0, mlp_3, block_1);
                }

                // block_2
                {
                    // stride = 2
                    block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);

                    // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
                    // layer norm
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
                    // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
                    block_1 = ggml_norm(ctx0, block_1, eps);
                    block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
                    // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
                    // hardswish
                    ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);

                    // not sure the parameters is right for globalAvgPooling
                    block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
                    // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
                    // pointwise conv
                    block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
                    block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
                    block_1 = ggml_relu(ctx0, block_1);
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
                    block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
                    block_1 = ggml_hardsigmoid(ctx0, block_1);

                    // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
                    block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
                    block_1 = ggml_mul(ctx0, block_1_hw, block_1);

                    int w = block_1->ne[0], h = block_1->ne[1];
                    block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
                    block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
                    // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
                    block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
                    block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);


                    // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
                    block_1 = ggml_norm(ctx0, block_1, eps);
                    block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
                    block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
                    // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
                }
                embeddings = block_1;
            }
            else if (ctx->proj_type() == PROJECTOR_TYPE_LDPV2)
            {
                int n_patch = 24;
                ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
                mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
                mlp_0 = ggml_gelu(ctx0, mlp_0);
                ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
                mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
                // mlp_2 ne = [2048, 576, 1, 1]
                // // AVG Pool Layer 2*2, strides = 2
                mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
                // mlp_2 ne = [576, 2048, 1, 1]
                mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
                // mlp_2 ne [24, 24, 2048, 1]
                mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
                // weight ne = [3, 3, 2048, 1]
                ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
                peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
                peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
                mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
                peg_0 = ggml_add(ctx0, peg_0, mlp_2);
                peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
                embeddings = peg_0;
            }
            else {
                GGML_ABORT("fatal error");
            }
        }

        // glm projector
        else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE) {
            size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
            embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
            embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
            embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
            embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
            // GLU
            {
                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
                embeddings = ggml_norm(ctx0, embeddings, eps);
                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
                embeddings = ggml_gelu_inplace(ctx0, embeddings);
                ggml_tensor * x = embeddings;
                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
                x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
                embeddings = ggml_swiglu_split(ctx0, embeddings, x);
                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
            }
            // arrangement of BOI/EOI token embeddings
            // note: these embeddings are not present in text model, hence we cannot process them as text tokens
            // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53
            {
                embeddings = ggml_concat(ctx0, model.mm_glm_tok_boi, embeddings, 1); // BOI
                embeddings = ggml_concat(ctx0, embeddings, model.mm_glm_tok_eoi, 1); // EOI
            }
        }

        else {
            GGML_ABORT("llava: unknown projector type");
        }

        // build the graph
        ggml_build_forward_expand(gf, embeddings);

        return gf;
    }

    // whisper encoder with custom projector
    ggml_cgraph * build_whisper_enc() {
        const int n_frames = img.nx;
        const int n_pos    = n_frames / 2;
        GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);

        ggml_tensor * inp = build_inp_raw(1);

        // conv1d block
        {
            // convolution + gelu
            ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
            cur = ggml_add(ctx0, cur, model.conv1d_1_b);

            cur = ggml_gelu_erf(ctx0, cur);

            cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
            cur = ggml_add(ctx0, cur, model.conv1d_2_b);

            cur = ggml_gelu_erf(ctx0, cur);
            // transpose
            inp = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
            cb(inp, "after_conv1d", -1);
        }

        // sanity check (only check one layer, but it should be the same for all)
        GGML_ASSERT(model.layers[0].ln_1_w && model.layers[0].ln_1_b);
        GGML_ASSERT(model.layers[0].ln_2_w && model.layers[0].ln_2_b);
        GGML_ASSERT(model.layers[0].q_b);
        GGML_ASSERT(model.layers[0].v_b);
        GGML_ASSERT(!model.layers[0].k_b); // no bias for k
        GGML_ASSERT(model.post_ln_w && model.post_ln_b);

        ggml_tensor * pos_embd_selected = ggml_view_2d(
            ctx0, model.position_embeddings,
            model.position_embeddings->ne[0], n_pos,
            model.position_embeddings->nb[1], 0
        );
        ggml_tensor * cur = build_vit(
                                inp, n_pos,
                                NORM_TYPE_NORMAL,
                                hparams.ffn_op,
                                pos_embd_selected,
                                nullptr);

        cb(cur, "after_transformer", -1);

        if (model.audio_has_stack_frames()) {
            // StackAudioFrames
            // https://huggingface.co/fixie-ai/ultravox-v0_5-llama-3_2-1b/blob/main/ultravox_model.py
            int64_t stride = n_embd * hparams.proj_stack_factor;
            int64_t padded_len = GGML_PAD(ggml_nelements(cur), stride);
            int64_t pad = padded_len - ggml_nelements(cur);
            if (pad > 0) {
                cur = ggml_view_1d(ctx0, cur, ggml_nelements(cur), 0);
                cur = ggml_pad(ctx0, cur, pad, 0, 0, 0);
            }
            cur = ggml_view_2d(ctx0, cur, stride, padded_len / stride,
                                ggml_row_size(cur->type, stride), 0);
            cb(cur, "after_stacked", -1);
        }

        if (ctx->proj_type() == PROJECTOR_TYPE_ULTRAVOX) {
            // UltravoxProjector
            // pre-norm
            cur = ggml_rms_norm(ctx0, cur, 1e-6);
            cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);

            // ffn in
            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);

            // swiglu
            // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
            cur = ggml_swiglu_swapped(ctx0, cur);

            // mid-norm
            cur = ggml_rms_norm(ctx0, cur, 1e-6);
            cur = ggml_mul(ctx0, cur, model.mm_norm_mid_w);

            // ffn out
            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);

        } else if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2A) {
            // projector
            cur = ggml_mul_mat(ctx0, model.mm_fc_w, cur);
            cur = ggml_add(ctx0, cur, model.mm_fc_b);

        } else if (ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL) {
            // projector
            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
            cur = ggml_gelu_erf(ctx0, cur);
            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);

        } else {
            GGML_ABORT("%s: unknown projector type", __func__);
        }

        cb(cur, "projected", -1);

        ggml_build_forward_expand(gf, cur);

        return gf;
    }

private:
    //
    // utility functions
    //

    void cb(ggml_tensor * cur0, const char * name, int il) const {
        if (ctx->debug_graph) {
            ggml_tensor * cur = ggml_cpy(ctx0, cur0, ggml_dup_tensor(ctx0, cur0));
            std::string cur_name = il >= 0 ? std::string(name) + "_" + std::to_string(il) : name;
            ggml_set_name(cur, cur_name.c_str());
            ggml_set_output(cur);
            ggml_build_forward_expand(gf, cur);
            ctx->debug_print_tensors.push_back(cur);
        }
    }

    // build vision transformer (ViT) cgraph
    // this function should cover most of the models
    // if your model has specific features, you should probably duplicate this function
    ggml_tensor * build_vit(
                ggml_tensor * inp,
                int64_t n_pos,
                norm_type norm_t,
                ffn_op_type ffn_t,
                ggml_tensor * learned_pos_embd,
                std::function<ggml_tensor *(ggml_tensor *, const clip_layer &)> add_pos
            ) {
        if (learned_pos_embd) {
            inp = ggml_add(ctx0, inp, learned_pos_embd);
            cb(inp, "pos_embed", -1);
        }

        ggml_tensor * inpL = inp;

        // pre-layernorm
        if (model.pre_ln_w) {
            inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
            cb(inpL, "pre_ln", -1);
        }

        // loop over layers
        for (int il = 0; il < n_layer; il++) {
            auto & layer = model.layers[il];
            ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states

            // layernorm1
            cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
            cb(cur, "layer_inp_normed", il);

            // self-attention
            {
                ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur);
                if (layer.q_b) {
                    Qcur = ggml_add(ctx0, Qcur, layer.q_b);
                }

                ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
                if (layer.k_b) {
                    Kcur = ggml_add(ctx0, Kcur, layer.k_b);
                }

                ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
                if (layer.v_b) {
                    Vcur = ggml_add(ctx0, Vcur, layer.v_b);
                }

                if (layer.q_norm) {
                    Qcur = build_norm(Qcur, layer.q_norm, NULL, norm_t, eps, il);
                    cb(Qcur, "Qcur_norm", il);
                }

                if (layer.k_norm) {
                    Kcur = build_norm(Kcur, layer.k_norm, NULL, norm_t, eps, il);
                    cb(Kcur, "Kcur_norm", il);
                }

                Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
                Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
                Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);

                cb(Qcur, "Qcur", il);
                cb(Kcur, "Kcur", il);
                cb(Vcur, "Vcur", il);

                if (add_pos) {
                    Qcur = add_pos(Qcur, layer);
                    Kcur = add_pos(Kcur, layer);
                    cb(Qcur, "Qcur_pos", il);
                    cb(Kcur, "Kcur_pos", il);
                }

                cur = build_attn(layer.o_w, layer.o_b,
                    Qcur, Kcur, Vcur, nullptr, kq_scale, il);
                cb(cur, "attn_out", il);
            }

            if (layer.ls_1_w) {
                cur = ggml_mul(ctx0, cur, layer.ls_1_w);
                cb(cur, "attn_out_scaled", il);
            }

            // re-add the layer input, e.g., residual
            cur = ggml_add(ctx0, cur, inpL);

            inpL = cur; // inpL = residual, cur = hidden_states

            cb(cur, "ffn_inp", il);

            // layernorm2
            cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
            cb(cur, "ffn_inp_normed", il);

            // ffn
            cur = build_ffn(cur,
                layer.ff_up_w, layer.ff_up_b,
                layer.ff_gate_w, layer.ff_gate_b,
                layer.ff_down_w, layer.ff_down_b,
                ffn_t, il);

            cb(cur, "ffn_out", il);

            if (layer.ls_2_w) {
                cur = ggml_mul(ctx0, cur, layer.ls_2_w);
                cb(cur, "ffn_out_scaled", il);
            }

            // residual 2
            cur = ggml_add(ctx0, inpL, cur);
            cb(cur, "layer_out", il);

            inpL = cur;
        }

        if (ctx->model.audio_has_avgpool()) {
            ggml_tensor * cur = inpL;
            cur = ggml_transpose(ctx0, cur);
            cur = ggml_cont(ctx0, cur);
            cur = ggml_pool_1d(ctx0, cur, GGML_OP_POOL_AVG, 2, 2, 0);
            cur = ggml_transpose(ctx0, cur);
            cur = ggml_cont(ctx0, cur);
            inpL = cur;
        }

        // post-layernorm
        if (model.post_ln_w) {
            inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, -1);
        }
        return inpL;
    }

    // build the input after conv2d (inp_raw --> patches)
    // returns tensor with shape [n_embd, n_patches]
    ggml_tensor * build_inp() {
        ggml_tensor * inp_raw = build_inp_raw();
        ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
        inp = ggml_reshape_2d(ctx0, inp, n_patches, n_embd);
        inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
        if (model.patch_bias) {
            inp = ggml_add(ctx0, inp, model.patch_bias);
            cb(inp, "patch_bias", -1);
        }
        return inp;
    }

    ggml_tensor * build_inp_raw(int channels = 3) {
        ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, img.nx, img.ny, channels);
        ggml_set_name(inp_raw, "inp_raw");
        ggml_set_input(inp_raw);
        return inp_raw;
    }

    ggml_tensor * build_norm(
            ggml_tensor * cur,
            ggml_tensor * mw,
            ggml_tensor * mb,
            norm_type type,
            float norm_eps,
            int il) const {

        cur = type == NORM_TYPE_RMS
            ? ggml_rms_norm(ctx0, cur, norm_eps)
            : ggml_norm(ctx0, cur, norm_eps);

        if (mw || mb) {
            cb(cur, "norm", il);
        }

        if (mw) {
            cur = ggml_mul(ctx0, cur, mw);
            if (mb) {
                cb(cur, "norm_w", il);
            }
        }

        if (mb) {
            cur = ggml_add(ctx0, cur, mb);
        }

        return cur;
    }

    ggml_tensor * build_ffn(
            ggml_tensor * cur,
            ggml_tensor * up,
            ggml_tensor * up_b,
            ggml_tensor * gate,
            ggml_tensor * gate_b,
            ggml_tensor * down,
            ggml_tensor * down_b,
            ffn_op_type type_op,
            int il) const {

        ggml_tensor * tmp = up ? ggml_mul_mat(ctx0, up, cur) : cur;
        cb(tmp, "ffn_up", il);

        if (up_b) {
            tmp = ggml_add(ctx0, tmp, up_b);
            cb(tmp, "ffn_up_b", il);
        }

        if (gate) {
            cur = ggml_mul_mat(ctx0, gate, cur);
            cb(cur, "ffn_gate", il);

            if (gate_b) {
                cur = ggml_add(ctx0, cur, gate_b);
                cb(cur, "ffn_gate_b", il);
            }
        } else {
            cur = tmp;
        }

        // we only support parallel ffn for now
        switch (type_op) {
            case FFN_SILU:
                if (gate) {
                    cur = ggml_swiglu_split(ctx0, cur, tmp);
                    cb(cur, "ffn_swiglu", il);
                } else {
                    cur = ggml_silu(ctx0, cur);
                    cb(cur, "ffn_silu", il);
                } break;
            case FFN_GELU:
                if (gate) {
                    cur = ggml_geglu_split(ctx0, cur, tmp);
                    cb(cur, "ffn_geglu", il);
                } else {
                    cur = ggml_gelu(ctx0, cur);
                    cb(cur, "ffn_gelu", il);
                } break;
            case FFN_GELU_ERF:
                if (gate) {
                    cur = ggml_geglu_erf_split(ctx0, cur, tmp);
                    cb(cur, "ffn_geglu_erf", il);
                } else {
                    cur = ggml_gelu_erf(ctx0, cur);
                    cb(cur, "ffn_gelu_erf", il);
                } break;
            case FFN_GELU_QUICK:
                if (gate) {
                    cur = ggml_geglu_quick_split(ctx0, cur, tmp);
                    cb(cur, "ffn_geglu_quick", il);
                } else {
                    cur = ggml_gelu_quick(ctx0, cur);
                    cb(cur, "ffn_gelu_quick", il);
                } break;
        }

        if (down) {
            cur = ggml_mul_mat(ctx0, down, cur);
        }

        if (down_b) {
            cb(cur, "ffn_down", il);
        }

        if (down_b) {
            cur = ggml_add(ctx0, cur, down_b);
        }

        return cur;
    }

    ggml_tensor * build_attn(
            ggml_tensor * wo,
            ggml_tensor * wo_b,
            ggml_tensor * q_cur,
            ggml_tensor * k_cur,
            ggml_tensor * v_cur,
            ggml_tensor * kq_mask,
            float kq_scale,
            int il) const {
        // these nodes are added to the graph together so that they are not reordered
        // by doing so, the number of splits in the graph is reduced
        ggml_build_forward_expand(gf, q_cur);
        ggml_build_forward_expand(gf, k_cur);
        ggml_build_forward_expand(gf, v_cur);

        ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
        //cb(q, "q", il);

        ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
        //cb(k, "k", il);

        ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3);
        v = ggml_cont(ctx0, v);
        //cb(k, "v", il);

        ggml_tensor * cur;

        // TODO @ngxson : support flash attention
        {
            const auto n_tokens = q->ne[1];
            const auto n_head   = q->ne[2];
            // const auto n_kv     = k->ne[1]; // for flash attention

            ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
            // F32 may not needed for vision encoders?
            // ggml_mul_mat_set_prec(kq, GGML_PREC_F32);

            kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, 0.0f);

            ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
            cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
            cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
        }

        cb(cur, "kqv_out", il);

        if (wo) {
            cur = ggml_mul_mat(ctx0, wo, cur);
        }

        if (wo_b) {
            cur = ggml_add(ctx0, cur, wo_b);
        }

        return cur;
    }

    // implementation of the 2D RoPE without adding a new op in ggml
    // this is not efficient (use double the memory), but works on all backends
    // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
    static ggml_tensor * build_rope_2d(
        ggml_context * ctx0,
        ggml_tensor * cur,
        ggml_tensor * pos_a, // first half
        ggml_tensor * pos_b, // second half
        const float freq_base,
        const bool interleave_freq
    ) {
        const int64_t n_dim  = cur->ne[0];
        const int64_t n_head = cur->ne[1];
        const int64_t n_pos  = cur->ne[2];

        // for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos)
        // we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3
        // first half of cur will use 1e-0, 1e-2 (even)
        // second half of cur will use 1e-1, 1e-3 (odd)
        // the trick here is to rotate just half of n_dim, so inv_freq will automatically be even
        //  ^ don't ask me why, it's math! -2(2i) / n_dim == -2i / (n_dim/2)
        // then for the second half, we use freq_scale to shift the inv_freq
        //  ^ why? replace (2i) with (2i+1) in the above equation
        const float freq_scale_odd = interleave_freq
                                    ? std::pow(freq_base, (float)-2/n_dim)
                                    : 1.0;

        // first half
        ggml_tensor * first;
        {
            first = ggml_view_3d(ctx0, cur,
                n_dim/2, n_head, n_pos,
                ggml_row_size(cur->type, n_dim),
                ggml_row_size(cur->type, n_dim*n_head),
                0);
            first = ggml_rope_ext(
                ctx0,
                first,
                pos_a,      // positions
                nullptr,    // freq factors
                n_dim/2,    // n_dims
                0, 0, freq_base,
                1.0f, 0.0f, 1.0f, 0.0f, 0.0f
            );
        }

        // second half
        ggml_tensor * second;
        {
            second = ggml_view_3d(ctx0, cur,
                n_dim/2, n_head, n_pos,
                ggml_row_size(cur->type, n_dim),
                ggml_row_size(cur->type, n_dim*n_head),
                n_dim/2 * ggml_element_size(cur));
            second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors
            second = ggml_rope_ext(
                ctx0,
                second,
                pos_b,      // positions
                nullptr,    // freq factors
                n_dim/2,    // n_dims
                0, 0, freq_base,
                freq_scale_odd,
                0.0f, 1.0f, 0.0f, 0.0f
            );
        }

        cur = ggml_concat(ctx0, first, second, 0);
        return cur;
    }

};

static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
    GGML_ASSERT(imgs.entries.size() == 1 && "n_batch > 1 is not supported");
    clip_graph graph(ctx, *imgs.entries[0]);

    ggml_cgraph * res;

    switch (ctx->proj_type()) {
        case PROJECTOR_TYPE_GEMMA3:
        case PROJECTOR_TYPE_IDEFICS3:
            {
                res = graph.build_siglip();
            } break;
        case PROJECTOR_TYPE_PIXTRAL:
            {
                res = graph.build_pixtral();
            } break;
        case PROJECTOR_TYPE_QWEN2VL:
        case PROJECTOR_TYPE_QWEN25VL:
            {
                res = graph.build_qwen2vl();
            } break;
        case PROJECTOR_TYPE_MINICPMV:
            {
                res = graph.build_minicpmv();
            } break;
        case PROJECTOR_TYPE_INTERNVL:
            {
                res = graph.build_internvl();
            } break;
        case PROJECTOR_TYPE_LLAMA4:
            {
                res = graph.build_llama4();
            } break;
        case PROJECTOR_TYPE_ULTRAVOX:
        case PROJECTOR_TYPE_VOXTRAL:
        case PROJECTOR_TYPE_QWEN2A:
            {
                res = graph.build_whisper_enc();
            } break;
        default:
            {
                res = graph.build_llava();
            } break;
    }
    return res;
}

struct clip_model_loader {
    ggml_context_ptr ctx_meta;
    gguf_context_ptr ctx_gguf;

    std::string fname;

    size_t model_size = 0; // in bytes

    bool has_vision = false;
    bool has_audio  = false;

    // TODO @ngxson : we should not pass clip_ctx here, it should be clip_model
    clip_model_loader(const char * fname) : fname(fname) {
        struct ggml_context * meta = nullptr;

        struct gguf_init_params params = {
            /*.no_alloc = */ true,
            /*.ctx      = */ &meta,
        };

        ctx_gguf = gguf_context_ptr(gguf_init_from_file(fname, params));
        if (!ctx_gguf.get()) {
            throw std::runtime_error(string_format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
        }

        ctx_meta.reset(meta);

        const int n_tensors = gguf_get_n_tensors(ctx_gguf.get());

        // print gguf info
        {
            std::string name;
            get_string(KEY_NAME, name, false);
            std::string description;
            get_string(KEY_DESCRIPTION, description, false);
            LOG_INF("%s: model name:   %s\n",  __func__, name.c_str());
            LOG_INF("%s: description:  %s\n",  __func__, description.c_str());
            LOG_INF("%s: GGUF version: %d\n",  __func__, gguf_get_version(ctx_gguf.get()));
            LOG_INF("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx_gguf.get()));
            LOG_INF("%s: n_tensors:    %d\n",  __func__, n_tensors);
            LOG_INF("%s: n_kv:         %d\n",  __func__, (int)gguf_get_n_kv(ctx_gguf.get()));
            LOG_INF("\n");
        }

        // modalities
        {
            get_bool(KEY_HAS_VISION_ENC, has_vision, false);
            get_bool(KEY_HAS_AUDIO_ENC,  has_audio,  false);

            if (has_vision) {
                LOG_INF("%s: has vision encoder\n", __func__);
            }
            if (has_audio) {
                LOG_INF("%s: has audio encoder\n", __func__);
            }
        }

        // tensors
        {
            for (int i = 0; i < n_tensors; ++i) {
                const char * name = gguf_get_tensor_name(ctx_gguf.get(), i);
                const size_t offset = gguf_get_tensor_offset(ctx_gguf.get(), i);
                enum ggml_type type = gguf_get_tensor_type(ctx_gguf.get(), i);
                ggml_tensor * cur = ggml_get_tensor(meta, name);
                size_t tensor_size = ggml_nbytes(cur);
                model_size += tensor_size;
                LOG_DBG("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
                    __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
            }
        }
    }

    void load_hparams(clip_model & model, clip_modality modality) {
        auto & hparams = model.hparams;
        std::string log_ffn_op; // for logging

        // sanity check
        if (modality == CLIP_MODALITY_VISION) {
            GGML_ASSERT(has_vision);
        } else if (modality == CLIP_MODALITY_AUDIO) {
            GGML_ASSERT(has_audio);
        }
        model.modality = modality;


        // projector type
        std::string proj_type;
        {
            get_string(KEY_PROJ_TYPE, proj_type, false);
            if (!proj_type.empty()) {
                model.proj_type = clip_projector_type_from_string(proj_type);
            }
            if (model.proj_type == PROJECTOR_TYPE_UNKNOWN) {
                throw std::runtime_error(string_format("%s: unknown projector type: %s\n", __func__, proj_type.c_str()));
            }

            // correct arch for multimodal models
            if (model.proj_type == PROJECTOR_TYPE_QWEN25O) {
                model.proj_type = modality == CLIP_MODALITY_VISION
                                    ? PROJECTOR_TYPE_QWEN25VL
                                    : PROJECTOR_TYPE_QWEN2A;
            }
        }

        const bool is_vision = model.modality == CLIP_MODALITY_VISION;
        const bool is_audio  = model.modality == CLIP_MODALITY_AUDIO;

        // other hparams
        {
            const char * prefix = is_vision ? "vision" : "audio";
            get_u32(string_format(KEY_N_EMBD,         prefix), hparams.n_embd);
            get_u32(string_format(KEY_N_HEAD,         prefix), hparams.n_head);
            get_u32(string_format(KEY_N_FF,           prefix), hparams.n_ff);
            get_u32(string_format(KEY_N_BLOCK,        prefix), hparams.n_layer);
            get_u32(string_format(KEY_PROJ_DIM,       prefix), hparams.projection_dim);
            get_f32(string_format(KEY_LAYER_NORM_EPS, prefix), hparams.eps);

            if (is_vision) {
                get_u32(KEY_IMAGE_SIZE, hparams.image_size);
                get_u32(KEY_PATCH_SIZE, hparams.patch_size);
                get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy

            } else if (is_audio) {
                get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);

            } else {
                GGML_ASSERT(false && "unknown modality");
            }

            // for pinpoints, we need to convert it into a list of resolution candidates
            {
                std::vector<int> pinpoints;
                get_arr_int(KEY_IMAGE_GRID_PINPOINTS, pinpoints, false);
                if (!pinpoints.empty()) {
                    for (size_t i = 0; i < pinpoints.size(); i += 2) {
                        hparams.image_res_candidates.push_back({
                            pinpoints[i],
                            pinpoints[i+1],
                        });
                    }
                }
            }

            // default warmup value
            hparams.warmup_image_size = hparams.image_size;

            hparams.has_llava_projector = model.proj_type == PROJECTOR_TYPE_MLP
                                       || model.proj_type == PROJECTOR_TYPE_MLP_NORM
                                       || model.proj_type == PROJECTOR_TYPE_LDP
                                       || model.proj_type == PROJECTOR_TYPE_LDPV2;

            {
                bool use_gelu = false;
                bool use_silu = false;
                get_bool(KEY_USE_GELU, use_gelu, false);
                get_bool(KEY_USE_SILU, use_silu, false);
                if (use_gelu && use_silu) {
                    throw std::runtime_error(string_format("%s: both use_gelu and use_silu are set to true\n", __func__));
                }
                if (use_gelu) {
                    hparams.ffn_op = FFN_GELU;
                    log_ffn_op = "gelu";
                } else if (use_silu) {
                    hparams.ffn_op = FFN_SILU;
                    log_ffn_op = "silu";
                } else {
                    hparams.ffn_op = FFN_GELU_QUICK;
                    log_ffn_op = "gelu_quick";
                }
            }

            {
                std::string mm_patch_merge_type;
                get_string(KEY_MM_PATCH_MERGE_TYPE, mm_patch_merge_type, false);
                if (mm_patch_merge_type == "spatial_unpad") {
                    hparams.mm_patch_merge_type = PATCH_MERGE_SPATIAL_UNPAD;
                }
            }

            if (is_vision) {
                int idx_mean = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_MEAN);
                int idx_std  = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_STD);
                GGML_ASSERT(idx_mean >= 0 && "image_mean not found");
                GGML_ASSERT(idx_std >= 0  && "image_std not found");
                const float * mean_data = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_mean);
                const float * std_data  = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_std);
                for (int i = 0; i < 3; ++i) {
                    hparams.image_mean[i] = mean_data[i];
                    hparams.image_std[i]  = std_data[i];
                }
            }

            // Load the vision feature layer indices if they are explicitly provided;
            // if multiple vision feature layers are present, the values will be concatenated
            // to form the final visual features.
            // NOTE: gguf conversions should standardize the values of the vision feature layer to
            // be non-negative, since we use -1 to mark values as unset here.
            std::vector<int> vision_feature_layer;
            get_arr_int(KEY_FEATURE_LAYER, vision_feature_layer, false);
            // convert std::vector to std::unordered_set
            for (auto & layer : vision_feature_layer) {
                hparams.vision_feature_layer.insert(layer);
            }

            // model-specific params
            switch (model.proj_type) {
                case PROJECTOR_TYPE_MINICPMV:
                    {
                        if (hparams.minicpmv_version == 0) {
                            hparams.minicpmv_version = 2; // default to 2 if not set
                        }
                    } break;
                case PROJECTOR_TYPE_IDEFICS3:
                case PROJECTOR_TYPE_INTERNVL:
                    {
                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
                    } break;
                case PROJECTOR_TYPE_PIXTRAL:
                    {
                        hparams.rope_theta = 10000.0f;
                        hparams.warmup_image_size = hparams.patch_size * 8;
                        // Mistral Small 2506 needs 1024x1024 image size cap to prevent OOM
                        // ref: https://github.com/ggml-org/llama.cpp/issues/14310
                        hparams.image_size = 1024;
                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                    } break;
                case PROJECTOR_TYPE_GEMMA3:
                    {
                        // default value (used by all model sizes in gemma 3 family)
                        // number of patches for each **side** is reduced by a factor of 4
                        hparams.proj_scale_factor = 4;
                        // test model (tinygemma3) has a different value, we optionally read it
                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
                    } break;
                case PROJECTOR_TYPE_QWEN2VL:
                    {
                        // max image size = sqrt(max_pixels) = 3584
                        // ref: https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json
                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
                        hparams.image_size = 1024;
                        hparams.warmup_image_size = hparams.patch_size * 8;
                    } break;
                case PROJECTOR_TYPE_QWEN25VL:
                    {
                        // max image size = sqrt(max_pixels)
                        // https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
                        hparams.image_size = 1024;
                        hparams.warmup_image_size = hparams.patch_size * 8;
                        get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern);
                    } break;
                case PROJECTOR_TYPE_LLAMA4:
                    {
                        hparams.rope_theta = 10000.0f;
                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor);
                        set_llava_uhd_res_candidates(model, 3);
                    } break;
                case PROJECTOR_TYPE_ULTRAVOX:
                case PROJECTOR_TYPE_QWEN2A:
                case PROJECTOR_TYPE_VOXTRAL:
                    {
                        bool require_stack = model.proj_type == PROJECTOR_TYPE_ULTRAVOX ||
                                             model.proj_type == PROJECTOR_TYPE_VOXTRAL;
                        get_u32(KEY_A_PROJ_STACK_FACTOR, hparams.proj_stack_factor, require_stack);
                        if (hparams.n_mel_bins != 128) {
                            throw std::runtime_error(string_format("%s: only 128 mel bins are supported for ultravox\n", __func__));
                        }
                        hparams.ffn_op = FFN_GELU_ERF;
                        log_ffn_op = "gelu_erf"; // temporary solution for logging
                    } break;
                default:
                    break;
            }

            LOG_INF("%s: projector:          %s\n", __func__, proj_type.c_str());
            LOG_INF("%s: n_embd:             %d\n", __func__, hparams.n_embd);
            LOG_INF("%s: n_head:             %d\n", __func__, hparams.n_head);
            LOG_INF("%s: n_ff:               %d\n", __func__, hparams.n_ff);
            LOG_INF("%s: n_layer:            %d\n", __func__, hparams.n_layer);
            LOG_INF("%s: ffn_op:             %s\n", __func__, log_ffn_op.c_str());
            LOG_INF("%s: projection_dim:     %d\n", __func__, hparams.projection_dim);
            if (is_vision) {
                LOG_INF("\n--- vision hparams ---\n");
                LOG_INF("%s: image_size:         %d\n", __func__, hparams.image_size);
                LOG_INF("%s: patch_size:         %d\n", __func__, hparams.patch_size);
                LOG_INF("%s: has_llava_proj:     %d\n", __func__, hparams.has_llava_projector);
                LOG_INF("%s: minicpmv_version:   %d\n", __func__, hparams.minicpmv_version);
                LOG_INF("%s: proj_scale_factor:  %d\n", __func__, hparams.proj_scale_factor);
                LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
            } else if (is_audio) {
                LOG_INF("\n--- audio hparams ---\n");
                LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
                LOG_INF("%s: proj_stack_factor:  %d\n", __func__, hparams.proj_stack_factor);
            }
            LOG_INF("\n");
            LOG_INF("%s: model size:         %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0);
            LOG_INF("%s: metadata size:      %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0);
        }
    }

    void load_tensors(clip_ctx & ctx_clip) {
        auto & model = ctx_clip.model;
        auto & hparams = model.hparams;
        std::map<std::string, size_t> tensor_offset;
        std::vector<ggml_tensor *> tensors_to_load;

        // TODO @ngxson : support both audio and video in the future
        const char * prefix = model.modality == CLIP_MODALITY_AUDIO ? "a" : "v";

        // get offsets
        for (int64_t i = 0; i < gguf_get_n_tensors(ctx_gguf.get()); ++i) {
            const char * name = gguf_get_tensor_name(ctx_gguf.get(), i);
            tensor_offset[name] = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), i);
        }

        // create data context
        struct ggml_init_params params = {
            /*.mem_size =*/ static_cast<size_t>(gguf_get_n_tensors(ctx_gguf.get()) + 1) * ggml_tensor_overhead(),
            /*.mem_buffer =*/ NULL,
            /*.no_alloc =*/ true,
        };
        ctx_clip.ctx_data.reset(ggml_init(params));
        if (!ctx_clip.ctx_data) {
            throw std::runtime_error(string_format("%s: failed to init ggml context\n", __func__));
        }

        // helper function
        auto get_tensor = [&](const std::string & name, bool required = true) {
            ggml_tensor * cur = ggml_get_tensor(ctx_meta.get(), name.c_str());
            if (!cur && required) {
                throw std::runtime_error(string_format("%s: unable to find tensor %s\n", __func__, name.c_str()));
            }
            if (cur) {
                tensors_to_load.push_back(cur);
                // add tensors to context
                ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data.get(), cur);
                ggml_set_name(data_tensor, cur->name);
                cur = data_tensor;
            }
            return cur;
        };

        model.class_embedding = get_tensor(TN_CLASS_EMBD, false);

        model.pre_ln_w = get_tensor(string_format(TN_LN_PRE, prefix, "weight"), false);
        model.pre_ln_b = get_tensor(string_format(TN_LN_PRE, prefix, "bias"),   false);

        model.post_ln_w = get_tensor(string_format(TN_LN_POST, prefix, "weight"), false);
        model.post_ln_b = get_tensor(string_format(TN_LN_POST, prefix, "bias"),   false);

        model.patch_bias = get_tensor(TN_PATCH_BIAS, false);
        model.patch_embeddings_0 = get_tensor(TN_PATCH_EMBD,   false);
        model.patch_embeddings_1 = get_tensor(TN_PATCH_EMBD_1, false);

        model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, prefix), false);

        // layers
        model.layers.resize(hparams.n_layer);
        for (int il = 0; il < hparams.n_layer; ++il) {
            auto & layer = model.layers[il];
            layer.k_w    = get_tensor(string_format(TN_ATTN_K,      prefix, il, "weight"));
            layer.q_w    = get_tensor(string_format(TN_ATTN_Q,      prefix, il, "weight"));
            layer.v_w    = get_tensor(string_format(TN_ATTN_V,      prefix, il, "weight"));
            layer.o_w    = get_tensor(string_format(TN_ATTN_OUTPUT, prefix, il, "weight"));
            layer.k_norm = get_tensor(string_format(TN_ATTN_K_NORM, prefix, il, "weight"), false);
            layer.q_norm = get_tensor(string_format(TN_ATTN_Q_NORM, prefix, il, "weight"), false);
            layer.ln_1_w = get_tensor(string_format(TN_LN_1,        prefix, il, "weight"), false);
            layer.ln_2_w = get_tensor(string_format(TN_LN_2,        prefix, il, "weight"), false);
            layer.ls_1_w = get_tensor(string_format(TN_LS_1,        prefix, il, "weight"), false); // no bias
            layer.ls_2_w = get_tensor(string_format(TN_LS_2,        prefix, il, "weight"), false); // no bias

            layer.k_b    = get_tensor(string_format(TN_ATTN_K,      prefix, il, "bias"), false);
            layer.q_b    = get_tensor(string_format(TN_ATTN_Q,      prefix, il, "bias"), false);
            layer.v_b    = get_tensor(string_format(TN_ATTN_V,      prefix, il, "bias"), false);
            layer.o_b    = get_tensor(string_format(TN_ATTN_OUTPUT, prefix, il, "bias"), false);
            layer.ln_1_b = get_tensor(string_format(TN_LN_1,        prefix, il, "bias"), false);
            layer.ln_2_b = get_tensor(string_format(TN_LN_2,        prefix, il, "bias"), false);

            // ffn
            layer.ff_up_w   = get_tensor(string_format(TN_FFN_UP,   prefix, il, "weight"));
            layer.ff_up_b   = get_tensor(string_format(TN_FFN_UP,   prefix, il, "bias"),   false);
            layer.ff_gate_w = get_tensor(string_format(TN_FFN_GATE, prefix, il, "weight"), false);
            layer.ff_gate_b = get_tensor(string_format(TN_FFN_GATE, prefix, il, "bias"),   false);
            layer.ff_down_w = get_tensor(string_format(TN_FFN_DOWN, prefix, il, "weight"));
            layer.ff_down_b = get_tensor(string_format(TN_FFN_DOWN, prefix, il, "bias"),   false);

            // some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here
            // note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check!
            if (layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd) {
                // swap up and down weights
                ggml_tensor * tmp = layer.ff_up_w;
                layer.ff_up_w = layer.ff_down_w;
                layer.ff_down_w = tmp;
                // swap up and down biases
                tmp = layer.ff_up_b;
                layer.ff_up_b = layer.ff_down_b;
                layer.ff_down_b = tmp;
            }
        }

        switch (model.proj_type) {
            case PROJECTOR_TYPE_MLP:
            case PROJECTOR_TYPE_MLP_NORM:
                {
                    // LLaVA projection
                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"), false);
                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"), false);
                    // Yi-type llava
                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"), false);
                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
                    // missing in Yi-type llava
                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"), false);
                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
                    // Yi-type llava
                    model.mm_3_w = get_tensor(string_format(TN_LLAVA_PROJ, 3, "weight"), false);
                    model.mm_3_b = get_tensor(string_format(TN_LLAVA_PROJ, 3, "bias"), false);
                    model.mm_4_w = get_tensor(string_format(TN_LLAVA_PROJ, 4, "weight"), false);
                    model.mm_4_b = get_tensor(string_format(TN_LLAVA_PROJ, 4, "bias"), false);
                    if (model.mm_3_w) {
                        // TODO: this is a hack to support Yi-type llava
                        model.proj_type = PROJECTOR_TYPE_MLP_NORM;
                    }
                    model.image_newline = get_tensor(TN_IMAGE_NEWLINE, false);
                } break;
            case PROJECTOR_TYPE_LDP:
                {
                    // MobileVLM projection
                    model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
                    model.mm_model_mlp_1_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "bias"));
                    model.mm_model_mlp_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
                    model.mm_model_mlp_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
                    model.mm_model_block_1_block_0_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
                    model.mm_model_block_1_block_0_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
                    model.mm_model_block_1_block_0_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
                    model.mm_model_block_1_block_1_fc1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight"));
                    model.mm_model_block_1_block_1_fc1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias"));
                    model.mm_model_block_1_block_1_fc2_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight"));
                    model.mm_model_block_1_block_1_fc2_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias"));
                    model.mm_model_block_1_block_2_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
                    model.mm_model_block_1_block_2_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
                    model.mm_model_block_1_block_2_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
                    model.mm_model_block_2_block_0_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
                    model.mm_model_block_2_block_0_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
                    model.mm_model_block_2_block_0_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
                    model.mm_model_block_2_block_1_fc1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight"));
                    model.mm_model_block_2_block_1_fc1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias"));
                    model.mm_model_block_2_block_1_fc2_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight"));
                    model.mm_model_block_2_block_1_fc2_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias"));
                    model.mm_model_block_2_block_2_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
                    model.mm_model_block_2_block_2_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
                    model.mm_model_block_2_block_2_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
                } break;
            case PROJECTOR_TYPE_LDPV2:
                {
                    // MobilVLM_V2 projection
                    model.mm_model_mlp_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
                    model.mm_model_mlp_0_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "bias"));
                    model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "weight"));
                    model.mm_model_mlp_2_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "bias"));
                    model.mm_model_peg_0_w = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "weight"));
                    model.mm_model_peg_0_b = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "bias"));
                } break;
            case PROJECTOR_TYPE_MINICPMV:
                {
                    // model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
                    model.mm_model_pos_embed_k = get_tensor(TN_MINICPMV_POS_EMBD_K);
                    model.mm_model_query = get_tensor(TN_MINICPMV_QUERY);
                    model.mm_model_proj = get_tensor(TN_MINICPMV_PROJ);
                    model.mm_model_kv_proj = get_tensor(TN_MINICPMV_KV_PROJ);
                    model.mm_model_attn_q_w = get_tensor(string_format(TN_MINICPMV_ATTN, "q", "weight"));
                    model.mm_model_attn_k_w = get_tensor(string_format(TN_MINICPMV_ATTN, "k", "weight"));
                    model.mm_model_attn_v_w = get_tensor(string_format(TN_MINICPMV_ATTN, "v", "weight"));
                    model.mm_model_attn_q_b = get_tensor(string_format(TN_MINICPMV_ATTN, "q", "bias"));
                    model.mm_model_attn_k_b = get_tensor(string_format(TN_MINICPMV_ATTN, "k", "bias"));
                    model.mm_model_attn_v_b = get_tensor(string_format(TN_MINICPMV_ATTN, "v", "bias"));
                    model.mm_model_attn_o_w = get_tensor(string_format(TN_MINICPMV_ATTN, "out", "weight"));
                    model.mm_model_attn_o_b = get_tensor(string_format(TN_MINICPMV_ATTN, "out", "bias"));
                    model.mm_model_ln_q_w = get_tensor(string_format(TN_MINICPMV_LN, "q", "weight"));
                    model.mm_model_ln_q_b = get_tensor(string_format(TN_MINICPMV_LN, "q", "bias"));
                    model.mm_model_ln_kv_w = get_tensor(string_format(TN_MINICPMV_LN, "kv", "weight"));
                    model.mm_model_ln_kv_b = get_tensor(string_format(TN_MINICPMV_LN, "kv", "bias"));
                    model.mm_model_ln_post_w = get_tensor(string_format(TN_MINICPMV_LN, "post", "weight"));
                    model.mm_model_ln_post_b = get_tensor(string_format(TN_MINICPMV_LN, "post", "bias"));
                } break;
            case PROJECTOR_TYPE_GLM_EDGE:
                {
                    model.mm_model_adapter_conv_w = get_tensor(string_format(TN_GLM_ADAPER_CONV, "weight"));
                    model.mm_model_adapter_conv_b = get_tensor(string_format(TN_GLM_ADAPER_CONV, "bias"));
                    model.mm_model_mlp_0_w = get_tensor(string_format(TN_GLM_ADAPTER_LINEAR, "weight"));
                    model.mm_model_ln_q_w = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1, "weight"));
                    model.mm_model_ln_q_b = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1, "bias"));
                    model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H, "weight"));
                    model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE, "weight"));
                    model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H, "weight"));
                    model.mm_glm_tok_boi = get_tensor(string_format(TN_TOK_GLM_BOI, "weight"));
                    model.mm_glm_tok_eoi = get_tensor(string_format(TN_TOK_GLM_EOI, "weight"));
                } break;
            case PROJECTOR_TYPE_QWEN2VL:
            case PROJECTOR_TYPE_QWEN25VL:
                {
                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                } break;
            case PROJECTOR_TYPE_GEMMA3:
                {
                    model.mm_input_proj_w = get_tensor(TN_MM_INP_PROJ);
                    model.mm_soft_emb_norm_w = get_tensor(TN_MM_SOFT_EMB_N);
                } break;
            case PROJECTOR_TYPE_IDEFICS3:
                {
                    model.projection = get_tensor(TN_MM_PROJECTOR);
                } break;
            case PROJECTOR_TYPE_PIXTRAL:
                {
                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
                    // [IMG_BREAK] token embedding
                    model.token_embd_img_break = get_tensor(TN_TOK_IMG_BREAK);
                    // for mistral small 3.1
                    model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
                    model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                } break;
            case PROJECTOR_TYPE_ULTRAVOX:
                {
                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                    model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"));
                    model.mm_norm_mid_w = get_tensor(string_format(TN_MM_NORM_MID, "weight"));
                } break;
            case PROJECTOR_TYPE_QWEN2A:
                {
                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
                    model.mm_fc_w = get_tensor(string_format(TN_MM_AUDIO_FC, "weight"));
                    model.mm_fc_b = get_tensor(string_format(TN_MM_AUDIO_FC, "bias"));
                } break;
            case PROJECTOR_TYPE_VOXTRAL:
                {
                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                } break;
            case PROJECTOR_TYPE_INTERNVL:
                {
                    model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
                    model.mm_0_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "bias"));
                    model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
                    model.mm_1_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "bias"));
                    model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
                    model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
                } break;
            case PROJECTOR_TYPE_LLAMA4:
                {
                    model.mm_model_proj    = get_tensor(TN_MM_PROJECTOR);
                    model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
                    model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "weight"));
                } break;
            default:
                GGML_ASSERT(false && "unknown projector type");
        }

        // load data
        {
            std::vector<uint8_t> read_buf;

            auto fin = std::ifstream(fname, std::ios::binary);
            if (!fin) {
                throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
            }

            // alloc memory and offload data
            ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend);
            ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft));
            ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
            for (auto & t : tensors_to_load) {
                ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
                const size_t offset = tensor_offset[t->name];
                fin.seekg(offset, std::ios::beg);
                if (!fin) {
                    throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name));
                }
                size_t num_bytes = ggml_nbytes(cur);
                if (ggml_backend_buft_is_host(buft)) {
                    // for the CPU and Metal backend, we can read directly into the tensor
                    fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
                } else {
                    // read into a temporary buffer first, then copy to device memory
                    read_buf.resize(num_bytes);
                    fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
                    ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                }
            }
            fin.close();

            LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
        }
    }

    void alloc_compute_meta(clip_ctx & ctx_clip) {
        const auto & hparams = ctx_clip.model.hparams;
        ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());

        // create a fake batch
        clip_image_f32_batch batch;
        clip_image_f32_ptr img(clip_image_f32_init());
        if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
            img->nx = hparams.warmup_image_size;
            img->ny = hparams.warmup_image_size;
        } else {
            img->nx = hparams.warmup_audio_size;
            img->ny = hparams.n_mel_bins;
        }
        batch.entries.push_back(std::move(img));

        ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch);
        ggml_backend_sched_reserve(ctx_clip.sched.get(), gf);

        for (size_t i = 0; i < ctx_clip.backend_ptrs.size(); ++i) {
            ggml_backend_t backend = ctx_clip.backend_ptrs[i];
            ggml_backend_buffer_type_t buft = ctx_clip.backend_buft[i];
            size_t size = ggml_backend_sched_get_buffer_size(ctx_clip.sched.get(), backend);
            if (size > 1) {
                LOG_INF("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                        ggml_backend_buft_name(buft),
                        size / 1024.0 / 1024.0);
            }
        }
    }

    void get_bool(const std::string & key, bool & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        output = gguf_get_val_bool(ctx_gguf.get(), i);
    }

    void get_i32(const std::string & key, int & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        output = gguf_get_val_i32(ctx_gguf.get(), i);
    }

    void get_u32(const std::string & key, int & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        output = gguf_get_val_u32(ctx_gguf.get(), i);
    }

    void get_f32(const std::string & key, float & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        output = gguf_get_val_f32(ctx_gguf.get(), i);
    }

    void get_string(const std::string & key, std::string & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        output = std::string(gguf_get_val_str(ctx_gguf.get(), i));
    }

    void get_arr_int(const std::string & key, std::vector<int> & output, bool required = true) {
        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
        if (i < 0) {
            if (required) throw std::runtime_error("Key not found: " + key);
            return;
        }
        int n = gguf_get_arr_n(ctx_gguf.get(), i);
        output.resize(n);
        const int32_t * values = (const int32_t *)gguf_get_arr_data(ctx_gguf.get(), i);
        for (int i = 0; i < n; ++i) {
            output[i] = values[i];
        }
    }

    void set_llava_uhd_res_candidates(clip_model & model, const int max_patches_per_side) {
        auto & hparams = model.hparams;
        for (int x = 1; x <= max_patches_per_side; x++) {
            for (int y = 1; y <= max_patches_per_side; y++) {
                if (x == 1 && y == 1) {
                    continue; // skip the first point
                }
                hparams.image_res_candidates.push_back(clip_image_size{
                    x*hparams.image_size,
                    y*hparams.image_size,
                });
            }
        }
    }
};

struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params) {
    g_logger_state.verbosity_thold = ctx_params.verbosity;
    clip_ctx * ctx_vision = nullptr;
    clip_ctx * ctx_audio = nullptr;

    try {
        clip_model_loader loader(fname);

        if (loader.has_vision) {
            ctx_vision = new clip_ctx(ctx_params);
            loader.load_hparams(ctx_vision->model, CLIP_MODALITY_VISION);
            loader.load_tensors(*ctx_vision);
            loader.alloc_compute_meta(*ctx_vision);
        }

        if (loader.has_audio) {
            ctx_audio = new clip_ctx(ctx_params);
            loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
            loader.load_tensors(*ctx_audio);
            loader.alloc_compute_meta(*ctx_audio);
        }

    } catch (const std::exception & e) {
        LOG_ERR("%s: failed to load model '%s': %s\n", __func__, fname, e.what());
        if (ctx_vision) {
            delete ctx_vision;
        }
        if (ctx_audio) {
            delete ctx_audio;
        }
        return {nullptr, nullptr};
    }

    return {ctx_vision, ctx_audio};
}

struct clip_image_size * clip_image_size_init() {
    struct clip_image_size * load_image_size = new struct clip_image_size();
    load_image_size->width = 448;
    load_image_size->height = 448;
    return load_image_size;
}

struct clip_image_u8 * clip_image_u8_init() {
    return new clip_image_u8();
}

struct clip_image_f32 * clip_image_f32_init() {
    return new clip_image_f32();
}

struct clip_image_f32_batch * clip_image_f32_batch_init() {
    return new clip_image_f32_batch();
}

unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny) {
    if (nx) *nx = img->nx;
    if (ny) *ny = img->ny;
    return img->buf.data();
}

void clip_image_size_free(struct clip_image_size * load_image_size) {
    if (load_image_size == nullptr) {
        return;
    }
    delete load_image_size;
}
void clip_image_u8_free(struct clip_image_u8  * img) { if (img) delete img; }
void clip_image_f32_free(struct clip_image_f32 * img) { if (img) delete img; }
void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { if (batch) delete batch; }
void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { if (batch) delete batch; }

size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch) {
    return batch->entries.size();
}

size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx) {
    if (idx < 0 || idx >= (int)batch->entries.size()) {
        LOG_ERR("%s: invalid index %d\n", __func__, idx);
        return 0;
    }
    return batch->entries[idx]->nx;
}

size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx) {
    if (idx < 0 || idx >= (int)batch->entries.size()) {
        LOG_ERR("%s: invalid index %d\n", __func__, idx);
        return 0;
    }
    return batch->entries[idx]->ny;
}

clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx) {
    if (idx < 0 || idx >= (int)batch->entries.size()) {
        LOG_ERR("%s: invalid index %d\n", __func__, idx);
        return nullptr;
    }
    return batch->entries[idx].get();
}

void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img) {
    img->nx = nx;
    img->ny = ny;
    img->buf.resize(3 * nx * ny);
    memcpy(img->buf.data(), rgb_pixels, img->buf.size());
}

// Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not
static void normalize_image_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst, const float mean[3], const float std[3]) {
    dst.nx = src.nx;
    dst.ny = src.ny;
    dst.buf.resize(src.buf.size());

    // TODO @ngxson : seems like this could be done more efficiently on cgraph
    for (size_t i = 0; i < src.buf.size(); ++i) {
        int c = i % 3; // rgb
        dst.buf[i] = (static_cast<float>(src.buf[i]) / 255.0f - mean[c]) / std[c];
    }
}

// set of tools to manupulate images
// in the future, we can have HW acceleration by allowing this struct to access 3rd party lib like imagick or opencv
struct image_manipulation {
    // Bilinear resize function
    static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) {
        dst.nx = target_width;
        dst.ny = target_height;
        dst.buf.resize(3 * target_width * target_height);

        float x_ratio = static_cast<float>(src.nx - 1) / target_width;
        float y_ratio = static_cast<float>(src.ny - 1) / target_height;

        for (int y = 0; y < target_height; y++) {
            for (int x = 0; x < target_width; x++) {
                float px = x_ratio * x;
                float py = y_ratio * y;
                int x_floor = static_cast<int>(px);
                int y_floor = static_cast<int>(py);
                float x_lerp = px - x_floor;
                float y_lerp = py - y_floor;

                for (int c = 0; c < 3; c++) {
                    float top = lerp(
                        static_cast<float>(src.buf[3 * (y_floor * src.nx + x_floor) + c]),
                        static_cast<float>(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]),
                        x_lerp
                    );
                    float bottom = lerp(
                        static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]),
                        static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]),
                        x_lerp
                    );
                    dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(lerp(top, bottom, y_lerp));
                }
            }
        }
    }

    // Bicubic resize function
    // part of image will be cropped if the aspect ratio is different
    static bool bicubic_resize(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
        const int nx = img.nx;
        const int ny = img.ny;

        dst.nx = target_width;
        dst.ny = target_height;
        dst.buf.resize(3 * target_width * target_height);

        float Cc;
        float C[5];
        float d0, d2, d3, a0, a1, a2, a3;
        int i, j, k, jj;
        int x, y;
        float dx, dy;
        float tx, ty;

        tx = (float)nx / (float)target_width;
        ty = (float)ny / (float)target_height;

        // Bicubic interpolation; adapted from ViT.cpp, inspired from :
        //    -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36
        //    -> https://en.wikipedia.org/wiki/Bicubic_interpolation

        for (i = 0; i < target_height; i++) {
            for (j = 0; j < target_width; j++) {
                x = (int)(tx * j);
                y = (int)(ty * i);

                dx = tx * j - x;
                dy = ty * i - y;

                for (k = 0; k < 3; k++) {
                    for (jj = 0; jj <= 3; jj++) {
                        d0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x - 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
                        d2 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
                        d3 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 2, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
                        a0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];

                        a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
                        a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
                        a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;

                        C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx;

                        d0 = C[0] - C[1];
                        d2 = C[2] - C[1];
                        d3 = C[3] - C[1];
                        a0 = C[1];
                        a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
                        a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
                        a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
                        Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy;

                        const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f);
                        dst.buf[(i * target_width + j) * 3 + k] = float(Cc2);
                    }
                }
            }
        }

        return true;
    }

    // llava-1.6 type of resize_and_pad
    // if the ratio is not 1:1, padding with pad_color will be applied
    // pad_color is single channel, default is 0 (black)
    static void resize_and_pad_image(const clip_image_u8 & image, clip_image_u8 & dst, const clip_image_size & target_resolution, std::array<uint8_t, 3> pad_color = {0, 0, 0}) {
        int target_width  = target_resolution.width;
        int target_height = target_resolution.height;

        float scale_w = static_cast<float>(target_width) / image.nx;
        float scale_h = static_cast<float>(target_height) / image.ny;

        int new_width, new_height;

        if (scale_w < scale_h) {
            new_width  = target_width;
            new_height = std::min(static_cast<int>(std::ceil(image.ny * scale_w)), target_height);
        } else {
            new_height = target_height;
            new_width  = std::min(static_cast<int>(std::ceil(image.nx * scale_h)), target_width);
        }

        clip_image_u8 resized_image;
        bicubic_resize(image, resized_image, new_width, new_height);

        clip_image_u8 padded_image;
        padded_image.nx = target_width;
        padded_image.ny = target_height;
        padded_image.buf.resize(3 * target_width * target_height);

        // Fill the padded image with the fill color
        for (size_t i = 0; i < padded_image.buf.size(); i += 3) {
            padded_image.buf[i]     = pad_color[0];
            padded_image.buf[i + 1] = pad_color[1];
            padded_image.buf[i + 2] = pad_color[2];
        }

        // Calculate padding offsets
        int pad_x = (target_width  - new_width)  / 2;
        int pad_y = (target_height - new_height) / 2;

        // Copy the resized image into the center of the padded buffer
        for (int y = 0; y < new_height; ++y) {
            for (int x = 0; x < new_width; ++x) {
                for (int c = 0; c < 3; ++c) {
                    padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c];
                }
            }
        }
        dst = std::move(padded_image);
    }

    static void crop_image(const clip_image_u8 & image, clip_image_u8 & dst, int x, int y, int w, int h) {
        dst.nx = w;
        dst.ny = h;
        dst.buf.resize(3 * w * h);

        for (int i = 0; i < h; ++i) {
            for (int j = 0; j < w; ++j) {
                int src_idx = 3 * ((y + i)*image.nx + (x + j));
                int dst_idx = 3 * (i*w + j);
                dst.buf[dst_idx]     = image.buf[src_idx];
                dst.buf[dst_idx + 1] = image.buf[src_idx + 1];
                dst.buf[dst_idx + 2] = image.buf[src_idx + 2];
            }
        }
    }

    // calculate the size of the **resized** image, while preserving the aspect ratio
    // the calculated size will be aligned to the nearest multiple of align_size
    // if H or W size is larger than max_dimension, it will be resized to max_dimension
    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int max_dimension) {
        if (inp_size.width <= 0 || inp_size.height <= 0 || align_size <= 0 || max_dimension <= 0) {
            return {0, 0};
        }

        float scale = std::min(1.0f, std::min(static_cast<float>(max_dimension) / inp_size.width,
                                              static_cast<float>(max_dimension) / inp_size.height));

        float target_width_f  = static_cast<float>(inp_size.width)  * scale;
        float target_height_f = static_cast<float>(inp_size.height) * scale;

        int aligned_width  = CLIP_ALIGN((int)target_width_f,  align_size);
        int aligned_height = CLIP_ALIGN((int)target_height_f, align_size);

        return {aligned_width, aligned_height};
    }

private:
    static inline int clip(int x, int lower, int upper) {
        return std::max(lower, std::min(x, upper));
    }

    // Linear interpolation between two points
    static inline float lerp(float s, float e, float t) {
        return s + (e - s) * t;
    }
};

/**
 * implementation of LLaVA-UHD:
 *  - https://arxiv.org/pdf/2403.11703
 *  - https://github.com/thunlp/LLaVA-UHD
 *  - https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
 *
 * overview:
 *   - an image always have a single overview (downscaled image)
 *   - an image can have 0 or multiple slices, depending on the image size
 *   - each slice can then be considered as a separate image
 *
 * for example:
 *
 * [overview] --> [slice 1] --> [slice 2]
 *           |                |
 *           +--> [slice 3] --> [slice 4]
 */
struct llava_uhd {
    struct slice_coordinates {
        int x;
        int y;
        clip_image_size size;
    };

    struct slice_instructions {
        clip_image_size overview_size; // size of downscaled image
        clip_image_size refined_size;  // size of image right before slicing (must be multiple of slice size)
        clip_image_size grid_size;     // grid_size.width * grid_size.height = number of slices
        std::vector<slice_coordinates> slices;
        bool padding_refined = false;  // if true, refine image will be padded to the grid size (e.g. llava-1.6)
    };

    static slice_instructions get_slice_instructions(struct clip_ctx * ctx, const clip_image_size & original_size) {
        slice_instructions res;
        const int patch_size      = clip_get_patch_size(ctx);
        const int slice_size      = clip_get_image_size(ctx);
        const int original_width  = original_size.width;
        const int original_height = original_size.height;

        const bool has_slices    = original_size.width > slice_size || original_size.height > slice_size;
        const bool has_pinpoints = !ctx->model.hparams.image_res_candidates.empty();

        if (!has_slices) {
            // skip slicing logic
            res.overview_size = clip_image_size{slice_size, slice_size};
            res.refined_size  = clip_image_size{0, 0};
            res.grid_size     = clip_image_size{0, 0};

            return res;
        }

        if (has_pinpoints) {
            // has pinpoints, use them to calculate the grid size (e.g. llava-1.6)
            auto refine_size = llava_uhd::select_best_resolution(
                original_size,
                ctx->model.hparams.image_res_candidates);
            res.overview_size   = clip_image_size{slice_size, slice_size};
            res.refined_size    = refine_size;
            res.grid_size       = clip_image_size{0, 0};
            res.padding_refined = true;

            LOG_DBG("%s: using pinpoints for slicing\n", __func__);
            LOG_DBG("%s: original size: %d x %d, overview size: %d x %d, refined size: %d x %d\n",
                    __func__, original_width, original_height,
                    res.overview_size.width, res.overview_size.height,
                    res.refined_size.width,  res.refined_size.height);

            for (int y = 0; y < refine_size.height; y += slice_size) {
                for (int x = 0; x < refine_size.width; x += slice_size) {
                    slice_coordinates slice;
                    slice.x = x;
                    slice.y = y;
                    slice.size.width  = std::min(slice_size, refine_size.width  - x);
                    slice.size.height = std::min(slice_size, refine_size.height - y);
                    res.slices.push_back(slice);
                    LOG_DBG("%s: slice %d: x=%d, y=%d, size=%dx%d\n",
                            __func__, (int)res.slices.size() - 1,
                            slice.x, slice.y, slice.size.width, slice.size.height);
                }
            }

            res.grid_size.height = refine_size.height / slice_size;
            res.grid_size.width  = refine_size.width  / slice_size;
            LOG_DBG("%s: grid size: %d x %d\n", __func__, res.grid_size.width, res.grid_size.height);

            return res;
        }

        // no pinpoints, dynamically calculate the grid size (e.g. minicpmv)

        auto best_size    = get_best_resize(original_size, slice_size, patch_size, !has_slices);
        res.overview_size = best_size;

        {
            const int max_slice_nums = 9; // TODO: this is only used by minicpmv, maybe remove it
            const float log_ratio = log((float)original_width / original_height);
            const float ratio = (float)original_width * original_height / (slice_size * slice_size);
            const int multiple = fmin(ceil(ratio), max_slice_nums);

            auto best_grid   = get_best_grid(max_slice_nums, multiple, log_ratio);
            auto refine_size = get_refine_size(original_size, best_grid, slice_size, patch_size, true);
            res.grid_size    = best_grid;
            res.refined_size = refine_size;

            LOG_DBG("%s: original size: %d x %d, overview size: %d x %d, refined size: %d x %d, grid size: %d x %d\n",
                    __func__, original_width, original_height,
                    res.overview_size.width, res.overview_size.height,
                    res.refined_size.width, res.refined_size.height,
                    res.grid_size.width, res.grid_size.height);

            int width  = refine_size.width;
            int height = refine_size.height;
            int grid_x = int(width  / best_grid.width);
            int grid_y = int(height / best_grid.height);
            for (int patches_y = 0,                    ic = 0;
                    patches_y < refine_size.height && ic < best_grid.height;
                    patches_y += grid_y,              ic += 1) {
                for (int patches_x = 0,                   jc = 0;
                        patches_x < refine_size.width && jc < best_grid.width;
                        patches_x += grid_x,             jc += 1) {
                    slice_coordinates slice;
                    slice.x = patches_x;
                    slice.y = patches_y;
                    slice.size.width  = grid_x;
                    slice.size.height = grid_y;
                    res.slices.push_back(slice);
                    LOG_DBG("%s: slice %d: x=%d, y=%d, size=%dx%d\n",
                            __func__, (int)res.slices.size() - 1,
                            slice.x, slice.y, slice.size.width, slice.size.height);
                }
            }
        }

        return res;
    }

    static std::vector<clip_image_u8_ptr> slice_image(const clip_image_u8 * img, const slice_instructions & inst) {
        std::vector<clip_image_u8_ptr> output;

        // resize to overview size
        clip_image_u8_ptr resized_img(clip_image_u8_init());
        image_manipulation::bicubic_resize(*img, *resized_img, inst.overview_size.width, inst.overview_size.height);
        output.push_back(std::move(resized_img));
        if (inst.slices.empty()) {
            // no slices, just return the resized image
            return output;
        }

        // resize to refined size
        clip_image_u8_ptr refined_img(clip_image_u8_init());
        if (inst.padding_refined) {
            image_manipulation::resize_and_pad_image(*img, *refined_img, inst.refined_size);
        } else {
            image_manipulation::bilinear_resize(*img, *refined_img, inst.refined_size.width, inst.refined_size.height);
        }

        // create slices
        for (const auto & slice : inst.slices) {
            int x = slice.x;
            int y = slice.y;
            int w = slice.size.width;
            int h = slice.size.height;

            clip_image_u8_ptr img_slice(clip_image_u8_init());
            image_manipulation::crop_image(*refined_img, *img_slice, x, y, w, h);
            output.push_back(std::move(img_slice));
        }

        return output;
    }

private:
    static clip_image_size get_best_resize(const clip_image_size & original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
        int width  = original_size.width;
        int height = original_size.height;
        if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
            float r = static_cast<float>(width) / height;
            height  = static_cast<int>(scale_resolution / std::sqrt(r));
            width   = static_cast<int>(height * r);
        }
        clip_image_size res;
        res.width  = ensure_divide(width,  patch_size);
        res.height = ensure_divide(height, patch_size);
        return res;
    }

    static clip_image_size resize_maintain_aspect_ratio(const clip_image_size & orig, const clip_image_size & target_max) {
        float scale_width  = static_cast<float>(target_max.width)  / orig.width;
        float scale_height = static_cast<float>(target_max.height) / orig.height;
        float scale = std::min(scale_width, scale_height);
        return clip_image_size{
            static_cast<int>(orig.width  * scale),
            static_cast<int>(orig.height * scale),
        };
    }

    /**
     * Selects the best resolution from a list of possible resolutions based on the original size.
     *
     * For example, when given a list of resolutions:
     *  - 100x100
     *  - 200x100
     *  - 100x200
     *  - 200x200
     *
     * And an input image of size 111x200, then 100x200 is the best fit (least wasted resolution).
     *
     * @param original_size The original size of the image
     * @param possible_resolutions A list of possible resolutions
     * @return The best fit resolution
     */
    static clip_image_size select_best_resolution(const clip_image_size & original_size, const std::vector<clip_image_size> & possible_resolutions) {
        clip_image_size best_fit;
        int min_wasted_area = std::numeric_limits<int>::max();
        int max_effective_resolution = 0;

        for (const clip_image_size & candidate : possible_resolutions) {
            auto target_size = resize_maintain_aspect_ratio(original_size, candidate);
            int effective_resolution = std::min(
                target_size.width * target_size.height,
                original_size.width * original_size.height);
            int wasted_area = (candidate.width * candidate.height) - effective_resolution;

            if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_area < min_wasted_area)) {
                max_effective_resolution = effective_resolution;
                min_wasted_area = wasted_area;
                best_fit = candidate;
            }

            LOG_DBG("%s: candidate: %d x %d, target: %d x %d, wasted: %d, effective: %d\n", __func__, candidate.width, candidate.height, target_size.width, target_size.height, wasted_area, effective_resolution);
        }

        return best_fit;
    }

    static int ensure_divide(int length, int patch_size) {
        return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
    }

    static clip_image_size get_refine_size(const clip_image_size & original_size, const clip_image_size & grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
        int width  = original_size.width;
        int height = original_size.height;
        int grid_x = grid.width;
        int grid_y = grid.height;

        int refine_width  = ensure_divide(width, grid_x);
        int refine_height = ensure_divide(height, grid_y);

        clip_image_size grid_size;
        grid_size.width  = refine_width  / grid_x;
        grid_size.height = refine_height / grid_y;

        auto best_grid_size  = get_best_resize(grid_size, scale_resolution, patch_size, allow_upscale);
        int best_grid_width  = best_grid_size.width;
        int best_grid_height = best_grid_size.height;

        clip_image_size refine_size;
        refine_size.width  = best_grid_width  * grid_x;
        refine_size.height = best_grid_height * grid_y;
        return refine_size;
    }

    static clip_image_size get_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
        std::vector<int> candidate_split_grids_nums;
        for (int i : {multiple - 1, multiple, multiple + 1}) {
            if (i == 1 || i > max_slice_nums) {
                continue;
            }
            candidate_split_grids_nums.push_back(i);
        }

        std::vector<clip_image_size> candidate_grids;
        for (int split_grids_nums : candidate_split_grids_nums) {
            int m = 1;
            while (m <= split_grids_nums) {
                if (split_grids_nums % m == 0) {
                    candidate_grids.push_back(clip_image_size{m, split_grids_nums / m});
                }
                ++m;
            }
        }

        clip_image_size best_grid{1, 1};
        float min_error = std::numeric_limits<float>::infinity();
        for (const auto& grid : candidate_grids) {
            float error = std::abs(log_ratio - std::log(1.0 * grid.width / grid.height));
            if (error < min_error) {
                best_grid = grid;
                min_error = error;
            }
        }
        return best_grid;
    }
};

// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
// res_imgs memory is being allocated here, previous allocations will be freed if found
bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) {
    clip_image_size original_size{img->nx, img->ny};
    bool pad_to_square = true;
    auto & params = ctx->model.hparams;
    // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
    if (params.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD) {
        pad_to_square = false;
    }

    if (clip_is_minicpmv(ctx)) {
        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);

        for (size_t i = 0; i < imgs.size(); ++i) {
            // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
            clip_image_f32_ptr res(clip_image_f32_init());
            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
            res_imgs->entries.push_back(std::move(res));
        }

        res_imgs->grid_x = inst.grid_size.width;
        res_imgs->grid_y = inst.grid_size.height;
        return true;

    } else if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
        clip_image_u8 resized;
        auto patch_size = params.patch_size * 2;
        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, patch_size, params.image_size);
        image_manipulation::bicubic_resize(*img, resized, new_size.width, new_size.height);

        clip_image_f32_ptr img_f32(clip_image_f32_init());
        // clip_image_f32_ptr res(clip_image_f32_init());
        normalize_image_u8_to_f32(resized, *img_f32, params.image_mean, params.image_std);
        // res_imgs->data[0] = *res;
        res_imgs->entries.push_back(std::move(img_f32));
        return true;
    }
    else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE
            || ctx->proj_type() == PROJECTOR_TYPE_GEMMA3
            || ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3
            || ctx->proj_type() == PROJECTOR_TYPE_INTERNVL // TODO @ngxson : support dynamic resolution
    ) {
        clip_image_u8 resized_image;
        int sz = params.image_size;
        image_manipulation::resize_and_pad_image(*img, resized_image, {sz, sz});
        clip_image_f32_ptr img_f32(clip_image_f32_init());
        //clip_image_save_to_bmp(resized_image, "resized.bmp");
        normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
        res_imgs->entries.push_back(std::move(img_f32));
        return true;

    } else if (ctx->proj_type() == PROJECTOR_TYPE_PIXTRAL) {
        clip_image_u8 resized_image;
        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, params.patch_size, params.image_size);
        image_manipulation::bilinear_resize(*img, resized_image, new_size.width, new_size.height);
        clip_image_f32_ptr img_f32(clip_image_f32_init());
        normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
        res_imgs->entries.push_back(std::move(img_f32));
        return true;

    } else if (ctx->proj_type() == PROJECTOR_TYPE_LLAMA4) {
        GGML_ASSERT(!params.image_res_candidates.empty());
        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);

        for (size_t i = 0; i < imgs.size(); ++i) {
            clip_image_f32_ptr res(clip_image_f32_init());
            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
            res_imgs->entries.push_back(std::move(res));
        }

        res_imgs->grid_x = inst.grid_size.width;
        res_imgs->grid_y = inst.grid_size.height;
        return true;

    }

    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156

    clip_image_u8_ptr temp(clip_image_u8_init()); // we will keep the input image data here temporarily

    if (pad_to_square) {
        // for llava-1.5, we resize image to a square, and pad the shorter side with a background color
        // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
        const int longer_side = std::max(img->nx, img->ny);
        temp->nx = longer_side;
        temp->ny = longer_side;
        temp->buf.resize(3 * longer_side * longer_side);

        // background color in RGB from LLaVA (this is the mean rgb color * 255)
        const std::array<uint8_t, 3> pad_color = {122, 116, 104};

        // resize the image to the target_size
        image_manipulation::resize_and_pad_image(*img, *temp, clip_image_size{params.image_size, params.image_size}, pad_color);

        clip_image_f32_ptr res(clip_image_f32_init());
        normalize_image_u8_to_f32(*temp, *res, params.image_mean, params.image_std);
        res_imgs->entries.push_back(std::move(res));
        return true;

    } else if (!params.image_res_candidates.empty()) {
        // "spatial_unpad" with "anyres" processing for llava-1.6
        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);

        for (size_t i = 0; i < imgs.size(); ++i) {
            // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
            clip_image_f32_ptr res(clip_image_f32_init());
            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
            res_imgs->entries.push_back(std::move(res));
        }

        return true;

    }

    GGML_ASSERT(false && "Unknown image preprocessing type");
}

ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
    return ctx->model.image_newline;
}

void clip_free(clip_ctx * ctx) {
    if (ctx == nullptr) {
        return;
    }
    delete ctx;
}

// deprecated
size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
    const int32_t nx = ctx->model.hparams.image_size;
    const int32_t ny = ctx->model.hparams.image_size;
    return clip_embd_nbytes_by_img(ctx, nx, ny);
}

size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h) {
    clip_image_f32 img;
    img.nx = img_w;
    img.ny = img_h;
    return clip_n_output_tokens(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
}

int32_t clip_get_image_size(const struct clip_ctx * ctx) {
    return ctx->model.hparams.image_size;
}

int32_t clip_get_patch_size(const struct clip_ctx * ctx) {
    return ctx->model.hparams.patch_size;
}

int32_t clip_get_hidden_size(const struct clip_ctx * ctx) {
    return ctx->model.hparams.n_embd;
}

const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
    return ctx->model.hparams.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD ? "spatial_unpad" : "flat";
}

int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
    const auto & params = ctx->model.hparams;
    const int n_total = clip_n_output_tokens(ctx, img);
    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
        return img->nx / (params.patch_size * 2) + (int)(img->nx % params.patch_size > 0);
    }
    return n_total;
}

int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
    const auto & params = ctx->model.hparams;
    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
        return img->ny / (params.patch_size * 2) + (int)(img->ny % params.patch_size > 0);
    }
    return 1;
}

int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
    const auto & params = ctx->model.hparams;

    // only for models using fixed size square images
    int n_patches_sq = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);

    projector_type proj = ctx->proj_type();

    switch (proj) {
        case PROJECTOR_TYPE_MLP:
        case PROJECTOR_TYPE_MLP_NORM:
            {
                // do nothing
            } break;
        case PROJECTOR_TYPE_LDP:
        case PROJECTOR_TYPE_LDPV2:
        case PROJECTOR_TYPE_GLM_EDGE:
            {
                n_patches_sq /= 4;
                if (ctx->model.mm_glm_tok_boi) {
                    n_patches_sq += 2; // for BOI and EOI token embeddings
                }
            } break;
        case PROJECTOR_TYPE_MINICPMV:
            {
                if (params.minicpmv_version == 2) {
                    // MiniCPM-V 2.5
                    n_patches_sq = 96;
                } else if (params.minicpmv_version == 3) {
                    // MiniCPM-V 2.6
                    n_patches_sq = 64;
                } else if (params.minicpmv_version == 4) {
                    // MiniCPM-o 2.6
                    n_patches_sq = 64;
                } else if (params.minicpmv_version == 5) {
                    // MiniCPM-V 4.0
                    n_patches_sq = 64;
                } else {
                    GGML_ABORT("Unknown minicpmv version");
                }
            } break;
        case PROJECTOR_TYPE_QWEN2VL:
        case PROJECTOR_TYPE_QWEN25VL:
            {
                // dynamic size
                int patch_size = params.patch_size * 2;
                int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
                int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
                n_patches_sq = x_patch * y_patch;
            } break;
        case PROJECTOR_TYPE_GEMMA3:
            {
                int n_per_side = params.image_size / params.patch_size;
                int n_per_side_2d_pool = n_per_side / params.proj_scale_factor;
                n_patches_sq = n_per_side_2d_pool * n_per_side_2d_pool;
            } break;
        case PROJECTOR_TYPE_IDEFICS3:
        case PROJECTOR_TYPE_INTERNVL:
            {
                // both W and H are divided by proj_scale_factor
                n_patches_sq /= (params.proj_scale_factor * params.proj_scale_factor);
            } break;
        case PROJECTOR_TYPE_PIXTRAL:
            {
                // dynamic size
                int n_merge = params.spatial_merge_size;
                int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
                int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
                n_patches_sq = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
            } break;
        case PROJECTOR_TYPE_LLAMA4:
            {
                int scale_factor = ctx->model.hparams.proj_scale_factor;
                n_patches_sq /= (scale_factor * scale_factor);
            } break;
        case PROJECTOR_TYPE_VOXTRAL:
        case PROJECTOR_TYPE_ULTRAVOX:
        case PROJECTOR_TYPE_QWEN2A:
            {
                n_patches_sq = img->nx;

                const int proj_stack_factor = ctx->model.hparams.proj_stack_factor;
                if (ctx->model.audio_has_stack_frames()) {
                    GGML_ASSERT(proj_stack_factor > 0);
                    const int n_len = CLIP_ALIGN(n_patches_sq, proj_stack_factor);
                    n_patches_sq = n_len / proj_stack_factor;
                }

                // whisper downscales input token by half after conv1d
                n_patches_sq /= 2;

                if (ctx->model.audio_has_avgpool()) {
                    // divide by 2 because of nn.AvgPool1d(2, stride=2)
                    n_patches_sq /= 2;
                }
            } break;
        default:
            GGML_ABORT("unsupported projector type");
    }

    return n_patches_sq;
}

static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
    assert(embed_dim % 2 == 0);
    int H = pos.size();
    int W = pos[0].size();

    std::vector<float> omega(embed_dim / 2);
    for (int i = 0; i < embed_dim / 2; ++i) {
        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
    }

    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
    for (int h = 0; h < H; ++h) {
        for (int w = 0; w < W; ++w) {
            for (int d = 0; d < embed_dim / 2; ++d) {
                float out_value = pos[h][w] * omega[d];
                emb[h][w][d] = sin(out_value);
                emb[h][w][d + embed_dim / 2] = cos(out_value);
            }
        }
    }

    return emb;
}

static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
    assert(embed_dim % 2 == 0);
    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)

    int H = emb_h.size();
    int W = emb_h[0].size();
    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));

    for (int h = 0; h < H; ++h) {
        for (int w = 0; w < W; ++w) {
            for (int d = 0; d < embed_dim / 2; ++d) {
                emb[h][w][d] = emb_h[h][w][d];
                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
            }
        }
    }
    return emb;
}

static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
    int grid_h_size = image_size.first;
    int grid_w_size = image_size.second;

    std::vector<float> grid_h(grid_h_size);
    std::vector<float> grid_w(grid_w_size);

    for (int i = 0; i < grid_h_size; ++i) {
        grid_h[i] = static_cast<float>(i);
    }
    for (int i = 0; i < grid_w_size; ++i) {
        grid_w[i] = static_cast<float>(i);
    }

    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
    for (int h = 0; h < grid_h_size; ++h) {
        for (int w = 0; w < grid_w_size; ++w) {
            grid[h][w] = grid_w[w];
        }
    }
    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
    for (int h = 0; h < grid_h_size; ++h) {
        for (int w = 0; w < grid_w_size; ++w) {
            grid_2d[0][h][w] = grid_h[h];
            grid_2d[1][h][w] = grid_w[w];
        }
    }

    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);

    int H = image_size.first;
    int W = image_size.second;
    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
    for (int h = 0; h < H; ++h) {
        for (int w = 0; w < W; ++w) {
            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
        }
    }

    return pos_embed_2d;
}

bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
    clip_image_f32_batch imgs;
    clip_image_f32_ptr img_copy(clip_image_f32_init());
    *img_copy = *img;
    imgs.entries.push_back(std::move(img_copy));

    return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
}

bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, float * vec) {
    const clip_image_f32_batch & imgs = *imgs_c_ptr;
    int batch_size = imgs.entries.size();

    // TODO @ngxson : implement batch size > 1 as a loop
    //                we don't need true batching support because the cgraph will gonna be big anyway
    if (batch_size != 1) {
        return false; // only support batch size of 1
    }

    // build the inference graph
    ctx->debug_print_tensors.clear();
    ggml_backend_sched_reset(ctx->sched.get());
    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
    ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);

    // set inputs
    const auto & model   = ctx->model;
    const auto & hparams = model.hparams;

    const int image_size_width  = imgs.entries[0]->nx;
    const int image_size_height = imgs.entries[0]->ny;

    const int patch_size    = hparams.patch_size;
    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
    const int n_pos = num_patches + (model.class_embedding ? 1 : 0);
    const int pos_w = image_size_width  / patch_size;
    const int pos_h = image_size_height / patch_size;

    const bool use_window_attn = hparams.n_wa_pattern > 0; // for qwen2.5vl

    auto get_inp_tensor = [&gf](const char * name) {
        ggml_tensor * inp = ggml_graph_get_tensor(gf, name);
        if (inp == nullptr) {
            GGML_ABORT("Failed to get tensor %s", name);
        }
        if (!(inp->flags & GGML_TENSOR_FLAG_INPUT)) {
            GGML_ABORT("Tensor %s is not an input tensor", name);
        }
        return inp;
    };

    auto set_input_f32 = [&get_inp_tensor](const char * name, std::vector<float> & values) {
        ggml_tensor * cur = get_inp_tensor(name);
        GGML_ASSERT(cur->type == GGML_TYPE_F32);
        GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size());
        ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur));
    };

    auto set_input_i32 = [&get_inp_tensor](const char * name, std::vector<int32_t> & values) {
        ggml_tensor * cur = get_inp_tensor(name);
        GGML_ASSERT(cur->type == GGML_TYPE_I32);
        GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size());
        ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur));
    };

    // set input pixel values
    if (!imgs.is_audio) {
        size_t nelem = 0;
        for (const auto & img : imgs.entries) {
            nelem += img->nx * img->ny * 3;
        }
        std::vector<float> inp_raw(nelem);

        // layout of data (note: the channel dim is unrolled to better visualize the layout):
        //
        // ┌──W──┐
        // │     H │  channel = R
        // ├─────┤ │
        // │     H │  channel = G
        // ├─────┤ │
        // │     H │  channel = B
        // └─────┘ │
        //   ──────┘ x B

        for (size_t i = 0; i < imgs.entries.size(); i++) {
            const int nx = imgs.entries[i]->nx;
            const int ny = imgs.entries[i]->ny;
            const int n = nx * ny;

            for (int b = 0; b < batch_size; b++) {
                float * batch_entry = inp_raw.data() + b * (3*n);
                for (int y = 0; y < ny; y++) {
                    for (int x = 0; x < nx; x++) {
                        size_t base_src = 3*(y * nx + x); // idx of the first channel
                        size_t base_dst =    y * nx + x;  // idx of the first channel
                        batch_entry[      base_dst] = imgs.entries[b]->buf[base_src    ];
                        batch_entry[1*n + base_dst] = imgs.entries[b]->buf[base_src + 1];
                        batch_entry[2*n + base_dst] = imgs.entries[b]->buf[base_src + 2];
                    }
                }
            }
        }
        set_input_f32("inp_raw", inp_raw);

    } else {
        // audio input
        GGML_ASSERT(imgs.entries.size() == 1);
        const auto & mel_inp = imgs.entries[0];
        const int n_step = mel_inp->nx;
        const int n_mel  = mel_inp->ny;
        std::vector<float> inp_raw(n_step * n_mel);
        std::memcpy(inp_raw.data(), mel_inp->buf.data(), n_step * n_mel * sizeof(float));
        set_input_f32("inp_raw", inp_raw);
    }

    // set input per projector
    switch (ctx->model.proj_type) {
        case PROJECTOR_TYPE_MINICPMV:
            {
                // inspired from siglip:
                //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
                //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
                std::vector<int32_t> positions(pos_h * pos_w);
                int bucket_coords_h[1024];
                int bucket_coords_w[1024];
                for (int i = 0; i < pos_h; i++){
                    bucket_coords_h[i] = std::floor(70.0*i/pos_h);
                }
                for (int i = 0; i < pos_w; i++){
                    bucket_coords_w[i] = std::floor(70.0*i/pos_w);
                }
                for (int i = 0, id = 0; i < pos_h; i++){
                    for (int j = 0; j < pos_w; j++){
                        positions[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
                    }
                }
                set_input_i32("positions", positions);

                // inspired from resampler of Qwen-VL:
                //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
                //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
                int embed_dim = clip_n_mmproj_embd(ctx);

                // TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos?
                auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));

                std::vector<float> pos_embed(embed_dim * pos_w * pos_h);
                for(int i = 0; i < pos_w * pos_h; ++i){
                    for(int j = 0; j < embed_dim; ++j){
                        pos_embed[i * embed_dim + j] = pos_embed_t[i][j];
                    }
                }

                set_input_f32("pos_embed", pos_embed);
            } break;
        case PROJECTOR_TYPE_QWEN2VL:
            {
                const int merge_ratio = 2;
                const int pw = image_size_width  / patch_size;
                const int ph = image_size_height / patch_size;
                std::vector<int> positions(n_pos * 4);
                int ptr = 0;
                for (int y = 0; y < ph; y += merge_ratio) {
                    for (int x = 0; x < pw; x += merge_ratio) {
                        for (int dy = 0; dy < 2; dy++) {
                            for (int dx = 0; dx < 2; dx++) {
                                positions[                  ptr] = y + dy;
                                positions[    num_patches + ptr] = x + dx;
                                positions[2 * num_patches + ptr] = y + dy;
                                positions[3 * num_patches + ptr] = x + dx;
                                ptr++;
                            }
                        }
                    }
                }

                set_input_i32("positions", positions);
            } break;
        case PROJECTOR_TYPE_QWEN25VL:
            {
                // pw * ph = number of tokens output by ViT after apply patch merger
                // ipw * ipw = number of vision token been processed inside ViT
                const int merge_ratio = 2;
                const int pw  = image_size_width  / patch_size / merge_ratio;
                const int ph  = image_size_height / patch_size / merge_ratio;
                const int ipw = image_size_width  / patch_size;
                const int iph = image_size_height / patch_size;

                std::vector<int> idx    (ph * pw);
                std::vector<int> inv_idx(ph * pw);

                if (use_window_attn) {
                    const int attn_window_size = 112;
                    const int grid_window = attn_window_size / patch_size / merge_ratio;
                    int dst = 0;
                    // [num_vision_tokens, num_vision_tokens] attention mask tensor
                    std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest());
                    int mask_row = 0;

                    for (int y = 0; y < ph; y += grid_window) {
                        for (int x = 0; x < pw; x += grid_window) {
                            const int win_h = std::min(grid_window, ph - y);
                            const int win_w = std::min(grid_window, pw - x);
                            const int dst_0 = dst;
                            // group all tokens belong to the same window togather (to a continue range)
                            for (int dy = 0; dy < win_h; dy++) {
                                for (int dx = 0; dx < win_w; dx++) {
                                    const int src = (y + dy) * pw + (x + dx);
                                    GGML_ASSERT(src < (int)idx.size());
                                    GGML_ASSERT(dst < (int)inv_idx.size());
                                    idx    [src] = dst;
                                    inv_idx[dst] = src;
                                    dst++;
                                }
                            }

                            for (int r=0; r < win_h * win_w * merge_ratio * merge_ratio; r++) {
                                int row_offset = mask_row * (ipw * iph);
                                std::fill(
                                    mask.begin() + row_offset + (dst_0 * merge_ratio * merge_ratio),
                                    mask.begin() + row_offset + (dst   * merge_ratio * merge_ratio),
                                    0.0);
                                mask_row++;
                            }
                        }
                    }

                    set_input_i32("window_idx",     idx);
                    set_input_i32("inv_window_idx", inv_idx);
                    set_input_f32("window_mask",    mask);
                } else {
                    for (int i = 0; i < ph * pw; i++) {
                        idx[i] = i;
                    }
                }

                const int mpow = merge_ratio * merge_ratio;
                std::vector<int> positions(n_pos * 4);

                int ptr = 0;
                for (int y = 0; y < iph; y += merge_ratio) {
                    for (int x = 0; x < ipw; x += merge_ratio) {
                        for (int dy = 0; dy < 2; dy++) {
                            for (int dx = 0; dx < 2; dx++) {
                                auto remap = idx[ptr / mpow];
                                remap = (remap * mpow) + (ptr % mpow);

                                positions[                  remap] = y + dy;
                                positions[    num_patches + remap] = x + dx;
                                positions[2 * num_patches + remap] = y + dy;
                                positions[3 * num_patches + remap] = x + dx;
                                ptr++;
                            }
                        }
                    }
                }

                set_input_i32("positions", positions);
            } break;
        case PROJECTOR_TYPE_PIXTRAL:
            {
                // set the 2D positions
                int n_patches_per_col = image_size_width / patch_size;
                std::vector<int> pos_data(n_pos);
                // dimension H
                for (int i = 0; i < n_pos; i++) {
                    pos_data[i] = i / n_patches_per_col;
                }
                set_input_i32("pos_h", pos_data);
                // dimension W
                for (int i = 0; i < n_pos; i++) {
                    pos_data[i] = i % n_patches_per_col;
                }
                set_input_i32("pos_w", pos_data);
            } break;
        case PROJECTOR_TYPE_GLM_EDGE:
        {
            // llava and other models
            std::vector<int32_t> positions(n_pos);
            for (int i = 0; i < n_pos; i++) {
                positions[i] = i;
            }
            set_input_i32("positions", positions);
        } break;
        case PROJECTOR_TYPE_MLP:
        case PROJECTOR_TYPE_MLP_NORM:
        case PROJECTOR_TYPE_LDP:
        case PROJECTOR_TYPE_LDPV2:
            {
                // llava and other models
                std::vector<int32_t> positions(n_pos);
                for (int i = 0; i < n_pos; i++) {
                    positions[i] = i;
                }
                set_input_i32("positions", positions);

                // The patches vector is used to get rows to index into the embeds with;
                // we should skip dim 0 only if we have CLS to avoid going out of bounds
                // when retrieving the rows.
                int patch_offset = model.class_embedding ? 1 : 0;
                std::vector<int32_t> patches(num_patches);
                for (int i = 0; i < num_patches; i++) {
                    patches[i] = i + patch_offset;
                }
                set_input_i32("patches", patches);
            } break;
        case PROJECTOR_TYPE_GEMMA3:
        case PROJECTOR_TYPE_IDEFICS3:
        case PROJECTOR_TYPE_INTERNVL:
        case PROJECTOR_TYPE_QWEN2A:
        case PROJECTOR_TYPE_ULTRAVOX:
        case PROJECTOR_TYPE_VOXTRAL:
            {
                // do nothing
            } break;
        case PROJECTOR_TYPE_LLAMA4:
            {
                // set the 2D positions
                int n_patches_per_col = image_size_width / patch_size;
                std::vector<int> pos_data(num_patches + 1, 0); // +1 for the [CLS] token
                // last pos is always kept 0, it's for CLS
                // dimension H
                for (int i = 0; i < num_patches; i++) {
                    pos_data[i] = (i / n_patches_per_col) + 1;
                }
                set_input_i32("pos_h", pos_data);
                // dimension W
                for (int i = 0; i < num_patches; i++) {
                    pos_data[i] = (i % n_patches_per_col) + 1;
                }
                set_input_i32("pos_w", pos_data);
            } break;
        default:
            GGML_ABORT("Unknown projector type");
    }

    // ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
    ggml_backend_dev_t dev = ggml_backend_get_device(ctx->backend_cpu);
    ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
    if (reg) {
        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
        if (ggml_backend_set_n_threads_fn) {
            ggml_backend_set_n_threads_fn(ctx->backend_cpu, n_threads);
        }
    }

    auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
    if (status != GGML_STATUS_SUCCESS) {
        LOG_ERR("%s: ggml_backend_sched_graph_compute failed with error %d\n", __func__, status);
        return false;
    }

    // print debug nodes
    if (ctx->debug_graph) {
        LOG_INF("\n\n---\n\n");
        LOG_INF("\n\nDebug graph:\n\n");
        for (ggml_tensor * t : ctx->debug_print_tensors) {
            std::vector<uint8_t> data(ggml_nbytes(t));
            ggml_backend_tensor_get(t, data.data(), 0, ggml_nbytes(t));
            print_tensor_shape(t);
            print_tensor_data(t, data.data(), 3);
        }
    }

    // the last node is the embedding tensor
    ggml_tensor * embeddings = ggml_graph_node(gf, -1);

    // sanity check (only support batch size of 1 for now)
    const int n_tokens_out = embeddings->ne[1];
    const int expected_n_tokens_out = clip_n_output_tokens(ctx, imgs.entries[0].get());
    if (n_tokens_out != expected_n_tokens_out) {
        LOG_ERR("%s: expected output %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out);
        GGML_ABORT("Invalid number of output tokens");
    }

    // copy the embeddings to the location passed by the user
    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));

    return true;
}

int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
    const auto & hparams = ctx->model.hparams;
    switch (ctx->model.proj_type) {
        case PROJECTOR_TYPE_LDP:
            return ctx->model.mm_model_block_1_block_2_1_b->ne[0];
        case PROJECTOR_TYPE_LDPV2:
            return ctx->model.mm_model_peg_0_b->ne[0];
        case PROJECTOR_TYPE_MLP:
        case PROJECTOR_TYPE_PIXTRAL:
            return ctx->model.mm_2_w->ne[1];
        case PROJECTOR_TYPE_MLP_NORM:
            return ctx->model.mm_3_b->ne[0];
        case PROJECTOR_TYPE_MINICPMV:
            if (hparams.minicpmv_version == 2) {
                // MiniCPM-V 2.5
                return 4096;
            } else if (hparams.minicpmv_version == 3) {
                // MiniCPM-V 2.6
                return 3584;
            } else if (hparams.minicpmv_version == 4) {
                // MiniCPM-o 2.6
                return 3584;
            } else if (hparams.minicpmv_version == 5) {
                // MiniCPM-V 4.0
                return 2560;
            }
            GGML_ABORT("Unknown minicpmv version");
        case PROJECTOR_TYPE_GLM_EDGE:
            return ctx->model.mm_model_mlp_3_w->ne[1];
        case PROJECTOR_TYPE_QWEN2VL:
        case PROJECTOR_TYPE_QWEN25VL:
            return ctx->model.mm_1_b->ne[0];
        case PROJECTOR_TYPE_GEMMA3:
            return ctx->model.mm_input_proj_w->ne[0];
        case PROJECTOR_TYPE_IDEFICS3:
            return ctx->model.projection->ne[1];
        case PROJECTOR_TYPE_ULTRAVOX:
        case PROJECTOR_TYPE_VOXTRAL:
            return ctx->model.mm_2_w->ne[1];
        case PROJECTOR_TYPE_INTERNVL:
            return ctx->model.mm_3_w->ne[1];
        case PROJECTOR_TYPE_LLAMA4:
            return ctx->model.mm_model_proj->ne[1];
        case PROJECTOR_TYPE_QWEN2A:
            return ctx->model.mm_fc_w->ne[1];
        default:
            GGML_ABORT("Unknown projector type");
    }
}

int clip_is_minicpmv(const struct clip_ctx * ctx) {
    if (ctx->proj_type() == PROJECTOR_TYPE_MINICPMV) {
        return ctx->model.hparams.minicpmv_version;
    }
    return 0;
}

bool clip_is_glm(const struct clip_ctx * ctx) {
    return ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE;
}

bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
    return ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL
        || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL;
}

bool clip_is_llava(const struct clip_ctx * ctx) {
    return ctx->model.hparams.has_llava_projector;
}

bool clip_is_gemma3(const struct clip_ctx * ctx) {
    return ctx->proj_type() == PROJECTOR_TYPE_GEMMA3;
}

bool clip_has_vision_encoder(const struct clip_ctx * ctx) {
    return ctx->model.modality == CLIP_MODALITY_VISION;
}

bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
    return ctx->model.modality == CLIP_MODALITY_AUDIO;
}

bool clip_has_whisper_encoder(const struct clip_ctx * ctx) {
    return ctx->proj_type() == PROJECTOR_TYPE_ULTRAVOX
        || ctx->proj_type() == PROJECTOR_TYPE_QWEN2A
        || ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL;
}

bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
    clip_image_f32 clip_img;
    clip_img.buf.resize(h * w * 3);
    for (int i = 0; i < h*w*3; i++)
    {
        clip_img.buf[i] = img[i];
    }
    clip_img.nx = w;
    clip_img.ny = h;
    clip_image_encode(ctx, n_threads, &clip_img, vec);
    return true;
}

//
// API used internally with mtmd
//

projector_type clip_get_projector_type(const struct clip_ctx * ctx) {
    return ctx->proj_type();
}

void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel) {
    clip_image_f32 * audio = new clip_image_f32;
    audio->nx = n_frames;
    audio->ny = n_mel;
    audio->buf.resize(n_frames * n_mel);
    std::memcpy(audio->buf.data(), mel, n_frames * n_mel * sizeof(float));

    batch->entries.push_back(clip_image_f32_ptr(audio));
    batch->is_audio = true;
}