server : add VSCode's Github Copilot Chat support (#12896 )

* server : add VSCode's Github Copilot Chat support * cont : update handler name
rpc : Set cache directory in rpc-server.cpp on FreeBSD (#12903 )
2026-07-14 16:35:55 +02:00 · 2025-04-11 23:37:41 +03:00 · 2025-04-11 22:04:14 +02:00 · 2025-04-11 21:47:52 +02:00 · 2025-04-11 21:45:44 +02:00 · 2025-04-11 15:32:14 +02:00
27 changed files with 1419 additions and 591 deletions
@@ -36,13 +36,13 @@ jobs:
      matrix:
        config:
          # Multi-stage build
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: true}
-          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
-          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true }
+          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
    steps:
      - name: Check out the repo
        uses: actions/checkout@v4
@@ -97,6 +97,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
 - [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
+- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e)
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
@@ -259,7 +260,9 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)

-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from Hugging Face by using this CLI argument: `-hf <user>/<model>[:quant]`
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
+
+By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.

 After downloading a model, use the CLI tools to run it locally - see below.

@@ -41,6 +41,11 @@ COMMON_CMAKE_ARGS=(
    -DGGML_OPENMP=${GGML_OPENMP}
 )

+XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
+MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
+MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
+echo "Detected Xcode version: $XCODE_VERSION"
+
 check_required_tool() {
    local tool=$1
    local install_message=$2
@@ -325,21 +330,28 @@ combine_static_libraries() {

    # Platform-specific post-processing for device builds
    if [[ "$is_simulator" == "false" ]]; then
-        if command -v vtool &>/dev/null; then
+        if command -v xcrun vtool &>/dev/null; then
            case "$platform" in
                "ios")
                    echo "Marking binary as a framework binary for iOS..."
-                    vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
+                    xcrun vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                    ;;
                "visionos")
                    echo "Marking binary as a framework binary for visionOS..."
-                    vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
+                    if [[ "$MAJOR_VERSION" -gt 16 ]] || [[ "$MAJOR_VERSION" -eq 16 && "$MINOR_VERSION" -gt 2 ]]; then
+                        echo "Xcode version greater than 16.2, using visionOS."
+                        VISION_OS_BUILD_VERSION="visionos"
+                    else
+                        echo "Xcode version less than or equal to 16.2, using xros."
+                        VISION_OS_BUILD_VERSION="xros"
+                    fi
+                    xcrun vtool -set-build-version ${VISION_OS_BUILD_VERSION} ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                    ;;
                "tvos")
                    echo "Marking binary as a framework binary for tvOS..."
-                    vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
+                    xcrun vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                    ;;
            esac
@@ -228,12 +228,13 @@ static bool common_download_file_single(const std::string & url, const std::stri
    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);

+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
    // Check if hf-token or bearer-token was specified
    if (!bearer_token.empty()) {
        std::string auth_header = "Authorization: Bearer " + bearer_token;
        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
    }
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);

 #if defined(_WIN32)
    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
@@ -544,7 +545,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
    curl_slist_ptr http_headers;
    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
+
+    std::string model_endpoint = get_model_endpoint();
+
+    std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag;
    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
@@ -659,13 +663,8 @@ static void common_params_handle_model(
                }
            }

-            std::string hf_endpoint = "https://huggingface.co/";
-            const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-            if (hf_endpoint_env) {
-                hf_endpoint = hf_endpoint_env;
-                if (hf_endpoint.back() != '/') hf_endpoint += '/';
-            }
-            model.url = hf_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file;
+            std::string model_endpoint = get_model_endpoint();
+            model.url = model_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file;
            // make sure model path is present (for caching purposes)
            if (model.path.empty()) {
                // this is to avoid different repo having same file name, or same file name in different subdirs
@@ -1622,7 +1622,7 @@ static common_chat_params common_chat_templates_apply_jinja(
    }

    // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
-    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
+    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null() && params.tools.is_array() && params.json_schema.is_null()) {
        return common_chat_params_init_hermes_2_pro(tmpl, params);
    }

@@ -830,7 +830,7 @@ std::string fs_get_cache_directory() {
    if (getenv("LLAMA_CACHE")) {
        cache_directory = std::getenv("LLAMA_CACHE");
    } else {
-#ifdef __linux__
+#if defined(__linux__) || defined(__FreeBSD__)
        if (std::getenv("XDG_CACHE_HOME")) {
            cache_directory = std::getenv("XDG_CACHE_HOME");
        } else {
@@ -840,7 +840,9 @@ std::string fs_get_cache_directory() {
        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
        cache_directory = std::getenv("LOCALAPPDATA");
-#endif // __linux__
+#else
+#  error Unknown architecture
+#endif
        cache_directory = ensure_trailing_slash(cache_directory);
        cache_directory += "llama.cpp";
    }
@@ -1027,6 +1029,19 @@ struct common_init_result common_init_from_params(common_params & params) {
    return iparams;
 }

+std::string get_model_endpoint() {
+    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
+    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
+    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
+    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
+    std::string model_endpoint = "https://huggingface.co/";
+    if (endpoint_env) {
+        model_endpoint = endpoint_env;
+        if (model_endpoint.back() != '/') model_endpoint += '/';
+    }
+    return model_endpoint;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
    llama_clear_adapter_lora(ctx);
    for (auto & la : lora) {
@@ -543,6 +543,8 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);

+std::string                   get_model_endpoint();
+
 //
 // Batch utils
 //
@@ -735,6 +735,9 @@ class Model:
        if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406":
            # ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
            res = "llama4"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"

        if res is None:
            logger.warning("\n")
@@ -1750,7 +1753,7 @@ class LlamaModel(Model):

                low_freq_wavelen = old_context_len / low_freq_factor
                high_freq_wavelen = old_context_len / high_freq_factor
-                assert low_freq_wavelen != high_freq_wavelen
+                # assert low_freq_wavelen != high_freq_wavelen # Errors for Llama4

                rope_factors = []
                for freq in freqs:
@@ -1806,10 +1809,6 @@ class Llama4Model(LlamaModel):
        self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
-        name = name.replace("language_model.", "")
-        name = name.replace("feed_forward.", "mlp.") # a bit hacky for now
-        name = name.replace(".router.weight", ".gate.weight") # a bit hacky for now
-
        # split the gate_up into gate and up
        if "gate_up_proj" in name:
            name_up = name.replace("gate_up_proj", "up_proj.weight")
@@ -4901,6 +4900,22 @@ class JaisModel(Model):
        self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)


+@Model.register("Glm4ForCausalLM")
+class Glm4Model(Model):
+    model_arch = gguf.MODEL_ARCH.GLM4
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+
+
@Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(Model):
    model_arch = gguf.MODEL_ARCH.CHATGLM
@@ -5592,7 +5607,6 @@ def main() -> None:
    with torch.inference_mode():
        output_type = ftype_map[args.outtype]
        model_architecture = hparams["architectures"][0]
-
        try:
            model_class = Model.from_model_architecture(model_architecture)
        except NotImplementedError:
@@ -114,6 +114,7 @@ models = [
    {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
    {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
    {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
+    {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
 ]


@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "gguf.h"
+#include "clip.h"

 #include "clip.h"

@@ -202,23 +203,31 @@ static void clip_log_internal(enum ggml_log_level level, const char * format, ..
 // cpp wrappers
 //

+// wrapper for clip_image_size
+struct clip_image_size_deleter {
+    void operator()(clip_image_size * val) { clip_image_size_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
+// wrapper for clip_image_u8
 struct clip_image_u8_deleter {
    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
 };
+typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;

+// wrapper for clip_image_f32
 struct clip_image_f32_deleter {
    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
 };
+typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;

-struct clip_image_f32_batch_deleter {
-    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+struct clip_image_u8_batch {
+    std::vector<clip_image_u8_ptr> entries;
 };

-typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
-typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
-typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
-
-// TODO @ngxson : we're currently having a naming clash between struct clip_image_size and function clip_image_size()
+struct clip_image_f32_batch {
+    std::vector<clip_image_f32_ptr> entries;
+};

 //
 // common utils
@@ -315,58 +315,47 @@ struct clip_ctx {
    bool use_gelu = false;
    bool use_silu = false;

-    struct gguf_context * ctx_gguf = nullptr;
-    struct ggml_context * ctx_data = nullptr;
+    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_t buf  = nullptr;
+    ggml_backend_ptr backend;
+    ggml_backend_ptr backend_cpu;
+    ggml_backend_buffer_ptr buf;

    ggml_backend_sched_ptr sched;

-    struct clip_image_size * load_image_size = nullptr;
+    clip_image_size load_image_size;

    clip_ctx(clip_context_params & ctx_params) {
-        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-        backend     = ctx_params.use_gpu
+        backend_cpu = ggml_backend_ptr(ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr));
+        backend     = ggml_backend_ptr(ctx_params.use_gpu
                        ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
-                        : nullptr;
+                        : 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));
+            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend.get()));
+            backend_ptrs.push_back(backend.get());
+            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend.get()));
        } else {
-            backend = backend_cpu;
+            backend = std::move(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));
+        backend_ptrs.push_back(backend_cpu.get());
+        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu.get()));

        sched.reset(
            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false)
        );
    }
-
-    ~clip_ctx() {
-        ggml_free(ctx_data);
-        gguf_free(ctx_gguf);
-        ggml_backend_buffer_free(buf);
-        ggml_backend_free(backend);
-        if (backend_cpu != backend) {
-            ggml_backend_free(backend_cpu);
-        }
-        clip_image_size_free(load_image_size);
-    }
 };

-static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
    const auto & model = ctx->vision_model;
    const auto & hparams = model.hparams;

@@ -382,7 +371,7 @@ static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_im
    const int n_layer              = hparams.n_layer;
    const float eps                = hparams.eps;

-    GGML_ASSERT(imgs->size == 1); // batch_size == 1
+    GGML_ASSERT(imgs.entries.size() == 1); // batch_size == 1

    struct ggml_init_params params = {
        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
@@ -390,7 +379,9 @@ static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_im
        /*.no_alloc   =*/ true,
    };

-    struct ggml_context * ctx0 = ggml_init(params);
+    ggml_context_ptr ctx0_ptr(ggml_init(params));
+    auto ctx0 = ctx0_ptr.get();
+
    struct ggml_cgraph * gf = ggml_new_graph(ctx0);

    // input raw
@@ -512,12 +503,10 @@ static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_im
    // build the graph
    ggml_build_forward_expand(gf, embeddings);

-    ggml_free(ctx0);
-
    return gf;
 }

-static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) {
    if (!ctx->has_vision_encoder) {
        LOG_ERR("This gguf file seems to have no vision encoder\n");
        return nullptr;
@@ -530,23 +519,20 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
    int image_size_width  = image_size;
    int image_size_height = image_size;
    if (ctx->has_minicpmv_projector) {
-        if (load_image_size == nullptr) {
-            load_image_size = clip_image_size_init();
-        }
-        LOG_DBG("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
-        image_size_width  = load_image_size->width;
-        image_size_height = load_image_size->height;
+        LOG_DBG("%s: %d %d\n", __func__, load_image_size.width, load_image_size.height);
+        image_size_width  = load_image_size.width;
+        image_size_height = load_image_size.height;
        if (is_inf) {
-            image_size_width  = imgs->data->nx;
-            image_size_height = imgs->data->ny;
+            image_size_width  = imgs.entries[0]->nx;
+            image_size_height = imgs.entries[0]->ny;
        }
    }
    else if (ctx->has_qwen2vl_merger) {
        // use the image's native resolution when image is avaible
        if (is_inf) {
        // if (imgs->data->nx && imgs->data->ny) {
-            image_size_width  = imgs->data->nx;
-            image_size_height = imgs->data->ny;
+            image_size_width  = imgs.entries[0]->nx;
+            image_size_height = imgs.entries[0]->ny;
        }
    }
    const int patch_size           = hparams.patch_size;
@@ -561,7 +547,7 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
    const float eps                = hparams.eps;
    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};

-    const int batch_size = imgs->size;
+    const int batch_size = imgs.entries.size();

    if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) {
        GGML_ASSERT(batch_size == 1);
@@ -573,7 +559,9 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
        /*.no_alloc   =*/ true,
    };

-    struct ggml_context * ctx0 = ggml_init(params);
+    ggml_context_ptr ctx0_ptr(ggml_init(params));
+    auto ctx0 = ctx0_ptr.get();
+
    struct ggml_cgraph * gf = ggml_new_graph(ctx0);

    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
@@ -1061,7 +1049,7 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
            }
        } else {
-            GGML_ABORT("fatel error");
+            GGML_ABORT("fatal error");
        }
    }
    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
@@ -1081,12 +1069,10 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
    // build the graph
    ggml_build_forward_expand(gf, embeddings);

-    ggml_free(ctx0);
-
    return gf;
 }

-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) {
    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
        return clip_image_build_graph_siglip(ctx, imgs);
    } else {
@@ -1257,7 +1243,7 @@ struct clip_model_loader {
            /*.mem_buffer =*/ NULL,
            /*.no_alloc =*/ true,
        };
-        ctx_clip.ctx_data = ggml_init(params);
+        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__));
        }
@@ -1271,7 +1257,7 @@ struct clip_model_loader {
            if (cur) {
                tensors_to_load.push_back(cur);
                // add tensors to context
-                struct ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data, cur);
+                struct ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data.get(), cur);
                ggml_set_name(data_tensor, cur->name);
                cur = data_tensor;
            }
@@ -1442,11 +1428,11 @@ struct clip_model_loader {
            }

            // alloc memory and offload data
-            ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend);
-            ctx_clip.buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data, buft);
-            ggml_backend_buffer_set_usage(ctx_clip.buf, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+            ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend.get());
+            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) {
-                struct ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data, t->name);
+                struct 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) {
@@ -1471,10 +1457,20 @@ struct clip_model_loader {

    void alloc_compute_meta() {
        ctx_clip.buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+
+        // create a fake batch
        clip_image_f32_batch batch;
-        batch.size = 1;
-        batch.data = nullptr;
-        ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, &batch, nullptr, false);
+        clip_image_f32_ptr img(clip_image_f32_init());
+        clip_image_size image_size;
+        image_size.width  = clip_get_image_size(&ctx_clip);
+        image_size.height = clip_get_image_size(&ctx_clip);
+        int n_patches = clip_get_image_size(&ctx_clip) / image_size.width;
+        img->nx = n_patches;
+        img->ny = n_patches;
+        img->buf.resize(n_patches * image_size.width * image_size.height * 3);
+        batch.entries.push_back(std::move(img));
+
+        ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch, image_size, false);
        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];
@@ -1575,11 +1571,11 @@ struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_p
 }

 void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
-    ctx_clip->load_image_size = load_image_size;
+    ctx_clip->load_image_size = *load_image_size; // copy
 }

 struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) {
-    return ctx_clip->load_image_size;
+    return &ctx_clip->load_image_size;
 }

 struct clip_image_size * clip_image_size_init() {
@@ -1597,6 +1593,10 @@ 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;
@@ -1609,19 +1609,37 @@ void clip_image_size_free(struct clip_image_size * load_image_size) {
    }
    delete load_image_size;
 }
-void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
-void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
-    if (batch->size > 0) {
-        delete[] batch->data;
-        batch->size = 0;
-    }
+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();
 }
-void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
-    if (batch->size > 0) {
-        delete[] batch->data;
-        batch->size = 0;
+
+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) {
@@ -1695,14 +1713,15 @@ static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int ta
 }

 // 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());
+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());

-    for (size_t i = 0; i < src->buf.size(); ++i) {
+    // 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];
+        dst.buf[i] = (static_cast<float>(src.buf[i]) / 255.0f - mean[c]) / std[c];
    }
 }

@@ -1710,7 +1729,7 @@ inline int clip(int x, int lower, int upper) {
    return std::max(lower, std::min(x, upper));
 }

-static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) {
+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;

@@ -1848,13 +1867,13 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int> & or
    return best_fit;
 }

-static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & image, int patch_size) {
-    std::vector<clip_image_u8*> patches;
+static std::vector<clip_image_u8_ptr> divide_to_patches_u8(const clip_image_u8 & image, int patch_size) {
+    std::vector<clip_image_u8_ptr> patches;
    int width = image.nx;
    int height = image.ny;
    for (int i = 0; i < height; i += patch_size) {
        for (int j = 0; j < width; j += patch_size) {
-            clip_image_u8 *patch = clip_image_u8_init();
+            clip_image_u8_ptr patch(clip_image_u8_init());
            patch->nx = std::min(patch_size, width - j);
            patch->ny = std::min(patch_size, height - i);
            patch->buf.resize(3 * patch->nx * patch->ny);
@@ -1865,7 +1884,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
                    }
                }
            }
-            patches.push_back(patch);
+            patches.push_back(std::move(patch));
        }
    }
    return patches;
@@ -1946,7 +1965,7 @@ static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int mul
 //    -> 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
-static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
+static std::vector<std::vector<clip_image_u8_ptr>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
    const std::pair<int, int> original_size={img->nx,img->ny};
    const int original_width = img->nx;
    const int original_height = img->ny;
@@ -1954,30 +1973,30 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
    const int multiple = fmin(ceil(ratio), max_slice_nums);

-    std::vector<std::vector<clip_image_u8 *>> images;
+    std::vector<std::vector<clip_image_u8_ptr>> images;
    LOG_DBG("%s: multiple %d\n", __func__, multiple);
-    images.push_back(std::vector<clip_image_u8 *>());
+    images.push_back(std::vector<clip_image_u8_ptr>());

    if (multiple <= 1) {
        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
-        clip_image_u8 * source_image = clip_image_u8_init();
+        clip_image_u8_ptr source_image(clip_image_u8_init());
        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
        // source_image = image.resize(best_size, Image.Resampling.BICUBIC)
-        images[images.size()-1].push_back(source_image);
+        images.back().push_back(std::move(source_image));
    }
    else if (multiple > 1) {
        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
-        clip_image_u8 * source_image = clip_image_u8_init();
+        clip_image_u8_ptr source_image(clip_image_u8_init());
        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
        // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
        LOG_DBG("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
-        images[images.size()-1].push_back(source_image);
+        images.back().push_back(std::move(source_image));

        std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
        LOG_DBG("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);

        auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
-        clip_image_u8 * refine_image = clip_image_u8_init();
+        clip_image_u8_ptr refine_image(clip_image_u8_init());
        bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);

        LOG_DBG("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
@@ -1988,9 +2007,9 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
        int grid_x = int(width / best_grid.first);
        int grid_y = int(height / best_grid.second);
        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
-            images.push_back(std::vector<clip_image_u8 *>());
+            images.push_back(std::vector<clip_image_u8_ptr>());
            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
-                clip_image_u8 * patch = clip_image_u8_init();
+                clip_image_u8_ptr patch(clip_image_u8_init());
                patch->nx = grid_x;
                patch->ny = grid_y;
                patch->buf.resize(3 * patch->nx * patch->ny);
@@ -2003,10 +2022,9 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
                        patch->buf[j+2] = refine_image->buf[i+2];
                    }
                }
-                images[images.size()-1].push_back(patch);
+                images.back().push_back(std::move(patch));
            }
        }
-        clip_image_u8_free(refine_image);
    }
    return images;
 }
@@ -2014,8 +2032,8 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
 int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
    const int max_slice_nums=9;
    const int scale_resolution=448;
-    const int original_width = ctx_clip->load_image_size->width;
-    const int original_height = ctx_clip->load_image_size->height;
+    const int original_width = ctx_clip->load_image_size.width;
+    const int original_height = ctx_clip->load_image_size.height;
    const float log_ratio = log(1.0*original_width/original_height);
    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
    const int multiple = fmin(ceil(ratio), max_slice_nums);
@@ -2025,64 +2043,44 @@ int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {

 // 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, clip_image_f32_batch * res_imgs) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) {

-    if(clip_is_minicpmv(ctx)){
+    if (clip_is_minicpmv(ctx)) {
        int max_slice_nums = 9;
-        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img, max_slice_nums);
-        res_imgs->size = 0;
-        for (size_t i = 0; i < imgs.size(); ++i){
-            res_imgs->size += imgs[i].size();
-        }
-        res_imgs->data = new clip_image_f32[res_imgs->size];
-        int idx = 0;
+        std::vector<std::vector<clip_image_u8_ptr>> imgs = uhd_slice_image(img, max_slice_nums);
        for (size_t i = 0; i < imgs.size(); ++i) {
            for (size_t j = 0; j < imgs[i].size(); ++j) {
                LOG_DBG("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
-                clip_image_f32 * res = clip_image_f32_init();
-                normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
-                res_imgs->data[idx++] = *res;
-                clip_image_f32_free(res);
-            }
-        }
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            for (size_t j = 0; j < imgs[i].size(); ++j) {
-                if (imgs[i][j] != nullptr) {
-                    clip_image_u8_free(imgs[i][j]);
-                }
+                clip_image_f32_ptr res(clip_image_f32_init());
+                normalize_image_u8_to_f32(*imgs[i][j], *res, ctx->image_mean, ctx->image_std);
+                res_imgs->entries.push_back(std::move(res));
            }
        }
        return true;
    }
    else if (ctx->has_qwen2vl_merger) {
-        clip_image_u8 * resized = clip_image_u8_init();
-        auto patch_size = clip_patch_size(ctx) * 2;
+        clip_image_u8 resized;
+        auto patch_size = clip_get_patch_size(ctx) * 2;
        int nx = ceil((float)img->nx / patch_size) * patch_size;
        int ny = ceil((float)img->ny / patch_size) * patch_size;
-        bicubic_resize(*img, *resized, nx, ny);
+        bicubic_resize(*img, resized, nx, ny);

-        res_imgs->data = new clip_image_f32[1];
-        // clip_image_f32 * res = clip_image_f32_init();
-        normalize_image_u8_to_f32(resized, res_imgs->data, ctx->image_mean, ctx->image_std);
+        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, ctx->image_mean, ctx->image_std);
        // res_imgs->data[0] = *res;
-        res_imgs->size = 1;
-
-        // clip_image_f32_free(res);
-        clip_image_u8_free(resized);
+        res_imgs->entries.push_back(std::move(img_f32));
        return true;
    }

    if (ctx->has_glm_projector || ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
-        res_imgs->size = 1;
-        res_imgs->data = new clip_image_f32[res_imgs->size];
        clip_image_u8 resized_image;
        int32_t sz=ctx->vision_model.hparams.image_size;
        bicubic_resize(*img, resized_image,sz,sz);
-        clip_image_f32 * res = clip_image_f32_init();
+        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, res, ctx->image_mean, ctx->image_std);
-        res_imgs->data[0] = *res;
-        clip_image_f32_free(res);
+        normalize_image_u8_to_f32(resized_image, *img_f32, ctx->image_mean, ctx->image_std);
+        res_imgs->entries.push_back(std::move(img_f32));
        return true;
    }

@@ -2097,16 +2095,12 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
        pad_to_square = false;
    }
    // free the previous res_imgs if any set
-    if (res_imgs->size > 0) {
-        clip_image_f32_batch_free(res_imgs);
-    }
-    res_imgs->data = nullptr;
-    res_imgs->size = 0;
+    res_imgs->entries.clear();

    // 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 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
+    clip_image_u8_ptr temp(clip_image_u8_init()); // we will keep the input image data here temporarily
    if (pad_to_square && img->nx != img->ny) {
        int longer_side = std::max(img->nx, img->ny);
        temp->nx = longer_side;
@@ -2149,28 +2143,18 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
            //     clip_image_u8_free(temp2);
            // }

-            std::vector<clip_image_u8 *> patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6)
+            std::vector<clip_image_u8_ptr> patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6)

-            clip_image_u8 *image_original_resize = clip_image_u8_init();
+            clip_image_u8_ptr image_original_resize(clip_image_u8_init());
            // bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
            bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
-            patches.insert(patches.begin(), image_original_resize);
-            // clip_image_f32_batch_init(patches.size());
-            res_imgs->size = patches.size();
-            res_imgs->data = new clip_image_f32[res_imgs->size];
-            int num=0;
-            for (auto& patch : patches) {
-                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
-                num++;
+            patches.insert(patches.begin(), std::move(image_original_resize));
+            for (auto & patch : patches) {
+                clip_image_f32_ptr res(clip_image_f32_init());
+                normalize_image_u8_to_f32(*patch, *res, ctx->image_mean, ctx->image_std);
+                res_imgs->entries.push_back(std::move(res));
            }

-            for (size_t i = 0; i < patches.size(); i++) {
-                // LOG_DBG("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
-                clip_image_u8_free(patches[i]);
-            }
-
-            clip_image_u8_free(temp);
-
            return true;
        } else {
            temp->nx = img->nx;
@@ -2186,7 +2170,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli

    const int nx2 = ctx->vision_model.hparams.image_size;
    const int ny2 = ctx->vision_model.hparams.image_size;
-    clip_image_f32 * res = clip_image_f32_init();
+    clip_image_f32_ptr res(clip_image_f32_init());
    res->nx = nx2;
    res->ny = ny2;
    res->buf.resize(3 * nx2 * ny2);
@@ -2238,7 +2222,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
            }
        }
    }
-    clip_image_u8_free(temp);

    // {
    //     clip_image_u8 * temp2 = clip_image_u8_init();
@@ -2248,10 +2231,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
    // }
    // res_imgs.push_back(res);

-    res_imgs->size = 1;
-    res_imgs->data = new clip_image_f32[res_imgs->size];
-    res_imgs->data[0] = *res;
-    clip_image_f32_free(res);
+    res_imgs->entries.push_back(std::move(res));

    return true;
 }
@@ -2279,15 +2259,15 @@ size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w
    return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }

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

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

-int32_t clip_hidden_size(const struct clip_ctx * ctx) {
+int32_t clip_get_hidden_size(const struct clip_ctx * ctx) {
    return ctx->vision_model.hparams.hidden_size;
 }

@@ -2434,19 +2414,23 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
        return false;
    }

-    clip_image_f32_batch imgs{};
-    imgs.size = 1;
-    imgs.data = img;
+    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, float * 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;
+
    if (!ctx->has_vision_encoder) {
        LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__);
        return false;
    }

-    int batch_size = imgs->size;
+    int batch_size = imgs.entries.size();
    if (ctx->has_llava_projector) {
        GGML_ASSERT(batch_size == 1); // TODO: support multiple images
    }
@@ -2473,25 +2457,22 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    int image_size_width  = image_size;
    int image_size_height = image_size;
    if (ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger) {
-        image_size_width  = imgs->data[0].nx;
-        image_size_height = imgs->data[0].ny;
+        image_size_width  = imgs.entries[0]->nx;
+        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 num_positions = num_patches + (model.class_embedding ? 1 : 0);
-    if(ctx->load_image_size==nullptr){
-        ctx->load_image_size= clip_image_size_init();
-    }
-    const int pos_w = ctx->load_image_size->width/patch_size;
-    const int pos_h = ctx->load_image_size->height/patch_size;
+    const int pos_w = ctx->load_image_size.width / patch_size;
+    const int pos_h = ctx->load_image_size.height / patch_size;

    {
        struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
        float * data = (float *)malloc(ggml_nbytes(inp_raw));

-        for (size_t i = 0; i < imgs->size; i++) {
-            const int nx = imgs->data[i].nx;
-            const int ny = imgs->data[i].ny;
+        for (size_t i = 0; i < imgs.entries.size(); i++) {
+            const int nx = imgs.entries[i]->nx;
+            const int ny = imgs.entries[i]->ny;
            if (!(ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger)) {
                GGML_ASSERT(nx == image_size && ny == image_size);
            }
@@ -2502,7 +2483,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                for (int k = 0; k < 3; k++) {
                    for (int y = 0; y < ny; y++) {
                        for (int x = 0; x < nx; x++) {
-                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs.entries[b]->buf[3 * (y * nx + x) + k];
                        }
                    }
                }
@@ -2629,7 +2610,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
        }
    }

-    ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
+    ggml_backend_cpu_set_n_threads(ctx->backend_cpu.get(), n_threads);

    auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
    if (status != GGML_STATUS_SUCCESS) {
@@ -2662,8 +2643,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
        /* verbosity */ GGML_LOG_LEVEL_ERROR,
    });

-    const auto & ctx_src = ctx_clip->ctx_gguf;
-    const auto & ctx_data = ctx_clip->ctx_data;
+    const auto & ctx_src = ctx_clip->ctx_gguf.get();
+    const auto & ctx_data = ctx_clip->ctx_data.get();

    auto * ctx_out = gguf_init_empty();
    gguf_set_kv(ctx_out, ctx_src);
@@ -30,15 +30,8 @@ struct clip_image_size {
    int height;
 };

-struct clip_image_u8_batch {
-    struct clip_image_u8 * data;
-    size_t size;
-};
-
-struct clip_image_f32_batch {
-    struct clip_image_f32 * data;
-    size_t size;
-};
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;

 struct clip_context_params {
    bool use_gpu;
@@ -55,9 +48,9 @@ CLIP_API void clip_free(struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);

-CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx);

 // TODO: should be enum, not string
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
@@ -73,9 +66,10 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
 CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);

-CLIP_API struct clip_image_size * clip_image_size_init();
-CLIP_API struct clip_image_u8  * clip_image_u8_init ();
-CLIP_API struct clip_image_f32 * clip_image_f32_init();
+CLIP_API struct clip_image_size      * clip_image_size_init();
+CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava

 // nx, ny are the output image dimensions
 CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
@@ -86,6 +80,12 @@ CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);

+// use for accessing underlay data of clip_image_f32_batch
+CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
+CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
+CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
+CLIP_API clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+
 /**
 * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
 * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
@@ -10,6 +10,7 @@
 #include <cstring>
 #include <limits>
 #include <vector>
+#include <memory>

 #if defined(LLAVA_LOG_OFF)
 #   define LOG_INF(...)
@@ -45,6 +46,17 @@ struct clip_image_grid_shape {
    int second;
 };

+// convenience cpp wrapper
+struct clip_image_f32_batch_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
+
+struct clip_image_size_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
 /**
 * Selects the best resolution from a list of possible resolutions based on the original size.
 *
@@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
        struct ggml_context * ctx;
    } model;

-    const int32_t image_size = clip_image_size(ctx_clip);
-    const int32_t patch_size = clip_patch_size(ctx_clip);
+    const int32_t image_size = clip_get_image_size(ctx_clip);
+    const int32_t patch_size = clip_get_patch_size(ctx_clip);

    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)

@@ -246,12 +258,9 @@ static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size)

 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
-    clip_image_f32_batch img_res_v;
-    img_res_v.size = 0;
-    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
        LOG_ERR("%s: unable to preprocess image\n", __func__);
-        delete[] img_res_v.data;
        return false;
    }

@@ -259,66 +268,72 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli

    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);

+    const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
+
    if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
        std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        struct clip_image_size * load_image_size = clip_image_size_init();
+        image_embd_v.resize(n_imgs);
+        clip_image_size load_image_size;

-        for (size_t i = 0; i < img_res_v.size; i++) {
+        for (size_t i = 0; i < n_imgs; i++) {
            const int64_t t_img_enc_step_start_us = ggml_time_us();
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            int patch_size=14;
-            load_image_size->width = img_res_v.data[i].nx;
-            load_image_size->height = img_res_v.data[i].ny;
-            clip_add_load_image_size(ctx_clip, load_image_size);
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            int patch_size = 14;
+            load_image_size.width = nx;
+            load_image_size.height = ny;
+            clip_add_load_image_size(ctx_clip, &load_image_size);

            bool encoded = false;
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
            if (clip_is_qwen2vl(ctx_clip)) {
-                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
            }
            else {
-                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]);
            }

            if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                return false;
            }
            const int64_t t_img_enc_steop_batch_us = ggml_time_us();
-            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
        }
        const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);

        int n_img_pos_out = 0;
        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
            std::memcpy(
                image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                image_embd_v[i],
-                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
+                clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
        }
        *n_img_pos = n_img_pos_out;
        for (size_t i = 0; i < image_embd_v.size(); i++) {
            free(image_embd_v[i]);
        }
        image_embd_v.clear();
-        load_image_size->width = img->nx;
-        load_image_size->height = img->ny;
-        clip_add_load_image_size(ctx_clip, load_image_size);
-        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
+        load_image_size.width = img->nx;
+        load_image_size.height = img->ny;
+        clip_add_load_image_size(ctx_clip, &load_image_size);
+        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height);
    }
    else if (clip_is_glm(ctx_clip)){
        struct clip_image_size * load_image_size = clip_image_size_init();
-        load_image_size->width = img_res_v.data[0].nx;
-        load_image_size->height = img_res_v.data[0].ny;
+        load_image_size->width  = clip_image_f32_batch_nx(img_res_v.get(), 0);
+        load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
        clip_add_load_image_size(ctx_clip, load_image_size);

-        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
-        int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+        int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
        *n_img_pos = (pos * pos + 2);
        if (!encoded){
            LOG_ERR("Unable to encode image \n");
@@ -328,8 +343,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
        // flat / default llava-1.5 type embedding
        *n_img_pos = clip_n_patches(ctx_clip);
-        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
-        delete[] img_res_v.data;
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
        if (!encoded) {
            LOG_ERR("Unable to encode image\n");

@@ -340,17 +355,18 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
        // spatial_unpad llava-1.6 type embedding
        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
        std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        image_embd_v.resize(n_imgs);
+        for (size_t i = 0; i < n_imgs; i++) {
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
            if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                return false;
            }
        }
        const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);

        const int32_t * image_grid = clip_image_grid(ctx_clip);
        const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
@@ -360,12 +376,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
        }

-        // free all img_res_v - not needed anymore
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
-
-        const int32_t image_size = clip_image_size(ctx_clip);
+        const int32_t image_size = clip_get_image_size(ctx_clip);

        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);

@@ -41,14 +41,14 @@ struct mtmd_context {
 };

 struct mtmd_image_tokens_data {
-    clip_image_f32_batch_ptr batch_f32; // preprocessed image patches
+    clip_image_f32_batch batch_f32; // preprocessed image patches
 };

 struct mtmd_image_tokens {
    uint32_t nx; // number of tokens in x direction
    uint32_t ny; // number of tokens in y direction
    uint32_t n_tokens() const { return nx * ny; }
-    clip_image_f32_batch_ptr batch_f32; // preprocessed image patches
+    clip_image_f32_batch batch_f32; // preprocessed image patches
 };

 mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
@@ -141,8 +141,8 @@ mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
            std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);

            // preprocess image
-            clip_image_f32_batch_ptr batch_f32(new clip_image_f32_batch);
-            bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), batch_f32.get());
+            clip_image_f32_batch batch_f32;
+            bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32);
            if (!ok) {
                LOG_ERR("Unable to preprocess image\n");
                return nullptr;
@@ -181,7 +181,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
    bool ok = clip_image_batch_encode(
        ctx->ctx_clip,
        ctx->n_threads,
-        image_tokens->batch_f32.get(),
+        &image_tokens->batch_f32,
        ctx->image_embd_v.data());
    return ok ? 0 : 1;
 }
@@ -126,7 +126,7 @@ static std::string fs_get_cache_directory() {
    if (getenv("LLAMA_CACHE")) {
        cache_directory = std::getenv("LLAMA_CACHE");
    } else {
-#ifdef __linux__
+#if defined(__linux__) || defined(__FreeBSD__)
        if (std::getenv("XDG_CACHE_HOME")) {
            cache_directory = std::getenv("XDG_CACHE_HOME");
        } else {
@@ -136,7 +136,9 @@ static std::string fs_get_cache_directory() {
        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
        cache_directory = std::getenv("LOCALAPPDATA");
-#endif // __linux__
+#else
+#  error Unknown architecture
+#endif
        cache_directory = ensure_trailing_slash(cache_directory);
        cache_directory += "llama.cpp";
    }
@@ -697,8 +697,10 @@ class LlamaData {
        std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
        std::string              url;

+        std::string model_endpoint = get_model_endpoint();
+
        if (pos == std::string::npos) {
-            auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/");
+            auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/");
            hfr                             = model_name;

            nlohmann::json manifest;
@@ -713,7 +715,7 @@ class LlamaData {
            hff = model.substr(pos + 1);
        }

-        url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
+        url = model_endpoint + hfr + "/resolve/main/" + hff;

        return download(url, bn, true, headers);
    }
@@ -3907,6 +3907,21 @@ int main(int argc, char ** argv) {
        res_ok(res, {{ "success", true }});
    };

+    const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+        json data = {
+            {
+                "template", common_chat_templates_source(ctx_server.chat_templates.get()),
+            },
+            {
+                "model_info", {
+                    { "llama.context_length", ctx_server.slots.back().n_ctx, },
+                }
+            },
+        };
+
+        res_ok(res, data);
+    };
+
    // handle completion-like requests (completion, chat, infill)
    // we can optionally provide a custom format for partial results and final results
    const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
@@ -4471,6 +4486,7 @@ int main(int argc, char ** argv) {
    svr->Get ("/metrics",             handle_metrics);
    svr->Get ("/props",               handle_props);
    svr->Post("/props",               handle_props_change);
+    svr->Post("/api/show",            handle_api_show);
    svr->Get ("/models",              handle_models); // public endpoint (no API key check)
    svr->Get ("/v1/models",           handle_models); // public endpoint (no API key check)
    svr->Post("/completion",          handle_completions); // legacy
@@ -2,6 +2,13 @@
 #define GGML_SYCL_ELEMENTWISE_HPP

 #include "common.hpp"
+#include "ggml.h"
+#include <limits.h>
+
+template <typename T>
+T neg_infinity() {
+    return -std::numeric_limits<T>::infinity();
+}

 static __dpct_inline__ float op_repeat(const float a, const float b) {
    return b;
@@ -24,6 +31,19 @@ static __dpct_inline__ float op_div(const float a, const float b) {
    return a / b;
 }

+template<typename T>
+struct typed_data {
+    const T * src;
+    T * dst;
+};
+
+template<typename T>
+typed_data<T> cast_data(ggml_tensor * dst) {
+    return {
+        /* .src = */ static_cast<const T *>(dst->src[0]->data),
+        /* .dst = */ static_cast<T *>(dst->data)
+    };
+}

 void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

@@ -65,6 +85,10 @@ void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

 void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

+void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+// ---------
+
 void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

 void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -1617,17 +1617,6 @@ static void scale_f32(const float * x, float * dst, const float scale, const int
    dst[i] = scale * x[i];
 }

-static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k,
-                      const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-
-    if (i >= k) {
-        return;
-    }
-
-    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
-}

 template <typename Ti, typename To>
 static  void pool2d_nchw_kernel(
@@ -1768,18 +1757,6 @@ static void scale_f32_sycl(const float *x, float *dst, const float scale,
        });
 }

-static void clamp_f32_sycl(const float *x, float *dst, const float min,
-                           const float max, const int k,
-                           queue_ptr stream) {
-    const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            clamp_f32(x, dst, min, max, k, item_ct1);
-        });
-}

 static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols,
                              const int nrows, queue_ptr stream) {
@@ -2258,26 +2235,6 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor *dst
    SYCL_CHECK(0);
 }

-inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
-    float *       dst_dd  = static_cast<float *>(dst->data);
-
-    float min;
-    float max;
-    memcpy(&min, dst->op_params, sizeof(float));
-    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
-
-    clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(dst->src[0]), ctx.stream());
-    /*
-    DPCT1010:88: SYCL uses exceptions to report errors and does not use the
-    error codes. The call was replaced with 0. You need to rewrite this code.
-    */
-    SYCL_CHECK(0);
-}
-
 static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
    static bool peer_access_enabled = false;

@@ -3218,10 +3175,6 @@ static void ggml_sycl_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
    ggml_sycl_op_scale(ctx, dst);
 }

-static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_op_clamp(ctx, dst);
-}
-
 static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    ggml_sycl_op_diag_mask_inf(ctx, dst);
 }
@@ -3700,7 +3653,8 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_

 #ifdef GGML_SYCL_GRAPH
    if (!g_ggml_sycl_disable_graph) {
-        if (!sycl_ctx->exec_graph && !dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_graph)) {
+        const bool graph_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_limited_graph);
+        if (!graph_support) {
            GGML_SYCL_DEBUG("[SYCL-GRAPH] can not use graphs on device:%d\n", sycl_ctx->device);
            ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
            return GGML_STATUS_SUCCESS;
@@ -3711,8 +3665,10 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
        ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
        model_sycl_graph.end_recording();

-        if (!sycl_ctx->exec_graph) {
-            auto exec_graph = model_sycl_graph.finalize({sycl_ex::property::graph::updatable{}});
+        const bool graph_update_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_graph);
+        if (!sycl_ctx->exec_graph || !graph_update_support) {
+            auto exec_graph = graph_update_support ? model_sycl_graph.finalize(sycl_ex::property::graph::updatable{}) :
+                                                     model_sycl_graph.finalize();
            sycl_ctx->exec_graph = std::make_unique<
                sycl_ex::command_graph<sycl_ex::graph_state::executable>>(exec_graph);
        } else {
@@ -3900,7 +3856,11 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                case GGML_UNARY_OP_GELU_QUICK:
                case GGML_UNARY_OP_TANH:
                case GGML_UNARY_OP_EXP:
-                    return ggml_is_contiguous(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32);
+#if defined (GGML_SYCL_F16)
+                    return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type);
+#else
+                    return ggml_is_contiguous(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) && (op->type == op->src[0]->type);
+#endif
                default:
                    return false;
            }
@@ -4022,13 +3982,18 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_SUB:
        case GGML_OP_MUL:
        case GGML_OP_DIV:
+            return (op->src[0]->type == GGML_TYPE_F32);
        case GGML_OP_SQR:
        case GGML_OP_SQRT:
        case GGML_OP_SIN:
        case GGML_OP_COS:
        case GGML_OP_CLAMP:
        case GGML_OP_LOG:
-            return (op->src[0]->type == GGML_TYPE_F32);
+#if defined (GGML_SYCL_F16)
+            return ((op->type == GGML_TYPE_F32 || op->type == GGML_SYCL_F16) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_SYCL_F16) && (op->type == op->src[0]->type));
+#else
+            return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type);
+#endif
        case GGML_OP_NORM:
        case GGML_OP_RMS_NORM:
        case GGML_OP_L2_NORM:
@@ -280,6 +280,7 @@ class MODEL_ARCH(IntEnum):
    DEEPSEEK         = auto()
    DEEPSEEK2        = auto()
    CHATGLM          = auto()
+    GLM4             = auto()
    BITNET           = auto()
    T5               = auto()
    T5ENCODER        = auto()
@@ -487,6 +488,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.DEEPSEEK:         "deepseek",
    MODEL_ARCH.DEEPSEEK2:        "deepseek2",
    MODEL_ARCH.CHATGLM:          "chatglm",
+    MODEL_ARCH.GLM4:             "glm4",
    MODEL_ARCH.BITNET:           "bitnet",
    MODEL_ARCH.T5:               "t5",
    MODEL_ARCH.T5ENCODER:        "t5encoder",
@@ -1561,6 +1563,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.GLM4 : [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
    MODEL_ARCH.BITNET: [
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
@@ -13,7 +13,7 @@ class TensorNameMap:
            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
            "transformer.word_embeddings",               # falcon
            "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414
            "tok_embeddings",                            # llama-pth
            "embeddings.word_embeddings",                # bert nomic-bert
            "language_model.embedding.word_embeddings",  # persimmon
@@ -30,6 +30,7 @@ class TensorNameMap:
            "rwkv.embeddings",                           # rwkv6
            "model.embeddings",                          # rwkv7
            "model.word_embeddings",                     # bailingmoe
+            "language_model.model.embed_tokens",         # llama4
        ),

        # Token type embeddings
@@ -67,6 +68,7 @@ class TensorNameMap:
            "output_layer",              # chatglm
            "head",                      # rwkv
            "head.out",                  # wavtokenizer
+            "language_model.lm_head",    # llama4
        ),

        # Output norm
@@ -89,6 +91,7 @@ class TensorNameMap:
            "rwkv.ln_out",                             # rwkv6
            "model.ln_out",                            # rwkv7
            "backbone.final_layer_norm",               # wavtokenizer
+            "language_model.model.norm",               # llama4
        ),

        # Rope frequencies
@@ -130,6 +133,7 @@ class TensorNameMap:
            "transformer.layers.{bid}.attn_norm",                   # openelm
            "rwkv.blocks.{bid}.ln1",                                # rwkv6
            "model.layers.{bid}.ln1",                               # rwkv7
+            "language_model.model.layers.{bid}.input_layernorm",    # llama4
        ),

        # Attention norm 2
@@ -169,6 +173,7 @@ class TensorNameMap:
            "model.layers.{bid}.attention.wq",                           # internlm2
            "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
            "transformer.h.{bid}.attn.attention.q_proj",                 # exaone
+            "language_model.model.layers.{bid}.self_attn.q_proj",        # llama4
        ),

        # Attention key
@@ -183,6 +188,7 @@ class TensorNameMap:
            "model.layers.{bid}.attention.wk",                         # internlm2
            "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
            "transformer.h.{bid}.attn.attention.k_proj",               # exaone
+            "language_model.model.layers.{bid}.self_attn.k_proj",      # llama4
        ),

        # Attention value
@@ -196,6 +202,7 @@ class TensorNameMap:
            "model.layers.{bid}.attention.wv",                           # internlm2
            "transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok
            "transformer.h.{bid}.attn.attention.v_proj",                 # exaone
+            "language_model.model.layers.{bid}.self_attn.v_proj",        # llama4
        ),

        # Attention output
@@ -222,6 +229,7 @@ class TensorNameMap:
            "encoder.layers.{bid}.self_attention.dense",                    # chatglm
            "transformer.layers.{bid}.attn.out_proj",                       # openelm
            "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
+            "language_model.model.layers.{bid}.self_attn.o_proj",           # llama4
        ),

        # Attention output norm
@@ -233,7 +241,8 @@ class TensorNameMap:
        ),

        MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2    # ge
+            "model.layers.{bid}.post_self_attn_layernorm",     # glm-4-0414
        ),

        # Rotary embeddings
@@ -259,6 +268,7 @@ class TensorNameMap:
            "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
            "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
            "transformer.layers.{bid}.ffn_norm",                             # openelm
+            "language_model.model.layers.{bid}.post_attention_layernorm",    # llama4
        ),

        # Post feed-forward norm
@@ -269,6 +279,7 @@ class TensorNameMap:
        # Post feed-forward norm
        MODEL_TENSOR.FFN_POST_NORM: (
            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
+            "model.layers.{bid}.post_mlp_layernorm", # glm-4-0414
        ),

        MODEL_TENSOR.FFN_GATE_INP: (
@@ -278,6 +289,7 @@ class TensorNameMap:
            "transformer.decoder_layer.{bid}.router",           # Grok
            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
+            "language_model.model.layers.{bid}.feed_forward.router", # llama4
        ),

        MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -306,7 +318,7 @@ class TensorNameMap:
            "h.{bid}.mlp.c_fc",                                       # gpt2
            "transformer.h.{bid}.mlp.fc1",                            # phi2
            "model.layers.{bid}.mlp.fc1",                             # phi2
-            "model.layers.{bid}.mlp.gate_up_proj",                    # phi3
+            "model.layers.{bid}.mlp.gate_up_proj",                    # phi3 glm-4-0414
            "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
            "model.layers.{bid}.feed_forward.w3",                     # internlm2
            "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
@@ -315,6 +327,7 @@ class TensorNameMap:
            "model.layers.{bid}.residual_mlp.w3",                     # arctic
            "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
            "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
+            "language_model.model.layers.{bid}.feed_forward.up_proj", # llama4
        ),

        MODEL_TENSOR.FFN_UP_EXP: (
@@ -323,11 +336,13 @@ class TensorNameMap:
            "transformer.blocks.{bid}.ffn.experts.mlp.v1",    # dbrx
            "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
            "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
+            "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
        ),

        MODEL_TENSOR.FFN_UP_SHEXP: (
            "model.layers.{bid}.mlp.shared_expert.up_proj",  # qwen2moe
            "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2
+            "language_model.model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4
        ),

        # AWQ-activation gate
@@ -348,6 +363,7 @@ class TensorNameMap:
            "transformer.h.{bid}.mlp.linear_1",           # refact
            "model.layers.{bid}.residual_mlp.w1",         # arctic
            "transformer.h.{bid}.mlp.c_fc_0",             # exaone
+            "language_model.model.layers.{bid}.feed_forward.gate_proj", # llama4
        ),

        MODEL_TENSOR.FFN_GATE_EXP: (
@@ -356,11 +372,13 @@ class TensorNameMap:
            "transformer.blocks.{bid}.ffn.experts.mlp.w1",    # dbrx
            "model.layers.{bid}.mlp.experts.gate_proj",       # qwen2moe olmoe (merged)
            "model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
+            "language_model.model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
        ),

        MODEL_TENSOR.FFN_GATE_SHEXP: (
            "model.layers.{bid}.mlp.shared_expert.gate_proj",  # qwen2moe
            "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2
+            "language_model.model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4
        ),

        # Feed-forward down
@@ -389,6 +407,7 @@ class TensorNameMap:
            "encoder.layer.{bid}.mlp.down_layer",                     # jina-bert-v2
            "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
            "model.layers.h.{bid}.mlp.c_proj",                        # exaone
+            "language_model.model.layers.{bid}.feed_forward.down_proj", # llama4
        ),

        MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -398,11 +417,13 @@ class TensorNameMap:
            "model.layers.{bid}.mlp.experts.down_proj",          # qwen2moe olmoe (merged)
            "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
            "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
+            "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
        ),

        MODEL_TENSOR.FFN_DOWN_SHEXP: (
            "model.layers.{bid}.mlp.shared_expert.down_proj",  # qwen2moe
            "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
+            "language_model.model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
        ),

        MODEL_TENSOR.ATTN_Q_NORM: (
@@ -54,6 +54,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_DEEPSEEK,         "deepseek"         },
    { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
    { LLM_ARCH_CHATGLM,          "chatglm"          },
+    { LLM_ARCH_GLM4,             "glm4"             },
    { LLM_ARCH_BITNET,           "bitnet"           },
    { LLM_ARCH_T5,               "t5"               },
    { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -1152,6 +1153,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
        },
    },
+    {
+        LLM_ARCH_GLM4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
    {
        LLM_ARCH_BITNET,
        {
@@ -58,6 +58,7 @@ enum llm_arch {
    LLM_ARCH_DEEPSEEK,
    LLM_ARCH_DEEPSEEK2,
    LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
    LLM_ARCH_BITNET,
    LLM_ARCH_T5,
    LLM_ARCH_T5ENCODER,
@@ -256,6 +257,8 @@ enum llm_tensor {
    LLM_TENSOR_ATTN_Q_NORM,
    LLM_TENSOR_ATTN_K_NORM,
    LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
    LLM_TENSOR_SSM_IN,
    LLM_TENSOR_SSM_CONV1D,
    LLM_TENSOR_SSM_X,
@@ -1205,6 +1205,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                    default: type = LLM_TYPE_UNKNOWN;
                }
            } break;
+        case LLM_ARCH_GLM4:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 40: type = LLM_TYPE_9B; break;
+                    case 61: type = LLM_TYPE_32B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
        case LLM_ARCH_BITNET:
            {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -3476,6 +3485,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                    }
                } break;
+            case LLM_ARCH_GLM4:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        }
+
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+
+                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
            case LLM_ARCH_NEMOTRON:
                {
                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -10854,6 +10902,157 @@ struct llm_build_chatglm : public llm_graph_context {
    }
 };

+struct llm_build_glm4 : public llm_graph_context {
+    llm_build_glm4(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = nullptr;
+                ggml_tensor * Kcur = nullptr;
+                ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv == nullptr) {
+                    Qcur = build_lora_mm(model.layers[il].wq, cur);
+                    if (model.layers[il].bq) {
+                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    }
+                    Kcur = build_lora_mm(model.layers[il].wk, cur);
+                    if (model.layers[il].bk) {
+                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    }
+                    Vcur = build_lora_mm(model.layers[il].wv, cur);
+                    if (model.layers[il].bv) {
+                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    }
+                } else {
+                    cur = build_lora_mm(model.layers[il].wqkv, cur);
+                    cb(cur, "wqkv", il);
+                    if (model.layers[il].bqkv) {
+                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                        cb(cur, "bqkv", il);
+                    }
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Post-attention norm (new!)
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Add the input (residual connection after post-attention norm)
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                // Pre-MLP norm
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                // MLP
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
+
+                // Post-MLP norm
+                cur = build_norm(cur,
+                        model.layers[il].ffn_post_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "post_mlp_norm", il);
+            }
+
+            // Add residual connection after post-MLP norm
+            inpL = ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        // Final norm
+        cur = build_norm(inpL,
+                model.output_norm,
+                NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // Output projection
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_nemotron : public llm_graph_context {
    llm_build_nemotron(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
        const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -12735,6 +12934,10 @@ llm_graph_result_ptr llama_model::build_graph(
            {
                llm = std::make_unique<llm_build_chatglm>(*this, params, gf);
            } break;
+        case LLM_ARCH_GLM4:
+            {
+                llm = std::make_unique<llm_build_glm4>(*this, params, gf);
+            } break;
        case LLM_ARCH_BITNET:
            {
                llm = std::make_unique<llm_build_bitnet>(*this, params, gf);
@@ -12932,6 +13135,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_PLM:
        case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GLM4:
        case LLM_ARCH_GRANITE:
        case LLM_ARCH_GRANITE_MOE:
        case LLM_ARCH_CHAMELEON:
@@ -1572,6 +1572,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                pre_type = LLAMA_VOCAB_PRE_TYPE_PORO;
                clean_spaces = false;
            } else if (
+                tokenizer_pre == "glm4" ||
                tokenizer_pre == "chatglm-bpe") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_CHATGLM4;
                special_bos_id = LLAMA_TOKEN_NULL;
@@ -569,6 +569,7 @@ static void test_template_output_parsers() {
    {
        // Not supported yet
        auto tmpls = read_templates("models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja");
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
    }
    {
@@ -665,6 +666,7 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja");
        std::vector<std::string> end_tokens{ "<|im_end|>" };

+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
        assert_equals(
            COMMON_CHAT_FORMAT_HERMES_2_PRO,
@@ -793,6 +795,7 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja");
        std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };

+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
                      common_chat_templates_apply(tmpls.get(), inputs_tools_builtin).format);
@@ -815,6 +818,7 @@ static void test_template_output_parsers() {
        std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };

        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);

        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
@@ -824,6 +828,8 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/meetkai-functionary-medium-v3.1.jinja");
        std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };

+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY,
+                      common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
                      common_chat_templates_apply(tmpls.get(), inputs_tools).format);

@@ -851,6 +857,7 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja");
        std::vector<std::string>   end_tokens{ "<|eot_id|>" };

+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_FIREFUNCTION_V2, common_chat_templates_apply(tmpls.get(), inputs_tools).format);

        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
@@ -862,6 +869,7 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja");
        std::vector<std::string>   end_tokens{ "<｜end▁of▁sentence｜>" };

+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);

@@ -891,6 +899,7 @@ static void test_template_output_parsers() {
        auto tmpls = read_templates("models/templates/llama-cpp-deepseek-r1.jinja");
        std::vector<std::string>   end_tokens{ "<｜end▁of▁sentence｜>" };

+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
Author	SHA1	Message	Date
Georgi Gerganov	c94085df28	server : add VSCode's Github Copilot Chat support (#12896 ) * server : add VSCode's Github Copilot Chat support * cont : update handler name	2025-04-11 23:37:41 +03:00
yuri@FreeBSD	e8a62631b3	rpc : Set cache directory in rpc-server.cpp on FreeBSD (#12903 )	2025-04-11 22:04:14 +02:00
Olivier Chafik	b6930ebc42	`tool-call`: fix non-tool-calling grammar crashes w/ Qwen / Hermes 2 templates (#12900 ) * `tool-call`: don't call common_chat_params_init_hermes_2_pro when there aren't tools (or when there's a schema) * test all chat formats w/o tools	2025-04-11 21:47:52 +02:00
yuri@FreeBSD	68b08f36d0	common : Define cache directory on FreeBSD (#12892 )	2025-04-11 21:45:44 +02:00
Ewan Crawford	578754b315	sycl: Support sycl_ext_oneapi_limited_graph (#12873 ) The current usage of the SYCL-Graph extension checks for the `sycl_ext_oneapi_graph` device aspect. However, it is also possible to support `sycl_ext_oneapi_limied_graph` devices that don't support update	2025-04-11 15:32:14 +02:00
tastelikefeet	b2034c2b55	contrib: support modelscope community (#12664 ) * support download from modelscope * support login * remove comments * add arguments * fix code * fix win32 * test passed * fix readme * revert readme * change to MODEL_ENDPOINT * revert tail line * fix readme * refactor model endpoint * remove blank line * fix header * fix as comments * update comment * update readme --------- Co-authored-by: tastelikefeet <yuze.zyz@alibaba-inc/com>	2025-04-11 14:01:56 +02:00
Yuxuan Zhang	06bb53ad9b	llama-model : add Glm4Model implementation for GLM-4-0414 (#12867 ) * GLM-4-0414 * use original one * Using with tensor map * fix bug * change order * change order * format with flask8	2025-04-11 12:10:10 +02:00
Xuan-Son Nguyen	0c50923944	clip : use smart pointer (⚠️ breaking change) (#12869 ) * clip : use smart pointers * fix warmup * add forward declaration * misisng include * fix include (2) * composite * simplify batch ptr * fix conflict	2025-04-11 12:09:39 +02:00
Akarshan Biswas	fccf9cae83	SYCL: Add fp16 type support to unary op kernels (#12788 ) * SYCL: Add fp16 support to some elementwise OP kernels * remove comment ggml-ci * Use static_cast directly * remove not needed cast from tanh * Use static cast and remove unneeded castings * Adjust device_support_op for unary OPs * Use cast_data and typed_data struct to deduplicate casting code	2025-04-11 16:03:50 +08:00
Daniel Han	ec6c09d0fa	convert : Llama4 RoPE fix (#12889 )	2025-04-11 09:49:09 +02:00
R0CKSTAR	8ac9f5d765	ci : Replace freediskspace to free_disk_space in docker.yml (#12861 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2025-04-11 09:26:17 +02:00
Daniel Bevenius	12e9158f25	xcf : add check for visionos build version (#12854 ) This commit adds a check for the visionos build version used with vtool in build-xcframework.sh. The script now checks the Xcode version and determines whether to use "xros" or "visionos" for the build version. This commit also uses xcrun for the vtool so that the version of vtool in xcode command line tools is used instead of the one in the system path. Refs: https://github.com/ggml-org/whisper.cpp/pull/2994#issuecomment-2773292223	2025-04-11 09:24:34 +02:00
Xuan-Son Nguyen	5b1f13cb64	convert : proper tensor name mapping for llama4 (#12870 ) * Llama-4 mapping * remove hacky renaming --------- Co-authored-by: Daniel Han <danielhanchen@gmail.com>	2025-04-11 09:23:37 +02:00