common: fix issue with regex_escape routine on windows (#14133)

* ggml-cpu: Factor out feature detection build from x86

* ggml-cpu: Add ARM feature detection and scoring

This is analogous to cpu-feats-x86.cpp. However, to detect compile-time
activation of features, we rely on GGML_USE_<FEAT> which need to be set
in cmake, instead of GGML_<FEAT> that users would set for x86.

This is because on ARM, users specify features with GGML_CPU_ARM_ARCH,
rather than with individual flags.

* ggml-cpu: Implement GGML_CPU_ALL_VARIANTS for ARM

Like x86, however to pass around arch flags within cmake, we use
GGML_INTERNAL_<FEAT> as we don't have GGML_<FEAT>.

Some features are optional, so we may need to build multiple backends
per arch version (armv8.2_1, armv8.2_2, ...), and let the scoring
function sort out which one can be used.

* ggml-cpu: Limit ARM GGML_CPU_ALL_VARIANTS to Linux for now

The other platforms will need their own specific variants.

This also fixes the bug that the the variant-building branch was always
being executed as the else-branch of GGML_NATIVE=OFF. The branch is
moved to an elseif-branch which restores the previous behavior.

.github/labeler.yml

@@ -86,3 +86,10 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
+
+Ascend NPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-cann.h
+            - ggml/src/ggml-cann/**
+            - docs/backend/CANN.md

.github/workflows/build-linux-cross.yml

@@ -231,3 +231,116 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-cpu-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-vulkan-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu \
+                  libvulkan-dev:loong64
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -306,6 +306,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
+          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -687,8 +688,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
+          - build: 'cpu-x64 (static)'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
@@ -839,12 +840,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-2022-cmake-cuda:
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -878,7 +879,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/release.yml

@@ -131,8 +131,9 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
+          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
+          # - build: 'arm64'
+          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -159,6 +160,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -207,6 +211,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -373,11 +380,11 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -405,7 +412,7 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/server.yml

@@ -180,7 +180,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -159,6 +159,11 @@ if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 #
 # build the library
 #

Makefile

@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
 endif
 
 # warnings
@@ -970,7 +970,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/repack.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \

README.md

@@ -3,6 +3,7 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
@@ -28,6 +29,30 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
+## Quick start
+
+Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
+
+- Install `llama.cpp` using [brew, nix or winget](docs/install.md)
+- Run with Docker - see our [Docker documentation](docs/docker.md)
+- Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
+- Build from source by cloning this repository - check out [our build guide](docs/build.md)
+
+Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
+
+Example command:
+
+```sh
+# Use a local model file
+llama-cli -m my_model.gguf
+
+# Or download and run a model directly from Hugging Face
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+
+# Launch OpenAI-compatible API server
+llama-server -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -230,6 +255,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
+
 ## Supported backends
 
 | Backend | Target devices |
@@ -246,16 +272,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
-## Building the project
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
-
-- Clone this repository and build locally, see [how to build](docs/build.md)
-- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
-- Use a Docker image, see [documentation for Docker](docs/docker.md)
-- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
-
 ## Obtaining and quantizing models
 
 The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
@@ -263,7 +279,11 @@ 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](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), 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]`. For example:
+
+```sh
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
 
 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/`.
 

ci/run.sh

@@ -46,7 +46,20 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
+
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
+        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
+        else
+            echo "Warning: Using fallback CUDA architectures"
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
+        fi
+    else
+        echo "Error: nvidia-smi not found, cannot build with CUDA"
+        exit 1
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

common/CMakeLists.txt

@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,7 +18,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
         endif()
     endif()
 
@@ -42,7 +42,7 @@ add_custom_command(
             -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
             -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
             -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
-    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    WORKING_DIRECTORY "${PROJECT_SOURCE_DIR}"
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM
 )

common/arg.cpp

@@ -2869,6 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
             else { throw std::invalid_argument("invalid value"); }
         }

common/chat.cpp

@@ -82,10 +82,10 @@ json common_chat_msg::to_json_oaicompat() const
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg) {
     std::vector<common_chat_msg_diff> diffs;
-    // if (previous_msg.reasoning_content != current.reasoning_content) {
-    //     auto & diff = diffs.emplace_back();
-    //     diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, current.reasoning_content);
-    // }
+    if (previous_msg.reasoning_content != new_msg.reasoning_content) {
+        auto & diff = diffs.emplace_back();
+        diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, new_msg.reasoning_content);
+    }
     if (previous_msg.content != new_msg.content) {
         auto & diff = diffs.emplace_back();
         diff.content_delta = string_diff(previous_msg.content, new_msg.content);
@@ -385,9 +385,9 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
 
 template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
-    // if (!diff.reasoning_content_delta.empty()) {
-    //     delta["reasoning_content"] = msg.reasoning_content;
-    // }
+    if (!diff.reasoning_content_delta.empty()) {
+        delta["reasoning_content"] = diff.reasoning_content_delta;
+    }
     if (!diff.content_delta.empty()) {
         delta["content"] = diff.content_delta;
     }
@@ -598,6 +598,7 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
+        case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }

common/chat.h

@@ -70,7 +70,7 @@ struct common_chat_msg {
 };
 
 struct common_chat_msg_diff {
-    // std::string reasoning_content_delta;
+    std::string reasoning_content_delta;
     std::string content_delta;
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;

common/common.cpp

@@ -466,7 +466,7 @@ size_t string_find_partial_stop(const std::string_view & str, const std::string_
 
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$0");
+    return std::regex_replace(s, special_chars, "\\$&");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -934,7 +934,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -1041,7 +1041,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_self_clear(lctx);
+        llama_memory_clear(llama_get_memory(lctx), true);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);

common/common.h

@@ -215,7 +215,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {

common/speculative.cpp

@@ -144,6 +144,8 @@ llama_tokens common_speculative_gen_draft(
     auto & smpl   = spec->smpl;
     auto & prompt = spec->prompt;
 
+    auto * mem = llama_get_memory(ctx);
+
     int reuse_i = 0;
     int reuse_n = 0;
 
@@ -173,7 +175,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(mem, false);
 
         prompt.clear();
     } else {
@@ -192,14 +194,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem, 0, 0, reuse_i);
+            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
+            llama_memory_seq_rm (mem, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -556,8 +556,11 @@ class TextModel(ModelBase):
             logger.info(f"gguf: experts used count = {n_experts_used}")
 
         if (head_dim := self.hparams.get("head_dim")) is not None:
-            self.gguf_writer.add_key_length(head_dim)
-            self.gguf_writer.add_value_length(head_dim)
+            # Workaround for incorrect AutoConfig value for DeepSeekV3 (is set correctly in DeepSeekV2Model class)
+            # https://github.com/huggingface/transformers/blob/19224c3642705c5b6988c9f5f4251f83323d05ae/src/transformers/models/deepseek_v3/configuration_deepseek_v3.py#L210
+            if self.hparams.get("model_type") != "deepseek_v3":
+                self.gguf_writer.add_key_length(head_dim)
+                self.gguf_writer.add_value_length(head_dim)
 
         self.gguf_writer.add_file_type(self.ftype)
         logger.info(f"gguf: file type = {self.ftype}")
@@ -3709,8 +3712,7 @@ class BertModel(TextModel):
         self._try_set_pooling_type()
 
         if self.cls_out_labels:
-            key_name = gguf.Keys.Classifier.OUTPUT_LABELS.format(arch = gguf.MODEL_ARCH_NAMES[self.model_arch])
-            self.gguf_writer.add_array(key_name, [v for k, v in sorted(self.cls_out_labels.items())])
+            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -4799,25 +4801,6 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.intermediate_size = self.hparams["intermediate_size"]
-
-    def get_tensors(self):
-        for name, data in super().get_tensors():
-            if 'gated_layer' in name:
-                d1 = data[:self.intermediate_size, :]
-                name1 = name.replace('gated_layers', 'gated_layers_w')
-                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
-                d2 = data[self.intermediate_size:, :]
-                name2 = name.replace('gated_layers', 'gated_layers_v')
-                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
-                yield name1, d1
-                yield name2, d2
-                continue
-
-            yield name, data
-
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -4833,14 +4816,6 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "bert.", remove the prefix
-        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-        if name.startswith("bert."):
-            name = name[5:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
 
 @ModelBase.register("OpenELMForCausalLM")
 class OpenELMModel(TextModel):

docs/backend/CANN.md

@@ -8,6 +8,7 @@
  - [DataType Supports](#datatype-supports)
  - [Docker](#docker)
  - [Linux](#linux)
+ - [Environment variable setup](#environment-variable-setup)
  - [TODO](#todo)
 
 
@@ -290,5 +291,24 @@ Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang
 
 We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers from Huawei Technologies Co., Ltd for their help during the code development and pull request.
 
+## Environment variable setup
+
+### GGML_CANN_ASYNC_MODE
+
+Enables asynchronous operator submission. Disabled by default.
+
+### GGML_CANN_MEM_POOL
+
+Specifies the memory pool management strategy:
+
+- vmm: Utilizes a virtual memory manager pool. If hardware support for VMM is unavailable, falls back to the legacy (leg) memory pool.
+
+- prio: Employs a priority queue-based memory pool management.
+- leg: Uses a fixed-size buffer pool.
+
+### GGML_CANN_DISABLE_BUF_POOL_CLEAN
+
+Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
+
 ## TODO
 - Support more models and data types.

docs/build.md

@@ -1,5 +1,9 @@
 # Build llama.cpp locally
 
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
 **To get the Code:**
 
 ```bash

docs/install.md

@@ -1,28 +1,42 @@
 # Install pre-built version of llama.cpp
 
-## Homebrew
+| Install via | Windows | Mac | Linux |
+|-------------|---------|-----|-------|
+| Winget      | ✅      |      |      |
+| Homebrew    |         | ✅   | ✅   |
+| MacPorts    |         | ✅   |      |
+| Nix         |         | ✅   | ✅   |
 
-On Mac and Linux, the homebrew package manager can be used via
+## Winget (Windows)
+
+```sh
+winget install llama.cpp
+```
+
+The package is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/issues/8188
+
+## Homebrew (Mac and Linux)
 
 ```sh
 brew install llama.cpp
 ```
+
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
-## MacPorts
+## MacPorts (Mac)
 
 ```sh
 sudo port install llama.cpp
 ```
-see also: https://ports.macports.org/port/llama.cpp/details/
 
-## Nix
+See also: https://ports.macports.org/port/llama.cpp/details/
 
-On Mac and Linux, the Nix package manager can be used via
+## Nix (Mac and Linux)
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
+
 For flake enabled installs.
 
 Or
@@ -34,13 +48,3 @@ nix-env --file '<nixpkgs>' --install --attr llama-cpp
 For non-flake enabled installs.
 
 This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-## Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-
-```sh
-flox install llama-cpp
-```
-
-Flox follows the nixpkgs build of llama.cpp.

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/embedding/embedding.cpp

@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -236,9 +236,24 @@ int main(int argc, char ** argv) {
                 LOG("\n");
             }
         } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            const uint32_t n_cls_out = llama_model_n_cls_out(model);
+            std::vector<std::string> cls_out_labels;
+
+            for (uint32_t i = 0; i < n_cls_out; i++) {
+                const char * label = llama_model_cls_label(model, i);
+                const std::string label_i(label == nullptr ? "" : label);
+                cls_out_labels.emplace_back(label_i.empty() ? std::to_string(i) : label_i);
+            }
+
             for (int j = 0; j < n_embd_count; j++) {
-                // NOTE: if you change this log - update the tests in ci/run.sh
-                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                for (uint32_t i = 0; i < n_cls_out; i++) {
+                    // NOTE: if you change this log - update the tests in ci/run.sh
+                    if (n_cls_out == 1) {
+                        LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                    } else {
+                        LOG("rerank score %d: %8.3f [%s]\n", j, emb[j * n_embd + i], cls_out_labels[i].c_str());
+                    }
+                }
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding

examples/gritlm/gritlm.cpp

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
+    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
 }

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_self_clear(context)
+        llama_memory_clear(llama_get_memory(context), true)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/lookahead/lookahead.cpp

@@ -60,6 +60,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // Tokenize the prompt
@@ -94,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+        llama_memory_seq_cp(mem, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -427,17 +429,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
+        llama_memory_seq_rm(mem, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_self_seq_keep(ctx, seq_id_best);
-            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_memory_seq_keep(mem, seq_id_best);
+            llama_memory_seq_cp  (mem, seq_id_best, 0, -1, -1);
+            llama_memory_seq_rm  (mem, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+                llama_memory_seq_cp(mem, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -181,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
+        llama_memory_seq_rm(llama_get_memory(ctx), 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/parallel/parallel.cpp

@@ -194,6 +194,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
@@ -259,7 +261,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+            llama_memory_seq_cp(mem, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -286,9 +288,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_self_seq_rm(ctx, i, -1, -1);
+                llama_memory_seq_rm(mem, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                llama_memory_seq_cp(mem, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -447,8 +449,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
+                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,6 +126,8 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
+    auto * mem = llama_get_memory(ctx);
+
     // fill the KV cache
     for (int i = 0; i < n_ctx; i += n_batch) {
         if (i > 0 && n_grp > 1) {
@@ -133,10 +135,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_self_seq_add(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_self_seq_div(ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_memory_seq_add(mem, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_memory_seq_div(mem, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -166,10 +168,10 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+        llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+        llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -195,10 +197,10 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+            llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+            llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
     }
 

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_process(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), false);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/save-load-state/save-load-state.cpp

@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_self_clear(ctx3);
+        llama_memory_clear(llama_get_memory(ctx3), true);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");

examples/speculative-simple/speculative-simple.cpp

@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -142,6 +142,8 @@ int main(int argc, char ** argv) {
         }
     }
 
+    auto * mem_tgt = llama_get_memory(ctx_tgt);
+    auto * mem_dft = llama_get_memory(ctx_dft);
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -420,14 +422,14 @@ int main(int argc, char ** argv) {
             {
                 LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
 
-                llama_kv_self_seq_keep(ctx_dft, s_keep);
-                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_dft, 0);
+                llama_memory_seq_keep(mem_dft, s_keep);
+                llama_memory_seq_cp  (mem_dft, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_dft, 0);
 
-                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_self_seq_keep(ctx_tgt, s_keep);
-                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_tgt, 0);
+                llama_memory_seq_rm  (mem_tgt, s_keep, n_past_tgt, -1);
+                llama_memory_seq_keep(mem_tgt, s_keep);
+                llama_memory_seq_cp  (mem_tgt, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +446,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_memory_seq_rm(mem_dft, 0, n_past_dft, -1);
             // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode(ctx_dft, batch_dft);
 
@@ -503,8 +505,8 @@ int main(int argc, char ** argv) {
                     if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
-                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
+                        llama_memory_seq_cp(mem_dft, s, n_seq_cur, -1, -1);
 
                         // all previous tokens from this branch are now also part of the new branch
                         for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -585,9 +587,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_self_seq_keep(ctx_tgt, 0);
+            llama_memory_seq_keep(mem_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_memory_seq_cp(mem_tgt, 0, s, -1, -1);
             }
 
             // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());

ggml/CMakeLists.txt

@@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
 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_REPACK       "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})
@@ -137,7 +137,7 @@ set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
 set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
-if (WIN32)
+if (MINGW)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 

ggml/src/CMakeLists.txt

@@ -125,7 +125,6 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 
@@ -213,6 +212,7 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
+add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -270,17 +270,23 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        foreach (feat NATIVE
+                      SSE42
+                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                      AMX_TILE AMX_INT8 AMX_BF16)
+            set(GGML_${feat} OFF)
+        endforeach()
 
-    foreach (feat ${ARGN})
-        set(GGML_${feat} ON)
-    endforeach()
+        foreach (feat ${ARGN})
+            set(GGML_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -290,6 +296,8 @@ ggml_add_backend(CPU)
 if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    elseif (GGML_CPU_ARM_ARCH)
+        message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
@@ -303,8 +311,20 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             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_SYSTEM_ARCH STREQUAL "ARM" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
+        # Many of these features are optional so we build versions with popular
+        # combinations and name the backends based on the version they were
+        # first released with
+        ggml_add_cpu_backend_variant(armv8.0_1)
+        ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+        ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+        ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+        ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+        ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+        ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+        ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
     else()
-        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-cann/common.h

@@ -37,6 +37,7 @@
 #include <thread>
 #include <unistd.h>
 #include <functional>
+#include <optional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -103,6 +104,9 @@ const ggml_cann_device_info& ggml_cann_info();
 void ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
+std::optional<std::string> get_env(const std::string& name);
+bool parse_bool(const std::string& value);
+
 /**
  * @brief Abstract base class for memory pools used by CANN.
  */
@@ -354,7 +358,8 @@ struct ggml_backend_cann_context {
         : 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);
+
+        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }

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

@@ -31,6 +31,8 @@
 #include <mutex>
 #include <queue>
 #include <chrono>
+#include <unordered_set>
+#include <optional>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -93,6 +95,26 @@ int32_t ggml_cann_get_device() {
     return id;
 }
 
+/**
+ * @brief Get the value of the specified environment variable (name).
+ *        if not empty, return a std::string object
+ */
+std::optional<std::string> get_env(const std::string& name) {
+    const char* val = std::getenv(name.c_str());
+    if (!val) return std::nullopt;
+    std::string res = std::string(val);
+    std::transform(res.begin(), res.end(), res.begin(), ::tolower);
+    return res;
+}
+
+/**
+ * @brief Verify whether the environment variable is a valid value.
+ */
+bool parse_bool(const std::string& value) {
+    std::unordered_set<std::string> valid_values = {"on", "1", "yes", "y", "enable", "true"};
+    return valid_values.find(value) != valid_values.end();
+}
+
 /**
  * @brief Initialize the CANN device information.
  *
@@ -214,7 +236,7 @@ struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
      * @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;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -410,7 +432,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf(int device) : device(device) {
-        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -731,16 +753,18 @@ 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) {
-    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) {
+    std::string mem_pool_type = get_env("GGML_CANN_MEM_POOL").value_or("");
+
+    if (mem_pool_type == "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));
     }
+
+    if (ggml_cann_info().devices[device].vmm && mem_pool_type != "leg") {
+        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));
+    }
+
     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));
 }

ggml/src/ggml-common.h

@@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,3 +1,17 @@
+function(ggml_add_cpu_backend_features cpu_name arch)
+    # The feature detection code is compiled as a separate target so that
+    # it can be built without the architecture flags
+    # Since multiple variants of the CPU backend may be included in the same
+    # build, using set_source_files_properties() to set the arch flags is not possible
+    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
+    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
+endfunction()
+
 function(ggml_add_cpu_backend_variant_impl tag_name)
     if (tag_name)
         set(GGML_CPU_NAME ggml-cpu-${tag_name})
@@ -10,14 +24,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.cpp
-        ggml-cpu/ggml-cpu-aarch64.h
-        ggml-cpu/ggml-cpu-hbm.cpp
-        ggml-cpu/ggml-cpu-hbm.h
-        ggml-cpu/ggml-cpu-quants.c
-        ggml-cpu/ggml-cpu-quants.h
-        ggml-cpu/ggml-cpu-traits.cpp
-        ggml-cpu/ggml-cpu-traits.h
+        ggml-cpu/repack.cpp
+        ggml-cpu/repack.h
+        ggml-cpu/hbm.cpp
+        ggml-cpu/hbm.h
+        ggml-cpu/quants.c
+        ggml-cpu/quants.h
+        ggml-cpu/traits.cpp
+        ggml-cpu/traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -84,6 +98,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (GGML_SYSTEM_ARCH STREQUAL "ARM")
         message(STATUS "ARM detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/arm/quants.c
+            ggml-cpu/arch/arm/repack.cpp
+            )
+
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
         else()
@@ -138,6 +157,49 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
+                elseif(GGML_CPU_ALL_VARIANTS)
+                    if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+                        # Begin with the lowest baseline
+                        set(ARM_MCPU "armv8-a")
+                        set(ARCH_TAGS "")
+                        set(ARCH_DEFINITIONS "")
+
+                        # When a feature is selected, bump the MCPU to the first
+                        # version that supported it
+                        if (GGML_INTERNAL_DOTPROD)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                        endif()
+                        if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                        endif()
+                        if (GGML_INTERNAL_SVE)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                        endif()
+                        if (GGML_INTERNAL_MATMUL_INT8)
+                            set(ARM_MCPU "armv8.6-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                        endif()
+                        if (GGML_INTERNAL_SVE2)
+                            set(ARM_MCPU "armv8.6-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                        endif()
+                        if (GGML_INTERNAL_SME)
+                            set(ARM_MCPU "armv9.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                        endif()
+
+                        list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                        ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
+                    endif()
                 endif()
             endif()
 
@@ -167,6 +229,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/x86/quants.c
+            ggml-cpu/arch/x86/repack.cpp
+            )
+
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -296,21 +363,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
                 message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
             endif()
-
-            # The feature detection code is compiled as a separate target so that
-            # it can be built without the architecture flags
-            # Since multiple variants of the CPU backend may be included in the same
-            # build, using set_source_files_properties() to set the arch flags is not possible
-            set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
-            target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
-            set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
         if (GGML_NATIVE)
             if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
                 file(READ "/proc/cpuinfo" POWER10_M)
@@ -318,7 +375,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
             endif()
 
-            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
@@ -337,6 +395,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
         message(STATUS "loongarch64 detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
+
         list(APPEND ARCH_FLAGS -march=loongarch64)
         if (GGML_LASX)
             list(APPEND ARCH_FLAGS -mlasx)
@@ -346,6 +406,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
         message(STATUS "riscv64 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/riscv/quants.c
+            ggml-cpu/arch/riscv/repack.cpp
+            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -357,6 +421,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
         string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
@@ -380,12 +445,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_VXE)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
         endif()
+    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
+        message(STATUS "Wasm detected")
+        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(STATUS "Unknown architecture")
+        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
+        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
     endif()
 
-    if (GGML_CPU_AARCH64)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    if (GGML_CPU_REPACK)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
     endif()
 
     if (GGML_CPU_KLEIDIAI)

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

@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,7 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>

ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp

@@ -0,0 +1,94 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__aarch64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM (1 << 13)
+#endif
+
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1 << 23)
+#endif
+
+struct aarch64_features {
+    // has_neon not needed, aarch64 has NEON guaranteed
+    bool has_dotprod     = false;
+    bool has_fp16_va     = false;
+    bool has_sve         = false;
+    bool has_sve2        = false;
+    bool has_i8mm        = false;
+    bool has_sme         = false;
+
+    aarch64_features() {
+#if defined(__linux__)
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_sve     = !!(hwcap & HWCAP_SVE);
+        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
+        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+        has_sme     = !!(hwcap2 & HWCAP2_SME);
+#elif defined(__APPLE__)
+        int oldp = 0;
+        size_t size = sizeof(oldp);
+
+        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
+            has_dotprod = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
+            has_i8mm = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
+            has_sme = static_cast<bool>(oldp);
+        }
+
+        // Apple apparently does not implement SVE yet
+#endif
+    }
+};
+
+static int ggml_backend_cpu_aarch64_score() {
+    int score = 1;
+    aarch64_features af;
+
+#ifdef GGML_USE_DOTPROD
+    if (!af.has_dotprod) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
+    if (!af.has_fp16_va) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_USE_SVE
+    if (!af.has_sve) { return 0; }
+    score += 1<<3;
+#endif
+#ifdef GGML_USE_MATMUL_INT8
+    if (!af.has_i8mm) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_USE_SVE2
+    if (!af.has_sve2) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_USE_SME
+    if (!af.has_sme) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
+
+# endif // defined(__aarch64__)

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -0,0 +1,396 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+            for (int l = 0; l < nb; l++) {
+                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
+                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
+                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
+                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
+                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
+                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
+                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
+                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
+                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
+
+                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
+                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                // vector version needs Zvfhmin extension
+                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float b_scales[8] = {
+                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                };
+                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
+                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
+            }
+            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
+        }
+        return;
+    }
+
+#endif
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                for (int l = 0; l < nb; l++) {
+                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                    // vector version needs Zvfhmin extension
+                    const float a_scales[4] = {
+                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                    };
+                    const float b_scales[8] = {
+                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    };
+                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+
+                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
+                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
+                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
+                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l0;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l0 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
+                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
+                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
+                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
+                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l1;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l1 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
+                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
+                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
+                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
+                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l2;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l2 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
+                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
+                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
+                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
+                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l3;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l3 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
+                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
+                    }
+                }
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
+            }
+        }
+
+        return;
+    }
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}

ggml/src/ggml-cpu/common.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
 

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -1,8 +0,0 @@
-#pragma once
-
-#include "ggml-cpu-traits.h"
-#include "ggml.h"
-
-// GGML internal header
-
-ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

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

@@ -506,3 +506,28 @@ void ggml_barrier(struct ggml_threadpool * tp);
 #ifdef __cplusplus
 }
 #endif
+
+#define GGML_DO_PRAGMA_(x) _Pragma (#x)
+#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
+// Note for Apple targets:
+// - clang: aliases are not supported on darwin
+// - all native kernels need to be implemented in both x86 and arm files
+// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
+# define GGML_WEAK_ALIAS(name, alias)
+#elif defined(__GNUC__)
+// GCC/Clang on *nix
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
+#elif defined(_MSC_VER) && defined(_WIN64)
+// MSVC
+// Note: C name mangling varies across different calling conventions
+// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
+#elif defined(_MSC_VER) && defined(WIN32)
+// ref: https://github.com/ggml-org/whisper.cpp/pull/3239#issuecomment-2958224591
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:_" #name "=_" #alias))
+#else
+# error "Unsupported compiler for GGML_WEAK_ALIAS"
+#endif
+
+#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)

ggml/src/ggml-cpu/ggml-cpu-quants.h

@@ -1,63 +0,0 @@
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML CPU internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-#ifdef __cplusplus
-}
-#endif

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

@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"

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

@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -596,8 +596,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });

ggml/src/ggml-cpu/hbm.cpp

@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "ggml-cpu-hbm.h"
+#include "hbm.h"
 
 // buffer type HBM
 

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

@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"
 

ggml/src/ggml-cpu/ops.cpp

@@ -8132,8 +8132,8 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
         #define WKV_VECTOR_SIZE 4
     #endif
 
-    int wkv_vector_size;
     #ifdef WKV_VECTOR_SIZE
+        int wkv_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             wkv_vector_size = svcntw();
         #else
@@ -8348,8 +8348,8 @@ static void ggml_compute_forward_gla_f32(
         #define GLA_VECTOR_SIZE 4
     #endif
 
-    int gla_vector_size;
     #ifdef GLA_VECTOR_SIZE
+        int gla_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             gla_vector_size = svcntw();
         #else

ggml/src/ggml-cpu/quants.h

@@ -0,0 +1,116 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy,  size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#if defined(GGML_CPU_GENERIC)
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,119 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+// Workaround for clang:
+// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
+// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif // !defined(__clang__)
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(GGML_CPU_GENERIC)
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

ggml/src/ggml-cpu/traits.cpp

@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"

ggml/src/ggml-cuda/common.cuh

@@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
             KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
         }

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

@@ -615,9 +615,8 @@ static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemsetAsync(ctx->dev_ptr, value, buffer->size, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
@@ -1144,7 +1143,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
     const char * src_ptr = (const char *) src->data;
     char       * dst_ptr = (char       *) dst;
 
@@ -1427,8 +1425,6 @@ static void ggml_cuda_op_mul_mat(
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
     ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
     ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
 
@@ -1750,7 +1746,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.

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

@@ -4766,6 +4766,8 @@ static bool ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    const int is_q = ggml_is_quantized(src1->type) ? 1 : 0;
+
                     // 2*(2*ncpsg + nqptg)*(nsg)
                     // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
                     //
@@ -4773,7 +4775,7 @@ static bool ggml_metal_encode_node(
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(2*ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -4810,9 +4812,9 @@ static bool ggml_metal_encode_node(
                     // and store the soft_max values and the mask
                     //
                     // ne00*(nsg)
-                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
+                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {

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

@@ -3328,14 +3328,12 @@ kernel void kernel_flash_attn_ext(
     constexpr short NW  = N_SIMDWIDTH;
     constexpr short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
 
-    const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
-    const short T  = DK + 2*TS; // shared memory size per query in (half)
+    const short TS = nsg*SH;      // shared memory size per query in (s_t == float)
+    const short T  = 2*DK + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*DK); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*DK); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*DK); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*DK); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +                0*DK); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +                0*DK); // same as above but in q4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + 2*Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
     threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
@@ -3354,7 +3352,7 @@ kernel void kernel_flash_attn_ext(
             if (iq1 + j < args.ne01) {
                 sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*DK4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = 0;
             }
         }
     }
@@ -3548,20 +3546,20 @@ kernel void kernel_flash_attn_ext(
 
             // O = diag(ms)*O
             {
-                s8x8_t mm;
-                simdgroup_load(mm, ss + 2*C, TS, 0, false);
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    simdgroup_multiply(lo[i], mm, lo[i]);
+                    simdgroup_multiply(lo[i], ms, lo[i]);
                 }
             }
 
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/8; ++cc) {
-                    s8x8_t ms;
-                    simdgroup_load(ms, ss + 8*cc, TS, 0, false);
+                    s8x8_t vs;
+                    simdgroup_load(vs, ss + 8*cc, TS, 0, false);
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
@@ -3572,7 +3570,7 @@ kernel void kernel_flash_attn_ext(
                             v8x8_t mv;
                             simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20
 
-                            simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
+                            simdgroup_multiply_accumulate(lo[i], vs, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < DV16; ii += 4) {
@@ -3593,10 +3591,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
                                 if (ii + tx < DV16) {
@@ -3611,10 +3609,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             }
                         }
@@ -3624,93 +3622,89 @@ kernel void kernel_flash_attn_ext(
         }
 
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
-        for (short j = 0; j < Q; ++j) {
-            if (tiisg == 0) {
-                ss[j*TS + 0] = S[j];
-                ss[j*TS + 1] = M[j];
-            }
+        for (short j = tiisg; j < Q; j += NW) {
+            ss[j*TS + 0] = S[j];
+            ss[j*TS + 1] = M[j];
         }
     }
 
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup float  * so  = (threadgroup float  *) (shmem_f16 + 0*DK); // reuse query data for accumulation
+    threadgroup float4 * so4 = (threadgroup float4 *) (shmem_f16 + 0*DK);
+
+    // store result to shared memory in F32
+    if (sgitg == 0) {
+        for (short i = 0; i < DV8; ++i) {
+            //simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            simdgroup_float8x8 t(1.0f);
+            simdgroup_multiply(t, lo[i], t);
+            simdgroup_store(t, so + i*8, DV, 0, false);
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        float S = { 0.0f };
-        float M = { -__FLT_MAX__/2 };
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // each simdgroup stores its output to shared memory, reusing sq
         if (sgitg == sg) {
-            for (short i = 0; i < DV8; ++i) {
-                simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            for (short j = tiisg; j < Q; j += NW) {
+                const float S0 = ss[j*TS - 1*SH + 0];
+                const float S1 = ss[j*TS        + 0];
+
+                const float M0 = ss[j*TS - 1*SH + 1];
+                const float M1 = ss[j*TS        + 1];
+
+                const float M = max(M0, M1);
+
+                float ms0 = exp(M0 - M);
+                float ms1 = exp(M1 - M);
+
+                const float S = S0*ms0 + S1*ms1;
+
+                ss[j*TS + 0] = S;
+                ss[j*TS + 1] = M;
+
+                ss[j*TS + 2*C + j - 1*SH] = ms0;
+                ss[j*TS + 2*C + j       ] = ms1;
             }
-        }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // the first simdgroup accumulates the results from the other simdgroups
-        if (sgitg == 0) {
-            for (short j = 0; j < Q; ++j) {
-                const float S0 = ss[j*TS +         0];
-                const float S1 = ss[j*TS + sg*SH + 0];
-
-                const float M0 = ss[j*TS +         1];
-                const float M1 = ss[j*TS + sg*SH + 1];
-
-                M = max(M0, M1);
-
-                const float ms0 = exp(M0 - M);
-                const float ms1 = exp(M1 - M);
-
-                S = S0*ms0 + S1*ms1;
-
-                if (tiisg == 0) {
-                    ss[j*TS + 0] = S;
-                    ss[j*TS + 1] = M;
-
-                    ss[j*TS + 2*C + j        ] = ms0;
-                    ss[j*TS + 2*C + j + sg*SH] = ms1;
-                }
-            }
+            //simdgroup_barrier(mem_flags::mem_threadgroup);
 
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
             {
                 s8x8_t ms0;
                 s8x8_t ms1;
 
-                simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
-                simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
+                simdgroup_load(ms0, ss + 2*C - 1*SH, TS, 0, false);
+                simdgroup_load(ms1, ss + 2*C,        TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    o8x8_t t;
+                    simdgroup_float8x8 t;
 
                     simdgroup_load    (t, so + i*8, DV, 0, false);
-                    simdgroup_multiply(t, ms1, t);
+                    simdgroup_multiply(t, ms0, t);
 
-                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                    simdgroup_multiply_accumulate(t, ms1, lo[i], t);
+                    simdgroup_store(t, so + i*8, DV, 0, false);
                 }
             }
         }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // store result to shared memory (reuse sq)
-    if (sgitg == 0) {
-        for (short i = 0; i < DV8; ++i) {
-            simdgroup_store(lo[i], so + i*8, DV, 0, false);
-        }
-    }
-
-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*(nsg-1)*SH + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
-    if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
-            const float S = ss[j*TS + 0];
+    for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
+        const float S = 1.0f/sf[j*TS + 0];
 
-            for (short i = tiisg; i < DV4; i += NW) {
-                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S;
-            }
+        device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
+
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[i] = (float4) so4[j*DV4 + i]*S;
         }
     }
 }
@@ -3719,12 +3713,22 @@ kernel void kernel_flash_attn_ext(
 //       template to be able to explore different combinations
 //
 #define FA_TYPES \
-    half,  half4,   simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    float,          simdgroup_float8x8, \
-    float,          simdgroup_float8x8, \
-    half,  half4,   simdgroup_half8x8
+    float,  float4,    simdgroup_float8x8, \
+    half,   half4x4,   simdgroup_half8x8,  \
+    half,   half4x4,   simdgroup_half8x8,  \
+    float,             simdgroup_float8x8, \
+    float,             simdgroup_float8x8, \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
@@ -3739,15 +3743,15 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 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>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
-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>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, 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_BF, 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_BF, 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_BF, 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>;
@@ -3801,6 +3805,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_at
 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
+#undef FA_TYPES_BF
 
 template<
     typename q4_t,  // query types in shared memory
@@ -3847,12 +3852,12 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                  0*DK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                  0*DK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 + sgitg*SH       + Q*DK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 + sgitg*SH       + Q*DK); // same as above but in s4_t
-    threadgroup float * sm  = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
-    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + sgitg*DV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*DK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*DK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*DK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*DK); // same as above but in s4_t
+    threadgroup float * sm  = (threadgroup float *) (shmem_f16 +   sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
+    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*DV       + Q*T);  // scratch buffer for the results
 
     // store the result for all queries in local memory (the O matrix from the paper)
     o4_t lo[DV4/NL];
@@ -4157,7 +4162,7 @@ kernel void kernel_flash_attn_ext_vec(
            half4,  \
     float,         \
     float, float4, \
-           half4
+           float4
 
 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;
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -80,6 +80,7 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_id_q4_0_f32_8x_flat
     mul
     norm
     relu
@@ -95,6 +96,12 @@ set(GGML_OPENCL_KERNELS
     sub
     sum_rows
     transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    pad
+    repeat
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -315,6 +315,13 @@ struct ggml_backend_opencl_context {
     cl_program program_softmax_4_f16;
     cl_program program_argsort_f32_i32;
     cl_program program_sum_rows_f32;
+    cl_program program_repeat;
+    cl_program program_pad;
+    cl_program program_tanh;
+    cl_program program_upscale;
+    cl_program program_concat;
+    cl_program program_tsembd;
+    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
@@ -351,6 +358,16 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
+    cl_kernel kernel_repeat;
+    cl_kernel kernel_pad;
+    cl_kernel kernel_tanh_f32_nd;
+    cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_upscale;
+    cl_kernel kernel_upscale_bilinear;
+    cl_kernel kernel_concat_f32_contiguous;
+    cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_timestep_embedding;
+    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -1097,6 +1114,166 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // repeat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "repeat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("repeat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_repeat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_repeat = clCreateKernel(backend_ctx->program_repeat, "kernel_repeat", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: repeat kernel source not found or empty. Repeat operations will not be available.\n");
+            backend_ctx->program_repeat = nullptr;
+            backend_ctx->kernel_repeat = nullptr;
+        }
+    }
+
+    // pad
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "pad.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("pad.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_pad =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_pad = clCreateKernel(backend_ctx->program_pad, "kernel_pad", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: pad kernel source not found or empty. Pad operations will not be available.\n");
+            backend_ctx->program_pad = nullptr;
+            backend_ctx->kernel_pad = nullptr;
+        }
+    }
+
+    // tanh
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tanh.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tanh.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tanh =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_tanh_f32_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_tanh_f16_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: tanh kernel source not found or empty. Tanh operation will not be available.\n");
+            backend_ctx->program_tanh = nullptr;
+            backend_ctx->kernel_tanh_f32_nd = nullptr;
+            backend_ctx->kernel_tanh_f16_nd = nullptr;
+        }
+    }
+
+    // upscale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "upscale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("upscale.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_upscale =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_upscale = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale", &err), err));
+            if (backend_ctx->program_upscale) {
+                 cl_int err_bilinear;
+                 backend_ctx->kernel_upscale_bilinear = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale_bilinear", &err_bilinear);
+                 if (err_bilinear != CL_SUCCESS) {
+                    GGML_LOG_WARN("ggml_opencl: kernel_upscale_bilinear not found in upscale.cl. Bilinear upscale will not be available. Error: %d\n", err_bilinear);
+                    backend_ctx->kernel_upscale_bilinear = nullptr;
+                 }
+            } else {
+                backend_ctx->kernel_upscale_bilinear = nullptr;
+            }
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: upscale kernel source not found or empty. Upscale operations will not be available.\n");
+            backend_ctx->program_upscale = nullptr;
+            backend_ctx->kernel_upscale = nullptr;
+            backend_ctx->kernel_upscale_bilinear = nullptr;
+        }
+    }
+
+    // concat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "concat.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("concat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_concat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+            CL_CHECK((backend_ctx->kernel_concat_f32_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_contiguous", &err), err));
+            CL_CHECK((backend_ctx->kernel_concat_f32_non_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_non_contiguous", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: concat kernel source not found or empty. Concat operations will not be available.\n");
+            backend_ctx->program_concat = nullptr;
+            backend_ctx->kernel_concat_f32_contiguous = nullptr;
+            backend_ctx->kernel_concat_f32_non_contiguous = nullptr;
+        }
+    }
+
+    // timestep_embedding
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tsembd.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("tsembd.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tsembd =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_timestep_embedding = clCreateKernel(backend_ctx->program_tsembd, "kernel_timestep_embedding", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: timestep_embedding kernel source not found or empty. This op will not be available.\n");
+            backend_ctx->program_tsembd = nullptr;
+            backend_ctx->kernel_timestep_embedding = nullptr;
+        }
+    }
+
+    // mul_mv_id_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat, "kernel_mul_mv_id_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1863,7 +2040,12 @@ static bool ggml_backend_opencl_cpy_tensor_async(ggml_backend_t backend, const g
 }
 
 static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+
+    cl_event evt;
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, 0, nullptr, &evt));
+    CL_CHECK(clWaitForEvents(1, &evt));
+    CL_CHECK(clReleaseEvent(evt));
 }
 
 // Syncronizes the 'backend_ctx's device with others so that commands
@@ -1976,9 +2158,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU_QUICK:
-                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 case GGML_UNARY_OP_SIGMOID:
                     return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_TANH:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
                 default:
                     return false;
             }
@@ -1988,6 +2173,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
             return true;
+        case GGML_OP_REPEAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
+        case GGML_OP_PAD:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   op->src[0]->ne[3] == 1 && op->ne[3] == 1;
+        case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONCAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_MUL_MAT:
@@ -2000,6 +2196,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
+        case GGML_OP_MUL_MAT_ID:
+            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+                if (op->src[1]->type == GGML_TYPE_F32) {
+                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+                }
+            }
+            return false;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -2052,7 +2255,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
     /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
     /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
-    /* .synchronize             = */ NULL,  /* ggml_backend_opencl_synchronize */
+    /* .synchronize             = */ ggml_backend_opencl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_update       = */ NULL,
@@ -4108,6 +4311,536 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0,
 #endif
 }
 
+static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_tanh_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_tanh_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_tanh");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0]; const int ne01 = src0->ne[1]; const int ne02 = src0->ne[2]; const int ne03 = src0->ne[3];
+    const cl_ulong nb00 = src0->nb[0]; const cl_ulong nb01 = src0->nb[1]; const cl_ulong nb02 = src0->nb[2]; const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0]; const int ne11 = dst->ne[1]; const int ne12 = dst->ne[2]; const int ne13 = dst->ne[3];
+    const cl_ulong nb10 = dst->nb[0]; const cl_ulong nb11 = dst->nb[1]; const cl_ulong nb12 = dst->nb[2]; const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == src0->type);
+
+    UNUSED(src1_shape_def);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_repeat == nullptr) {
+        GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int src0_ne0 = src0->ne[0]; const int src0_ne1 = src0->ne[1]; const int src0_ne2 = src0->ne[2]; const int src0_ne3 = src0->ne[3];
+    const cl_ulong src0_nb0 = src0->nb[0]; const cl_ulong src0_nb1 = src0->nb[1]; const cl_ulong src0_nb2 = src0->nb[2]; const cl_ulong src0_nb3 = src0->nb[3];
+
+    const int dst_ne0 = dst->ne[0]; const int dst_ne1 = dst->ne[1]; const int dst_ne2 = dst->ne[2]; const int dst_ne3 = dst->ne[3];
+    const cl_ulong dst_nb0 = dst->nb[0]; const cl_ulong dst_nb1 = dst->nb[1]; const cl_ulong dst_nb2 = dst->nb[2]; const cl_ulong dst_nb3 = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_repeat;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &src0_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &src0_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &src0_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &src0_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong),  &src0_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong),  &src0_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &src0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &src0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &dst_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &dst_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &dst_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dst_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &dst_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &dst_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &dst_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &dst_nb3));
+
+    size_t gws0 = dst_ne1 > 0 ? (size_t)dst_ne1 : 1;
+    size_t gws1 = dst_ne2 > 0 ? (size_t)dst_ne2 : 1;
+    size_t gws2 = dst_ne3 > 0 ? (size_t)dst_ne3 : 1;
+
+    size_t global_work_size[] = { gws0, gws1, gws2 };
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_pad == nullptr) {
+        GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int s_ne0 = src0->ne[0];
+    const int s_ne1 = src0->ne[1];
+    const int s_ne2 = src0->ne[2];
+
+    const int d_ne0 = dst->ne[0];
+    const int d_ne1 = dst->ne[1];
+    const int d_ne2 = dst->ne[2];
+
+    cl_kernel kernel = backend_ctx->kernel_pad;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &s_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &s_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &s_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne2));
+
+    size_t lws0 = 64;
+    size_t gws0 = (( (size_t)d_ne0 + lws0 - 1 ) / lws0) * lws0;
+
+    size_t global_work_size[] = { gws0, (size_t)d_ne1, (size_t)d_ne2 };
+    size_t local_work_size[]  = { lws0, 1, 1 };
+
+    size_t * local_work_size_ptr = local_work_size;
+     if (d_ne0 % lws0 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    cl_kernel kernel = nullptr;
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        kernel = backend_ctx->kernel_upscale;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: nearest upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        kernel = backend_ctx->kernel_upscale_bilinear;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: bilinear upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else {
+        GGML_LOG_WARN("%s: unsupported upscale mode %d, skipping OpenCL execution.\n", __func__, mode);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne00_src = src0->ne[0];
+    const int ne01_src = src0->ne[1];
+
+    const int ne10_dst = dst->ne[0];
+    const int ne11_dst = dst->ne[1];
+    const int ne12_dst = dst->ne[2];
+    const int ne13_dst = dst->ne[3];
+
+    const float sf0 = (float)dst->ne[0] / src0->ne[0];
+    const float sf1 = (float)dst->ne[1] / src0->ne[1];
+    const float sf2 = (float)dst->ne[2] / src0->ne[2];
+    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00_src));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01_src));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+    }
+
+
+    size_t dst_total_elements = (size_t)ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    if (dst_total_elements == 0) {
+        return;
+    }
+    size_t global_work_size[] = { dst_total_elements, 1, 1 };
+    size_t local_work_size_pref = 256;
+    size_t local_work_size[] = { MIN(local_work_size_pref, dst_total_elements), 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (dst_total_elements % local_work_size[0] != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], 1, 1};
+    size_t profiling_lws[3] = {local_work_size_ptr ? local_work_size[0] : 0, 1, 1};
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_concat_f32_contiguous == nullptr || backend_ctx->kernel_concat_f32_non_contiguous == nullptr) {
+        GGML_LOG_WARN("%s: concat kernels not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra0_cl = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1_cl = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad_cl = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra0_cl->offset + src0->view_offs;
+    cl_ulong off_src1 = extra1_cl->offset + src1->view_offs;
+    cl_ulong off_dst  = extrad_cl->offset + dst->view_offs;
+
+    const int32_t dim = ((const int32_t *) dst->op_params)[0];
+    GGML_ASSERT(dim >= 0 && dim <= 3);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+        if (dim == 3) {
+
+            size_t nbytes_src0 = ggml_nbytes(src0);
+            size_t nbytes_src1 = ggml_nbytes(src1);
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra0_cl->data_device, extrad_cl->data_device,
+                                         off_src0, off_dst, nbytes_src0, 0, NULL, NULL));
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra1_cl->data_device, extrad_cl->data_device,
+                                         off_src1, off_dst + nbytes_src0, nbytes_src1, 0, NULL, NULL));
+        } else {
+
+            cl_kernel kernel = backend_ctx->kernel_concat_f32_contiguous;
+            size_t global_work_size[3];
+
+            for (int i3 = 0; i3 < dst->ne[3]; ++i3) {
+                cl_ulong current_off_src0 = off_src0 + (i3 * src0->nb[3]);
+                cl_ulong current_off_src1 = off_src1 + (i3 * src1->nb[3]);
+                cl_ulong current_off_dst  = off_dst  + (i3 * dst->nb[3]);
+
+                int d_ne00 = src0->ne[0]; int d_ne01 = src0->ne[1]; int d_ne02 = src0->ne[2];
+                int d_ne10 = src1->ne[0]; int d_ne11 = src1->ne[1]; int d_ne12 = src1->ne[2];
+                int d_ne0  = dst->ne[0];  int d_ne1  = dst->ne[1];  int d_ne2  = dst->ne[2];
+
+                CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &current_off_src0));
+                CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &current_off_src1));
+                CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &current_off_dst));
+                CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &d_ne00));
+                CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne01));
+                CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne02));
+                CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne10));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &d_ne11));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &d_ne12));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &d_ne0));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &d_ne1));
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &d_ne2));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dim));
+
+                global_work_size[0] = d_ne0;
+                global_work_size[1] = d_ne1;
+                global_work_size[2] = d_ne2;
+
+                CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+            }
+        }
+    } else {
+        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
+
+        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+
+        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+
+        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
+
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_src1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
+
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),    &nb03));
+
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),    &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),    &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
+
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
+        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong),    &d_nb3));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &dim));
+
+        size_t global_work_size_nc[] = { d_ne1 > 0 ? (size_t)d_ne1 : 1,
+                                         d_ne2 > 0 ? (size_t)d_ne2 : 1,
+                                         d_ne3 > 0 ? (size_t)d_ne3 : 1 };
+
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size_nc, NULL, 0, NULL, NULL));
+    }
+}
+
+static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_timestep_embedding == nullptr) {
+        GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int logical_dim = dst->op_params[0];
+    const int max_period  = dst->op_params[1];
+    const int dst_nb1_bytes = dst->nb[1];
+
+    cl_kernel kernel = backend_ctx->kernel_timestep_embedding;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &dst_nb1_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &logical_dim));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &max_period));
+
+    size_t gws0 = (size_t)(((logical_dim + 1) / 2) + 1);
+
+    size_t gws1 = (size_t)src0->ne[0];
+
+    size_t global_work_size[] = {gws0, gws1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, &evt)); // Pass 2 for 2D problem
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], global_work_size[1], 1};
+    size_t profiling_lws[3] = {0,0,0}; // Reflects NULL LWS
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, NULL)); // Pass 2 for 2D problem
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4828,6 +5561,136 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     }
 }
 
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+#endif
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int ne20 = src2->ne[0];
+    const int ne21 = src2->ne[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    const int r2 = ne12/ne02;
+    const int r3 = ne13/ne03;
+    const int dst_rows = ne20*ne21; // ne20 = n_used_experts, ne21 = n_rows
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int sgs   = 32; // subgroup size
+    int nsg   = 1;  // number of subgroups
+    int nrows = 1;  // number of row in src1
+    int ndst  = 4;  // number of values produced by each subgroup
+
+    cl_kernel kernel;
+
+    // subgroup mat vec
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0: {
+            kernel = backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &r3));
+
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");;
+    }
+
+    int _ne1 = 1;
+    int ne123 = dst_rows;
+
+    size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123};
+    size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -5667,6 +6530,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_sigmoid;
                     break;
+                case GGML_UNARY_OP_TANH:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_tanh;
+                    break;
                 default:
                     return false;
             } break;
@@ -5694,12 +6563,48 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_group_norm;
             break;
+                case GGML_OP_REPEAT:
+             if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_repeat;
+            break;
+        case GGML_OP_PAD:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_pad(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_UPSCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_upscale(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_CONCAT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_concat;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_timestep_embedding(backend, tensor->src[0], tensor);
+            return true;
         case GGML_OP_MUL_MAT:
             if (!any_on_device && !ggml_cl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
                 return false;
             }
             func = ggml_cl_mul_mat;
             break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul_mat_id;
+            break;
         case GGML_OP_SCALE:
             if (!any_on_device) {
                 return false;

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

@@ -0,0 +1,109 @@
+kernel void kernel_concat_f32_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
+    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
+    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
+    int dim
+) {
+    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
+    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
+    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+
+    int i0 = get_global_id(0); // Index along dst's 0th dimension
+    int i1 = get_global_id(1); // Index along dst's 1st dimension
+    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+
+    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
+        return;
+    }
+
+    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
+    ulong src_idx;
+
+    if (dim == 0) {
+        if (i0 < d_ne00) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 1) {
+        if (i1 < d_ne01) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 2) {
+        if (i2 < d_ne02) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+
+            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    }
+}
+
+kernel void kernel_concat_f32_non_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+
+    long ne00, long ne01, long ne02, long ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
+
+    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
+    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
+    int dim
+) {
+    global const char * src0_base = p_src0 + off_src0;
+    global const char * src1_base = p_src1 + off_src1;
+    global char * dst_base        = p_dst + off_dst;
+
+    long current_i1 = get_global_id(0); // Index for dst_dim_1
+    long current_i2 = get_global_id(1); // Index for dst_dim_2
+    long current_i3 = get_global_id(2); // Index for dst_dim_3
+
+    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
+        return;
+    }
+
+    global const float * x_val_ptr;
+    global float * y_val_ptr;
+
+    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
+        bool use_src0;
+        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
+
+        if (dim == 0) {
+            use_src0 = (current_i0 < ne00);
+            if (!use_src0) { s_i0 = current_i0 - ne00; }
+        } else if (dim == 1) {
+            use_src0 = (current_i1 < ne01);
+            if (!use_src0) { s_i1 = current_i1 - ne01; }
+        } else if (dim == 2) {
+            use_src0 = (current_i2 < ne02);
+            if (!use_src0) { s_i2 = current_i2 - ne02; }
+        } else { // dim == 3
+            use_src0 = (current_i3 < ne03);
+            if (!use_src0) { s_i3 = current_i3 - ne03; }
+        }
+
+        if (use_src0) {
+            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+        } else {
+            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+        }
+
+        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
+        *y_val_ptr = *x_val_ptr;
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl

@@ -0,0 +1,283 @@
+#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
+
+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];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float 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);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * 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 ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    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;
+    float8 sumf = 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 += 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 += 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 += 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 += 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 += 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 += 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 += 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 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        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)
+    );
+
+    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;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

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

@@ -0,0 +1,30 @@
+kernel void kernel_pad(
+        global const void * src0_ptr,
+        ulong src0_offset,
+        global void * dst_ptr,
+        ulong dst_offset,
+        int s_ne0, int s_ne1, int s_ne2,
+        int d_ne0, int d_ne1, int d_ne2
+) {
+    global const float * src0 = (global const float *)((global const char *)src0_ptr + src0_offset);
+    global float * dst = (global float *)((global char *)dst_ptr + dst_offset);
+
+    int nidx   = get_global_id(0);
+    int idx_d1 = get_group_id(1);
+    int idx_d2 = get_group_id(2);
+
+    if (nidx >= d_ne0) {
+        return;
+    }
+
+    int dst_el_offset = nidx + idx_d1 * d_ne0 + idx_d2 * d_ne0 * d_ne1;
+
+    bool in_src_bounds = (nidx < s_ne0) && (idx_d1 < s_ne1) && (idx_d2 < s_ne2);
+
+    if (in_src_bounds) {
+        int src_el_offset = nidx + idx_d1 * s_ne0 + idx_d2 * s_ne0 * s_ne1;
+        dst[dst_el_offset] = src0[src_el_offset];
+    } else {
+        dst[dst_el_offset] = 0.0f;
+    }
+}

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

@@ -0,0 +1,39 @@
+kernel void kernel_repeat(
+    global const char * src0_data_in,
+    global       char * dst_data_in,
+    ulong src0_offset,
+    ulong dst_offset,
+    int src0_ne0, int src0_ne1, int src0_ne2, int src0_ne3,
+    ulong src0_nb0, ulong src0_nb1, ulong src0_nb2, ulong src0_nb3,
+    int dst_ne0, int dst_ne1, int dst_ne2, int dst_ne3,
+    ulong dst_nb0, ulong dst_nb1, ulong dst_nb2, ulong dst_nb3
+) {
+    global const char * src0_data = src0_data_in + src0_offset;
+    global       char * dst_data  = dst_data_in + dst_offset;
+
+    const int d3 = get_global_id(2);
+    const int d2 = get_global_id(1);
+    const int d1 = get_global_id(0);
+
+    if (d3 >= dst_ne3 || d2 >= dst_ne2 || d1 >= dst_ne1) {
+        return;
+    }
+
+    const int s3 = d3 % src0_ne3;
+    const int s2 = d2 % src0_ne2;
+    const int s1 = d1 % src0_ne1;
+
+    const global char * p_src0_slice = src0_data + (ulong)s3*src0_nb3 + (ulong)s2*src0_nb2 + (ulong)s1*src0_nb1;
+    global char * p_dst_slice  = dst_data  + (ulong)d3*dst_nb3 + (ulong)d2*dst_nb2 + (ulong)d1*dst_nb1;
+
+    for (int d0 = 0; d0 < dst_ne0; ++d0) {
+        // Determine source index for dimension 0 based on tiling/broadcasting.
+        const int s0 = d0 % src0_ne0;
+
+        const global char * restrict current_src_el_ptr = p_src0_slice + (ulong)s0*src0_nb0;
+        global char * restrict current_dst_el_ptr  = p_dst_slice  + (ulong)d0*dst_nb0;
+        for (int k = 0; k < src0_nb0; ++k) {
+            current_dst_el_ptr[k] = current_src_el_ptr[k];
+        }
+    }
+}

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

@@ -0,0 +1,63 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#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
+
+kernel void kernel_tanh_f32_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    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 i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_tanh_f16_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    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 i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}

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

@@ -0,0 +1,48 @@
+kernel void kernel_timestep_embedding(
+    global const void * p_timesteps,
+    ulong off_timesteps,
+    global void * p_dst,
+    ulong off_dst,
+    int dst_nb1_bytes,
+    int logical_dim,
+    int max_period
+) {
+    int local_i;
+    int local_j;
+    int local_half_dim;
+    float local_timestep_val;
+    float local_freq;
+    float local_arg;
+    global float * local_embed_data_ptr;
+    global const float * local_timesteps_input_ptr;
+    global float * local_dst_output_base_ptr;
+
+    local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps);
+    local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst);
+
+    local_i = get_global_id(1);
+    local_j = get_global_id(0);
+
+    local_half_dim = logical_dim / 2;
+    local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
+
+    if (logical_dim % 2 != 0 && local_j == ((logical_dim + 1) / 2)) {
+        local_embed_data_ptr[logical_dim] = 0.0f;
+    }
+
+    if (local_j >= local_half_dim) {
+        return;
+    }
+
+    local_timestep_val = local_timesteps_input_ptr[local_i];
+
+    if (local_half_dim == 0) {
+        local_freq = 1.0f;
+    } else {
+        local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim);
+    }
+
+    local_arg = local_timestep_val * local_freq;
+    local_embed_data_ptr[local_j] = cos(local_arg);
+    local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg);
+}

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

@@ -0,0 +1,121 @@
+kernel void kernel_upscale(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne10,
+    int ne11,
+    int ne12,
+    int ne13,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10 * ne11 * ne12 * ne13;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = index / (ne10 * ne11 * ne12);
+
+    int i00 = (int)(i10 / sf0);
+    int i01 = (int)(i11 / sf1);
+    int i02 = (int)(i12 / sf2);
+    int i03 = (int)(i13 / sf3);
+
+    ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00;
+    global const float * src_element_ptr = (global const float *)(src_base + offset_src_element);
+
+    dst_base[index] = *src_element_ptr;
+}
+
+kernel void kernel_upscale_bilinear(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne00_src,
+    int ne01_src,
+    int ne10_dst,
+    int ne11_dst,
+    int ne12_dst,
+    int ne13_dst,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10_dst = index % ne10_dst;
+    int i11_dst = (index / ne10_dst) % ne11_dst;
+    int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    int i02_src = (int)(i12_dst / sf2);
+    int i03_src = (int)(i13_dst / sf3);
+
+    const float pixel_offset = 0.5f;
+
+    float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    long y0_src = (long)floor(y_src_f);
+    long y1_src = y0_src + 1;
+
+    y0_src = max(0L, min(y0_src, (long)ne01_src - 1));
+    y1_src = max(0L, min(y1_src, (long)ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    long x0_src = (long)floor(x_src_f);
+    long x1_src = x0_src + 1;
+
+    x0_src = max(0L, min(x0_src, (long)ne00_src - 1));
+    x1_src = max(0L, min(x1_src, (long)ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst_base[index] = result;
+}

ggml/src/ggml-quants.c

@@ -2425,8 +2425,6 @@ void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK4_NL == 0);
     const int64_t nb = k / QK4_NL;

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -53,6 +53,9 @@ struct socket_t {
     }
 };
 
+// macro for nicer error messages on server crash
+#define RPC_STATUS_ASSERT(x) if (!(x)) GGML_ABORT("Remote RPC server crashed or returned malformed response")
+
 // all RPC structures must be packed
 #pragma pack(push, 1)
 // ggml_tensor is serialized into rpc_tensor
@@ -425,7 +428,7 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 static bool check_server_version(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_hello_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.major != RPC_PROTO_MAJOR_VERSION || response.minor > RPC_PROTO_MINOR_VERSION) {
         fprintf(stderr, "RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
         return false;
@@ -481,7 +484,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_free_buffer_req request = {ctx->remote_ptr};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     delete ctx;
 }
 
@@ -493,7 +496,7 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
     rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
     rpc_msg_buffer_get_base_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     ctx->base_ptr = reinterpret_cast<void *>(response.base_ptr);
     return ctx->base_ptr;
 }
@@ -545,7 +548,7 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
         request.tensor = serialize_tensor(tensor);
 
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0);
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
     }
     return GGML_STATUS_SUCCESS;
 }
@@ -560,7 +563,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
         request.hash = fnv_hash((const uint8_t*)data, size);
         rpc_msg_set_tensor_hash_rsp response;
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
         if (response.result) {
             // the server has the same data, no need to send it
             return;
@@ -573,7 +576,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
@@ -583,7 +586,7 @@ static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, con
     request.offset = offset;
     request.size = size;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
@@ -601,7 +604,7 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
     request.dst = serialize_tensor(dst);
     rpc_msg_copy_tensor_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.result;
 }
 
@@ -609,7 +612,7 @@ static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
@@ -635,7 +638,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
     rpc_msg_alloc_buffer_rsp response;
     auto sock = get_socket(buft_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
@@ -650,7 +653,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
 static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_alignment_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.alignment;
 }
 
@@ -662,7 +665,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ
 static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_max_size_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.max_size;
 }
 
@@ -683,7 +686,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty
 
         rpc_msg_get_alloc_size_rsp response;
         bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
 
         return response.alloc_size;
     } else {
@@ -761,7 +764,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
     rpc_msg_graph_compute_rsp response;
     auto sock = get_socket(rpc_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return (enum ggml_status)response.result;
 }
 
@@ -835,7 +838,7 @@ bool ggml_backend_is_rpc(ggml_backend_t backend) {
 static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
     rpc_msg_get_device_memory_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     *free = response.free_mem;
     *total = response.total_mem;
 }

ggml/src/ggml-sycl/common.hpp

@@ -149,8 +149,6 @@ typedef sycl::float2 dfloat2;
 
 #define MMVQ_MAX_BATCH_SIZE  8
 
-static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 static int g_all_sycl_device_count = -1;
 static bool g_ggml_backend_sycl_buffer_type_initialized = false;
 

ggml/src/ggml-sycl/convert.cpp

@@ -265,6 +265,17 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+        [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
+}
+
 template <typename dst_t>
 static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
                                         dpct::queue_ptr stream) {
@@ -530,7 +541,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_sycl;
         case GGML_TYPE_IQ1_M:
@@ -587,7 +602,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_sycl;
         case GGML_TYPE_IQ1_M:

ggml/src/ggml-sycl/cpy.cpp

@@ -1,8 +1,12 @@
 #include "cpy.hpp"
 
 #include <float.h>
+#include <string>
 
 #include "dequantize.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
+#include "ggml.h"
 
 static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) {
@@ -116,6 +120,15 @@ static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
     }
 }
 
+/* quantized type same copy */
+template<typename T>
+static void cpy_blck_q_q(const char * cxi, char * cdsti) {
+    const T * xi = (const T *) cxi;
+    T * dsti = (T *) cdsti;
+    *dsti = *xi;
+}
+
+
 static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *) (cdsti);
 
@@ -311,6 +324,34 @@ template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const
     }
 }
 
+
+template <typename T, int qk>
+static void cpy_q_q(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,
+                      const sycl::nd_item<3> & item_ct1) {
+    const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03      = i / (ne00 * ne01 * ne02);
+    const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
+    const int i00      = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00;
+    const int x_offset = (i00 / qk) * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03;
+
+
+    const int i13        = i / (ne10 * ne11 * ne12);
+    const int i12        = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11);
+    const int i11        = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10;
+    const int i10        = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10;
+    const int dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13;
+
+    cpy_blck_q_q<T>(cx + x_offset, cdst + dst_offset);
+}
+
 template <cpy_kernel_t cpy_blck, int qk>
 static void cpy_f32_q(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,
@@ -322,6 +363,7 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00
         return;
     }
 
+
     const int i03      = i / (ne00 * ne01 * ne02);
     const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
     const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
@@ -615,6 +657,70 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
     }
 }
 
+static void ggml_cpy_q8_0_q8_0(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, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_0_q5_0(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, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_1_q5_1(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, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_0_q4_0(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, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_1_q4_1(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, queue_ptr stream) {
+
+   const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+   stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
     scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
@@ -632,8 +738,10 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
 
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
-
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+    if ((src0->type == src1->type) && (ggml_is_contiguous(src0) && ggml_is_contiguous(src1))) {
+        GGML_SYCL_DEBUG("%s: memcpy path\n", __func__);
+        main_stream->memcpy(src1_ddc, src0_ddc, ggml_nbytes(src0));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
@@ -684,6 +792,16 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
         ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                  nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_q8_0_q8_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_q5_0_q5_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_Q5_1) {
+        ggml_cpy_q5_1_q5_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_q4_0_q4_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_q4_1_q4_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else {
         GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type),
                        ggml_type_name(src1->type));

ggml/src/ggml-sycl/dequantize.hpp

@@ -538,6 +538,38 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_block_q6_K_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                          const sycl::nd_item<3> & item_ct1, int64_t n_blocks) {
+    const int64_t ib = item_ct1.get_group(2);
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t ip  = tid / 32;       // ip is 0 or 1
+    const int64_t il  = tid - 32 * ip;  // 0...32
+    const int64_t is  = 8 * ip + il / 16;
+
+    const uint8_t *   base_ptr           = static_cast<const uint8_t *>(vx);
+    const auto        ql_offset          = ib * (QK_K / 2);
+    const auto        qh_offset          = (QK_K / 2) * n_blocks + (QK_K / 4) * ib;
+    const auto        base_scales_offset = (QK_K / 2) * n_blocks + (QK_K / 4) * n_blocks + (QK_K / 16) * ib;
+    const auto        base_d_offset      = ((QK_K / 2) + (QK_K / 4) + (QK_K / 16)) * n_blocks;
+    const uint8_t *   ql_ptr             = base_ptr + ql_offset;
+    const uint8_t *   qh_ptr             = base_ptr + qh_offset;
+    const uint8_t *   scales_ptr         = base_ptr + base_scales_offset;
+    const ggml_half * d                  = (const ggml_half *) (base_ptr + base_d_offset) + ib;
+
+    dst_t * y = yy + ib * QK_K + 128 * ip + il;
+
+    const uint8_t * ql = ql_ptr + 64 * ip + il;
+    const uint8_t   qh = *(qh_ptr + 32 * ip + il);
+    const int8_t *  sc = reinterpret_cast<const int8_t *>(scales_ptr + is);
+
+    y[0]  = *d * sc[0] * ((int8_t) ((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = *d * sc[2] * ((int8_t) ((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = *d * sc[4] * ((int8_t) ((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = *d * sc[6] * ((int8_t) ((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
 template<typename dst_t>
 static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                      const sycl::nd_item<3> &item_ct1,

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -354,7 +354,8 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         assert(tensor->view_src->buffer->buft == buffer->buft);
         return GGML_STATUS_SUCCESS;
     }
-    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K) && !g_ggml_sycl_disable_optimize) {
+    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K || tensor->type == GGML_TYPE_Q6_K) &&
+        !g_ggml_sycl_disable_optimize) {
         ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
         tensor->extra                 = extra;
         ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
@@ -2989,6 +2990,7 @@ inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
         case GGML_TYPE_Q4_0:
             return true;
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
             return !g_ggml_sycl_prioritize_dmmv;
         default:
             return false;
@@ -3008,6 +3010,7 @@ inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
             return true;
         default:
             return false;
@@ -3092,6 +3095,50 @@ static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, d
     sycl::free(tmp_buf, *stream);
 }
 
+static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
+    GGML_ASSERT(size % sizeof(block_q6_K) == 0);
+    GGML_ASSERT(offset % sizeof(block_q6_K) == 0);
+
+    const int nblocks = size / sizeof(block_q6_K);
+
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+
+    auto *       ql_ptr     = data_device;
+    auto *       qh_ptr     = ql_ptr + (QK_K / 2) * nblocks;
+    auto *       scales_ptr = qh_ptr + (QK_K / 4) * nblocks;
+    sycl::half * dm_ptr     = (sycl::half *) (scales_ptr + (QK_K / 16) * nblocks);
+
+    stream
+        ->parallel_for(nblocks,
+                       [=](auto i) {
+                           const block_q6_K * x  = (const block_q6_K *) tmp_buf;
+                           const int          ib = i;
+
+                           const uint8_t * ql              = x[ib].ql;
+                           const uint8_t * qh              = x[ib].qh;
+                           uint8_t *       base_ql_ptr     = ql_ptr + (QK_K / 2) * ib;
+                           uint8_t *       base_qh_ptr     = qh_ptr + (QK_K / 4) * ib;
+                           uint8_t *       base_scales_ptr = scales_ptr + (QK_K / 16) * ib;
+
+                           for (int j = 0; j < QK_K / 2; ++j) {
+                               base_ql_ptr[j] = ql[j];
+                           }
+                           for (int j = 0; j < QK_K / 4; ++j) {
+                               base_qh_ptr[j] = qh[j];
+                           }
+
+                           for (int j = 0; j < QK_K / 16; ++j) {
+                               base_scales_ptr[j] = x[ib].scales[j];
+                           }
+
+                           dm_ptr[ib] = x[ib].d;
+                       })
+        .wait_and_throw();
+
+    sycl::free(tmp_buf, *stream);
+}
+
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
     uint8_t * data_device = (uint8_t *) src0->data;
     size_t ncols = src0->ne[0];
@@ -3105,6 +3152,9 @@ static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
         case GGML_TYPE_Q4_K:
             reorder_qw_q4_k(data_device, size, 0, stream);
             break;
+        case GGML_TYPE_Q6_K:
+            reorder_qw_q6_k(data_device, size, 0, stream);
+            break;
         default:
             GGML_ABORT("reorder_qw() called with unsupported type");
             break;
@@ -4226,6 +4276,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
+                if (src0_type == src1_type && (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) && src0_type != GGML_TYPE_BF16) {
+                    return true;
+                }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
@@ -4271,6 +4324,21 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
+                if(src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
                 return false;
             }
         case GGML_OP_CONCAT:

ggml/src/ggml-sycl/mmvq.cpp

@@ -31,11 +31,10 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
 
     float partial_sum = 0.0f;
     for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
-        const int ibx       = row * blocks_per_row + i;  // x block index
-        // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits
-        const int bx_offset = block_type::get_block_offset(ibx);
-        const int d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+        const int ibx = row * blocks_per_row + i;  // x block index
 
+        const auto         bx_offset      = block_type::get_block_offset(ibx, nblocks);
+        const auto         d_offset       = block_type::get_d_offset(nrows, ncols, ibx);
         // Y block index that aligns with ibx
         const int iby = i * block_type::block_to_q8_1_ratio();
         const int8_t* q8_1_quant_ptr = (const int8_t*)vy + iby * QK8_1;
@@ -46,7 +45,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
             // x block quant index when casting the quants to int
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs, nblocks);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
         }
     }
 
@@ -785,6 +784,24 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
+static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+                                               const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int        block_num_y   = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
+    });
+}
 static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -1070,7 +1087,14 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
             case GGML_TYPE_Q6_K:
-                mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q6_k_q8_1_sycl\n");
+                    reorder_mul_mat_vec_q6_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q6_k_q8_1_sycl\n");
+                    mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
                 break;
             case GGML_TYPE_IQ1_S:
                 mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);

ggml/src/ggml-sycl/quants.hpp

@@ -14,12 +14,13 @@
 #ifndef GGML_SYCL_QUANTS_HPP
 #define GGML_SYCL_QUANTS_HPP
 
+#include <utility>
+
 #include "ggml-common.h"
 #include "ggml.h"
 
 namespace ggml_sycl_reordered {
 
-
 // The reordered block moves quants (qs) and  scales(d) to two
 // uniform regions of memory that is contiguous in the same tensor.
 // What this means is that instead of having:
@@ -32,7 +33,6 @@ namespace ggml_sycl_reordered {
 
 template <ggml_type type> struct block_q_t;
 
-
 // qk number of weights / quants in a block
 // qr number of weights in a byte (described as 'before dequantization')
 //    for quantization types that has low and high bits split, qr is calculated with
@@ -47,10 +47,12 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
         static constexpr uint32_t vdr_mmvq = 2;
     };
 
-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }
 
-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
-        return (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half);
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        return { (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half), 0 };
     }
 
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
@@ -64,20 +66,46 @@ template <> struct block_q_t<GGML_TYPE_Q4_K> {
         static constexpr uint32_t vdr_mmvq = 2;
     };
 
-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }
 
-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
         auto nblocks = (nrows * (ncols / traits::qk));
-        return (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2));
+        return { nblocks * (QK_K / 2),
+                 (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2)) };
     }
 
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 
     constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
-
-    constexpr size_t get_dm_offset(int nblocks) { return get_total_qs_bytes(nblocks) + nblocks * K_SCALE_SIZE; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q6_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI6_K;
+        static constexpr uint32_t qr       = QR6_K;
+        static constexpr uint32_t vdr_mmvq = 1;
+    };
+
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
+        auto low_bits_index  = block_index * (traits::qk / traits::qr);
+        // the index of high bits it's after all low bits
+        auto high_bits_index = n_blocks * (QK_K / 2) + (block_index * (QK_K / 4));
+        return { low_bits_index, high_bits_index };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / traits::qk));
+        auto total_qs_bytes = nblocks * (QK_K / 2) + nblocks * (QK_K / 4);
+        auto block_scales   = total_qs_bytes + block_index * (QK_K / 16);
+        auto sb_scale       = total_qs_bytes + nblocks * (QK_K / 16);
+        return { block_scales, sb_scale };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
 }  // namespace ggml_sycl_reordered
 
 #endif  // GGML_SYCL_QUANTS_HPP

ggml/src/ggml-sycl/vecdotq.hpp

@@ -284,10 +284,11 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
         return d4 * (sumi * ds8f.x() - (8 * q4_0_traits::vdr_mmvq / q4_0_traits::qi) * ds8f.y());
     }
 
-    __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const int8_t* q8_1_quant_ptr, const sycl::half2* q8_1_ds, const int & iqs, int /* nblocks */) {
-        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
-        const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset.first;
+        const ggml_half d = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset.first));
         int             v[q4_0_traits::vdr_mmvq];
         int             u[2 * q4_0_traits::vdr_mmvq];
 
@@ -346,15 +347,15 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
     using q4_k_traits = typename q4_k_block::traits;
 
-    float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const int8_t* q8_1_quant_ptr, const sycl::half2* q8_1_ds, const int & iqs, int nblocks) {
-        const int ib = ibx_offset / (QK_K / 2);
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);
 
         const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
-        const uint8_t *    qs             = base + ibx_offset;
-        const int          total_qs_bytes = nblocks * (QK_K / 2);
-        const uint8_t *    scs            = base + total_qs_bytes + ib * K_SCALE_SIZE;
-        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset);
+        const uint8_t *    qs             = base + ibx_offset.first;
+        const uint8_t *    scs            = base + d_offset.first + ib * K_SCALE_SIZE;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
 
         const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
         const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
@@ -395,6 +396,66 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     }
 };
 
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q6_K;
+
+    using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
+    using q6_k_traits = typename q6_k_block::traits;
+
+    __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int vl, const int vh, const int * __restrict__ u,
+                                                      const int8_t * __restrict__ scales, const float d,
+                                                      const float * __restrict__ d8) {
+        float sumf = 0.0f;
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            const int sc = scales[4 * i];
+
+            const int vil = (vl >> (4 * i)) & 0x0F0F0F0F;
+
+            const int vih = ((vh >> (4 * i)) << 4) & 0x30303030;
+
+            const int vi = dpct::vectorized_binary<sycl::char4>((vil | vih), 0x20202020,
+                                                                dpct::sub_sat());  // vi = (vil | vih) - 32
+
+            sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc);                        // SIMD dot product
+        }
+
+        return d * sumf;
+    }
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr, const sycl::half2 * q8_1_ds,
+                     const int iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);
+
+        const uint8_t *   base   = static_cast<const uint8_t *>(vbq);
+        const uint8_t *   ql     = base + ibx_offset.first;
+        const uint8_t *   qh     = base + ibx_offset.second;
+        const int8_t *    scales = reinterpret_cast<const int8_t *>(base + d_offset.first);
+        const ggml_half * d      = (const ggml_half *) (base + d_offset.second) + ib;
+
+        const int bq8_offset   = 2 * QR6_K * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 4);
+        const int scale_offset = (QI6_K / 4) * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 8);
+        const int vh_shift     = 2 * ((iqs % (QI6_K / 2)) / (QI6_K / 4));
+
+        const int vl = get_int_from_uint8(ql, iqs);
+        const int vh = get_int_from_uint8(qh, (QI6_K / 4) * (iqs / (QI6_K / 2)) + iqs % (QI6_K / 4)) >> vh_shift;
+
+        const int8_t * scs = scales + scale_offset;
+
+        int   u[QR6_K];
+        float d8[QR6_K];
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            u[i] = get_int_from_int8_aligned(q8_1_quant_ptr + (bq8_offset + 2 * i) * QK8_1, iqs % QI8_1);
+            const sycl::half2 ds_values = *(q8_1_ds + bq8_offset + 2 * i);
+            d8[i]                       = ds_values[0];
+        }
+        return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scs, *d, d8);
+    }
+};
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 

ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp

@@ -0,0 +1,98 @@
+#version 450
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};     // dst - result      [KL, Cout]
+
+layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+} p;
+
+
+uint32_t Cout_idx = gl_WorkGroupID.x;
+const uint32_t bs = gl_WorkGroupSize.x;
+uint32_t tid = gl_LocalInvocationID.x;
+// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
+uint32_t tmp_len = bs*p.s0+p.K;
+shared D_TYPE tmp[4096];
+
+uint splitWork(uint workSize){
+    return (bs + workSize -1) / bs;
+}
+
+void main(){
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        if(idx < tmp_len){
+            tmp[idx] = 0.0;
+        }
+    }
+
+    uint32_t L_blocks = splitWork(p.L);
+    for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
+        if(L_block_id > 0){
+            barrier();
+            // Shift values in tmp to the current processing window
+            for(int i = 0; i < splitWork(tmp_len); i++){
+                uint32_t idx = i*bs+tid;
+                if(idx >= bs*p.s0 && idx < tmp_len){
+                    tmp[idx-bs*p.s0] = tmp[idx];
+                    tmp[idx] = 0.0;
+                }else if(idx >= p.K && idx < bs*p.s0){
+                    tmp[idx] = 0.0;
+                }
+            }
+        }
+        barrier();
+
+        // Save contributions of the block to tmp
+        uint32_t L_idx = L_block_id*bs + tid;
+        for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
+            D_TYPE dp = 0.0;
+            for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
+                A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
+                if(L_idx < p.L){
+                    B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
+                    dp = fma(elemKrn, elemInp, dp);
+                }
+            }
+            tmp[tid*p.s0 + K_idx] += dp;
+            barrier();
+        }
+
+        // Save the computed values except the last block that can have different size
+        uint32_t KLb_idx = L_block_id*bs*p.s0;
+        if(L_block_id < L_blocks-1){
+            for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
+                uint32_t sh_idx = p.s0*tid+s0_idx;
+                uint32_t KL_idx = KLb_idx+sh_idx;
+                if(KL_idx < p.KL){
+                    data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
+                }
+            }
+        }
+    }
+
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
+        if(KL_idx < p.KL){
+            data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -622,6 +622,8 @@ void process_shaders() {
 
     string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"},  {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+
     string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

gguf-py/gguf/gguf_writer.py

@@ -935,6 +935,9 @@ class GGUFWriter:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
+        self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
+
     # for vision models
 
     def add_clip_has_vision_encoder(self, value: bool) -> None:

gguf-py/gguf/tensor_mapping.py

@@ -333,7 +333,9 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
-            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2 (split up/gate, no longer used)
+            "encoder.layer.{bid}.mlp.gated_layers",                   # jina-bert-v2 (GEGLU)
+            "encoder.layer.{bid}.mlp.up_gated_layer",                 # jina-v2-code (GEGLU)
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
@@ -370,7 +372,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
-            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2 (split up/gate, no longer used)
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone