ggml : add SSE 4.2 and x64 base variant for CPUs without AVX (#12871)

* ggml : add SSE 4.2 variant for CPUs without AVX

* ggml : add x64 base ABI variant

.devops/cuda.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/intel.Dockerfile

@@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -1,4 +1,4 @@
-ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
+ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10
 
 FROM ascendai/cann:$ASCEND_VERSION AS build
 
@@ -6,7 +6,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN yum install -y gcc g++ cmake make
+RUN yum install -y gcc g++ cmake make libcurl-devel
 ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
 ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
 ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}

.devops/musa.Dockerfile

@@ -35,7 +35,7 @@ COPY . .
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/rocm.Dockerfile

@@ -17,8 +17,8 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # gfx906 is deprecated
 #check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
 
-#ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
-ARG ROCM_DOCKER_ARCH=gfx1100
+ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
+#ARG ROCM_DOCKER_ARCH=gfx1100
 
 # Set nvcc architectured
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .
 
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -16,7 +16,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.github/workflows/build.yml

@@ -601,8 +601,9 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-  build-linux-cross:
-    uses: ./.github/workflows/build-linux-cross.yml
+# Disabled for now due to sporadic issue syncing.
+#  build-linux-cross:
+#    uses: ./.github/workflows/build-linux-cross.yml
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -1766,16 +1767,17 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
+        shell: bash -el {0}
     strategy:
       matrix:
+        arch: [x86, aarch64]
         cann:
-          - '8.0.rc3.beta1-910b-openeuler22.03-py3.10'
+          - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
         device:
           - 'ascend910b3'
         build:
           - 'Release'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     container: ascendai/cann:${{ matrix.cann }}
     steps:
       - name: Checkout
@@ -1784,7 +1786,7 @@ jobs:
       - name: Dependencies
         run: |
           yum update -y
-          yum install -y git gcc gcc-c++ make cmake
+          yum install -y git gcc gcc-c++ make cmake libcurl-devel
 
       - name: Build
         run: |

.github/workflows/docker.yml

@@ -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: false}
-          - { 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

Makefile

@@ -780,10 +780,6 @@ ifdef GGML_HIP
 
 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
-ifdef GGML_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # GGML_HIP_UMA
-
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas

README.md

@@ -9,13 +9,6 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-> [!IMPORTANT]
-> New `llama.cpp` package location: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pkgs/container/llama.cpp)
->
-> Update your container URLs to: `ghcr.io/ggml-org/llama.cpp`
->
-> More info: https://github.com/ggml-org/llama.cpp/discussions/11801
-
 ## Recent API changes
 
 - [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
@@ -23,6 +16,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
 - **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
 - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
@@ -104,6 +98,7 @@ 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)
@@ -247,6 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |
 
 ## Building the project
 
@@ -265,7 +261,9 @@ 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.
 

build-xcframework.sh

@@ -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

common/arg.cpp

@@ -163,6 +163,8 @@ struct common_hf_file_res {
 #   if !defined(PATH_MAX)
 #   define PATH_MAX MAX_PATH
 #   endif
+#elif defined(_AIX)
+#include <sys/limits.h>
 #else
 #include <sys/syslimits.h>
 #endif
@@ -226,12 +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
@@ -542,7 +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);
@@ -657,13 +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
@@ -2725,7 +2726,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(

common/chat.cpp

@@ -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);
     }
 

common/common.cpp

@@ -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__) || defined(_AIX)
         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) {

common/common.h

@@ -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
 //

convert_hf_to_gguf_update.py

@@ -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", },
 ]
 
 

convert_lora_to_gguf.py

@@ -24,7 +24,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 # reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import LazyTorchTensor, Model
+from convert_hf_to_gguf import LazyTorchTensor, ModelBase
 
 logger = logging.getLogger("lora-to-gguf")
 
@@ -340,11 +340,11 @@ if __name__ == '__main__':
             sys.exit(1)
     else:
         logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = Model.load_hparams(dir_base_model)
+        hparams = ModelBase.load_hparams(dir_base_model)
 
     with torch.inference_mode():
         try:
-            model_class = Model.from_model_architecture(hparams["architectures"][0])
+            model_class = ModelBase.from_model_architecture(hparams["architectures"][0])
         except NotImplementedError:
             logger.error(f"Model {hparams['architectures'][0]} is not supported")
             sys.exit(1)

docs/backend/SYCL.md

@@ -425,13 +425,13 @@ Examples:
 - Use device 0:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
 *Notes:*
@@ -697,13 +697,13 @@ Examples:
 - Use device 0:
 
 ```
-build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```
-build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
 
 

docs/build.md

@@ -259,8 +259,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
       cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
-  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
-  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
@@ -296,6 +294,10 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
 
+### Unified Memory
+
+On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
 ## Vulkan
 
 **Windows**

examples/llava/CMakeLists.txt

@@ -1,3 +1,5 @@
+# llava (legacy)
+
 add_library(llava OBJECT
             llava.cpp
             llava.h
@@ -22,27 +24,46 @@ if (BUILD_SHARED_LIBS)
     install(TARGETS llava_shared LIBRARY)
 endif()
 
+# mtmd
+
+add_library(mtmd OBJECT
+            mtmd.cpp
+            mtmd.h
+            clip.cpp
+            clip.h
+            clip-impl.h
+            )
+
+target_link_libraries(mtmd PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(mtmd PUBLIC .)
+target_include_directories(mtmd PRIVATE ../..)
+target_include_directories(mtmd PRIVATE ../../common) # for stb_image.h
+
+target_compile_features(mtmd PRIVATE cxx_std_17)
+
+add_library(mtmd_static STATIC $<TARGET_OBJECTS:mtmd>)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(mtmd PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(mtmd_shared SHARED $<TARGET_OBJECTS:mtmd>)
+    target_link_libraries(mtmd_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS mtmd_shared LIBRARY)
+endif()
+
 if (NOT MSVC)
     target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
+    target_compile_options(mtmd PRIVATE -Wno-cast-qual) # stb_image.h
 endif()
 
 if(TARGET BUILD_INFO)
     add_dependencies(llava BUILD_INFO)
+    add_dependencies(mtmd BUILD_INFO)
 endif()
 
-set(TARGET llama-llava-cli)
-add_executable(${TARGET} llava-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-set(TARGET llama-minicpmv-cli)
-add_executable(${TARGET} minicpmv-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+add_executable(llama-llava-cli    deprecation-warning.cpp)
+add_executable(llama-gemma3-cli   deprecation-warning.cpp)
+add_executable(llama-minicpmv-cli deprecation-warning.cpp)
 
 set(TARGET llama-qwen2vl-cli)
 add_executable(${TARGET} qwen2vl-cli.cpp)
@@ -51,11 +72,11 @@ install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
-set(TARGET llama-gemma3-cli)
-add_executable(${TARGET} gemma3-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
+set(TARGET llama-mtmd-cli)
+add_executable(${TARGET} mtmd-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-llava-clip-quantize-cli)

examples/llava/clip-impl.h

@@ -1,5 +1,8 @@
 #include "ggml.h"
 #include "gguf.h"
+#include "clip.h"
+
+#include "clip.h"
 
 #include <climits>
 #include <cstdarg>
@@ -7,6 +10,7 @@
 #include <map>
 #include <sstream>
 #include <vector>
+#include <memory>
 
 // Internal header for clip.cpp
 
@@ -46,7 +50,6 @@
 // tensor name constants
 //
 
-#define TN_TOKEN_EMBD      "%s.token_embd.weight"
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
@@ -62,8 +65,6 @@
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
-#define TN_TEXT_PROJ       "text_projection.weight"
-#define TN_VIS_PROJ        "visual_projection.weight"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
@@ -120,6 +121,23 @@ static projector_type clip_projector_type_from_string(const std::string & str) {
     return PROJECTOR_TYPE_UNKNOWN;
 }
 
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
 //
 // logging
 //
@@ -178,6 +196,36 @@ static void clip_log_internal(enum ggml_log_level level, const char * format, ..
 #define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
 #define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
 
+//
+// 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_u8_batch {
+    std::vector<clip_image_u8_ptr> entries;
+};
+
+struct clip_image_f32_batch {
+    std::vector<clip_image_f32_ptr> entries;
+};
+
 //
 // common utils
 //
@@ -214,6 +262,20 @@ static void string_replace_all(std::string & s, const std::string & search, cons
     s = std::move(builder);
 }
 
+// split string by a `std::string delim` instead of `char delim`
+static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t pos = 0;
+    std::string token;
+    while ((pos = s.find(delimiter)) != std::string::npos) {
+        token = s.substr(0, pos);
+        tokens.push_back(token);
+        s.erase(0, pos + delimiter.length());
+    }
+    tokens.push_back(s);
+    return tokens;
+}
+
 //
 // gguf utils
 //
@@ -271,3 +333,9 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
     }
 }
+
+//
+// API used internally with mtmd
+//
+
+projector_type clip_get_projector_type(const struct clip_ctx * ctx);

examples/llava/clip.h

@@ -30,19 +30,13 @@ 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_f32;
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;
 
 struct clip_context_params {
     bool use_gpu;
-    ggml_log_level verbosity;
+    enum ggml_log_level verbosity;
 };
 
 // deprecated, use clip_init
@@ -55,9 +49,9 @@ 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);
@@ -73,15 +67,26 @@ 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);
+
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 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 struct 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
@@ -106,6 +111,8 @@ CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
 CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
 

examples/llava/deprecation-warning.cpp

@@ -0,0 +1,22 @@
+#include <cstdio>
+#include <string>
+
+int main(int argc, char** argv) {
+    std::string filename = "main";
+    if (argc >= 1) {
+        filename = argv[0];
+    }
+
+    // Get only the program name from the full path
+    size_t pos = filename.find_last_of("/\\");
+    if (pos != std::string::npos) {
+        filename = filename.substr(pos+1);
+    }
+
+    fprintf(stdout, "\n");
+    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
+    fprintf(stdout, "Please use 'llama-mtmd-cli' instead.\n");
+    fprintf(stdout, "\n");
+
+    return EXIT_FAILURE;
+}

examples/llava/llava-cli.cpp

@@ -1,332 +0,0 @@
-#include "arg.h"
-#include "base64.hpp"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-    return true;
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
-static const char* IMG_BASE64_TAG_END = "\">";
-
-static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
-    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
-    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
-}
-
-static bool prompt_contains_image(const std::string& prompt) {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    return (begin != std::string::npos);
-}
-
-// replaces the base64 image tag in the prompt with `replacement`
-static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
-    size_t img_base64_str_start, img_base64_str_end;
-    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
-    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
-        return NULL;
-    }
-
-    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
-    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
-    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
-
-    auto required_bytes = base64::required_encode_size(base64_str.size());
-    auto img_bytes = std::vector<unsigned char>(required_bytes);
-    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
-
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
-    if (!embed) {
-        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
-        return NULL;
-    }
-
-    return embed;
-}
-
-static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    if (begin == std::string::npos || end == std::string::npos) {
-        return prompt;
-    }
-    auto pre = prompt.substr(0, begin);
-    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
-    return pre + replacement + post;
-}
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void print_usage(int, char ** argv) {
-    LOG("\n example usage:\n");
-    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
-
-    // load and preprocess the image
-    llava_image_embed * embed = NULL;
-    auto prompt = params->prompt;
-    if (prompt_contains_image(prompt)) {
-        if (!params->image.empty()) {
-            LOG_INF("using base64 encoded image instead of command line image path\n");
-        }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
-        if (!embed) {
-            LOG_ERR("%s: can't load image from prompt\n", __func__);
-            return NULL;
-        }
-        params->prompt = remove_image_from_prompt(prompt);
-    } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
-        if (!embed) {
-            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
-            return NULL;
-        }
-    }
-
-    return embed;
-}
-
-static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
-    int n_past = 0;
-
-    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
-
-    std::string system_prompt, user_prompt;
-    size_t image_pos = prompt.find("<image>");
-    if (image_pos != std::string::npos) {
-        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
-        system_prompt = prompt.substr(0, image_pos);
-        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    } else {
-        // llava-1.5 native mode
-        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
-        user_prompt = prompt + "\nASSISTANT:";
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
-    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-
-    // generate the response
-
-    LOG("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    if (!smpl) {
-        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
-        exit(1);
-    }
-
-    std::string response = "";
-    for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-        response += tmp;
-        if (strcmp(tmp, "</s>") == 0) break;
-        if (strstr(tmp, "###")) break; // Yi-VL behavior
-        LOG("%s", tmp);
-        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
-        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
-        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
-
-        fflush(stdout);
-    }
-
-    common_sampler_free(smpl);
-    LOG("\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->ctx_clip = ctx_clip;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        print_usage(argc, argv);
-        return 1;
-    }
-
-    auto * model = llava_init(&params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
-        return 1;
-    }
-
-    if (prompt_contains_image(params.prompt)) {
-        auto * ctx_llava = llava_init_context(&params, model);
-
-        auto * image_embed = load_image(ctx_llava, &params, "");
-
-        // process the prompt
-        process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-        llama_perf_context_print(ctx_llava->ctx_llama);
-        llava_image_embed_free(image_embed);
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    } else {
-        for (auto & image : params.image) {
-            auto * ctx_llava = llava_init_context(&params, model);
-
-            auto * image_embed = load_image(ctx_llava, &params, image);
-            if (!image_embed) {
-                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
-                return 1;
-            }
-
-            // process the prompt
-            process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-            llama_perf_context_print(ctx_llava->ctx_llama);
-            llava_image_embed_free(image_embed);
-            ctx_llava->model = NULL;
-            llava_free(ctx_llava);
-        }
-    }
-
-    llama_model_free(model);
-
-    return 0;
-}

examples/llava/llava.cpp

@@ -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);
 

examples/llava/minicpmv-cli.cpp

@@ -1,354 +0,0 @@
-#include "arg.h"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <algorithm>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-#include <iostream> // TODO: remove me
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    if (params->n_ctx < 2048) {
-        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
-        LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
-        ctx_params.n_ctx = 2048;
-    } else {
-        ctx_params.n_ctx = params->n_ctx;
-    }
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-    struct clip_context_params clip_params = {
-        /* use_gpu */   params->n_gpu_layers != 0,
-        /* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
-    };
-    auto * ctx_clip = clip_init(clip_path, clip_params);
-    return ctx_clip;
-}
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-}
-
-static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
-    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
-    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
-
-    auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
-    slice_embed->embed = image_embed;
-    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
-    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
-    llava_image_embed_free(slice_embed);
-}
-
-static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
-    std::string system_prompt;
-    int idx = 0;
-    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (has_minicpmv_projector == 2) {
-        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
-    }
-    else if (has_minicpmv_projector == 3) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    else if (has_minicpmv_projector == 4) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
-    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-    if (num_image_embeds > 1) {
-        if (has_minicpmv_projector == 2) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-            eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-        }
-        else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-        }
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
-    auto * ctx_clip = clip_init_context(params);
-    auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
-    if (!embeds) {
-        LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
-        return NULL;
-    }
-
-    // process the prompt
-    if (params->prompt.empty() && params->interactive == false) {
-        LOG_ERR("prompt should be given or interactive mode should be on");
-        return NULL;
-    }
-
-    auto * model = llava_init(params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
-        return NULL;
-    }
-    const int64_t t_llava_init_start_us = ggml_time_us();
-    auto * ctx_llava = llava_init_context(params, model);
-    ctx_llava->ctx_clip = ctx_clip;
-    const int64_t t_llava_init_end_us = ggml_time_us();
-    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
-    LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
-
-    const int64_t t_process_image_start_us = ggml_time_us();
-    process_image(ctx_llava, embeds, params, n_past);
-    const int64_t t_process_image_end_us = ggml_time_us();
-    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
-    LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
-
-    llava_image_embed_free(embeds);
-    return ctx_llava;
-}
-
-static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
-    std::string user_prompt = prompt;
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (!is_first) {
-        if (has_minicpmv_projector == 2) {
-            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 3) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 4) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-    if (has_minicpmv_projector == 2) {
-        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 3) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 4) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-
-    // generate the response
-
-    LOG_INF("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    return smpl;
-}
-
-static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
-
-    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-    return tmp;
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty())) {
-        show_additional_info(argc, argv);
-        return 1;
-    }
-
-    for (auto & image : params.image) {
-        int n_past = 0;
-        auto * ctx_llava = minicpmv_init(&params, image, n_past);
-
-        if (!params.prompt.empty()) {
-            LOG("<user>%s\n", params.prompt.c_str());
-            LOG("<assistant>");
-            auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
-            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-            std::string response;
-            bool have_tmp = false;
-            for (int i = 0; i < max_tgt_len; i++) {
-                const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                response += tmp;
-                if (strcmp(tmp, "</s>") == 0){
-                    if (!have_tmp) {
-                        continue;
-                    }
-                    break;
-                }
-                if (strstr(tmp, "###")) break; // Yi-VL behavior
-                have_tmp = true;
-                printf("%s", tmp);
-                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-
-                fflush(stdout);
-            }
-            common_sampler_free(smpl);
-        }else {
-            while (true) {
-                LOG("<user>");
-                std::string prompt;
-                std::getline(std::cin, prompt);
-                LOG("<assistant>");
-                auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
-                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-                std::string response;
-                for (int i = 0; i < max_tgt_len; i++) {
-                    const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                    response += tmp;
-                    if (strcmp(tmp, "</s>") == 0) break;
-                    printf("%s", tmp);// mistral llava-1.6
-                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-                    fflush(stdout);
-                }
-                common_sampler_free(smpl);
-            }
-        }
-        printf("\n");
-        llama_perf_context_print(ctx_llava->ctx_llama);
-
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    }
-
-    return 0;
-}

examples/llava/mtmd-cli.cpp

@@ -2,11 +2,11 @@
 #include "log.h"
 #include "common.h"
 #include "sampling.h"
-#include "clip.h"
-#include "stb_image.h"
 #include "llama.h"
 #include "ggml.h"
 #include "console.h"
+#include "chat.h"
+#include "mtmd.h"
 
 #include <vector>
 #include <limits.h>
@@ -28,15 +28,16 @@ static bool g_is_generating = false;
 
 /**
  * Please note that this is NOT a production-ready stuff.
- * It is a playground for trying Gemma 3 vision capabilities.
+ * It is a playground for trying multimodal support in llama.cpp.
  * For contributors: please keep this code simple and easy to understand.
  */
 
 static void show_additional_info(int /*argc*/, char ** argv) {
     LOG(
-        "Experimental CLI for using Gemma 3 vision model\n\n"
+        "Experimental CLI for multimodal\n\n"
         "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
         "  -m and --mmproj are required\n"
+        "  -hf user/repo can replace both -m and --mmproj in most cases\n"
         "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
         argv[0]
     );
@@ -56,38 +57,77 @@ static void sigint_handler(int signo) {
 }
 #endif
 
-struct gemma3_context {
-    struct clip_ctx    * ctx_clip = NULL;
-    common_init_result   llama_init;
+struct mtmd_cli_context {
+    mtmd_context_ptr ctx_vision;
+    common_init_result llama_init;
 
     llama_model       * model;
     llama_context     * lctx;
     const llama_vocab * vocab;
     llama_batch         batch;
+    int                 n_batch;
+
+    // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
+    // so here we don't need to keep track of chat history
+    common_chat_templates_ptr tmpls;
+
+    // support for legacy templates (models not having EOT token)
+    llama_tokens antiprompt_tokens;
 
     int n_threads    = 1;
     llama_pos n_past = 0;
 
-    gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) {
+    mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
         model = llama_init.model.get();
         lctx = llama_init.context.get();
         vocab = llama_model_get_vocab(model);
         n_threads = params.cpuparams.n_threads;
         batch = llama_batch_init(params.n_batch, 0, 1);
-        init_clip_model(params);
+        n_batch = params.n_batch;
+
+        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
+            LOG_ERR("Model does not have chat template.\n");
+            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
+            LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            exit(1);
+        }
+
+        tmpls = common_chat_templates_init(model, params.chat_template);
+        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja).c_str());
+
+        init_vision_context(params);
+
+        // load antiprompt tokens for legacy templates
+        if (params.chat_template == "vicuna") {
+            antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
+        } else if (params.chat_template == "deepseek") {
+            antiprompt_tokens = common_tokenize(lctx, "###", false, true);
+        }
     }
 
-    void init_clip_model(common_params & params) {
+    void init_vision_context(common_params & params) {
         const char * clip_path = params.mmproj.path.c_str();
-        ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
-        if (!ctx_clip) {
-            LOG_ERR("Failed to load CLIP model from %s\n", clip_path);
+        ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
+            /* use_gpu */   true,
+            /* timings */   true,
+            /* n_threads */ params.cpuparams.n_threads,
+            /* verbosity */ GGML_LOG_LEVEL_INFO,
+        }));
+        if (!ctx_vision.get()) {
+            LOG_ERR("Failed to load vision model from %s\n", clip_path);
             exit(1);
         }
     }
 
-    ~gemma3_context() {
-        clip_free(ctx_clip);
+    bool check_antiprompt(const llama_tokens & generated_tokens) {
+        if (antiprompt_tokens.empty() || generated_tokens.size() < antiprompt_tokens.size()) {
+            return false;
+        }
+        return std::equal(
+            generated_tokens.end() - antiprompt_tokens.size(),
+            generated_tokens.end(),
+            antiprompt_tokens.begin()
+        );
     }
 };
 
@@ -124,78 +164,8 @@ struct decode_embd_batch {
     }
 };
 
-static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
-    llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
-    common_batch_clear(ctx.batch);
-    for (llama_token & t : tokens) {
-        common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false);
-    }
-    if (logits_last) {
-        ctx.batch.logits[ctx.batch.n_tokens - 1] = true;
-    }
-    // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
-    if (llama_decode(ctx.lctx, ctx.batch)) {
-        LOG_ERR("Failed to decode text\n");
-        return 1;
-    }
-    return 0;
-}
-
-static int eval_image(gemma3_context & ctx, std::string & fname) {
-    std::vector<float> image_embd_v;
-    int n_embd = llama_model_n_embd(ctx.model);
-    int n_tokens = 256;
-    image_embd_v.resize(n_tokens * n_embd);
-
-    bool ok;
-    struct clip_image_u8 * img_u8 = clip_image_u8_init();
-    ok = clip_image_load_from_file(fname.c_str(), img_u8);
-    if (!ok) {
-        LOG_ERR("Unable to load image %s\n", fname.c_str());
-        clip_image_u8_free(img_u8);
-        return 2; // non-fatal error
-    }
-
-    clip_image_f32_batch batch_f32;
-    ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32);
-    if (!ok) {
-        LOG_ERR("Unable to preprocess image\n");
-        clip_image_f32_batch_free(&batch_f32);
-        clip_image_u8_free(img_u8);
-        return 1;
-    }
-
-    int64_t t0 = ggml_time_ms();
-    LOG("Encoding image %s\n", fname.c_str());
-    ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data());
-    if (!ok) {
-        LOG_ERR("Unable to encode image\n");
-        clip_image_f32_batch_free(&batch_f32);
-        clip_image_u8_free(img_u8);
-        return 1;
-    }
-    LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
-
-    clip_image_f32_batch_free(&batch_f32);
-    clip_image_u8_free(img_u8);
-
-    // decode image embeddings
-    int64_t t1 = ggml_time_ms();
-    eval_text(ctx, "<start_of_image>");
-    llama_set_causal_attn(ctx.lctx, false);
-    decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0);
-    if (llama_decode(ctx.lctx, batch_img.batch)) {
-        LOG_ERR("failed to decode image\n");
-        return 1;
-    }
-    ctx.n_past += n_tokens;
-    llama_set_causal_attn(ctx.lctx, true);
-    eval_text(ctx, "<end_of_image>");
-    LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
-    return 0;
-}
-
-static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
+static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
+    llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
         if (i > n_predict || !g_is_generating) {
             printf("\n");
@@ -203,9 +173,10 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
         }
 
         llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        generated_tokens.push_back(token_id);
         common_sampler_accept(smpl, token_id, true);
 
-        if (llama_vocab_is_eog(ctx.vocab, token_id)) {
+        if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
             printf("\n");
             break; // end of generation
         }
@@ -224,6 +195,46 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
     return 0;
 }
 
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
+    std::vector<mtmd_bitmap> bitmaps;
+
+    common_chat_templates_inputs tmpl_inputs;
+    tmpl_inputs.messages = {msg};
+    tmpl_inputs.add_generation_prompt = true;
+    tmpl_inputs.use_jinja = false; // jinja is buggy here
+    auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
+    LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
+
+    for (auto & fname : images_fname) {
+        mtmd_bitmap bitmap;
+        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
+            LOG_ERR("Unable to load image %s\n", fname.c_str());
+            return 2; // image not found
+        }
+        bitmaps.push_back(std::move(bitmap));
+    }
+
+    mtmd_input_text text;
+    text.text          = formatted_chat.prompt;
+    text.add_special   = add_bos;
+    text.parse_special = true;
+    mtmd_input_chunks chunks;
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
+    if (res != 0) {
+        LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
+        return 1;
+    }
+
+    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
+        LOG_ERR("Unable to eval prompt\n");
+        return 1;
+    }
+
+    ctx.n_past += mtmd_helper_get_n_tokens(chunks);
+
+    return 0;
+}
+
 int main(int argc, char ** argv) {
     ggml_time_init();
 
@@ -241,7 +252,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    gemma3_context ctx(params);
+    mtmd_cli_context ctx(params);
     printf("%s: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
@@ -265,21 +276,15 @@ int main(int argc, char ** argv) {
 #endif
     }
 
-    if (eval_text(ctx, "<bos>")) {
-        return 1;
-    }
-
     if (is_single_turn) {
         g_is_generating = true;
-        if (eval_text(ctx, "<start_of_turn>user\n")) {
-            return 1;
+        if (params.prompt.find("<__image__>") == std::string::npos) {
+            params.prompt += " <__image__>";
         }
-        for (auto & fname : params.image) {
-            if (eval_image(ctx, fname)) {
-                return 1;
-            }
-        }
-        if (eval_text(ctx, params.prompt + "<end_of_turn><start_of_turn>model\n", true)) {
+        common_chat_msg msg;
+        msg.role = "user";
+        msg.content = params.prompt;
+        if (eval_message(ctx, msg, params.image, true)) {
             return 1;
         }
         if (generate_response(ctx, smpl, n_predict)) {
@@ -293,9 +298,9 @@ int main(int argc, char ** argv) {
         LOG("\n   /quit or /exit   exit the program");
         LOG("\n");
 
-        if (eval_text(ctx, "<start_of_turn>user\n")) {
-            return 1;
-        }
+        bool is_first_msg = true;
+        std::vector<std::string> images_fname;
+        std::string content;
 
         while (true) {
             g_is_generating = false;
@@ -320,26 +325,33 @@ int main(int argc, char ** argv) {
             g_is_generating = true;
             if (line.find("/image") == 0) {
                 std::string image = line.substr(7);
-                int res = eval_image(ctx, image);
-                if (res == 2) {
-                    continue; // image not found
-                }
-                if (res) {
-                    return 1;
-                }
+                images_fname.push_back(string_strip(image));
+                content += "<__image__>";
+                continue;
+            } else {
+                content += line;
+            }
+            common_chat_msg msg;
+            msg.role = "user";
+            msg.content = content;
+            int ret = eval_message(ctx, msg, images_fname, is_first_msg);
+            if (ret == 2) {
+                // non-fatal error
+                images_fname.clear();
+                content.clear();
                 continue;
             }
-            if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
+            if (ret) {
                 return 1;
             }
             if (generate_response(ctx, smpl, n_predict)) {
                 return 1;
             }
-            if (eval_text(ctx, "<end_of_turn><start_of_turn>user\n")) {
-                return 1;
-            }
+            images_fname.clear();
+            content.clear();
+            is_first_msg = false;
         }
     }
-
+    llama_perf_context_print(ctx.lctx);
     return 0;
 }

examples/llava/mtmd.cpp

@@ -0,0 +1,586 @@
+#include "clip.h"
+#include "clip-impl.h"
+#include "mtmd.h"
+
+#include "llama.h"
+
+#include <algorithm>
+#include <cerrno>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
+#include <vector>
+
+// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
+// models not having it (llava-1.6) will process embeddings without any special tokens in-between
+enum mtmd_slice_tmpl {
+    MTMD_SLICE_TMPL_NONE,
+    MTMD_SLICE_TMPL_MINICPMV_2_5,
+    MTMD_SLICE_TMPL_MINICPMV_2_6,
+    // TODO @ngxson : add support for idefics (SmolVLM)
+};
+
+struct mtmd_context {
+    struct clip_ctx * ctx_clip;
+    const struct llama_model * text_model;
+    std::vector<float> image_embd_v; // image embedding vector
+
+    bool print_timings;
+    int n_threads;
+    std::string image_marker;
+
+    // for minicpmv, we need special tokens in-between slices
+    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
+    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
+    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
+    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
+
+    // TODO @ngxson : add timings
+
+    mtmd_context(const char * mmproj_fname,
+                   const llama_model * text_model,
+                   const mtmd_context_params & ctx_params) :
+        print_timings(ctx_params.print_timings),
+        n_threads    (ctx_params.n_threads),
+        image_marker (ctx_params.image_marker)
+    {
+        clip_context_params ctx_clip_params;
+        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
+        ctx_clip_params.verbosity = ctx_params.verbosity;
+        ctx_clip = clip_init(mmproj_fname, ctx_clip_params);
+        if (!ctx_clip) {
+            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
+        }
+        this->text_model = text_model;
+
+        GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
+
+        int minicpmv_version = clip_is_minicpmv(ctx_clip);
+        if (minicpmv_version == 2) {
+            // minicpmv 2.5 format:
+            // <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_5;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_slices_start  = lookup_token("<slice>");
+            tok_slices_end    = lookup_token("</slice>");
+            tok_sli_img_start = tok_ov_img_start;
+            tok_sli_img_end   = tok_ov_img_end;
+            tok_row_end       = lookup_token("\n");
+
+        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
+            // minicpmv 2.6 format:
+            // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_sli_img_start = lookup_token("<slice>");
+            tok_sli_img_end   = lookup_token("</slice>");
+            tok_row_end       = lookup_token("\n");
+
+        } else if (minicpmv_version != 0) {
+            GGML_ASSERT(false && "unsupported minicpmv version");
+        }
+    }
+
+    ~mtmd_context() {
+        clip_free(ctx_clip);
+    }
+
+private:
+    llama_token lookup_token(const std::string & token_text) {
+        const llama_vocab * vocab = llama_model_get_vocab(text_model);
+        const int n_vocab = llama_vocab_n_tokens(vocab);
+        for (int i = 0; i < n_vocab; i++) {
+            if (token_to_piece(vocab, i, true) == token_text) {
+                return i;
+            }
+        }
+        return LLAMA_TOKEN_NULL;
+    }
+
+    std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
+        std::string piece;
+        piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
+        const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+        if (n_chars < 0) {
+            piece.resize(-n_chars);
+            int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+            GGML_ASSERT(check == -n_chars);
+        } else {
+            piece.resize(n_chars);
+        }
+        return piece;
+    }
+};
+
+struct mtmd_image_tokens_data {
+    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 batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
+};
+
+mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+        const struct llama_model * text_model,
+        const struct mtmd_context_params ctx_params) {
+    try {
+        return new mtmd_context(mmproj_fname, text_model, ctx_params);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return nullptr;
+    }
+}
+
+void mtmd_free(mtmd_context * ctx) {
+    if (ctx) {
+        delete ctx;
+    }
+}
+
+// copied from common_tokenize
+static std::vector<llama_token> mtmd_tokenize_text_internal(
+    const struct llama_vocab * vocab,
+           const std::string & text,
+                        bool   add_special,
+                        bool   parse_special) {
+    // upper limit for the number of tokens
+    int n_tokens = text.length() + 2 * add_special;
+    std::vector<llama_token> result(n_tokens);
+    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+    return result;
+}
+
+int32_t mtmd_tokenize(mtmd_context * ctx,
+                        std::vector<mtmd_input_chunk> & output,
+                        const mtmd_input_text & text,
+                        const std::vector<mtmd_bitmap> & bitmaps) {
+    auto vocab = llama_model_get_vocab(ctx->text_model);
+
+    std::string prompt_modified(text.text);
+    std::string marker_modified(ctx->image_marker);
+    // a bit hacky here, but works for now
+    // for some models, we need to add prefix and suffix to the image embeddings
+    if (clip_is_gemma3(ctx->ctx_clip)) {
+        // gemma 3
+        // <start_of_image> ... (image embeddings) ... <end_of_image>
+        marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+    }
+
+    // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
+    // for glm-edge, we don't need to add because the tokens are already in the returned embeddings
+
+    // TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
+
+    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
+    output.clear();
+    output.reserve(parts.size());
+
+    size_t i_img = 0;
+
+    // utility for adding raw tokens
+    auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
+        mtmd_input_chunk chunk{
+            MTMD_INPUT_CHUNK_TYPE_TEXT,
+            std::move(tokens),
+            {},
+        };
+        output.emplace_back(std::move(chunk));
+    };
+
+    // utility for splitting batch of multiple images into chunks of batch having single images
+    auto split_batch_to_chunk = [&ctx](clip_image_f32_batch && batch_f32, const std::string & id) {
+        std::vector<mtmd_input_chunk> chunks;
+
+        for (auto & entry : batch_f32.entries) {
+            mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+            image_tokens->nx = clip_n_patches(ctx->ctx_clip);
+            image_tokens->ny = 1;
+            image_tokens->batch_f32.entries.push_back(std::move(entry));
+            image_tokens->id = id;
+
+            mtmd_input_chunk chunk{
+                MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                {},
+                std::move(image_tokens),
+            };
+            chunks.emplace_back(std::move(chunk));
+        }
+
+        return chunks;
+    };
+
+    for (const auto & part : parts) {
+        //printf("tokenizing part: %s\n", part.c_str());
+        bool add_bos = &parts.front() == &part;
+        auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
+        if (tokens.empty()) {
+            continue;
+        }
+        mtmd_input_chunk chunk{
+            MTMD_INPUT_CHUNK_TYPE_TEXT,
+            std::move(tokens),
+            {},
+        };
+        output.emplace_back(std::move(chunk));
+
+        if (&parts.back() != &part) {
+            // add image token to middle of 2 parts
+
+            if (i_img >= bitmaps.size()) {
+                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
+                return 1;
+            }
+
+            // convert mtmd_bitmap to clip_image_u8
+            clip_image_u8_ptr img_u8(clip_image_u8_init());
+            img_u8->nx = bitmaps[i_img].nx;
+            img_u8->ny = bitmaps[i_img].ny;
+            img_u8->buf.resize(bitmaps[i_img].data.size());
+            std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
+            clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
+
+            // preprocess image
+            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 2;
+            }
+
+            if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
+                // split batch into chunks of single images
+                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
+                GGML_ASSERT(chunks.size() > 0);
+
+                // add overview image
+                add_text_chunk({ctx->tok_ov_img_start});
+                output.emplace_back(std::move(chunks.front()));
+                chunks.erase(chunks.begin());
+                add_text_chunk({ctx->tok_ov_img_end});
+
+                // add slices
+                if (!chunks.empty()) {
+                    clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
+                    int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
+                    int n_row = (int)chunks.size() / n_col;
+                    GGML_ASSERT(n_row * n_col == (int)chunks.size());
+                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_slices_start});
+                    }
+                    for (int y = 0; y < n_row; y++) {
+                        for (int x = 0; x < n_col; x++) {
+                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
+                                add_text_chunk({ctx->tok_sli_img_start});
+                            }
+                            output.emplace_back(std::move(chunks[y * n_col + x]));
+                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
+                                add_text_chunk({ctx->tok_sli_img_end});
+                            }
+                        }
+                        if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
+                            add_text_chunk({ctx->tok_row_end});
+                        }
+                    }
+                    if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_slices_end});
+                    }
+                }
+
+            } else {
+                mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+                image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
+                image_tokens->ny = 1; // TODO
+                image_tokens->batch_f32 = std::move(batch_f32);
+                image_tokens->id = bitmaps[i_img].id; // optional
+
+                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
+                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
+                LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
+
+                if (clip_is_glm(ctx->ctx_clip)) {
+                    // glm-edge
+                    image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
+                }
+
+                mtmd_input_chunk chunk{
+                    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                    {},
+                    std::move(image_tokens),
+                };
+                output.emplace_back(std::move(chunk));
+            }
+
+            i_img++; // move to next image
+        }
+    }
+
+    return 0;
+}
+
+void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
+    if (image_tokens) {
+        delete image_tokens;
+    }
+}
+
+size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->n_tokens();
+}
+
+size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->nx;
+}
+
+size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->ny;
+}
+
+std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->id;
+}
+
+int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
+    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
+    bool ok = false;
+
+    // only effective for minicpmv and qwen2vl, other models will ignore load_image_size
+    {
+        clip_image_size slice_size{
+            image_tokens->batch_f32.entries[0]->nx,
+            image_tokens->batch_f32.entries[0]->ny};
+        clip_add_load_image_size(ctx->ctx_clip, &slice_size);
+    }
+
+    if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
+        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
+        const auto & entries = image_tokens->batch_f32.entries;
+        for (size_t i = 0; i < entries.size(); i++) {
+            int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
+            ok = clip_image_encode(
+                ctx->ctx_clip,
+                ctx->n_threads,
+                entries[i].get(),
+                ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
+        }
+    } else {
+        ok = clip_image_batch_encode(
+            ctx->ctx_clip,
+            ctx->n_threads,
+            &image_tokens->batch_f32,
+            ctx->image_embd_v.data());
+    }
+
+    return ok ? 0 : 1;
+}
+
+float * mtmd_get_output_embd(mtmd_context * ctx) {
+    return ctx->image_embd_v.data();
+}
+
+size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
+    size_t n_tokens = 0;
+    for (auto & chunk : chunks) {
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            n_tokens += chunk.tokens_text.size();
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            n_tokens += chunk.tokens_image->n_tokens();
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_tokens;
+}
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
+int32_t mtmd_helper_eval(mtmd_context * ctx,
+        llama_context * lctx,
+        mtmd_input_chunks & chunks,
+        llama_pos pos0,
+        llama_seq_id seq_id,
+        int32_t n_batch) {
+    int32_t ret;
+    llama_pos n_past = pos0;
+    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+
+    for (auto & chunk : chunks) {
+        bool is_last = &chunk == &chunks.back();
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            text_batch.n_tokens = chunk.tokens_text.size();
+            size_t i = 0;
+            while (i < chunk.tokens_text.size()) { // split into batches
+                for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
+                    text_batch.token   [i]    = chunk.tokens_text[i];
+                    text_batch.pos     [i]    = n_past++;
+                    text_batch.n_seq_id[i]    = 1;
+                    text_batch.seq_id  [i][0] = seq_id;
+                    text_batch.logits  [i]    = false;
+                }
+                if (is_last) {
+                    // always get logits for last input chunk
+                    text_batch.logits[text_batch.n_tokens - 1] = true;
+                }
+                ret = llama_decode(lctx, text_batch);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode text\n");
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
+            }
+
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
+            GGML_ASSERT(chunk.tokens_image != nullptr);
+            int64_t t0 = ggml_time_ms();
+            if (ctx->print_timings) {
+                LOG_INF("encoding image or slice...\n");
+            }
+            ret = mtmd_encode(ctx, chunk.tokens_image.get());
+            if (ret != 0) {
+                LOG_ERR("failed to encode image\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            if (ctx->print_timings) {
+                LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+            }
+
+            int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
+            int32_t i_batch = 0;
+            int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
+            float * embd = mtmd_get_output_embd(ctx);
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, false);
+                // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+            }
+
+            while (i_batch < n_img_batches) { // split into batches
+                int32_t pos_offset = i_batch*n_batch;
+                int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+                float * embd_batch = embd + pos_offset*n_mmproj_embd;
+                decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
+
+                printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+
+                int64_t t1 = ggml_time_ms();
+                ret = llama_decode(lctx, batch_img.batch);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode image\n");
+                    llama_set_causal_attn(lctx, true); // restore causal attn
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
+
+                if (ctx->print_timings) {
+                    LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
+                }
+
+                i_batch++;
+                n_past += n_tokens_batch;
+            }
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, true);
+            }
+
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+
+    llama_batch_free(text_batch);
+    return 0;
+}
+
+int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) {
+    clip_image_u8_ptr img_u8(clip_image_u8_init());
+    bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
+    if (!ok) {
+        LOG_ERR("Unable to load image from buffer\n");
+        return 1;
+    }
+    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
+    output.data.resize(output.nx * output.ny * 3);
+    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
+    return 0;
+}
+
+int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) {
+    clip_image_u8_ptr img_u8(clip_image_u8_init());
+    bool ok = clip_image_load_from_file(fname, img_u8.get());
+    if (!ok) {
+        LOG_ERR("Unable to load image %s\n", fname);
+        return 1;
+    }
+    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
+    output.data.resize(output.nx * output.ny * 3);
+    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
+    return 0;
+}
+
+bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
+    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
+    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return true;
+    }
+    return false;
+}
+
+void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
+    mtmd_image_tokens_free(val);
+}

examples/llava/mtmd.h

@@ -0,0 +1,161 @@
+#ifndef MTMD_H
+#define MTMD_H
+
+#include "ggml.h"
+#include "llama.h"
+#include "clip.h"
+
+#include <vector>
+#include <cinttypes>
+#include <memory>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define MTMD_API __declspec(dllexport)
+#        else
+#            define MTMD_API __declspec(dllimport)
+#        endif
+#    else
+#        define MTMD_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define MTMD_API
+#endif
+
+#ifdef __cplusplus
+
+enum mtmd_input_chunk_type {
+    MTMD_INPUT_CHUNK_TYPE_TEXT,
+    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+};
+
+struct mtmd_context;
+struct mtmd_image_tokens;
+
+// represents raw image data, layout is RGBRGBRGB...
+// length of data must be nx * ny * 3
+struct mtmd_bitmap {
+    uint32_t nx;
+    uint32_t ny;
+    std::vector<unsigned char> data;
+    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
+};
+
+struct mtmd_image_tokens_deleter {
+    void operator()(mtmd_image_tokens * val); // forward declaration
+};
+using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
+
+struct mtmd_input_chunk {
+    mtmd_input_chunk_type type;
+    std::vector<llama_token> tokens_text;
+    mtmd_image_tokens_ptr tokens_image;
+};
+
+using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
+
+struct mtmd_context_params {
+    bool use_gpu = true;
+    bool print_timings = true;
+    int n_threads = 4;
+    enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO;
+    const char * image_marker = "<__image__>";
+};
+
+struct mtmd_input_text {
+    std::string text;
+    bool add_special;
+    bool parse_special;
+};
+
+// initialize the mtmd context
+// return nullptr on failure
+MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+                                                const llama_model * text_model,
+                                                const mtmd_context_params ctx_params);
+
+MTMD_API void mtmd_free(mtmd_context * ctx);
+
+// tokenize an input text prompt and an image
+// the prompt must have the input image marker (default: "<__image__>") in it
+// the marker will be replaced with the image tokens
+// for example:
+//   "here is an image: <__image__>\ndescribe it in detail."
+//   this will gives 3 chunks:
+//   1. "here is an image: <start_of_image>"
+//   2. (image tokens)
+//   3. "<end_of_image>\ndescribe it in detail."
+// number of bitmaps must be equal to the number of image markers in the prompt
+// this function is thread-safe (shared ctx)
+// return values:
+//   0 on success
+//   1 on number of images not matching the number of markers
+//   2 on image preprocessing error
+MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
+                                std::vector<mtmd_input_chunk> & output,
+                                const mtmd_input_text & text,
+                                const std::vector<mtmd_bitmap> & bitmaps);
+
+// access mtmd_image_tokens
+MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
+MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
+MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
+
+// returns 0 on success
+MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
+                            const mtmd_image_tokens * image_tokens);
+
+// get output embeddings from the last encode pass
+MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
+
+// whether we need to set non-causal mask before llama_decode
+MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
+
+
+
+//
+// helper functions (can be implemented based on other functions)
+//
+
+// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
+MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
+
+// helper function that automatically:
+// 1. run llama_decode() on text chunks
+// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
+// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
+// otherwise, returns 0 on success
+MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
+                                llama_context * lctx,
+                                mtmd_input_chunks & chunks,
+                                llama_pos pos0,
+                                llama_seq_id seq_id,
+                                int32_t n_batch);
+
+// helper function to construct a mtmd_bitmap from a file
+// returns 0 on success
+// this function is thread-safe
+MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output);
+
+// helper function to construct a mtmd_bitmap from a buffer
+// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
+// returns 0 on success
+// this function is thread-safe
+MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output);
+
+// convenient unique_ptr wrappers
+struct mtmd_context_deleter {
+    void operator()(mtmd_context * val) { mtmd_free(val); }
+};
+using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
+
+#else
+
+static_assert(false && "C header is not yet supported by this library");
+
+#endif
+
+#endif

examples/llava/tests.sh

@@ -17,26 +17,30 @@ cd $PROJ_ROOT
 
 arr_bin=()
 arr_hf=()
+arr_tmpl=() # chat template
 
 add_test() {
     local bin=$1
     local hf=$2
+    local tmpl=${3:-""} # default to empty string if not provided
     arr_bin+=("$bin")
     arr_hf+=("$hf")
+    arr_tmpl+=("$tmpl")
 }
 
-add_test "llama-gemma3-cli"   "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
-add_test "llama-llava-cli"    "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
-add_test "llama-llava-cli"    "second-state/Llava-v1.5-7B-GGUF:Q2_K"
-add_test "llama-llava-cli"    "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"
-add_test "llama-llava-cli"    "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
-add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
+add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
+add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
+add_test "llama-mtmd-cli"  "second-state/Llava-v1.5-7B-GGUF:Q2_K"            "vicuna"
+add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"           "vicuna"
+add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
+add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
+add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
 add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
 
+# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"  # this model has broken chat template, not usable
+
 ###############
 
 cmake --build build -j --target "${arr_bin[@]}"
@@ -46,12 +50,20 @@ arr_res=()
 for i in "${!arr_bin[@]}"; do
     bin="${arr_bin[$i]}"
     hf="${arr_hf[$i]}"
+    tmpl="${arr_tmpl[$i]}"
 
     echo "Running test with binary: $bin and HF model: $hf"
     echo ""
     echo ""
 
-    output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty)
+    output=$(\
+        "$PROJ_ROOT/build/bin/$bin" \
+        -hf "$hf" \
+        --image $SCRIPT_DIR/test-1.jpeg \
+        -p "what is the publisher name of the newspaper?" \
+        --temp 0 -n 128 \
+        ${tmpl:+--chat-template "$tmpl"} \
+        2>&1 | tee /dev/tty)
 
     echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
 

examples/main/main.cpp

@@ -865,9 +865,22 @@ int main(int argc, char ** argv) {
                 console::set_display(console::reset);
                 display = true;
 
-                // Add tokens to embd only if the input buffer is non-empty
-                // Entering a empty line lets the user pass control back
-                if (buffer.length() > 1) {
+                if (buffer.empty()) { // Ctrl+D on empty line exits
+                    LOG("EOF by user\n");
+                    break;
+                }
+
+                if (buffer.back() == '\n') {
+                    // Implement #587:
+                    // If the user wants the text to end in a newline,
+                    // this should be accomplished by explicitly adding a newline by using \ followed by return,
+                    // then returning control by pressing return again.
+                    buffer.pop_back();
+                }
+
+                if (buffer.empty()) { // Enter key on empty line lets the user pass control back
+                    LOG_DBG("empty line, passing control back\n");
+                } else { // Add tokens to embd only if the input buffer is non-empty
                     // append input suffix if any
                     if (!params.input_suffix.empty() && !params.conversation_mode) {
                         LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
@@ -915,8 +928,6 @@ int main(int argc, char ** argv) {
 
                     n_remain -= line_inp.size();
                     LOG_DBG("n_remain: %d\n", n_remain);
-                } else {
-                    LOG_DBG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again

examples/quantize/quantize.cpp

@@ -9,6 +9,7 @@
 #include <fstream>
 #include <cmath>
 #include <cctype>
+#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -16,7 +17,7 @@ struct quant_option {
     std::string desc;
 };
 
-static const std::vector<struct quant_option> QUANT_OPTIONS = {
+static const std::vector<quant_option> QUANT_OPTIONS = {
     { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
     { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
     { "Q5_0",     LLAMA_FTYPE_MOSTLY_Q5_0,     " 5.21G, +0.1316 ppl @ Llama-3-8B",  },
@@ -105,7 +106,8 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
+    printf("       [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@@ -114,6 +116,8 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
+    printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@@ -244,6 +248,107 @@ static ggml_type parse_ggml_type(const char * arg) {
     return GGML_TYPE_COUNT;
 }
 
+// Allowed tensors for arbitrary quantization with --tensor-type option
+static const std::vector<std::string> ALLOWED_TENSOR_TYPE = {
+    "attn_k",
+    "attn_kv_a_mqa",
+    "attn_kv_b",
+    "attn_o",
+    "attn_output",
+    "attn_q",
+    "attn_q_a",
+    "attn_q_b",
+    "attn_qkv",
+    "attn_v",
+    "channel_mix_key",
+    "channel_mix_receptance",
+    "channel_mix_value",
+    "cls",
+    "cls.output",
+    "cross_attn_k",
+    "cross_attn_o",
+    "cross_attn_q",
+    "cross_attn_v",
+    "ffn_act",
+    "ffn_down",
+    "ffn_down_exps",
+    "ffn_down_shexp",
+    "ffn_gate",
+    "ffn_gate_exps",
+    "ffn_gate_shexp",
+    "ffn_up",
+    "ffn_up_exps",
+    "ffn_up_shexp",
+    "ssm_in",
+    "ssm_out",
+    "time_mix_gate",
+    "time_mix_key",
+    "time_mix_output",
+    "time_mix_receptance",
+    "time_mix_value",
+};
+
+// changes to this struct must be replicated in llama-quant.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
+static bool parse_tensor_type(const char * data, std::vector<tensor_quantization> & tensor_type) {
+    const char * sep = strchr(data, '=');
+    if (sep == nullptr) {
+        printf("\n%s: malformed tensor type '%s'\n\n", __func__, data);
+        return false;
+    }
+
+    const size_t tn_len = sep - data;
+    if (tn_len == 0) {
+        printf("\n%s: missing tensor name\n\n", __func__);
+        return false;
+    }
+
+    if (const size_t qt_len = strlen(sep); qt_len == 1) {
+        printf("\n%s: missing quantization type\n\n", __func__);
+        return false;
+    }
+
+    std::string tn(data, tn_len);
+    std::transform(tn.begin(), tn.end(), tn.begin(), tolower);
+    sep++;
+    const std::string qt(sep);
+
+    bool found = false;
+    for (const auto & allowed : ALLOWED_TENSOR_TYPE) {
+        std::string tensor;
+        tensor = tn.rfind('.') != std::string::npos ? tn.substr(tn.rfind('.') + 1) : tn;
+        // handle special case of cls.output
+        std::string cls_output = "cls.output";
+        if (tn.find(cls_output) != std::string::npos) {
+            tensor = "cls.output";
+        }
+        // check if an allowed tensor exists and it's at the end of the kv string
+        if (tensor == allowed) {
+            found = true;
+            break;
+        }
+    }
+    if (!found) {
+        printf("\n%s: invalid tensor name '%s'\n\n", __func__, tn.c_str());
+        return false;
+    }
+
+    if (parse_ggml_type(qt.c_str()) == GGML_TYPE_COUNT) {
+        printf("\n%s: invalid quantization type '%s'\n\n", __func__, qt.c_str());
+        return false;
+    }
+
+    tensor_quantization tqz;
+    tqz.name = tn;
+    tqz.quant = parse_ggml_type(qt.c_str());
+    tensor_type.emplace_back(std::move(tqz));
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -255,6 +360,7 @@ int main(int argc, char ** argv) {
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
     std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<tensor_quantization> tensor_types;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -277,6 +383,10 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--tensor-type") == 0) {
+            if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
@@ -361,6 +471,9 @@ int main(int argc, char ** argv) {
         kv_overrides.back().key[0] = 0;
         params.kv_overrides = &kv_overrides;
     }
+    if (!tensor_types.empty()) {
+        params.tensor_types = &tensor_types;
+    }
 
     llama_backend_init();
 

examples/rpc/rpc-server.cpp

@@ -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__) || defined(_AIX)
         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";
     }
@@ -295,7 +297,10 @@ int main(int argc, char * argv[]) {
         }
         cache_dir = cache_dir_str.c_str();
     }
-    printf("Starting RPC server\n");
+    printf("Starting RPC server v%d.%d.%d\n",
+           RPC_PROTO_MAJOR_VERSION,
+           RPC_PROTO_MINOR_VERSION,
+           RPC_PROTO_PATCH_VERSION);
     printf("  endpoint       : %s\n", endpoint.c_str());
     printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
     printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));

examples/run/run.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);
     }

examples/server/server.cpp

@@ -1552,29 +1552,30 @@ struct server_queue {
     std::condition_variable condition_tasks;
 
     // callback functions
-    std::function<void(server_task)> callback_new_task;
-    std::function<void(void)>        callback_update_slots;
+    std::function<void(server_task &&)> callback_new_task;
+    std::function<void(void)>           callback_update_slots;
 
     // Add a new task to the end of the queue
-    int post(server_task task, bool front = false) {
+    int post(server_task && task, bool front = false) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         GGML_ASSERT(task.id != -1);
         // if this is cancel task make sure to clean up pending tasks
         if (task.type == SERVER_TASK_TYPE_CANCEL) {
             cleanup_pending_task(task.id_target);
         }
-        QUE_DBG("new task, id = %d, front = %d\n", task.id, front);
+        const int task_id = task.id;
+        QUE_DBG("new task, id = %d, front = %d\n", task_id, front);
         if (front) {
             queue_tasks.push_front(std::move(task));
         } else {
             queue_tasks.push_back(std::move(task));
         }
         condition_tasks.notify_one();
-        return task.id;
+        return task_id;
     }
 
     // multi-task version of post()
-    int post(std::vector<server_task> & tasks, bool front = false) {
+    int post(std::vector<server_task> && tasks, bool front = false) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         for (auto & task : tasks) {
             if (task.id == -1) {
@@ -1596,7 +1597,7 @@ struct server_queue {
     }
 
     // Add a new task, but defer until one slot is available
-    void defer(server_task task) {
+    void defer(server_task && task) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         QUE_DBG("defer task, id = %d\n", task.id);
         queue_tasks_deferred.push_back(std::move(task));
@@ -1611,7 +1612,7 @@ struct server_queue {
     }
 
     // Register function to process a new task
-    void on_new_task(std::function<void(server_task)> callback) {
+    void on_new_task(std::function<void(server_task &&)> callback) {
         callback_new_task = std::move(callback);
     }
 
@@ -1660,7 +1661,7 @@ struct server_queue {
                     lock.unlock();
                     break;
                 }
-                server_task task = queue_tasks.front();
+                server_task task = std::move(queue_tasks.front());
                 queue_tasks.pop_front();
                 lock.unlock();
 
@@ -1705,6 +1706,8 @@ private:
 };
 
 struct server_response {
+    bool running = true;
+
     // for keeping track of all tasks waiting for the result
     std::unordered_set<int> waiting_task_ids;
 
@@ -1759,6 +1762,10 @@ struct server_response {
         while (true) {
             std::unique_lock<std::mutex> lock(mutex_results);
             condition_results.wait(lock, [&]{
+                if (!running) {
+                    SRV_DBG("%s : queue result stop\n", __func__);
+                    std::terminate(); // we cannot return here since the caller is HTTP code
+                }
                 return !queue_results.empty();
             });
 
@@ -1789,6 +1796,10 @@ struct server_response {
             }
 
             std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
+            if (!running) {
+                SRV_DBG("%s : queue result stop\n", __func__);
+                std::terminate(); // we cannot return here since the caller is HTTP code
+            }
             if (cr_res == std::cv_status::timeout) {
                 return nullptr;
             }
@@ -1818,6 +1829,12 @@ struct server_response {
             }
         }
     }
+
+    // terminate the waiting loop
+    void terminate() {
+        running = false;
+        condition_results.notify_all();
+    }
 };
 
 struct server_context {
@@ -1988,7 +2005,7 @@ struct server_context {
 
             slot.reset();
 
-            slots.push_back(slot);
+            slots.push_back(std::move(slot));
         }
 
         default_generation_settings_for_props = slots[0].to_json();
@@ -2089,7 +2106,7 @@ struct server_context {
         return true;
     }
 
-    bool launch_slot_with_task(server_slot & slot, const server_task & task) {
+    bool launch_slot_with_task(server_slot & slot, server_task && task) {
         slot.reset();
         slot.id_task       = task.id;
         slot.index         = task.index;
@@ -2097,10 +2114,10 @@ struct server_context {
         slot.params        = std::move(task.params);
         slot.prompt_tokens = std::move(task.prompt_tokens);
 
-        if (!are_lora_equal(task.params.lora, slot.lora)) {
+        if (!are_lora_equal(slot.params.lora, slot.lora)) {
             // if lora is changed, we cannot reuse cached tokens
             slot.cache_tokens.clear();
-            slot.lora = task.params.lora;
+            slot.lora = slot.params.lora;
         }
 
         bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
@@ -2531,10 +2548,10 @@ struct server_context {
             server_task task(SERVER_TASK_TYPE_CANCEL);
             task.id_target = id_task;
             queue_results.remove_waiting_task_id(id_task);
-            cancel_tasks.push_back(task);
+            cancel_tasks.push_back(std::move(task));
         }
         // push to beginning of the queue, so it has highest priority
-        queue_tasks.post(cancel_tasks, true);
+        queue_tasks.post(std::move(cancel_tasks), true);
     }
 
     // receive the results from task(s)
@@ -2621,7 +2638,7 @@ struct server_context {
     // Functions to process the task
     //
 
-    void process_single_task(server_task task) {
+    void process_single_task(server_task && task) {
         switch (task.type) {
             case SERVER_TASK_TYPE_COMPLETION:
             case SERVER_TASK_TYPE_INFILL:
@@ -2635,17 +2652,17 @@ struct server_context {
                     if (slot == nullptr) {
                         // if no slot is available, we defer this task for processing later
                         SRV_DBG("no slot is available, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
-                    if (!launch_slot_with_task(*slot, task)) {
+                    if (!launch_slot_with_task(*slot, std::move(task))) {
                         SRV_ERR("failed to launch slot with task, id_task = %d\n", task.id);
                         break;
                     }
@@ -2724,7 +2741,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2760,7 +2777,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2803,7 +2820,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2855,7 +2872,7 @@ struct server_context {
 
             server_task task(SERVER_TASK_TYPE_NEXT_RESPONSE);
             task.id = queue_tasks.get_new_id();
-            queue_tasks.post(task);
+            queue_tasks.post(std::move(task));
         }
 
         // apply context-shift if needed
@@ -3617,14 +3634,17 @@ int main(int argc, char ** argv) {
         }
 
         // request slots data using task queue
-        server_task task(SERVER_TASK_TYPE_METRICS);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task, true); // high-priority task
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_METRICS);
+            task.id = task_id;
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task), true); // high-priority task
+        }
 
         // get the result
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3653,16 +3673,17 @@ int main(int argc, char ** argv) {
         }
 
         // request slots data using task queue
-        server_task task(SERVER_TASK_TYPE_METRICS);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.metrics_reset_bucket = true;
-
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task, true); // high-priority task
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_METRICS);
+            task.id = task_id;
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task), true); // high-priority task
+        }
 
         // get the result
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3759,17 +3780,20 @@ int main(int argc, char ** argv) {
         }
         std::string filepath = params.slot_save_path + filename;
 
-        server_task task(SERVER_TASK_TYPE_SLOT_SAVE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id  = id_slot;
-        task.slot_action.filename = filename;
-        task.slot_action.filepath = filepath;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_SAVE);
+            task.id = task_id;
+            task.slot_action.slot_id  = id_slot;
+            task.slot_action.filename = filename;
+            task.slot_action.filepath = filepath;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3788,17 +3812,20 @@ int main(int argc, char ** argv) {
         }
         std::string filepath = params.slot_save_path + filename;
 
-        server_task task(SERVER_TASK_TYPE_SLOT_RESTORE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id  = id_slot;
-        task.slot_action.filename = filename;
-        task.slot_action.filepath = filepath;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_RESTORE);
+            task.id = task_id;
+            task.slot_action.slot_id  = id_slot;
+            task.slot_action.filename = filename;
+            task.slot_action.filepath = filepath;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3810,15 +3837,18 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_slots_erase = [&ctx_server, &res_error, &res_ok](const httplib::Request & /* req */, httplib::Response & res, int id_slot) {
-        server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id = id_slot;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
+            task.id = task_id;
+            task.slot_action.slot_id = id_slot;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3891,6 +3921,21 @@ 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](
@@ -3907,9 +3952,10 @@ int main(int argc, char ** argv) {
         }
 
         auto completion_id = gen_chatcmplid();
-        std::vector<server_task> tasks;
-
+        std::unordered_set<int> task_ids;
         try {
+            std::vector<server_task> tasks;
+
             const auto & prompt = data.at("prompt");
             // TODO: this log can become very long, put it behind a flag or think about a more compact format
             //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
@@ -3924,9 +3970,9 @@ int main(int argc, char ** argv) {
 
                 task.prompt_tokens    = std::move(tokenized_prompts[i]);
                 task.params           = server_task::params_from_json_cmpl(
-                                            ctx_server.ctx,
-                                            ctx_server.params_base,
-                                            data);
+                        ctx_server.ctx,
+                        ctx_server.params_base,
+                        data);
                 task.id_selected_slot = json_value(data, "id_slot", -1);
 
                 // OAI-compat
@@ -3934,18 +3980,18 @@ int main(int argc, char ** argv) {
                 task.params.oaicompat_cmpl_id         = completion_id;
                 // oaicompat_model is already populated by params_from_json_cmpl
 
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
+
+            task_ids = server_task::get_list_id(tasks);
+            ctx_server.queue_results.add_waiting_tasks(tasks);
+            ctx_server.queue_tasks.post(std::move(tasks));
         } catch (const std::exception & e) {
             res_error(res, format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
-        ctx_server.queue_results.add_waiting_tasks(tasks);
-        ctx_server.queue_tasks.post(tasks);
-
         bool stream = json_value(data, "stream", false);
-        const auto task_ids = server_task::get_list_id(tasks);
 
         if (!stream) {
             ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
@@ -4237,6 +4283,7 @@ int main(int argc, char ** argv) {
         // create and queue the task
         json responses = json::array();
         bool error = false;
+        std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
             for (size_t i = 0; i < tokenized_prompts.size(); i++) {
@@ -4249,28 +4296,27 @@ int main(int argc, char ** argv) {
                 // OAI-compat
                 task.params.oaicompat = oaicompat;
 
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
 
+            task_ids = server_task::get_list_id(tasks);
             ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(tasks);
-
-            // get the result
-            std::unordered_set<int> task_ids = server_task::get_list_id(tasks);
-
-            ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-                for (auto & res : results) {
-                    GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
-                    responses.push_back(res->to_json());
-                }
-            }, [&](const json & error_data) {
-                res_error(res, error_data);
-                error = true;
-            }, req.is_connection_closed);
-
-            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+            ctx_server.queue_tasks.post(std::move(tasks));
         }
 
+        // get the result
+        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
+            for (auto & res : results) {
+                GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
+                responses.push_back(res->to_json());
+            }
+        }, [&](const json & error_data) {
+            res_error(res, error_data);
+            error = true;
+        }, req.is_connection_closed);
+
+        ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+
         if (error) {
             return;
         }
@@ -4336,6 +4382,7 @@ int main(int argc, char ** argv) {
         // create and queue the task
         json responses = json::array();
         bool error = false;
+        std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
             std::vector<llama_tokens> tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
@@ -4345,26 +4392,24 @@ int main(int argc, char ** argv) {
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
                 task.prompt_tokens = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
 
+            task_ids = server_task::get_list_id(tasks);
             ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(tasks);
-
-            // get the result
-            std::unordered_set<int> task_ids = server_task::get_list_id(tasks);
-
-            ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-                for (auto & res : results) {
-                    GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
-                    responses.push_back(res->to_json());
-                }
-            }, [&](const json & error_data) {
-                res_error(res, error_data);
-                error = true;
-            }, req.is_connection_closed);
+            ctx_server.queue_tasks.post(std::move(tasks));
         }
 
+        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
+            for (auto & res : results) {
+                GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
+                responses.push_back(res->to_json());
+            }
+        }, [&](const json & error_data) {
+            res_error(res, error_data);
+            error = true;
+        }, req.is_connection_closed);
+
         if (error) {
             return;
         }
@@ -4400,14 +4445,19 @@ int main(int argc, char ** argv) {
             res_error(res, format_error_response("Request body must be an array", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
-        server_task task(SERVER_TASK_TYPE_SET_LORA);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SET_LORA);
+            task.id = task_id;
+            task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
+
+        // get the result
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -4455,6 +4505,7 @@ 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
@@ -4491,9 +4542,10 @@ int main(int argc, char ** argv) {
     svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
 
     // clean up function, to be called before exit
-    auto clean_up = [&svr]() {
+    auto clean_up = [&svr, &ctx_server]() {
         SRV_INF("%s: cleaning up before exit...\n", __func__);
         svr->stop();
+        ctx_server.queue_results.terminate();
         llama_backend_free();
     };
 
@@ -4534,7 +4586,7 @@ int main(int argc, char ** argv) {
 
     if (!ctx_server.load_model(params)) {
         clean_up();
-        // t.join(); // FIXME: see below
+        t.join();
         LOG_ERR("%s: exiting due to model loading error\n", __func__);
         return 1;
     }
@@ -4549,8 +4601,8 @@ int main(int argc, char ** argv) {
         common_chat_templates_source(ctx_server.chat_templates.get()),
         common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
 
-    ctx_server.queue_tasks.on_new_task([&ctx_server](const server_task & task) {
-        ctx_server.process_single_task(task);
+    ctx_server.queue_tasks.on_new_task([&ctx_server](server_task && task) {
+        ctx_server.process_single_task(std::move(task));
     });
 
     ctx_server.queue_tasks.on_update_slots([&ctx_server]() {
@@ -4582,7 +4634,7 @@ int main(int argc, char ** argv) {
     ctx_server.queue_tasks.start_loop();
 
     clean_up();
-    // t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
+    t.join();
 
     return 0;
 }

examples/server/tests/unit/test_embedding.py

@@ -49,6 +49,26 @@ def test_embedding_multiple():
         assert len(d['embedding']) > 1
 
 
+def test_embedding_multiple_with_fa():
+    server = ServerPreset.bert_bge_small_with_fa()
+    server.pooling = 'last'
+    server.start()
+    # one of these should trigger the FA branch (i.e. context size % 256 == 0)
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": [
+            "a "*253,
+            "b "*254,
+            "c "*255,
+            "d "*256,
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 4
+    for d in res.body['data']:
+        assert 'embedding' in d
+        assert len(d['embedding']) > 1
+
+
 @pytest.mark.parametrize(
     "input,is_multi_prompt",
     [

examples/server/tests/utils.py

@@ -323,6 +323,21 @@ class ServerPreset:
         server.server_embeddings = True
         return server
 
+    @staticmethod
+    def bert_bge_small_with_fa() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
+        server.model_alias = "bert-bge-small"
+        server.n_ctx = 1024
+        server.n_batch = 300
+        server.n_ubatch = 300
+        server.n_slots = 2
+        server.fa = True
+        server.seed = 42
+        server.server_embeddings = True
+        return server
+
     @staticmethod
     def tinyllama_infill() -> ServerProcess:
         server = ServerProcess()

examples/server/utils.hpp

@@ -3,7 +3,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "common/base64.hpp"
+#include "base64.hpp"
 
 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576

examples/server/webui/src/components/ChatScreen.tsx

@@ -1,4 +1,4 @@
-import { useEffect, useMemo, useRef, useState } from 'react';
+import { useEffect, useMemo, useState } from 'react';
 import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
 import ChatMessage from './ChatMessage';
 import { CanvasType, Message, PendingMessage } from '../utils/types';
@@ -6,6 +6,7 @@ import { classNames, cleanCurrentUrl, throttle } from '../utils/misc';
 import CanvasPyInterpreter from './CanvasPyInterpreter';
 import StorageUtils from '../utils/storage';
 import { useVSCodeContext } from '../utils/llama-vscode';
+import { useChatTextarea, ChatTextareaApi } from './useChatTextarea.ts';
 
 /**
  * A message display is a message node with additional information for rendering.
@@ -99,7 +100,8 @@ export default function ChatScreen() {
     canvasData,
     replaceMessageAndGenerate,
   } = useAppContext();
-  const textarea = useOptimizedTextarea(prefilledMsg.content());
+
+  const textarea: ChatTextareaApi = useChatTextarea(prefilledMsg.content());
 
   const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
   // TODO: improve this when we have "upload file" feature
@@ -248,14 +250,16 @@ export default function ChatScreen() {
         </div>
 
         {/* chat input */}
-        <div className="flex flex-row items-center pt-8 pb-6 sticky bottom-0 bg-base-100">
+        <div className="flex flex-row items-end pt-8 pb-6 sticky bottom-0 bg-base-100">
           <textarea
-            className="textarea textarea-bordered w-full"
+            // Default (mobile): Enable vertical resize, overflow auto for scrolling if needed
+            // Large screens (lg:): Disable manual resize, apply max-height for autosize limit
+            className="textarea textarea-bordered w-full resize-vertical lg:resize-none lg:max-h-48 lg:overflow-y-auto" // Adjust lg:max-h-48 as needed (e.g., lg:max-h-60)
             placeholder="Type a message (Shift+Enter to add a new line)"
             ref={textarea.ref}
+            onInput={textarea.onInput} // Hook's input handler (will only resize height on lg+ screens)
             onKeyDown={(e) => {
               if (e.nativeEvent.isComposing || e.keyCode === 229) return;
-              if (e.key === 'Enter' && e.shiftKey) return;
               if (e.key === 'Enter' && !e.shiftKey) {
                 e.preventDefault();
                 sendNewMessage();
@@ -263,7 +267,11 @@ export default function ChatScreen() {
             }}
             id="msg-input"
             dir="auto"
+            // Set a base height of 2 rows for mobile views
+            // On lg+ screens, the hook will calculate and set the initial height anyway
+            rows={2}
           ></textarea>
+
           {isGenerating(currConvId ?? '') ? (
             <button
               className="btn btn-neutral ml-2"
@@ -286,43 +294,3 @@ export default function ChatScreen() {
     </div>
   );
 }
-
-export interface OptimizedTextareaValue {
-  value: () => string;
-  setValue: (value: string) => void;
-  focus: () => void;
-  ref: React.RefObject<HTMLTextAreaElement>;
-}
-
-// This is a workaround to prevent the textarea from re-rendering when the inner content changes
-// See https://github.com/ggml-org/llama.cpp/pull/12299
-function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
-  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
-  const textareaRef = useRef<HTMLTextAreaElement>(null);
-
-  useEffect(() => {
-    if (textareaRef.current && savedInitValue) {
-      textareaRef.current.value = savedInitValue;
-      setSavedInitValue('');
-    }
-  }, [textareaRef, savedInitValue, setSavedInitValue]);
-
-  return {
-    value: () => {
-      return textareaRef.current?.value ?? savedInitValue;
-    },
-    setValue: (value: string) => {
-      if (textareaRef.current) {
-        textareaRef.current.value = value;
-      }
-    },
-    focus: () => {
-      if (textareaRef.current) {
-        // focus and move the cursor to the end
-        textareaRef.current.focus();
-        textareaRef.current.selectionStart = textareaRef.current.value.length;
-      }
-    },
-    ref: textareaRef,
-  };
-}

examples/server/webui/src/components/useChatTextarea.ts

@@ -0,0 +1,96 @@
+import { useEffect, useRef, useState, useCallback } from 'react';
+
+// Media Query for detecting "large" screens (matching Tailwind's lg: breakpoint)
+const LARGE_SCREEN_MQ = '(min-width: 1024px)';
+
+// Calculates and sets the textarea height based on its scrollHeight
+const adjustTextareaHeight = (textarea: HTMLTextAreaElement | null) => {
+  if (!textarea) return;
+
+  // Only perform auto-sizing on large screens
+  if (!window.matchMedia(LARGE_SCREEN_MQ).matches) {
+    // On small screens, reset inline height and max-height styles.
+    // This allows CSS (e.g., `rows` attribute or classes) to control the height,
+    // and enables manual resizing if `resize-vertical` is set.
+    textarea.style.height = ''; // Use 'auto' or '' to reset
+    textarea.style.maxHeight = '';
+    return; // Do not adjust height programmatically on small screens
+  }
+
+  const computedStyle = window.getComputedStyle(textarea);
+  // Get the max-height specified by CSS (e.g., from `lg:max-h-48`)
+  const currentMaxHeight = computedStyle.maxHeight;
+
+  // Temporarily remove max-height to allow scrollHeight to be calculated correctly
+  textarea.style.maxHeight = 'none';
+  // Reset height to 'auto' to measure the actual scrollHeight needed
+  textarea.style.height = 'auto';
+  // Set the height to the calculated scrollHeight
+  textarea.style.height = `${textarea.scrollHeight}px`;
+  // Re-apply the original max-height from CSS to enforce the limit
+  textarea.style.maxHeight = currentMaxHeight;
+};
+
+// Interface describing the API returned by the hook
+export interface ChatTextareaApi {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+  onInput: (event: React.FormEvent<HTMLTextAreaElement>) => void; // Input handler
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+// combined now with auto-sizing logic.
+export function useChatTextarea(initValue: string): ChatTextareaApi {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  // Effect to set initial value and height on mount or when initValue changes
+  useEffect(() => {
+    const textarea = textareaRef.current;
+    if (textarea) {
+      if (typeof savedInitValue === 'string' && savedInitValue.length > 0) {
+        textarea.value = savedInitValue;
+        // Call adjustTextareaHeight - it will check screen size internally
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+        setSavedInitValue(''); // Reset after applying
+      } else {
+        // Adjust height even if there's no initial value (for initial render)
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+      }
+    }
+  }, [textareaRef, savedInitValue]); // Depend on ref and savedInitValue
+
+  const handleInput = useCallback(
+    (event: React.FormEvent<HTMLTextAreaElement>) => {
+      // Call adjustTextareaHeight on every input - it will decide whether to act
+      adjustTextareaHeight(event.currentTarget);
+    },
+    []
+  );
+
+  return {
+    // Method to get the current value directly from the textarea
+    value: () => {
+      return textareaRef.current?.value ?? '';
+    },
+    // Method to programmatically set the value and trigger height adjustment
+    setValue: (value: string) => {
+      const textarea = textareaRef.current;
+      if (textarea) {
+        textarea.value = value;
+        // Call adjustTextareaHeight - it will check screen size internally
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        textareaRef.current.focus();
+      }
+    },
+    ref: textareaRef,
+    onInput: handleInput,
+  };
+}

examples/server/webui/src/utils/llama-vscode.ts

@@ -1,6 +1,6 @@
 import { useEffect, useState } from 'react';
 import { MessageExtraContext } from './types';
-import { OptimizedTextareaValue } from '../components/ChatScreen';
+import { ChatTextareaApi } from '../components/useChatTextarea.ts';
 
 // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
 // Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@@ -15,7 +15,7 @@ interface SetTextEvData {
  * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n  return 123' }, '*');
  */
 
-export const useVSCodeContext = (textarea: OptimizedTextareaValue) => {
+export const useVSCodeContext = (textarea: ChatTextareaApi) => {
   const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
     null
   );

examples/server_embd.py

@@ -15,7 +15,7 @@ async def main():
     model_url = "http://127.0.0.1:6900"
     responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
         url= f"{model_url}/embedding",
-        json= {"content": str(0)*1024}
+        json= {"content": "a "*1022}
     ) for i in range(n)])
 
     for response in responses:

examples/sycl/build.sh

@@ -8,10 +8,10 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DLLAMA_CURL=OFF # faster for long-prompt inference
 
 #for FP32
-cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=OFF
 
 #build example/main
 #cmake --build . --config Release --target main

ggml/CMakeLists.txt

@@ -107,6 +107,7 @@ message(DEBUG "INS_ENB             : ${INS_ENB}")
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
+option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
@@ -170,7 +171,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)

ggml/include/ggml-rpc.h

@@ -7,6 +7,9 @@
 extern "C" {
 #endif
 
+#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API

ggml/include/ggml.h

@@ -507,17 +507,12 @@ extern "C" {
 
         GGML_OP_UNARY,
 
-        GGML_OP_MAP_UNARY,
-        GGML_OP_MAP_BINARY,
-
-        GGML_OP_MAP_CUSTOM1_F32,
-        GGML_OP_MAP_CUSTOM2_F32,
-        GGML_OP_MAP_CUSTOM3_F32,
-
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
 
+        GGML_OP_CUSTOM,
+
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
@@ -1722,24 +1717,29 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    // nearest interpolate
+    enum ggml_scale_mode {
+        GGML_SCALE_MODE_NEAREST  = 0,
+        GGML_SCALE_MODE_BILINEAR = 1,
+    };
+
+    // interpolate
     // multiplies ne0 and ne1 by scale factor
-    // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   scale_factor);
+            int                   scale_factor,
+            enum ggml_scale_mode  mode);
 
-    // nearest interpolate
-    // nearest interpolate to specified dimensions
-    // used in tortoise.cpp
+    // interpolate
+    // interpolate scale to specified dimensions
     GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
-            int                   ne3);
+            int                   ne3,
+            enum ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1916,83 +1916,6 @@ extern "C" {
 
     // custom operators
 
-    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
-            struct ggml_context        * ctx,
-            struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun),
-        "use ggml_map_custom1 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
-            struct ggml_context        * ctx,
-            struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun),
-        "use ggml_map_custom1_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun),
-        "use ggml_map_custom2 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun),
-        "use ggml_map_custom2_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun),
-        "use ggml_map_custom1 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun),
-        "use ggml_map_custom1_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun),
-        "use ggml_map_custom2 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun),
-        "use ggml_map_custom2_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-            struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun),
-        "use ggml_map_custom3 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-            struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun),
-        "use ggml_map_custom3_inplace instead");
-
-    // custom operators v2
-
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
     typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
@@ -2048,6 +1971,30 @@ extern "C" {
             int                     n_tasks,
             void                  * userdata);
 
+    typedef void (*ggml_custom_op_t)(struct ggml_tensor * dst , int ith, int nth, void * userdata);
+
+    GGML_API struct ggml_tensor * ggml_custom_4d(
+            struct ggml_context * ctx,
+            enum ggml_type        type,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            struct ggml_tensor ** args,
+            int                   n_args,
+            ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
+    GGML_API struct ggml_tensor * ggml_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor ** args,
+            int                   n_args,
+            ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
     // loss function
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss(

ggml/src/CMakeLists.txt

@@ -267,6 +267,7 @@ function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
     foreach (feat NATIVE
+                  SSE42
                   AVX AVX2 BMI2 AVX_VNNI FMA F16C
                   AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
                   AMX_TILE AMX_INT8 AMX_BF16)
@@ -286,14 +287,16 @@ if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
     endif()
-    ggml_add_cpu_backend_variant(sandybridge    AVX)
-    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+    ggml_add_cpu_backend_variant(x64)
+    ggml_add_cpu_backend_variant(sse42        SSE42)
+    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
     if (NOT MSVC)
         # MSVC doesn't support AMX
-        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-cann/acl_tensor.cpp

@@ -41,6 +41,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT4;
         case GGML_TYPE_Q8_0:
             return ACL_INT8;
+        case GGML_TYPE_I64:
+            return ACL_INT64;
         default:
             return ACL_DT_UNDEFINED;
     }

ggml/src/ggml-cann/aclnn_ops.h

@@ -1,15 +1,4 @@
-#ifndef CANN_ACLNN_OPS
-#define CANN_ACLNN_OPS
-
 /**
- * @file    acl_tensor
- * @brief   This file contains related functions of ggml_tensor and acl_tensor.
- *          Contains conversion from ggml_tensor to acl_tensor, broadcast and other
- *          functions.
- * @author  hipudding <huafengchun@gmail.com>
- * @author  wangshuai09 <391746016@qq.com>
- * @date    July 15, 2024
- *
  * Copyright (c) 2023-2024 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
@@ -31,6 +20,10 @@
  * IN THE SOFTWARE.
  */
 
+#ifndef CANN_ACLNN_OPS
+#define CANN_ACLNN_OPS
+
+#include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
 #include <aclnnop/aclnn_exp.h>
@@ -50,6 +43,8 @@
 #include <aclnnop/aclnn_sqrt.h>
 #include <aclnnop/aclnn_sin.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_log.h>
+#include <aclnnop/aclnn_sign.h>
 #include "acl_tensor.h"
 #include "common.h"
 
@@ -483,8 +478,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *          operation is executed using the CANN backend for optimized performance.
  *
  * @param ctx The CANN context used for operations.
- * @param dst The destination tensor where the indices of the maximum values will be stored.
- *            dst->op is `GGML_OP_ARGMAX`.
+ * @param dst The destination tensor where the indices of the maximum values will
+ *            be stored. dst->op is `GGML_OP_ARGMAX`.
  */
 void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
@@ -599,6 +594,286 @@ void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
 void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
     aclTensor* acl_dst);
 
+/**
+ * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
+ * output tensor.
+ *
+ * This function checks whether broadcasting is needed between `src0` and `src1`.
+ * If broadcasting is required, it calculates the proper shapes and creates
+ * ACL tensors with broadcast parameters. Otherwise, it directly creates ACL tensors
+ * based on the original tensor shapes.
+ *
+ * @param src0     The first input tensor (reference shape).
+ * @param src1     The second input tensor (possibly broadcasted).
+ * @param dst      The destination/output tensor.
+ * @param acl_src0 Output pointer to the created ACL tensor corresponding to src0.
+ * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
+ * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
+ */
+void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
+    aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
+
+/**
+ * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
+ * tensor using the CANN backend.
+ *
+ * @details This function performs a 1D transposed convolution (also known as
+ * deconvolution) operation on the input tensor. The computed result is stored
+ * in the destination tensor `dst`. The operation is optimized using the CANN
+ * backend for improved performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the transposed convolution result
+ * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
+ */
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
+ * using the CANN backend.
+ *
+ * @details This function performs an element-wise ELU activation on the input
+ *          tensor.
+ *          The result is written to the destination tensor `dst` in-place.
+ *          The ELU function is defined as:
+ *
+ *          \text{ELU}(x) =
+ *          \begin{cases}
+ *          x, & \text{if } x > 0 \\
+ *          \alpha \left( \exp(x) - 1 \right), & \text{if } x \leq 0
+ *          \end{cases}
+ *
+ *          where α (alpha) is a hyperparameter, typically set to 1.0.
+ *          This operation is optimized using the CANN backend for high-performance
+ *          inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the ELU-activated result will be stored.
+ *            dst->op is expected to be `GGML_OP_ELU`.
+ */
+void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the mean of a ggml tensor element-wise using the CANN backend.
+ *
+ * @details This function calculates the element-wise mean of the input tensor.
+ *          The result is written to the destination tensor `dst`.
+ *          The mean is computed by averaging the values across the entire tensor.
+ *
+ *          This operation is optimized using the CANN backend for high-performance inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the mean result will be stored.
+ *            dst->op is expected to be `GGML_OP_MEAN`.
+ */
+void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies 1D reflect padding to a ggml tensor using the CANN backend.
+ *
+ * @details This function performs 1D reflect padding on the input tensor.
+ *          The amount of padding on each side is specified by parameters stored in `dst->op_params`.
+ *          The operation reflects the values at the borders of the tensor to generate the padded output.
+ *
+ *          This operation is optimized using the CANN backend for high-performance inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the padded result will be stored.
+ *            dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
+ */
+void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Counts the number of equal elements in two ggml tensors using the CANN backend.
+ *
+ * @details This function performs an element-wise comparison between two input tensors,
+ *          and counts the number of positions where the elements are equal. The result is
+ *          stored in the destination tensor `dst` as a scalar.
+ *
+ *          The operation is optimized using the CANN backend, making it suitable for
+ *          high-performance inference or training scenarios.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
+ */
+void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies the Step activation function to a ggml tensor using the CANN backend.
+ *
+ * @details This function applies a step function element-wise to the input tensor, where
+ *          each element is transformed to 1.0 if it is greater than 0, and 0.0 otherwise.
+ *          The result is stored in the destination tensor `dst`.
+ *
+ *          This operation is accelerated using the CANN backend to improve runtime performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            dst->op is expected to be `GGML_OP_STEP`.
+ */
+void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/*
+ * @brief A generic wrapper for ACL resources with custom deleter support.
+ */
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+
+/**
+ * @brief Trait structure used to define how to destroy a given ACL resource type.
+ *
+ * @tparam T ACL resource type.
+ */
+template<typename T>
+struct acl_resource_traits;
+
+/**
+ * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
+ */
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
+ */
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
+ */
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
+ */
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
+};
+
+/**
+ * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ *
+ * @tparam T ACL resource type.
+ * @param ptr Raw pointer to ACL resource.
+ * @return any_acl_resource Smart pointer that handles destruction.
+ */
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
+}
+
+/**
+ * @brief Registers multiple ACL resources into a vector for lifetime management.
+ *
+ * @tparam Args Variadic list of ACL resource types.
+ * @param vec Target vector to hold ACL resources.
+ * @param args Raw pointers to ACL resources.
+ */
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+    (vec.emplace_back(make_acl_resource(args)), ...);
+}
+
+/**
+ * @brief Task class that wraps the execution of an aclnn function call.
+ */
+class aclnn_task : public cann_task {
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
+};
+
+/**
+ * @brief Task class that releases ACL resources after usage.
+ */
+class release_resource_task : public cann_task {
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
+
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
+    std::vector<any_acl_resource> resource_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory copy operations.
+ */
+class async_memcpy_task : public cann_task {
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
+    aclrtMemcpyKind kind_;
+    aclrtStream stream_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory set operations.
+ */
+class async_memset_task : public cann_task {
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
+};
+
 /**
  * @brief Launches an asynchronous task using the memory allocator.
  *
@@ -616,45 +891,101 @@ void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
  * other task before this asynchronous task ends, because all tasks in the
  * same stream are executed in queue order.
  */
-#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...)                                                \
-    do {                                                                                     \
-        uint64_t        workspaceSize = 0;                                                   \
-        aclOpExecutor * executor;                                                            \
-        void *          workspaceAddr = nullptr;                                             \
-                                                                                             \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
-                                                                                             \
-        if (workspaceSize > 0) {                                                             \
-            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);             \
-            workspaceAddr = workspace_allocator.get();                                       \
-        }                                                                                    \
-        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream()));     \
+
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
     } while (0)
 
-
 /**
- * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one output tensor.
+ * @brief Registers and releases multiple ACL resources, optionally deferring the release
+ *        using a task.
  *
- * This function checks whether broadcasting is needed between `src0` and `src1`.
- * If broadcasting is required, it calculates the proper shapes and creates
- * ACL tensors with broadcast parameters. Otherwise, it directly creates ACL tensors
- * based on the original tensor shapes.
- *
- * @param src0     The first input tensor (reference shape).
- * @param src1     The second input tensor (possibly broadcasted).
- * @param dst      The destination/output tensor.
- * @param acl_src0 Output pointer to the created ACL tensor corresponding to src0.
- * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
- * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
+ * @tparam Args Types of the ACL resources.
+ * @param ctx Backend context which manages task submission and async mode.
+ * @param args Pointers to ACL resources to be released.
  */
-void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclTensor ** acl_src0,
-                        aclTensor ** acl_src1, aclTensor ** acl_dst);
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+    std::vector<any_acl_resource> resources;
+    register_acl_resources(resources, std::forward<Args>(args)...);
+    if(ctx.async_mode) {
+        auto task = std::make_unique<release_resource_task>(std::move(resources));
+        ctx.task_queue.submit_task(std::move(task));
+    }
+}
 
 /**
- * @brief Applies a element-wise operation to two input tensors using the CANN backend.
+ * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
  *
- * This templated function takes a binary operator and applies it to two source tensors
- * associated with the destination tensor. The function handles broadcasting as needed.
+ * @param ctx Backend context containing stream and async configuration.
+ * @param dst Destination memory address.
+ * @param src Source memory address.
+ * @param len Size of memory to copy (in bytes).
+ * @param kind Type of memory copy (host-to-device, device-to-host, etc).
+ */
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
+    }
+}
+
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx->async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
+        ctx->task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param buffer Memory buffer to be set.
+ * @param size Size of the memory buffer (in bytes).
+ * @param value Value to set in the buffer.
+ */
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
+    }
+}
+
+/**
+ * @brief Applies a element-wise operation to two input tensors using the CANN
+ * backend.
+ *
+ * This templated function takes a binary operator and applies it to two source
+ * tensors
+ * associated with the destination tensor. The function handles broadcasting as
+ * needed.
  *
  * @tparam binary_op A callable object (e.g., lambda or function pointer) representing
  *         the binary operation to be performed. It must take three arguments:
@@ -676,11 +1007,10 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
     binary_op(ctx, acl_src0, acl_src1, acl_dst);
 
-    ACL_CHECK(aclDestroyTensor(acl_src0));
-    ACL_CHECK(aclDestroyTensor(acl_src1));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
+
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
  *
@@ -690,7 +1020,6 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @tparam unary_op A callable with the signature:
  *         void(ggml_backend_cann_context&, aclTensor*, aclTensor*)
  *         where the first aclTensor is the source and the second is the destination.
- *
  * @param ctx The CANN backend context for managing resources and execution.
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
@@ -702,10 +1031,28 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
+/**
+ * @brief   Applies a unary operation to a ggml tensor using the CANN backend.
+ *
+ * @details This function performs a unary operation on the input tensor using
+ * a user-provided lambda or callable object `unary_op`, which accepts the CANN
+ * context and two ACL tensors (source and destination). Internally, this function
+ * creates ACL representations of the ggml tensors and invokes the unary operation.
+ * The result is stored in the destination tensor `dst`. This utility abstracts the
+ * common boilerplate of tensor conversion and cleanup when implementing unary ops.
+ *
+ * @param unary_op A callable that performs the unary operation using CANN APIs.
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            The source tensor is retrieved from `dst->src[0]`.
+ */
+void ggml_cann_unary_op(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
  * @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
  *
@@ -715,7 +1062,7 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
  *
  * Internally, the lambda will call:
  * @code
- * GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);
+ * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
  * @endcode
  *
  * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
@@ -723,13 +1070,14 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
  * @see ggml_cann_unary_op
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                         \
-    do {                                                         \
-        auto lambda = [](auto ctx, auto acl_src, auto acl_dst) { \
-            GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);  \
-        };                                                       \
-        ggml_cann_unary_op<lambda>(ctx, dst);                    \
-    }                                                            \
+#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_unary_op(lambda, ctx, dst);                         \
+    }                                                                 \
     while (0)
-
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -31,9 +31,16 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <unistd.h>
+#include <functional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
+#include "../ggml-impl.h"
 
 #define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
@@ -205,6 +212,127 @@ struct ggml_cann_pool_alloc {
     ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };
 
+/**
+ * @brief Function pointer type for ACLNN operator calls.
+ */
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+
+/**
+ * @brief Base class for all CANN tasks to be submitted to the task queue.
+ *
+ * Users should override the run_task() method with actual task logic.
+ */
+class cann_task {
+public:
+    virtual void run_task() {}
+};
+
+/**
+ * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
+ */
+class cann_task_queue {
+public:
+    /**
+     * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
+     *
+     * @param capacity Queue capacity. Must be a power of 2.
+     * @param device Target device ID (used for context setting).
+     */
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
+        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
+        mask_ = capacity_ - 1;
+    }
+
+    /**
+     * @brief Attempts to enqueue a task into the queue.
+     *
+     * @param item Unique pointer to the task.
+     * @return true if the task was successfully enqueued, false if the queue was full.
+     */
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
+        size_t next_tail = (tail_ + 1) & mask_;
+
+        if (next_tail == head_) {
+            return false;
+        }
+
+        buffer_[tail_] = std::move(item);
+        std::atomic_thread_fence(std::memory_order_release);
+        tail_ = next_tail;
+
+        return true;
+    }
+
+    /**
+     * @brief Submits a task to the queue, and starts the worker thread if not already running.
+     *
+     * @param task Task to be submitted.
+     */
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
+            std::this_thread::yield();
+            continue;
+        }
+
+        if (!running_) {
+            running_ = true;
+            thread_ = std::thread(&cann_task_queue::execute, this);
+        }
+
+    }
+
+    /**
+     * @brief Waits until the queue is completely empty and no tasks are being processed.
+     */
+    void wait() {
+        while (running_ && head_ != tail_) {
+            std::this_thread::yield();
+            continue;
+        }
+    }
+
+    /**
+     * @brief Stops the task queue and joins the worker thread.
+     */
+    void stop() {
+        running_ = false;
+        if (thread_.joinable()) {
+            thread_.join();
+        }
+    }
+
+private:
+    /**
+     * @brief Worker thread function that continuously dequeues and executes tasks.
+     */
+    void execute() {
+        ggml_cann_set_device(device_);
+
+        while (running_) {
+            if(head_ == tail_) {
+                std::this_thread::yield();
+                continue;
+            }
+
+            std::atomic_thread_fence(std::memory_order_acquire);
+            buffer_[head_]->run_task();
+            buffer_[head_].reset();
+            head_ = (head_ + 1) & mask_;
+        }
+    }
+
+    std::vector<std::unique_ptr<cann_task>> buffer_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -213,6 +341,8 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    cann_task_queue task_queue;
+    bool async_mode;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -221,9 +351,12 @@ struct ggml_backend_cann_context {
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
     }
 
     /**
@@ -231,6 +364,7 @@ struct ggml_backend_cann_context {
      */
     ~ggml_backend_cann_context() {
         ggml_cann_set_device(device);
+        task_queue.stop();
         if (copy_event != nullptr) {
             ACL_CHECK(aclrtDestroyEvent(copy_event));
         }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -29,6 +29,8 @@
 #include <cstdio>
 #include <cstring>
 #include <mutex>
+#include <queue>
+#include <chrono>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -119,9 +121,10 @@ static ggml_cann_device_info ggml_cann_init() {
         prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
         prop.location.id = id;
         prop.reserve = 0;
-        ACL_CHECK(aclrtMemGetAllocationGranularity(
+        err = aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
-            &info.devices[id].vmm_granularity));
+            &info.devices[id].vmm_granularity);
+        info.devices[id].vmm = err == ACL_SUCCESS;
 
         size_t free, total;
         ggml_backend_cann_get_device_memory(id, &free, &total);
@@ -148,11 +151,223 @@ const ggml_cann_device_info& ggml_cann_info() {
 
 //#define DEBUG_CANN_MALLOC
 /**
- * @brief A pool of CANN buffers(legacy).
+ * @brief A pool of CANN buffers(priority segment buffer).
  *
  * This class manages a pool of CANN buffers for a specific device.
  */
-struct ggml_cann_pool_leg : public ggml_cann_pool {
+struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer.
+        size_t size = 0;      ///< Size of the buffer.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
+
+        bool operator>(const ggml_cann_buffer& other) const {
+            return size > other.size;
+        }
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    std::unordered_map<void*, size_t> buffer_pool;
+    std::priority_queue<ggml_cann_buffer,
+                        std::vector<ggml_cann_buffer>,
+                        std::greater<>> free_buffers ;
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_buf_prio(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_buf_prio() {
+        ggml_cann_set_device(device);
+        for (auto& [b_ptr, b_size] : buffer_pool) {
+            aclrtFree(b_ptr);
+            pool_size -= b_size;
+        }
+        buffer_pool.clear();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        std::vector<ggml_cann_buffer> free_buffers_rest;
+        free_buffers_rest.reserve(free_buffers.size());
+        while (!free_buffers.empty()) {
+            auto b = free_buffers.top();
+            free_buffers.pop();
+
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    ptr = b.ptr;
+#ifdef DEBUG_CANN_MALLOC
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
+#endif
+                    break;
+                }
+            }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+                buffer_pool.erase(b.ptr);
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                continue;
+            }
+            free_buffers_rest.push_back(b);
+        }
+        for (ggml_cann_buffer &b : free_buffers_rest) {
+            free_buffers.push(std::move(b));
+        }
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+        if (ptr != nullptr) {
+            return ptr;
+        }
+
+        // allocate a new buffer if no buffer can be reused
+        ggml_cann_set_device(device);
+        ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: allocate %p, "
+            "pool_size = %5u MB, "
+            "size = %5u MB\n",
+            device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+            (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
+#endif
+        buffer_pool.emplace(ptr, size);
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
+        auto it = buffer_pool.find(ptr);
+        if (it == buffer_pool.end()) {
+            GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
+        }
+
+        auto now = std::chrono::steady_clock::now();
+        free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: return   %p, "
+            "pool_size = %5u MB\n",
+            device, ptr,
+            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers(segment buffer).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_buf : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
     /**
      * @brief The maximum number of buffers in the pool.
      */
@@ -163,12 +378,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      */
     int device;
 
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
     /**
      * @brief Structure representing a CANN buffer.
      */
     struct ggml_cann_buffer {
         void* ptr = nullptr;  ///< Pointer to the buffer memory.
         size_t size = 0;      ///< Size of the buffer.
+        bool used = false;    ///< Whether the buffer is currently in use.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
     };
 
     /**
@@ -186,17 +408,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      *
      * @param device The device ID to associate with this buffer pool.
      */
-    explicit ggml_cann_pool_leg(int device) : device(device) {}
+    explicit ggml_cann_pool_buf(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
 
     /**
      * @brief Destructor to free all buffers in the pool.
      */
-    ~ggml_cann_pool_leg() {
+    ~ggml_cann_pool_buf() {
         ggml_cann_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
             if (b.ptr != nullptr) {
-                ACL_CHECK(aclrtFree(b.ptr));
+                aclrtFree(b.ptr);
                 pool_size -= b.size;
             }
         }
@@ -212,63 +436,93 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
-        const size_t alignment = 128;
         size = GGML_PAD(size, alignment);
         if (size == 0) {
             size = alignment;
         }
-#ifdef DEBUG_CANN_MALLOC
-        int nnz = 0;
-        size_t max_size = 0;
-#endif
-        size_t best_diff = 1ull << 36;
-        int ibest = -1;
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        int i = 0;
+        for (; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
+            if (b.ptr == nullptr) {
+                break;
+            }
+            if (b.used) {
+                continue;
+            }
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    b.used = true;
+                    ptr = b.ptr;
 #ifdef DEBUG_CANN_MALLOC
-                ++nnz;
-                if (b.size > max_size) max_size = b.size;
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
 #endif
-                if (b.size >= size) {
-                    size_t diff = b.size - size;
-                    if (diff < best_diff) {
-                        best_diff = diff;
-                        ibest = i;
-                        if (!best_diff) {
-                            void* ptr = b.ptr;
-                            *actual_size = b.size;
-                            b.ptr = nullptr;
-                            b.size = 0;
-                            return ptr;
-                        }
-                    }
+                    break;
                 }
             }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                b.ptr = nullptr;
+            }
         }
-        if (ibest >= 0) {
-            ggml_cann_buffer& b = buffer_pool[ibest];
-            void* ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
+        if (ptr != nullptr) {
             return ptr;
         }
-        void* ptr;
-        ggml_cann_set_device(device);
-        ACL_CHECK(
-            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = size;
-        pool_size += size;
+
+        if (i < MAX_BUFFERS) {
+            // allocate a new buffer if no buffer can be reused
+            ggml_cann_buffer& b = buffer_pool[i];
+            ggml_cann_set_device(device);
+            ACL_CHECK(aclrtMalloc(&b.ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+            pool_size += size;
+            *actual_size = size;
+            b.size = size;
+            b.used = true;
+            if (i >= MAX_BUFFERS - 8) {
+                GGML_LOG_WARN("cann pool[%d]: slots almost full\n", device);
+            }
 #ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO(
-            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
-            "requested %u MB\n",
-            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
-            (uint32_t)(pool_size / 1024 / 1024),
-            (uint32_t)(size / 1024 / 1024));
+            GGML_LOG_INFO(
+                "cann pool[%d]: allocate %p, "
+                "pool_size = %5u MB, "
+                "size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
 #endif
-        return ptr;
+            return b.ptr;
+        }
+
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 
     /**
@@ -278,18 +532,24 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @param size Size of the buffer to free.
      */
     void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr == nullptr) {
-                b.ptr = ptr;
-                b.size = size;
-                return;
+            if (b.ptr != ptr) {
+                continue;
             }
+            b.used = false;
+            b.last_used = std::chrono::steady_clock::now();
+#ifdef DEBUG_CANN_MALLOC
+            GGML_LOG_INFO(
+                "cann pool[%d]: return   %p, "
+                "pool_size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+            return;
         }
-        // memory should always buffered. these memory may still needed by
-        // tasks in stream.
-        // TODO, fix me.
-        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 };
 
@@ -347,8 +607,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_vmm(int device)
-        : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+    : device(device) {
         auto dev = ggml_cann_info().devices[device];
         granularity = dev.vmm_granularity;
         max_size = dev.total_vram;
@@ -471,7 +730,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
+    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
+    if (enable_buf_prio) {
+        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
+    }
+    GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
 
 // cann buffer
@@ -1020,8 +1290,11 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
     ggml_cann_set_device(buft_ctx->device);
 
-    size = std::max(size, (size_t)1);
-
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
     void* dev_ptr;
     aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
     if (err != ACL_SUCCESS) {
@@ -1330,12 +1603,13 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                     GGML_CANN_CALL_UNARY_OP(Silu);
                     break;
                 case GGML_UNARY_OP_GELU_QUICK: {
-                        auto lambda = [](auto ctx, auto acl_src, auto acl_dst) {
-                            GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst);
-                        };
-                        ggml_cann_unary_op<lambda>(ctx, dst);
-                    }
-                    break;
+                    auto lambda = [](ggml_backend_cann_context& ctx,
+                        aclTensor* acl_src,
+                        aclTensor* acl_dst) {
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
+                    };
+                    ggml_cann_unary_op(lambda, ctx, dst);
+                } break;
                 case GGML_UNARY_OP_TANH:
                     GGML_CANN_CALL_UNARY_OP(Tanh);
                     break;
@@ -1354,6 +1628,15 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                 case GGML_UNARY_OP_EXP:
                     GGML_CANN_CALL_UNARY_OP(Exp);
                     break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_cann_elu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SGN:
+                    GGML_CANN_CALL_UNARY_OP(Sign);
+                    break;
+                case GGML_UNARY_OP_STEP:
+                    ggml_cann_step(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -1448,7 +1731,22 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             break;
         case GGML_OP_SIN:
             ggml_cann_unary_op<aclnn_sin>(ctx, dst);
-        break;
+            break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            ggml_cann_conv_transpose_1d(ctx, dst);
+            break;
+        case GGML_OP_LOG:
+            GGML_CANN_CALL_UNARY_OP(Log);
+            break;
+        case GGML_OP_MEAN:
+            ggml_cann_mean(ctx, dst);
+            break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_cann_pad_reflect_1d(ctx, dst);
+            break;
+        case GGML_OP_COUNT_EQUAL:
+            ggml_cann_count_equal(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -1491,12 +1789,11 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
+
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
+ * This function asynchronously sets tensor data in the CANN backend.
  *
  * @param backend Pointer to the CANN backend structure.
  * @param tensor Pointer to the tensor structure to set data for.
@@ -1511,23 +1808,28 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
                                                size_t size) {
     ggml_backend_cann_context *cann_ctx =
         (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+        "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
+    ggml_cann_async_memcpy(cann_ctx, (char *)tensor->data + offset, data, size,
+        ACL_MEMCPY_HOST_TO_DEVICE);
 }
 
+/**
+ * @brief Gets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously gets tensor data in the CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to get data from.
+ * @param data Pointer to the host data to copy from the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
 static void ggml_backend_cann_get_tensor_async(
     ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
@@ -1538,20 +1840,11 @@ static void ggml_backend_cann_get_tensor_async(
 
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
                 "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
+
+    ggml_cann_async_memcpy(cann_ctx, data, (char *)tensor->data + offset, size,
+        ACL_MEMCPY_DEVICE_TO_HOST);
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
 }
 
 /**
@@ -1611,6 +1904,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
 
+        // wait for task_queue empty to keep task order.
+        cann_ctx_src->task_queue.wait();
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
@@ -1638,9 +1933,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
+    cann_ctx->task_queue.wait();
     ggml_cann_set_device(cann_ctx->device);
-
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
@@ -1710,6 +2004,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_STEP:
                     return true;
                 default:
                     return false;
@@ -1721,6 +2018,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return true;
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
                     // only support contiguous for quantized types.
                     return ggml_is_contiguous(op->src[0]) &&
                             ggml_is_contiguous(op->src[1]);
@@ -1788,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 return false;
             }
 
+            if(!ggml_is_contiguous(op->src[0])){
+                return false;
+            }
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -1796,10 +2100,19 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
                 return false;
             }
+            if (op->op_params[0] != GGML_SCALE_MODE_NEAREST) {
+                return false;
+            }
             return true;
         }
         case GGML_OP_POOL_2D: {
             const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
             const int       k0   = opts[1];
             const int       k1   = opts[2];
             const int       p0   = opts[5];
@@ -1842,6 +2155,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_ARGMAX:
         case GGML_OP_COS:
         case GGML_OP_SIN:
+        case GGML_OP_CONV_TRANSPOSE_1D:
+        case GGML_OP_LOG:
+        case GGML_OP_MEAN:
+        case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_COUNT_EQUAL:
             return true;
         default:
             return false;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -222,7 +222,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             elseif (GGML_AVX)
                 list(APPEND ARCH_FLAGS /arch:AVX)
                 list(APPEND ARCH_DEFINITIONS GGML_AVX)
-            else ()
+            elseif (GGML_SSE42)
                 list(APPEND ARCH_FLAGS /arch:SSE4.2)
                 list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
@@ -237,8 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_NATIVE)
                 list(APPEND ARCH_FLAGS -march=native)
             else ()
-                list(APPEND ARCH_FLAGS -msse4.2)
-                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_SSE42)
+                    list(APPEND ARCH_FLAGS -msse4.2)
+                    list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                endif()
                 if (GGML_F16C)
                     list(APPEND ARCH_FLAGS -mf16c)
                     list(APPEND ARCH_DEFINITIONS GGML_F16C)

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -263,7 +263,7 @@ void test_x86_is() {
 static int ggml_backend_cpu_x86_score() {
     // FIXME: this does not check for OS support
 
-    int score = 0;
+    int score = 1;
     cpuid_x86 is;
 
 #ifdef GGML_FMA

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -323,8 +323,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #else
 #ifdef __POWER9_VECTOR__
 #include <altivec.h>
-#undef bool
-#define bool _Bool
 #else
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2027,41 +2027,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
-        case GGML_OP_MAP_UNARY:
-            {
-                ggml_unary_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_unary(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_BINARY:
-            {
-                ggml_binary_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_binary(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM1_F32:
-            {
-                ggml_custom1_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1_f32(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM2_F32:
-            {
-                ggml_custom2_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2_f32(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM3_F32:
-            {
-                ggml_custom3_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3_f32(params, tensor, fun);
-            }
-            break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2077,6 +2042,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_map_custom3(params, tensor);
             }
             break;
+        case GGML_OP_CUSTOM:
+            {
+                ggml_compute_forward_custom(params, tensor);
+            }
+            break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 ggml_compute_forward_cross_entropy_loss(params, tensor);
@@ -2328,11 +2298,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
-        case GGML_OP_MAP_UNARY:
-        case GGML_OP_MAP_BINARY:
-        case GGML_OP_MAP_CUSTOM1_F32:
-        case GGML_OP_MAP_CUSTOM2_F32:
-        case GGML_OP_MAP_CUSTOM3_F32:
             {
                 n_tasks = 1;
             } break;
@@ -2366,6 +2331,16 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     n_tasks = MIN(p.n_tasks, n_threads);
                 }
             } break;
+        case GGML_OP_CUSTOM:
+            {
+                struct ggml_custom_op_params p;
+                memcpy(&p, node->op_params, sizeof(p));
+                if (p.n_tasks == GGML_N_TASKS_MAX) {
+                    n_tasks = n_threads;
+                } else {
+                    n_tasks = MIN(p.n_tasks, n_threads);
+                }
+            } break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -425,6 +425,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
         }
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
+        case GGML_OP_GET_ROWS_BACK:
+            return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
         case GGML_OP_OUT_PROD:
             return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
                 src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;

ggml/src/ggml-cpu/ops.cpp

@@ -6351,24 +6351,72 @@ static void ggml_compute_forward_upscale_f32(
     const float sf2 = (float)ne2/src0->ne[2];
     const float sf3 = (float)ne3/src0->ne[3];
 
-    // TODO: optimize
+    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
 
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        const int64_t i03 = i3 / sf3;
-        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
-            const int64_t i02 = i2 / sf2;
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                const int64_t i01 = i1 / sf1;
-                for (int64_t i0 = 0; i0 < ne0; i0++) {
-                    const int64_t i00 = i0 / sf0;
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const int64_t i01 = i1 / sf1;
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const int64_t i00 = i0 / sf0;
 
-                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
+                        const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
 
-                    *y = *x;
+                        *y = *x;
+                    }
                 }
             }
         }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        // setting a pixel offset of 0 would replicate the behavior of pytorch interpolate with align_corners=True
+        const float pixel_offset = 0.5f;
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    int64_t y0 = (int64_t)floorf(y);
+                    int64_t y1 = y0 + 1;
+
+                    y0 = std::max(int64_t(0), std::min(y0, ne01 - 1));
+                    y1 = std::max(int64_t(0), std::min(y1, ne01 - 1));
+
+                    float dy = y - (float)y0;
+                    dy = std::max(0.0f, std::min(dy, 1.0f));
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        int64_t x0 = (int64_t)floorf(x);
+                        int64_t x1 = x0 + 1;
+
+                        x0 = std::max(int64_t(0), std::min(x0, ne00 - 1));
+                        x1 = std::max(int64_t(0), std::min(x1, ne00 - 1));
+
+                        float dx = x - (float)x0;
+                        dx = std::max(0.0f, std::min(dx, 1.0f));
+
+                        // fetch the four surrounding pixel values and interpolate
+                        const float a = *(const float *)((const char *)src0->data + x0*nb00 + y0*nb01 + i02*nb02 + i03*nb03);
+                        const float b = *(const float *)((const char *)src0->data + x1*nb00 + y0*nb01 + i02*nb02 + i03*nb03);
+                        const float c = *(const float *)((const char *)src0->data + x0*nb00 + y1*nb01 + i02*nb02 + i03*nb03);
+                        const float d = *(const float *)((const char *)src0->data + x1*nb00 + y1*nb01 + i02*nb02 + i03*nb03);
+
+                        const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
+
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else {
+        GGML_ABORT("unsupported upscale mode");
     }
 }
 
@@ -6721,8 +6769,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     ggml_vec_dot_t    const kq_vec_dot     = ggml_get_type_traits_cpu(k->type)->vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
 
-    GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type");
-    GGML_ASSERT(v_to_float   && "fattn: unsupported V-type");
+    GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
+    GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
@@ -6818,10 +6866,14 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                     vs = expf(s - M);
                 }
 
-                v_to_float(v_data, V32, DV);
-
                 // V += v*expf(s - M)
-                ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                if (v_to_float) {
+                    v_to_float(v_data, V32, DV);
+                    ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                } else {
+                    // V is F32
+                    ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs);
+                }
             }
 
             S = S*ms + vs; // scale and increment sum with partial sum
@@ -8264,152 +8316,6 @@ void ggml_compute_forward_rwkv_wkv7(
     }
 }
 
-// ggml_compute_forward_map_unary
-
-static void ggml_compute_forward_map_unary_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_unary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, dst));
-
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
-
-    for (int i = 0; i < n; i++) {
-        fun(nc,
-                (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])));
-    }
-}
-
-void ggml_compute_forward_map_unary(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_unary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_unary_f32(params, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_map_binary
-
-static void ggml_compute_forward_map_binary_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_binary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(src1));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
-
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
-
-    for (int i = 0; i < n; i++) {
-        fun(nc,
-                (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])),
-                (float *) ((char *) src1->data + i*(src1->nb[1])));
-    }
-}
-
-void ggml_compute_forward_map_binary(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_binary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_binary_f32(params, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_map_custom1
-
-void ggml_compute_forward_map_custom1_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom1_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a);
-}
-
-// ggml_compute_forward_map_custom2
-
-void ggml_compute_forward_map_custom2_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom2_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-    const ggml_tensor * b = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a, b);
-}
-
-// ggml_compute_forward_map_custom3
-
-void ggml_compute_forward_map_custom3_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom3_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-    const ggml_tensor * b = dst->src[1];
-    const ggml_tensor * c = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a, b, c);
-}
-
 // ggml_compute_forward_map_custom1
 
 void ggml_compute_forward_map_custom1(
@@ -8455,6 +8361,18 @@ void ggml_compute_forward_map_custom3(
     p.fun(dst, a, b, c, params->ith, params->nth, p.userdata);
 }
 
+// ggml_compute_forward_custom
+
+void ggml_compute_forward_custom(
+    const struct ggml_compute_params * params,
+          struct ggml_tensor * dst) {
+
+    struct ggml_custom_op_params p;
+    memcpy(&p, dst->op_params, sizeof(p));
+
+    p.fun(dst, params->ith, params->nth, p.userdata);
+}
+
 // ggml_compute_forward_cross_entropy_loss
 
 static void ggml_compute_forward_cross_entropy_loss_f32(

ggml/src/ggml-cpu/ops.h

@@ -96,29 +96,10 @@ void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params,
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_map_unary(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_unary_op_f32_t fun);
-void ggml_compute_forward_map_binary(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_binary_op_f32_t fun);
-void ggml_compute_forward_map_custom1_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom1_op_f32_t fun);
-void ggml_compute_forward_map_custom2_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom2_op_f32_t fun);
-void ggml_compute_forward_map_custom3_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom3_op_f32_t fun);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/simd-mappings.h

@@ -392,7 +392,11 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ?                   \
   vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \
   vec_extract_fp32_from_shortl(vec_xl(0, p))
-#define GGML_ENDIAN_BYTE(i) ((unsigned char *)&(uint16_t){1})[i]
+static inline unsigned char ggml_endian_byte(int i) {
+       uint16_t tmp_val = 1;
+       return ((unsigned char *)&tmp_val)[i];
+}
+#define GGML_ENDIAN_BYTE(i) ggml_endian_byte(i)
 #define GGML_F16_VEC_STORE(p, r, i)                             \
   if (i & 0x1)                                                  \
     vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)],  \
@@ -851,13 +855,17 @@ static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
         tmp[i] = GGML_FP16_TO_FP32(x[i]);
     }
 
-    return vec_xl(0, tmp);
+    // note: keep type-cast here to prevent compiler bugs
+    // see: https://github.com/ggml-org/llama.cpp/issues/12846
+    return vec_xl(0, (const float *)(tmp));
 }
 
 static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
     float arr[4];
 
-    vec_xst(y, 0, arr);
+    // note: keep type-cast here to prevent compiler bugs
+    // see: https://github.com/ggml-org/llama.cpp/issues/12846
+    vec_xst(y, 0, (float *)(arr));
 
     for (int i = 0; i < 4; i++) {
         x[i] = GGML_FP32_TO_FP16(arr[i]);

ggml/src/ggml-cuda/cpy.cu

@@ -10,6 +10,13 @@ static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
     *dsti = *xi;
 }
 
+static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
+
+    *dsti = *xi;
+}
+
 static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
     const float * xi = (const float *) cxi;
     half * dsti = (half *) cdsti;
@@ -386,6 +393,16 @@ static void ggml_cpy_f32_f32_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
+static void ggml_cpy_f32_bf16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+}
+
 static void ggml_cpy_f32_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -534,7 +551,7 @@ static void ggml_cpy_f16_f16_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -571,7 +588,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char ** dest_ptrs_d = nullptr;
     int graph_cpynode_index = -1;
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
@@ -581,6 +598,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
@@ -617,7 +636,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
 #endif
@@ -626,7 +645,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
 }
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
@@ -634,6 +654,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         return (void*) cpy_f32_f16<cpy_1_f32_f16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {

ggml/src/ggml-cuda/cpy.cuh

@@ -2,7 +2,7 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -96,31 +96,32 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
-    auto res = hipMallocManaged(ptr, size);
-    if (res == hipSuccess) {
-        // if error we "need" to know why...
-        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
-    }
-    return res;
-#else
-
-#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
         err = cudaMallocManaged(ptr, size);
+#if defined(GGML_USE_HIP)
+        if (err == hipSuccess) {
+            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+        }
+
+        // fall back to cudaMalloc if not supported (e.g. on Windows)
+        if (err == hipErrorNotSupported) {
+            static bool warned_unsupported = false;
+            if (!warned_unsupported) {
+                GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
+                warned_unsupported = true;
+            }
+
+            err = cudaMalloc(ptr, size);
+        }
+#endif // defined(GGML_USE_HIP)
     }
     else
     {
         err = cudaMalloc(ptr, size);
     }
     return err;
-#else
-    return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIP)
-
-#endif
 }
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -2491,7 +2492,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
         if (node->op == GGML_OP_MUL_MAT_ID) {
             use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
         }
 
@@ -3079,6 +3080,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
+                    return true;
+                }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
                     return true;
                 }
@@ -3213,6 +3217,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3232,6 +3237,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek MLA
+                return false;
+            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/vendors/hip.h

@@ -71,6 +71,8 @@
 #define cudaLaunchHostFunc hipLaunchHostFunc
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMallocManaged hipMallocManaged
+#define cudaMemAdvise hipMemAdvise
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync

ggml/src/ggml-hip/CMakeLists.txt

@@ -89,10 +89,6 @@ endif()
 
 add_compile_definitions(GGML_USE_HIP)
 
-if (GGML_HIP_UMA)
-    add_compile_definitions(GGML_HIP_UMA)
-endif()
-
 if (GGML_CUDA_FORCE_MMQ)
     add_compile_definitions(GGML_CUDA_FORCE_MMQ)
 endif()

ggml/src/ggml-impl.h

@@ -16,6 +16,14 @@
 #include <arm_sve.h>
 #endif // __ARM_FEATURE_SVE
 
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
 #if defined(__F16C__)
 #include <immintrin.h>
 #endif
@@ -140,8 +148,14 @@ struct ggml_map_custom2_op_params {
 
 struct ggml_map_custom3_op_params {
     ggml_custom3_op_t fun;
-    int n_tasks;
-    void * userdata;
+    int               n_tasks;
+    void            * userdata;
+};
+
+struct ggml_custom_op_params {
+    ggml_custom_op_t fun;
+    int              n_tasks;
+    void           * userdata;
 };
 
 // bitset
@@ -311,13 +325,6 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
 // for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
 //
 #if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-
-    // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-    //
-    //   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-    //
-    #include <arm_neon.h>
-
     #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
     #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
@@ -355,8 +362,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
     #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 
     static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        register float f;
-        register double d;
+        float f;
+        double d;
         __asm__(
             "mtfprd %0,%2\n"
             "xscvhpdp %0,%0\n"
@@ -368,8 +375,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
     }
 
     static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-        register double d;
-        register ggml_fp16_t r;
+        double d;
+        ggml_fp16_t r;
         __asm__( /* xscvdphp can work on double or single precision */
             "xscvdphp %0,%2\n"
             "mffprd %1,%0\n" :

ggml/src/ggml-metal/ggml-metal.m

@@ -354,6 +354,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,
@@ -362,6 +363,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,
@@ -370,6 +372,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,
@@ -378,6 +381,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,
@@ -386,6 +390,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,
@@ -394,6 +399,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,
@@ -402,6 +408,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,
@@ -430,6 +444,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_SET_I32,
     GGML_METAL_KERNEL_TYPE_SET_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
@@ -460,6 +481,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SQRT,
     GGML_METAL_KERNEL_TYPE_SIN,
     GGML_METAL_KERNEL_TYPE_COS,
+    GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
@@ -1011,6 +1033,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,         flash_attn_ext_f16_h192,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,  flash_attn_ext_f16_hk192_hv128,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,         flash_attn_ext_f16_h256,         has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,  flash_attn_ext_f16_hk576_hv512,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,         flash_attn_ext_bf16_h64,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,         flash_attn_ext_bf16_h80,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,         flash_attn_ext_bf16_h96,         has_simdgroup_mm && use_bfloat);
@@ -1019,6 +1042,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,        flash_attn_ext_bf16_h192,        has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, flash_attn_ext_bf16_hk192_hv128, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,        flash_attn_ext_bf16_h256,        has_simdgroup_mm && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512, flash_attn_ext_bf16_hk576_hv512, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,         flash_attn_ext_q4_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,         flash_attn_ext_q4_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,         flash_attn_ext_q4_0_h96,         has_simdgroup_mm);
@@ -1027,6 +1051,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,        flash_attn_ext_q4_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, flash_attn_ext_q4_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,        flash_attn_ext_q4_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512, flash_attn_ext_q4_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,         flash_attn_ext_q4_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,         flash_attn_ext_q4_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,         flash_attn_ext_q4_1_h96,         has_simdgroup_mm);
@@ -1035,6 +1060,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,        flash_attn_ext_q4_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, flash_attn_ext_q4_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,        flash_attn_ext_q4_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512, flash_attn_ext_q4_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,         flash_attn_ext_q5_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,         flash_attn_ext_q5_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,         flash_attn_ext_q5_0_h96,         has_simdgroup_mm);
@@ -1043,6 +1069,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,        flash_attn_ext_q5_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, flash_attn_ext_q5_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,        flash_attn_ext_q5_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512, flash_attn_ext_q5_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,         flash_attn_ext_q5_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,         flash_attn_ext_q5_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,         flash_attn_ext_q5_1_h96,         has_simdgroup_mm);
@@ -1051,6 +1078,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,        flash_attn_ext_q5_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, flash_attn_ext_q5_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,        flash_attn_ext_q5_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512, flash_attn_ext_q5_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,         flash_attn_ext_q8_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,         flash_attn_ext_q8_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,         flash_attn_ext_q8_0_h96,         has_simdgroup_mm);
@@ -1059,6 +1087,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,        flash_attn_ext_q8_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,      flash_attn_ext_vec_f16_h96,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,     flash_attn_ext_vec_bf16_h96,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,     flash_attn_ext_vec_q4_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,     flash_attn_ext_vec_q4_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,     flash_attn_ext_vec_q5_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,     flash_attn_ext_vec_q5_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,     flash_attn_ext_vec_q8_0_h96,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,     flash_attn_ext_vec_f16_h128,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,    flash_attn_ext_vec_bf16_h128,    has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,    flash_attn_ext_vec_q4_0_h128,    has_simdgroup_reduction);
@@ -1087,6 +1123,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,    flash_attn_ext_vec_q5_0_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,    flash_attn_ext_vec_q5_1_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,    flash_attn_ext_vec_q8_0_h256,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,     flash_attn_ext_vec_f16_hk576_hv512,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,    flash_attn_ext_vec_bf16_hk576_hv512,    has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,    flash_attn_ext_vec_q4_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,    flash_attn_ext_vec_q4_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,    flash_attn_ext_vec_q5_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,    flash_attn_ext_vec_q5_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,    flash_attn_ext_vec_q8_0_hk576_hv512,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32,                         set_f32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32,                         set_i32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                     cpy_f32_f32,                     true);
@@ -1117,6 +1160,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT,                            sqrt,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                             sin,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
@@ -1278,6 +1322,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_NEG:
                     return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -1334,8 +1379,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
             return false;
-        case GGML_OP_POOL_2D:
         case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
+        case GGML_OP_POOL_2D:
         case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -1345,6 +1391,16 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
+            if (op->src[0]->ne[0] == 32) {
+                // head size == 32 (e.g. bert-bge-small)
+                // TODO: not sure if it is worth adding kernels for this size
+                return false;
+            }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek sizes
+                // TODO: disabled for now, until optmized
+                return false;
+            }
             if (op->src[1]->type != op->src[2]->type) {
                 return false;
             }
@@ -1957,6 +2013,18 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_NEG:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NEG].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -3837,12 +3905,14 @@ static void ggml_metal_encode_node(
                 // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
                 //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 //       these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
-                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 192)) {
+                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
@@ -3865,6 +3935,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break;
@@ -3887,6 +3959,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break;
@@ -3909,6 +3983,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break;
@@ -3931,6 +4007,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break;
@@ -3953,6 +4031,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break;
@@ -3975,6 +4055,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break;
@@ -4004,6 +4086,24 @@ static void ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 96:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 128:
                             {
                                 switch (src1->type) {
@@ -4076,12 +4176,36 @@ static void ggml_metal_encode_node(
                                         }
                                 }
                             } break;
+                        case 576:
+                            {
+                                if (ne20 == 512) {
+                                    switch (src1->type) {
+                                        case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512].pipeline; break;
+                                        default:
+                                            {
+                                                GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                                GGML_LOG_ERROR("add template specialization for this type\n");
+                                                GGML_ABORT("add template specialization for this type");
+                                            }
+                                    }
+                                } else {
+                                    GGML_LOG_ERROR("unsupported size: %lld\n", ne20);
+                                    GGML_LOG_ERROR("add template specialization for this size\n");
+                                    GGML_ABORT("add template specialization for this size");
+                                }
+                            } break;
                         default:
-                                  {
-                                      GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
-                                      GGML_LOG_ERROR("add template specialization for this size\n");
-                                      GGML_ABORT("add template specialization for this size");
-                                  }
+                            {
+                                GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                GGML_LOG_ERROR("add template specialization for this size\n");
+                                GGML_ABORT("add template specialization for this size");
+                            }
                     }
                 }
 

ggml/src/ggml-metal/ggml-metal.metal

@@ -949,6 +949,13 @@ kernel void kernel_cos(
     dst[tpig] = cos(src0[tpig]);
 }
 
+kernel void kernel_neg(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = -src0[tpig];
+}
+
 kernel void kernel_sum_rows(
         device const float * src0,
         device       float * dst,
@@ -3546,6 +3553,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h128")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_h192")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 192>;
 template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
+template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
@@ -3556,6 +3564,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3566,6 +3575,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
@@ -3575,6 +3585,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
@@ -3584,6 +3595,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
@@ -3593,6 +3605,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
@@ -3602,6 +3615,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
 
@@ -3959,6 +3973,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 96, 96, 4>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  128, 128, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 128, 128, 4>;
@@ -3999,6 +4023,16 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_
 template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 256, 256, 4>;
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 256, 256, 4>;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_hk576_hv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  576, 512, 2>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 576, 512, 2>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
+
 #undef FA_TYPES
 
 template<typename T>

ggml/src/ggml-opencl/CMakeLists.txt

@@ -54,16 +54,41 @@ function(ggml_opencl_add_kernel KNAME)
 endfunction()
 
 set(GGML_OPENCL_KERNELS
-    ggml-opencl
-    ggml-opencl_mm
-    ggml-opencl_cvt
-    ggml-opencl_gemv_noshuffle
-    ggml-opencl_gemv_noshuffle_general
-    ggml-opencl_mul_mat_Ab_Bi_8x4
-    ggml-opencl_transpose_16
-    ggml-opencl_transpose_32
-    ggml-opencl_transpose_32_16
-    ggml-opencl_im2col
+    add
+    clamp
+    cpy
+    cvt
+    diag_mask_inf
+    gelu
+    gemv_noshuffle_general
+    gemv_noshuffle
+    get_rows
+    im2col_f32
+    im2col_f16
+    mul_mat_Ab_Bi_8x4
+    mul_mv_f16_f16
+    mul_mv_f16_f32_1row
+    mul_mv_f16_f32_l4
+    mul_mv_f16_f32
+    mul_mv_f32_f32
+    mul_mv_q4_0_f32
+    mul_mv_q4_0_f32_v
+    mul_mv_q4_0_f32_8x_flat
+    mul_mv_q4_0_f32_1d_8x_flat
+    mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q6_k
+    mul
+    norm
+    relu
+    rms_norm
+    rope
+    scale
+    silu
+    softmax_4_f32
+    softmax_4_f16
+    softmax_f32
+    softmax_f16
+    transpose
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

ggml/src/ggml-opencl/kernels/add.cl

@@ -0,0 +1,83 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add
+//------------------------------------------------------------------------------
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+kernel void kernel_add(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) + *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        global float4 * src0,
+        ulong  offset0,
+        global float4 * src1,
+        ulong  offset1,
+        global float4 * dst,
+        ulong  offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] + src1[idx1];
+}

ggml/src/ggml-opencl/kernels/clamp.cl

@@ -0,0 +1,20 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// clamp
+//------------------------------------------------------------------------------
+kernel void kernel_clamp(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float min,
+        float max
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = src0[get_global_id(0)] < min ?
+        min :
+        (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]);
+}

ggml/src/ggml-opencl/kernels/cpy.cl

@@ -0,0 +1,184 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// cpy
+//------------------------------------------------------------------------------
+
+kernel void kernel_cpy_f16_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f16_f32(
+        global half * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f16(
+        global float * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -1,39 +1,20 @@
 //------------------------------------------------------------------------------
-// This file is contains additional kernels for data conversion.
+// This file is contains kernels for data conversion.
 // These kernels are used when loading the model, so its performance is less
 // important.
 //------------------------------------------------------------------------------
-#ifdef cl_khr_fp16
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
 
 #ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
 #pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
 #define INTEL_GPU 1
 #define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
 #define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
 #elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
 #define ADRENO_GPU 1
 #define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
 #endif
 
 #define QK4_0                   32
@@ -66,13 +47,44 @@ struct block_q4_0
 };
 
 //------------------------------------------------------------------------------
-// mul_vec_q_n_f32_flat_noshuffle
-//
-// This variation uses flat arrays (struct of arrays, SOA) representation for
-// quant tensors. It also uses non shuffled bit order for weights.
-//
-// The shuffled version is kept in the original file because moving it here
-// seems to result in worse performance for adreno.
+// kernel_convert_block_q4_0
+// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_0(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_0_noshuffle
+// Flatten q4_0 weights and unshuffle the bits
 //------------------------------------------------------------------------------
 
 kernel void kernel_convert_block_q4_0_noshuffle(

ggml/src/ggml-opencl/kernels/diag_mask_inf.cl

@@ -0,0 +1,58 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// diag_mask_inf kernels
+//------------------------------------------------------------------------------
+kernel void kernel_diag_mask_inf(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i02 = get_global_id(2);
+    int i01 = get_global_id(1);
+    int i00 = get_global_id(0);
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    int i = 2*get_global_id(0);
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int i4 = 4*i;
+    int i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    int i01 = i4/(ne00);      i4 -= i01*ne00;
+    int i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        (&dst[i+1])[k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            (&dst[i])[k] = -INFINITY;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/gelu.cl

@@ -0,0 +1,62 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// gelu
+//------------------------------------------------------------------------------
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+
+kernel void kernel_gelu(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_quick(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}

ggml/src/ggml-opencl/kernels/get_rows.cl

@@ -0,0 +1,163 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+#define QK4_0                   32
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+
+//------------------------------------------------------------------------------
+// dequantize_q4_0_f32, dequantize_q4_0_f16
+//------------------------------------------------------------------------------
+void dequantize_q4_0_f32(global struct block_q4_0 * xb, short il, float16 * reg) {
+    global ushort * qs = ((global ushort *)xb + 1);
+    float d1 = il ? (xb->d / 16.h) : xb->d;
+    float d2 = d1 / 256.f;
+    float md = -8.h * xb->d;
+    ushort mask0 = il ? 0x00F0 : 0x000F;
+    ushort mask1 = mask0 << 8;
+
+    reg->s0 = d1 * (qs[0] & mask0) + md;
+    reg->s1 = d2 * (qs[0] & mask1) + md;
+
+    reg->s2 = d1 * (qs[1] & mask0) + md;
+    reg->s3 = d2 * (qs[1] & mask1) + md;
+
+    reg->s4 = d1 * (qs[2] & mask0) + md;
+    reg->s5 = d2 * (qs[2] & mask1) + md;
+
+    reg->s6 = d1 * (qs[3] & mask0) + md;
+    reg->s7 = d2 * (qs[3] & mask1) + md;
+
+    reg->s8 = d1 * (qs[4] & mask0) + md;
+    reg->s9 = d2 * (qs[4] & mask1) + md;
+
+    reg->sa = d1 * (qs[5] & mask0) + md;
+    reg->sb = d2 * (qs[5] & mask1) + md;
+
+    reg->sc = d1 * (qs[6] & mask0) + md;
+    reg->sd = d2 * (qs[6] & mask1) + md;
+
+    reg->se = d1 * (qs[7] & mask0) + md;
+    reg->sf = d2 * (qs[7] & mask1) + md;
+}
+
+
+//------------------------------------------------------------------------------
+// get_rows
+//------------------------------------------------------------------------------
+kernel void kernel_get_rows_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_q4_0(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    const int NL = 2;
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) {
+        float16 temp;
+        dequantize_q4_0_f32(
+            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02)) + ind/NL, ind%NL, &temp);
+        *(((global float16 *) ((global char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl

@@ -1,146 +0,0 @@
-#ifdef cl_khr_fp16
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
-
-#ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
-#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
-#define INTEL_GPU 1
-#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
-#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
-#elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
-#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
-#define ADRENO_GPU 1
-#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
-#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
-#endif
-
-kernel void kernel_im2col_f32(
-        global float * src1,
-        ulong offset1,
-        global float * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    // threadIdx.x + blockIdx.x * blockDim.x
-    long i = get_global_id(0);
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}
-
-kernel void kernel_im2col_f16(
-        global float * src1,
-        ulong offset1,
-        global half  * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    long i = get_global_id(0);
-
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global half*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl

@@ -1,26 +0,0 @@
-// 16-bit transpose, loading/storing a 4x4 tile of elements
-
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-
-kernel void kernel_transpose_16(
-    __read_only image1d_buffer_t input,
-    __write_only image1d_buffer_t output,
-    const uint rows,
-    const uint cols
-) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-
-    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
-    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
-    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
-    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
-
-    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
-    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32.cl

@@ -1,25 +0,0 @@
-// 32-bit transpose, loading/storing a 4x4 tile of elements
-
-kernel void kernel_transpose_32(
-    __read_only image1d_buffer_t input,
-    __write_only image1d_buffer_t output,
-    const uint rows,
-    const uint cols
-) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-
-    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
-    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
-    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
-    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
-
-    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
-    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32_16.cl

@@ -1,35 +0,0 @@
-// 32-bit transpose, loading/storing a 4x4 tile of elements
-// Only used for activations
-// converts to FP16
-// also adds zero padding for non multiple of 8 prompt lengths
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-
-kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
-    half4 temp1 = {0,0,0,0};
-    half4 temp2 = {0,0,0,0};
-    half4 temp3 = {0,0,0,0};
-
-    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
-        temp0 = read_imageh(input, (j_2+0)*cols+i);
-    }
-    if((j_2+1)*cols+i*4+3 < rows*cols*16){
-        temp1 = read_imageh(input, (j_2+1)*cols+i);
-    }
-    if((j_2+2)*cols+i*4+3 < rows*cols*16){
-        temp2 = read_imageh(input, (j_2+2)*cols+i);
-    }
-    if((j_2+3)*cols+i*4+3 < rows*cols*16){
-        temp3 = read_imageh(input, (j_2+3)*cols+i);
-    }
-
-    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
-    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-}

ggml/src/ggml-opencl/kernels/im2col_f16.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f16(
+        global float * src1,
+        ulong offset1,
+        global half  * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-opencl/kernels/im2col_f32.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f32(
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-opencl/kernels/mul.cl

@@ -0,0 +1,79 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// mul
+//------------------------------------------------------------------------------
+kernel void kernel_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F16 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3)
+{
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F16;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half * y = (global half *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (half) x[i] * (half) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half  * y  = (global half  *) (src1 + offset_src1);
+            global half4 * y4 = (global half4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (half) x4[i].s0 * y4[i].s0;
+                sumf += (half) x4[i].s1 * y4[i].s1;
+                sumf += (half) x4[i].s2 * y4[i].s2;
+                sumf += (half) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (half) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += convert_float(x[i]) * y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += convert_float(x4[i].s0) * y4[i].s0;
+                sumf += convert_float(x4[i].s1) * y4[i].s1;
+                sumf += convert_float(x4[i].s2) * y4[i].s2;
+                sumf += convert_float(x4[i].s3) * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl

@@ -0,0 +1,94 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_1row(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global half  * x = (global half  *) (src0 + offset_src0);
+    global float * y = (global float *) (src1 + offset_src1);
+
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        global half4  * x4 = (global half4  *) x;
+        global float4 * y4 = (global float4 *) y;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += (float) x4[i].s0 * y4[i].s0;
+            sumf += (float) x4[i].s1 * y4[i].s1;
+            sumf += (float) x4[i].s2 * y4[i].s2;
+            sumf += (float) x4[i].s3 * y4[i].s3;
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) {
+                all_sum += (float) x[i] * y[i];
+            }
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl

@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Assumes row size (ne00) is a multiple of 4
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nrows = ne11;
+    int r0 = get_group_id(0);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half4 * x4 = (global half4 *) (src0 + offset_src0);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+        global float4 * y4 = (global float4 *) (src1 + offset_src1);
+
+        float sumf = 0;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += convert_float(x4[i].s0) * y4[i].s0;
+            sumf += convert_float(x4[i].s1) * y4[i].s1;
+            sumf += convert_float(x4[i].s2) * y4[i].s2;
+            sumf += convert_float(x4[i].s3) * y4[i].s3;
+        }
+
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F32_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f32_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F32_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global float * x = (global float *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global float4 * x4 = (global float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (float) x4[i].s0 * y4[i].s0;
+                sumf += (float) x4[i].s1 * y4[i].s1;
+                sumf += (float) x4[i].s2 * y4[i].s2;
+                sumf += (float) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl

@@ -0,0 +1,192 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32
+//------------------------------------------------------------------------------
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_4_0_dot_y(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        private float * yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float2 acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+    for (int i = 0; i < 8; i+=2) {
+        acc.s0 += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc.s1 += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (sumy * -8.f + acc.s0 + acc.s1);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];       // src1 vector cache
+    float sumf[N_DST]={0.f};
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            sumf[row] += block_q_4_0_dot_y(x+ib+row*nb, sumy, yl, il);
+        }
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float tot[N_DST] = {
+        sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]),
+        sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])};
+    for (int row = 0; row < N_DST; ++row) {
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row];
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl

@@ -0,0 +1,307 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 16 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 16
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 16x output.
+// Eeach simdgroup outputs 16 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float16 sumf = (float16)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
+                             0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  0*nb*QK4_0/2, d + ib +  0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  1*nb*QK4_0/2, d + ib +  1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  2*nb*QK4_0/2, d + ib +  2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  3*nb*QK4_0/2, d + ib +  3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  4*nb*QK4_0/2, d + ib +  4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  5*nb*QK4_0/2, d + ib +  5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  6*nb*QK4_0/2, d + ib +  6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  7*nb*QK4_0/2, d + ib +  7*nb, sumy, yl, il);
+
+        sumf.s8 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  8*nb*QK4_0/2, d + ib +  8*nb, sumy, yl, il);
+        sumf.s9 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  9*nb*QK4_0/2, d + ib +  9*nb, sumy, yl, il);
+        sumf.sa += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 10*nb*QK4_0/2, d + ib + 10*nb, sumy, yl, il);
+        sumf.sb += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 11*nb*QK4_0/2, d + ib + 11*nb, sumy, yl, il);
+
+        sumf.sc += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 12*nb*QK4_0/2, d + ib + 12*nb, sumy, yl, il);
+        sumf.sd += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 13*nb*QK4_0/2, d + ib + 13*nb, sumy, yl, il);
+        sumf.se += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 14*nb*QK4_0/2, d + ib + 14*nb, sumy, yl, il);
+        sumf.sf += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 15*nb*QK4_0/2, d + ib + 15*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float16 tot = (float16)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7),
+
+        sub_group_reduce_add(sumf.s8), sub_group_reduce_add(sumf.s9),
+        sub_group_reduce_add(sumf.sa), sub_group_reduce_add(sumf.sb),
+        sub_group_reduce_add(sumf.sc), sub_group_reduce_add(sumf.sd),
+        sub_group_reduce_add(sumf.se), sub_group_reduce_add(sumf.sf)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+
+        if (first_row + 8 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 8] = tot.s8;
+        }
+        if (first_row + 9 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 9] = tot.s9;
+        }
+        if (first_row + 10 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 10] = tot.sa;
+        }
+        if (first_row + 11 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 11] = tot.sb;
+        }
+
+        if (first_row + 12 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 12] = tot.sc;
+        }
+        if (first_row + 13 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 13] = tot.sd;
+        }
+        if (first_row + 14 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 14] = tot.se;
+        }
+        if (first_row + 15 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 15] = tot.sf;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_16x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}