gguf-py : avoid requiring pyside6 for other scripts (#13036 )

- gguf-py : remove gguf-py/gguf/scripts/__init__.py because it's not needed Implicit namespaces are supported since Python 3.3 (https://peps.python.org/pep-0420/), and the entrypoints in pyproject.toml can directly refer to the main functions.
CUDA: fix logic for clearing padding with -ngl 0 (#13320 )
2026-05-07 01:24:24 +00:00 · 2025-05-05 22:27:31 -04:00 · 2025-05-05 22:32:13 +02:00 · 2025-05-05 22:12:19 +02:00 · 2025-05-05 16:03:31 +02:00 · 2025-05-05 16:02:55 +02:00
412 changed files with 22067 additions and 12857 deletions
--- a/.clang-tidy
+++ b/.clang-tidy
@@ -13,6 +13,7 @@ Checks: >
    -readability-magic-numbers,
    -readability-uppercase-literal-suffix,
    -readability-simplify-boolean-expr,
+    -readability-math-missing-parentheses,
    clang-analyzer-*,
    -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
    performance-*,
--- a/.devops/cpu.Dockerfile
+++ b/.devops/cpu.Dockerfile
@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .

 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
    elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
    else \
        echo "Unsupported architecture"; \
        exit 1; \
--- a/.devops/cuda.Dockerfile
+++ b/.devops/cuda.Dockerfile
@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
    export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
    cmake --build build --config Release -j$(nproc)

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

 RUN mkdir -p /app/lib && \
--- a/.devops/llama-cli-cann.Dockerfile
+++ b/.devops/llama-cli-cann.Dockerfile
@@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH

 RUN echo "Building with static libs" && \
    source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
    cmake --build build --config Release --target llama-cli

 # TODO: use image with NNRT
--- a/.devops/musa.Dockerfile
+++ b/.devops/musa.Dockerfile
@@ -35,7 +35,7 @@ COPY . .
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
--- a/.devops/rocm.Dockerfile
+++ b/.devops/rocm.Dockerfile
@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .

 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
    && cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib \
--- a/.devops/vulkan.Dockerfile
+++ b/.devops/vulkan.Dockerfile
@@ -16,7 +16,7 @@ WORKDIR /app

 COPY . .

-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
--- a/.editorconfig
+++ b/.editorconfig
@@ -21,15 +21,15 @@ indent_style = tab
 [prompts/*.txt]
 insert_final_newline = unset

-[examples/server/public/*]
+[tools/server/public/*]
 indent_size = 2

-[examples/server/public/deps_*]
+[tools/server/public/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset

-[examples/server/deps_*]
+[tools/server/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset
@@ -37,7 +37,7 @@ indent_size = unset
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab

-[examples/cvector-generator/*.txt]
+[tools/cvector-generator/*.txt]
 trim_trailing_whitespace = unset
 insert_final_newline = unset

--- a/.flake8
+++ b/.flake8
@@ -2,8 +2,9 @@
 max-line-length = 125
 ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503
 exclude =
-    # Do not traverse examples
+    # Do not traverse examples and tools
    examples,
+    tools,
    # Do not include package initializers
    __init__.py,
    # No need to traverse our git directory
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -45,7 +45,9 @@ build:
            - CMakePresets.json
 examples:
    - changed-files:
-        - any-glob-to-any-file: examples/**
+        - any-glob-to-any-file:
+            - examples/**
+            - tools/**
 devops:
    - changed-files:
        - any-glob-to-any-file:
@@ -70,7 +72,7 @@ android:
 server:
    - changed-files:
        - any-glob-to-any-file:
-            - examples/server/**
+            - tools/server/**
 ggml:
    - changed-files:
        - any-glob-to-any-file:
--- a/.github/workflows/bench.yml.disabled
+++ b/.github/workflows/bench.yml.disabled
@@ -27,10 +27,10 @@ on:
  push:
    branches:
      - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
  pull_request_target:
    types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
  schedule:
    -  cron: '04 2 * * *'

@@ -69,7 +69,7 @@ jobs:
      - name: Install python env
        id: pipenv
        run: |
-          cd examples/server/bench
+          cd tools/server/bench
          python3 -m venv venv
          source venv/bin/activate
          pip install -r requirements.txt
@@ -79,7 +79,7 @@ jobs:
        run: |
          wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
          tar xzf prometheus*.tar.gz --strip-components=1
-          ./prometheus --config.file=examples/server/bench/prometheus.yml &
+          ./prometheus --config.file=tools/server/bench/prometheus.yml &
          while ! nc -z localhost 9090; do
            sleep 0.1
          done
@@ -92,7 +92,7 @@ jobs:
      - name: Install k6 and xk6-sse
        id: k6_installation
        run: |
-          cd examples/server/bench
+          cd tools/server/bench
          go install go.k6.io/xk6/cmd/xk6@latest
          xk6 build master \
              --with github.com/phymbert/xk6-sse
@@ -116,7 +116,7 @@ jobs:
      - name: Download the dataset
        id: download_dataset
        run: |
-          cd examples/server/bench
+          cd tools/server/bench
          wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json

      - name: Server bench
@@ -126,7 +126,7 @@ jobs:
        run: |
          set -eux

-          cd examples/server/bench
+          cd tools/server/bench
          source venv/bin/activate
          python bench.py \
              --runner-label ${{ env.RUNNER_LABEL }} \
@@ -157,9 +157,9 @@ jobs:
          name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
          compression-level: 9
          path: |
-            examples/server/bench/*.jpg
-            examples/server/bench/*.json
-            examples/server/bench/*.log
+            tools/server/bench/*.jpg
+            tools/server/bench/*.json
+            tools/server/bench/*.log

      - name: Commit status
        uses: Sibz/github-status-action@v1
@@ -178,17 +178,17 @@ jobs:
        with:
          client_id: ${{secrets.IMGUR_CLIENT_ID}}
          path: |
-            examples/server/bench/prompt_tokens_seconds.jpg
-            examples/server/bench/predicted_tokens_seconds.jpg
-            examples/server/bench/kv_cache_usage_ratio.jpg
-            examples/server/bench/requests_processing.jpg
+            tools/server/bench/prompt_tokens_seconds.jpg
+            tools/server/bench/predicted_tokens_seconds.jpg
+            tools/server/bench/kv_cache_usage_ratio.jpg
+            tools/server/bench/requests_processing.jpg

      - name: Extract mermaid
        id: set_mermaid
        run: |
          set -eux

-          cd examples/server/bench
+          cd tools/server/bench
          PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
          echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
          echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV
--- a/.github/workflows/build-linux-cross.yml
+++ b/.github/workflows/build-linux-cross.yml
@@ -4,18 +4,25 @@ on:
  workflow_call:

 jobs:
-  ubuntu-latest-riscv64-cpu-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-cpu-cross:
+    runs-on: ubuntu-24.04

    steps:
      - uses: actions/checkout@v4
      - name: Setup Riscv
        run: |
          sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
          sudo apt-get install -y --no-install-recommends \
                  build-essential \
                  gcc-14-riscv64-linux-gnu \
@@ -27,6 +34,7 @@ jobs:
          cmake -B build -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=riscv64 \
@@ -40,21 +48,25 @@ jobs:

          cmake --build build --config Release -j $(nproc)

-  ubuntu-latest-riscv64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-vulkan-cross:
+    runs-on: ubuntu-24.04

    steps:
      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
      - name: Setup Riscv
        run: |
          sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
          sudo apt-get install -y --no-install-recommends \
                  build-essential \
                  glslc \
@@ -69,6 +81,7 @@ jobs:
                         -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=riscv64 \
@@ -82,21 +95,25 @@ jobs:

          cmake --build build --config Release -j $(nproc)

-  ubuntu-latest-arm64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-arm64-vulkan-cross:
+    runs-on: ubuntu-24.04

    steps:
      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
      - name: Setup Arm64
        run: |
          sudo dpkg --add-architecture arm64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
          sudo apt-get install -y --no-install-recommends \
                  build-essential \
                  glslc \
@@ -110,6 +127,7 @@ jobs:
                         -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=aarch64 \
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -633,6 +633,7 @@ jobs:
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=iOS \
@@ -669,6 +670,7 @@ jobs:
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=tvOS \
@@ -699,6 +701,7 @@ jobs:
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=visionOS \
@@ -739,6 +742,7 @@ jobs:
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_CURL=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
@@ -767,7 +771,7 @@ jobs:
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
          key: windows-msys2
-          variant: sccache
+          variant: ccache
          evict-old-files: 1d

      - name: Setup ${{ matrix.sys }}
@@ -810,26 +814,18 @@ jobs:
    strategy:
      matrix:
        include:
-          - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
-          - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
-          - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
-          - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
+          - build: 'cpu-x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
          - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
          - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
          - build: 'llvm-arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
-          - build: 'msvc-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
          - build: 'llvm-arm64-opencl-adreno'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+         # - build: 'kompute-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_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'

    steps:
      - name: Clone
@@ -842,7 +838,7 @@ jobs:
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
          key: windows-latest-cmake-${{ matrix.build }}
-          variant: sccache
+          variant: ccache
          evict-old-files: 1d

      - name: Clone Kompute submodule
@@ -918,39 +914,26 @@ jobs:
          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt

-      - name: Check AVX512F support
-        id: check_avx512f
-        if: ${{ matrix.build == 'avx512-x64' }}
-        continue-on-error: true
-        run: |
-          cd build
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
-          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
-          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
-          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
-
      - name: Test
        id: cmake_test
-        # not all machines have native AVX-512
-        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
        run: |
          cd build
          ctest -L main -C Release --verbose --timeout 900

-      - name: Test (Intel SDE)
-        id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
-          # for some weird reason windows tar doesn't like sde tar.xz
-          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
-          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
-          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
-          cd build
-          $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
-          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
+      # TODO: disabled for now, consider adding tests for all CPU variants instead
+      # - name: Test (Intel SDE)
+      #   id: cmake_test_sde
+      #   if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+      #   run: |
+      #     curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
+      #     # for some weird reason windows tar doesn't like sde tar.xz
+      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+      #     $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+      #     cd build
+      #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
+      #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900

      - name: Determine tag name
        id: tag
@@ -1035,7 +1018,7 @@ jobs:
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
          key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }}
-          variant: sccache
+          variant: ccache
          evict-old-files: 1d

      - name: Install Cuda Toolkit 11.7
@@ -1113,6 +1096,8 @@ jobs:
          cmake -S . -B build -G "Ninja Multi-Config" ^
            -DLLAMA_BUILD_SERVER=ON ^
            -DGGML_NATIVE=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=ON ^
            -DGGML_CUDA=ON ^
            -DGGML_RPC=ON ^
            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
@@ -1187,7 +1172,7 @@ jobs:
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
          key: windows-latest-cmake-sycl
-          variant: sccache
+          variant: ccache
          evict-old-files: 1d

      - name: Install
@@ -1417,6 +1402,7 @@ jobs:
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_CURL=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=iOS \
@@ -1766,16 +1752,17 @@ jobs:
    if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
    defaults:
      run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
+        shell: bash -el {0}
    strategy:
      matrix:
+        arch: [x86, aarch64]
        cann:
          - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
        device:
          - 'ascend910b3'
        build:
          - 'Release'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
    container: ascendai/cann:${{ matrix.cann }}
    steps:
      - name: Checkout
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -15,10 +15,10 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']

 env:
  LLAMA_LOG_COLORS: 1
@@ -74,7 +74,7 @@ jobs:
      - name: Tests dependencies
        id: test_dependencies
        run: |
-          pip install -r examples/server/tests/requirements.txt
+          pip install -r tools/server/tests/requirements.txt

      # Setup nodejs (to be used for verifying bundled index.html)
      - uses: actions/setup-node@v4
@@ -84,14 +84,14 @@ jobs:
      - name: WebUI - Install dependencies
        id: webui_lint
        run: |
-          cd examples/server/webui
+          cd tools/server/webui
          npm ci

      - name: WebUI - Check code format
        id: webui_format
        run: |
          git config --global --add safe.directory $(realpath .)
-          cd examples/server/webui
+          cd tools/server/webui
          git status

          npm run format
@@ -108,7 +108,7 @@ jobs:
        id: verify_server_index_html
        run: |
          git config --global --add safe.directory $(realpath .)
-          cd examples/server/webui
+          cd tools/server/webui
          git status

          npm run build
@@ -161,21 +161,21 @@ jobs:
        env:
          GITHUB_ACTIONS: "true"
        run: |
-          cd examples/server/tests
+          cd tools/server/tests
          ./tests.sh

      - name: Tests (sanitizers)
        id: server_integration_tests_sanitizers
        if: ${{ matrix.sanitizer != '' }}
        run: |
-          cd examples/server/tests
+          cd tools/server/tests
          LLAMA_SANITIZE=1 ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
-          cd examples/server/tests
+          cd tools/server/tests
          SLOW_TESTS=1 ./tests.sh


@@ -211,7 +211,7 @@ jobs:
      - name: Tests dependencies
        id: test_dependencies
        run: |
-          pip install -r examples/server/tests/requirements.txt
+          pip install -r tools/server/tests/requirements.txt

      - name: Copy Libcurl
        id: prepare_libcurl
@@ -224,7 +224,7 @@ jobs:
        id: server_integration_tests
        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
-          cd examples/server/tests
+          cd tools/server/tests
          $env:PYTHONIOENCODING = ":replace"
          pytest -v -x -m "not slow"

@@ -232,6 +232,6 @@ jobs:
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
-          cd examples/server/tests
+          cd tools/server/tests
          $env:SLOW_TESTS = "1"
          pytest -v -x
--- a/.gitignore
+++ b/.gitignore
@@ -96,11 +96,11 @@ perf-*.txt
 # Examples

 examples/jeopardy/results.txt
-examples/server/*.css.hpp
-examples/server/*.html.hpp
-examples/server/*.js.hpp
-examples/server/*.mjs.hpp
-examples/server/*.gz.hpp
+tools/server/*.css.hpp
+tools/server/*.html.hpp
+tools/server/*.js.hpp
+tools/server/*.mjs.hpp
+tools/server/*.gz.hpp
 !build_64.sh
 !examples/*.bat
 !examples/*/*.kts
@@ -110,7 +110,7 @@ examples/server/*.gz.hpp

 # Server Web UI temporary files
 node_modules
-examples/server/webui/dist
+tools/server/webui/dist

 # Python

--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -77,6 +77,7 @@ option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE

 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})

@@ -187,6 +188,10 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
    add_subdirectory(pocs)
 endif()

+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
+    add_subdirectory(tools)
+endif()
+
 #
 # install
 #
--- a/2
+++ b/2
@@ -2,7 +2,7 @@

 /ci/ @ggerganov
 /.devops/*.Dockerfile @ngxson
-/examples/server/ @ngxson
+/tools/server/ @ngxson
 /ggml/src/ggml-cuda/fattn* @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.* @JohannesGaessler
 /ggml/src/ggml-cuda/mmv.* @JohannesGaessler
--- a/96
+++ b/96
@@ -780,10 +780,6 @@ ifdef GGML_HIP

 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA

-ifdef GGML_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # GGML_HIP_UMA
-
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
@@ -1160,10 +1156,10 @@ $(LIB_COMMON_S): $(OBJ_COMMON)

 # Clean generated server assets
 clean-server-assets:
-	find examples/server -type f -name "*.js.hpp"   -delete
-	find examples/server -type f -name "*.mjs.hpp"  -delete
-	find examples/server -type f -name "*.css.hpp"  -delete
-	find examples/server -type f -name "*.html.hpp" -delete
+	find tools/server -type f -name "*.js.hpp"   -delete
+	find tools/server -type f -name "*.mjs.hpp"  -delete
+	find tools/server -type f -name "*.css.hpp"  -delete
+	find tools/server -type f -name "*.html.hpp" -delete

 # Clean rule
 clean: clean-server-assets
@@ -1183,7 +1179,7 @@ clean: clean-server-assets
 # Helper function that replaces .c, .cpp, and .cu file endings with .o:
 GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))

-llama-cli: examples/main/main.cpp \
+llama-cli: tools/main/main.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1196,7 +1192,7 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-run: examples/run/run.cpp \
+llama-run: tools/run/run.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1211,7 +1207,7 @@ llama-simple-chat: examples/simple-chat/simple-chat.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-tokenize: examples/tokenize/tokenize.cpp \
+llama-tokenize: tools/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1221,27 +1217,27 @@ llama-batched: examples/batched/batched.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-batched-bench: examples/batched-bench/batched-bench.cpp \
+llama-batched-bench: tools/batched-bench/batched-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-quantize: examples/quantize/quantize.cpp \
+llama-quantize: tools/quantize/quantize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp \
+llama-quantize-stats: tools/quantize-stats/quantize-stats.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-perplexity: examples/perplexity/perplexity.cpp \
+llama-perplexity: tools/perplexity/perplexity.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-imatrix: examples/imatrix/imatrix.cpp \
+llama-imatrix: tools/imatrix/imatrix.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1283,7 +1279,7 @@ llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/s
 	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-gguf-split: examples/gguf-split/gguf-split.cpp \
+llama-gguf-split: tools/gguf-split/gguf-split.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1293,7 +1289,7 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
+llama-cvector-generator: tools/cvector-generator/cvector-generator.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1303,12 +1299,12 @@ llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-bench: examples/llama-bench/llama-bench.cpp \
+llama-bench: tools/llama-bench/llama-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-export-lora: examples/export-lora/export-lora.cpp \
+llama-export-lora: tools/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1364,17 +1360,17 @@ llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 ifdef GGML_RPC
-rpc-server: examples/rpc/rpc-server.cpp \
+rpc-server: tools/rpc/rpc-server.cpp \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 endif # GGML_RPC

 llama-server: \
-	examples/server/server.cpp \
-	examples/server/utils.hpp \
-	examples/server/httplib.h \
-	examples/server/index.html.hpp \
-	examples/server/loading.html.hpp \
+	tools/server/server.cpp \
+	tools/server/utils.hpp \
+	tools/server/httplib.h \
+	tools/server/index.html.hpp \
+	tools/server/loading.html.hpp \
 	common/chat.cpp \
 	common/chat.h \
 	common/chat-template.hpp \
@@ -1382,10 +1378,10 @@ llama-server: \
 	common/minja.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Itools/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)

-# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-examples/server/%.hpp: examples/server/public/% FORCE Makefile
+# Portable equivalent of `cd tools/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+tools/server/%.hpp: tools/server/public/% FORCE Makefile
 	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
 		echo "unsigned char $${NAME}[] = {" && \
 		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
@@ -1398,36 +1394,36 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-libllava.a: examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+libllava.a: tools/mtmd/llava.cpp \
+	tools/mtmd/llava.h \
+	tools/mtmd/clip.cpp \
+	tools/mtmd/clip.h \
 	common/stb_image.h \
 	common/base64.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual

-llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-llava-cli: tools/mtmd/llava-cli.cpp \
+	tools/mtmd/llava.cpp \
+	tools/mtmd/llava.h \
+	tools/mtmd/clip.cpp \
+	tools/mtmd/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

-llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-minicpmv-cli: tools/mtmd/minicpmv-cli.cpp \
+	tools/mtmd/llava.cpp \
+	tools/mtmd/llava.h \
+	tools/mtmd/clip.cpp \
+	tools/mtmd/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

-llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-qwen2vl-cli: tools/mtmd/qwen2vl-cli.cpp \
+	tools/mtmd/llava.cpp \
+	tools/mtmd/llava.h \
+	tools/mtmd/clip.cpp \
+	tools/mtmd/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

@@ -1484,12 +1480,12 @@ tests/test-double-float: tests/test-double-float.cpp

 tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 tests/test-chat: tests/test-chat.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 tests/test-opt: tests/test-opt.cpp \
--- a/README.md
+++ b/README.md
@@ -16,8 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)

 ## Hot topics

- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
+- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
+- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
@@ -241,7 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |

 ## Building the project

@@ -275,9 +276,9 @@ The Hugging Face platform provides a variety of online tools for converting, qua
 - Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
 - Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)

-To learn more about model quantization, [read this documentation](examples/quantize/README.md)
+To learn more about model quantization, [read this documentation](tools/quantize/README.md)

-## [`llama-cli`](examples/main)
+## [`llama-cli`](tools/main)

 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.

@@ -340,7 +341,7 @@ To learn more about model quantization, [read this documentation](examples/quant
    </details>


-## [`llama-server`](examples/server)
+## [`llama-server`](tools/server)

 #### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.

@@ -410,7 +411,7 @@ To learn more about model quantization, [read this documentation](examples/quant
    </details>


-## [`llama-perplexity`](examples/perplexity)
+## [`llama-perplexity`](tools/perplexity)

 #### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.

@@ -435,10 +436,10 @@ To learn more about model quantization, [read this documentation](examples/quant

    </details>

-[^1]: [examples/perplexity/README.md](./examples/perplexity/README.md)
+[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md)
 [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)

-## [`llama-bench`](examples/llama-bench)
+## [`llama-bench`](tools/llama-bench)

 #### Benchmark the performance of the inference for various parameters.

@@ -459,7 +460,7 @@ To learn more about model quantization, [read this documentation](examples/quant

    </details>

-## [`llama-run`](examples/run)
+## [`llama-run`](tools/run)

 #### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].

@@ -503,8 +504,8 @@ To learn more about model quantization, [read this documentation](examples/quant

 ## Other documentation

- [main (cli)](examples/main/README.md)
- [server](examples/server/README.md)
+- [main (cli)](tools/main/README.md)
+- [server](tools/server/README.md)
 - [GBNF grammars](grammars/README.md)

 #### Development documentation
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -40,7 +40,8 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks

 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value
+* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.

 ### Multi-Tenant environments
--- a/build-xcframework.sh
+++ b/build-xcframework.sh
@@ -8,6 +8,7 @@ TVOS_MIN_OS_VERSION=16.4

 BUILD_SHARED_LIBS=OFF
 LLAMA_BUILD_EXAMPLES=OFF
+LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 GGML_METAL=ON
@@ -31,6 +32,7 @@ COMMON_CMAKE_ARGS=(
    -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
    -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
    -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
+    -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
    -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
    -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -187,8 +187,8 @@ function gg_run_test_scripts_debug {

    set -e

-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log

    set +e
 }
@@ -211,8 +211,8 @@ function gg_run_test_scripts_release {

    set -e

-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log

    set +e
 }
--- a/cmake/arm64-windows-msvc.cmake
+++ b/cmake/arm64-windows-msvc.cmake
@@ -1,6 +0,0 @@
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR arm64 )
-
-set( target arm64-pc-windows-msvc )
-set( CMAKE_C_COMPILER_TARGET   ${target} )
-set( CMAKE_CXX_COMPILER_TARGET ${target} )
--- a/cmake/build-info.cmake
+++ b/cmake/build-info.cmake
@@ -41,14 +41,20 @@ endif()

 if(MSVC)
    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
    execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND ${CMAKE_C_COMPILER} --version
        OUTPUT_VARIABLE OUT
        OUTPUT_STRIP_TRAILING_WHITESPACE
    )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
    set(BUILD_COMPILER ${OUT})
+
    execute_process(
        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
        OUTPUT_VARIABLE OUT
--- a/cmake/x64-windows-llvm.cmake
+++ b/cmake/x64-windows-llvm.cmake
@@ -3,9 +3,3 @@ set( CMAKE_SYSTEM_PROCESSOR x86_64 )

 set( CMAKE_C_COMPILER    clang )
 set( CMAKE_CXX_COMPILER  clang++ )
-
-set( arch_c_flags "-march=native" )
-
-set( CMAKE_C_FLAGS_INIT   "${arch_c_flags}" )
-set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" )
-
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -39,7 +39,9 @@ add_custom_command(
    COMMENT "Generating build details from Git"
    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${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}/.."
    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
    VERBATIM
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -38,6 +38,30 @@

 using json = nlohmann::ordered_json;

+std::initializer_list<enum llama_example> mmproj_examples = {
+    LLAMA_EXAMPLE_LLAVA,
+    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+};
+
+static std::string read_file(const std::string & fname) {
+    std::ifstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    file.close();
+    return content;
+}
+
+static void write_file(const std::string & fname, const std::string & content) {
+    std::ofstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    file << content;
+    file.close();
+}
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
    this->examples = std::move(examples);
    return *this;
@@ -157,6 +181,10 @@ struct common_hf_file_res {

 #ifdef LLAMA_USE_CURL

+bool common_has_curl() {
+    return true;
+}
+
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
@@ -189,11 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2

-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
    int remaining_attempts = max_attempts;

    while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);

        CURLcode res = curl_easy_perform(curl);
        if (res == CURLE_OK) {
@@ -204,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);

        remaining_attempts--;
+        if (remaining_attempts == 0) break;
        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
    }

@@ -222,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri
        return false;
    }

-    bool force_download = false;
-
    // Set the URL, allow to follow http redirection
    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
@@ -247,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri

    // If the file exists, check its JSON metadata companion file.
    std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
    std::string etag;
    std::string last_modified;

@@ -257,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
        if (metadata_in.good()) {
            try {
                metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                    etag = metadata.at("etag");
                }
@@ -272,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
                    last_modified = metadata.at("lastModified");
                }
            } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
            }
        }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
    } else {
        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
    }
@@ -287,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
    };

    common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist

+    // get ETag to see if the remote file has changed
    {
        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
@@ -316,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri
        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);

-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        // we only allow retrying once for HEAD requests
+        // this is for the use case of using running offline (no internet), retrying can be annoying
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
        if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
        }

        long http_code = 0;
        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
        }
    }

-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
+    // if head_request_ok is false, we don't have the etag or last-modified headers
+    // we leave should_download as-is, which is true if the file does not exist
+    if (head_request_ok) {
+        // check if ETag or Last-Modified headers are different
+        // if it is, we need to download the file again
        if (!etag.empty() && etag != headers.etag) {
            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
            should_download = true;
@@ -341,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
            should_download = true;
        }
    }
+
    if (should_download) {
        std::string path_temporary = path + ".downloadInProgress";
        if (file_exists) {
@@ -394,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
        // start the download
        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
        if (!was_perform_successful) {
            return false;
        }
@@ -415,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri
            {"etag", headers.etag},
            {"lastModified", headers.last_modified}
        });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+        write_file(metadata_path, metadata.dump(4));
+        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());

        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
            return false;
        }
+    } else {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
    }

    return true;
@@ -522,6 +553,50 @@ static bool common_download_model(
    return true;
 }

+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::vector<char> res_buffer;
+
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        auto data_vec = static_cast<std::vector<char> *>(data);
+        data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (params.timeout > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout);
+    }
+    if (params.max_size > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
+    }
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    for (const auto & header : params.headers) {
+        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
+    }
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        std::string error_msg = curl_easy_strerror(res);
+        throw std::runtime_error("error: cannot make GET request: " + error_msg);
+    }
+
+    long res_code;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+
+    return { res_code, std::move(res_buffer) };
+}
+
 /**
 * Allow getting the HF file from the HF repo with tag (like ollama), for example:
 * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
@@ -541,46 +616,48 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
    }

-    // fetch model info from Hugging Face Hub API
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
+    std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag;

-    std::string model_endpoint = get_model_endpoint();
-
-    std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
+    // headers
+    std::vector<std::string> headers;
+    headers.push_back("Accept: application/json");
    if (!bearer_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + bearer_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        headers.push_back("Authorization: Bearer " + bearer_token);
    }
    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    // User-Agent header is already set in common_remote_get_content, no need to set it here

-    CURLcode res = curl_easy_perform(curl.get());
+    // we use "=" to avoid clashing with other component, while still being allowed on windows
+    std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json";
+    string_replace_all(cached_response_fname, "/", "_");
+    std::string cached_response_path = fs_get_cache_file(cached_response_fname);

-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
+    // make the request
+    common_remote_params params;
+    params.headers = headers;
+    long res_code = 0;
+    std::string res_str;
+    bool use_cache = false;
+    try {
+        auto res = common_remote_get_content(url, params);
+        res_code = res.first;
+        res_str = std::string(res.second.data(), res.second.size());
+    } catch (const std::exception & e) {
+        LOG_WRN("error: failed to get manifest: %s\n", e.what());
+        LOG_WRN("try reading from cache\n");
+        // try to read from cache
+        try {
+            res_str = read_file(cached_response_path);
+            res_code = 200;
+            use_cache = true;
+        } catch (const std::exception & e) {
+            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        }
    }
+    std::string ggufFile;
+    std::string mmprojFile;

-    long res_code;
-    std::string ggufFile   = "";
-    std::string mmprojFile = "";
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
+    if (res_code == 200 || res_code == 304) {
        // extract ggufFile.rfilename in json, using regex
        {
            std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
@@ -597,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
                mmprojFile = match[1].str();
            }
        }
+        if (!use_cache) {
+            // if not using cached response, update the cache file
+            write_file(cached_response_path, res_str);
+        }
    } else if (res_code == 401) {
        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
    } else {
@@ -613,6 +694,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_

 #else

+bool common_has_curl() {
+    return false;
+}
+
 static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
    LOG_ERR("error: built without CURL, cannot download model from internet\n");
    return false;
@@ -635,17 +720,30 @@ static struct common_hf_file_res common_get_hf_file(const std::string &, const s
    return {};
 }

+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params &) {
+    if (!url.empty()) {
+        throw std::runtime_error("error: built without CURL, cannot download model from the internet");
+    }
+
+    return {};
+}
+
 #endif // LLAMA_USE_CURL

 //
 // utils
 //

-static void common_params_handle_model(
+struct handle_model_result {
+    bool found_mmproj = false;
+    common_params_model mmproj;
+};
+
+static handle_model_result common_params_handle_model(
        struct common_params_model & model,
        const std::string & bearer_token,
-        const std::string & model_path_default,
-        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+        const std::string & model_path_default) {
+    handle_model_result result;
    // handle pre-fill default model path and url based on hf_repo and hf_file
    {
        if (!model.hf_repo.empty()) {
@@ -657,7 +755,12 @@ static void common_params_handle_model(
                        exit(1); // built without CURL, error message already printed
                    }
                    model.hf_repo = auto_detected.repo;
-                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                    model.hf_file = auto_detected.ggufFile;
+                    if (!auto_detected.mmprojFile.empty()) {
+                        result.found_mmproj   = true;
+                        result.mmproj.hf_repo = model.hf_repo;
+                        result.mmproj.hf_file = auto_detected.mmprojFile;
+                    }
                } else {
                    model.hf_file = model.path;
                }
@@ -694,6 +797,8 @@ static void common_params_handle_model(
            exit(1);
        }
    }
+
+    return result;
 }

 const std::vector<ggml_type> kv_cache_types = {
@@ -827,16 +932,25 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
    }

-    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model(params.speculative.model, params.hf_token, "");
-    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
-
-    // allow --mmproj to be set from -hf
-    // assuming that mmproj is always in the same repo as text model
-    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
-        params.mmproj.hf_repo = params.model.hf_repo;
+    // handle model and download
+    {
+        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH);
+        if (params.no_mmproj) {
+            params.mmproj = {};
+        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
+            // optionally, handle mmproj model when -hf is specified
+            params.mmproj = res.mmproj;
+        }
+        // only download mmproj if the current example is using it
+        for (auto & ex : mmproj_examples) {
+            if (ctx_arg.ex == ex) {
+                common_params_handle_model(params.mmproj,    params.hf_token, "");
+                break;
+            }
+        }
+        common_params_handle_model(params.speculative.model, params.hf_token, "");
+        common_params_handle_model(params.vocoder.model,     params.hf_token, "");
    }
-    common_params_handle_model(params.mmproj,            params.hf_token, "", true);

    if (params.escape) {
        string_process_escapes(params.prompt);
@@ -968,7 +1082,6 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
        "llama-embedding",
        "llama-eval-callback",
        "llama-export-lora",
-        "llama-gbnf-validator",
        "llama-gen-docs",
        "llama-gguf",
        "llama-gguf-hash",
@@ -976,20 +1089,18 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
        "llama-gritlm",
        "llama-imatrix",
        "llama-infill",
-        "llama-llava-cli",
+        "llama-mtmd-cli",
        "llama-llava-clip-quantize-cli",
        "llama-lookahead",
        "llama-lookup",
        "llama-lookup-create",
        "llama-lookup-merge",
        "llama-lookup-stats",
-        "llama-minicpmv-cli",
        "llama-parallel",
        "llama-passkey",
        "llama-perplexity",
        "llama-q8dot",
        "llama-quantize",
-        "llama-quantize-stats",
        "llama-qwen2vl-cli",
        "llama-retrieval",
        "llama-run",
@@ -1078,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
        fprintf(stderr, "%s\n", ex.what());
        ctx_arg.params = params_org;
        return false;
+    } catch (std::exception & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        exit(1); // for other exceptions, we exit with status code 1
    }

    return true;
@@ -1378,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"-f", "--file"}, "FNAME",
        "a file containing the prompt (default: none)",
        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
+            params.prompt = read_file(value);
            // store the external file name in params
            params.prompt_file = value;
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
            if (!params.prompt.empty() && params.prompt.back() == '\n') {
                params.prompt.pop_back();
            }
@@ -1394,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"-sysf", "--system-prompt-file"}, "FNAME",
        "a file containing the system prompt (default: none)",
        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            params.system_prompt = read_file(value);
            if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
                params.system_prompt.pop_back();
            }
@@ -1823,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--grammar-file"}, "FNAME",
        "file to read grammar from",
        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sampling.grammar)
-            );
+            params.sampling.grammar = read_file(value);
        }
    ).set_sparam());
    add_opt(common_arg(
@@ -1841,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.sampling.grammar = json_schema_to_grammar(json::parse(value));
        }
    ).set_sparam());
+    add_opt(common_arg(
+        {"-jf", "--json-schema-file"}, "FILE",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::string schema;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(schema)
+            );
+            params.sampling.grammar = json_schema_to_grammar(json::parse(schema));
+        }
+    ).set_sparam());
    add_opt(common_arg(
        {"--pooling"}, "{none,mean,cls,last,rank}",
        "pooling type for embeddings, use model default if unspecified",
@@ -2096,18 +2211,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
    add_opt(common_arg(
        {"--mmproj"}, "FILE",
-        "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "path to a multimodal projector file. see tools/mtmd/README.md",
        [](common_params & params, const std::string & value) {
            params.mmproj.path = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--mmproj-url"}, "URL",
-        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "URL to a multimodal projector file. see tools/mtmd/README.md",
        [](common_params & params, const std::string & value) {
            params.mmproj.url = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj"},
+        "explicitly disable multimodal projector, useful when using -hf",
+        [](common_params & params) {
+            params.no_mmproj = true;
+        }
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj-offload"},
+        "do not offload multimodal projector to GPU",
+        [](common_params & params) {
+            params.mmproj_use_gpu = false;
+        }
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--image"}, "FILE",
        "path to an image file. use with multimodal models. Specify multiple times for batching",
@@ -2382,6 +2511,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    add_opt(common_arg(
        {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
        "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
+        "mmproj is also downloaded automatically if available. to disable, add --no-mmproj\n"
        "example: unsloth/phi-4-GGUF:q4_k_m\n"
        "(default: unused)",
        [](common_params & params, const std::string & value) {
@@ -2653,7 +2783,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
    add_opt(common_arg(
        {"--cache-reuse"}, "N",
-        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
+        string_format(
+            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
+        ),
        [](common_params & params, int value) {
            params.n_cache_reuse = value;
        }
@@ -2726,7 +2859,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.chat_template = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
    add_opt(common_arg(
        {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
        string_format(
@@ -2736,14 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
        ),
        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
+            params.chat_template = read_file(value);
        }
    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
    add_opt(common_arg(
--- a/common/arg.h
+++ b/common/arg.h
@@ -78,3 +78,12 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e

 // function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
+
+struct common_remote_params {
+    std::vector<std::string> headers;
+    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
+    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
+};
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);
--- a/common/common.h
+++ b/common/common.h
@@ -96,6 +96,7 @@ enum common_sampler_type {
    COMMON_SAMPLER_TYPE_XTC         = 8,
    COMMON_SAMPLER_TYPE_INFILL      = 9,
    COMMON_SAMPLER_TYPE_PENALTIES   = 10,
+    COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
 };

 // dimensionality reduction methods, used by cvector-generator
@@ -161,6 +162,7 @@ struct common_params_sampling {
    std::vector<enum common_sampler_type> samplers = {
        COMMON_SAMPLER_TYPE_PENALTIES,
        COMMON_SAMPLER_TYPE_DRY,
+        COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
        COMMON_SAMPLER_TYPE_TOP_K,
        COMMON_SAMPLER_TYPE_TYPICAL_P,
        COMMON_SAMPLER_TYPE_TOP_P,
@@ -340,8 +342,10 @@ struct common_params {

    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;

-    // multimodal models (see examples/llava)
+    // multimodal models (see tools/mtmd)
    struct common_params_model mmproj;
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
    std::vector<std::string> image; // path to image file(s)

    // embedding
@@ -412,8 +416,8 @@ struct common_params {
    int n_pca_batch = 100;
    int n_pca_iterations = 1000;
    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "tools/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";

    bool spm_infill = false; // suffix/prefix/middle pattern for infill

--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -16,6 +16,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
    auto has_max = max_items != std::numeric_limits<int>::max();

+    if (max_items == 0) {
+        return "";
+    }
    if (min_items == 0 && max_items == 1) {
        return item_rule + "?";
    }
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -229,51 +229,48 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                params.logit_bias.data()));

    if (params.mirostat == 0) {
-        if (params.top_n_sigma >= 0) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
-        } else {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_DRY:
-                        {
-                            std::vector<const char *> c_breakers;
-                            c_breakers.reserve(params.dry_sequence_breakers.size());
-                            for (const auto & str : params.dry_sequence_breakers) {
-                                c_breakers.push_back(str.c_str());
-                            }
-
-                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                case COMMON_SAMPLER_TYPE_DRY:
+                    {
+                        std::vector<const char *> c_breakers;
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
+                        for (const auto & str : params.dry_sequence_breakers) {
+                            c_breakers.push_back(str.c_str());
                        }
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_XTC:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    case COMMON_SAMPLER_TYPE_INFILL:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
-                        break;
-                    case COMMON_SAMPLER_TYPE_PENALTIES:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                    }
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
+                    break;
+                case COMMON_SAMPLER_TYPE_PENALTIES:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
            }
        }
        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@@ -475,6 +472,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
        case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's';
        case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
        case COMMON_SAMPLER_TYPE_XTC:         return 'x';
@@ -490,6 +488,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
        case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma";
        case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
        case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
@@ -504,6 +503,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
        { "dry",         COMMON_SAMPLER_TYPE_DRY },
        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
        { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
@@ -517,6 +517,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
    std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
@@ -552,6 +553,7 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -115,6 +115,7 @@ models = [
    {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
    {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
    {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
+    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
 ]


--- a/convert_lora_to_gguf.py
+++ b/convert_lora_to_gguf.py
@@ -24,7 +24,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf

 # reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import LazyTorchTensor, Model
+from convert_hf_to_gguf import LazyTorchTensor, ModelBase

 logger = logging.getLogger("lora-to-gguf")

@@ -340,11 +340,11 @@ if __name__ == '__main__':
            sys.exit(1)
    else:
        logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = Model.load_hparams(dir_base_model)
+        hparams = ModelBase.load_hparams(dir_base_model)

    with torch.inference_mode():
        try:
-            model_class = Model.from_model_architecture(hparams["architectures"][0])
+            model_class = ModelBase.from_model_architecture(hparams["architectures"][0])
        except NotImplementedError:
            logger.error(f"Model {hparams['architectures'][0]} is not supported")
            sys.exit(1)
--- a/docs/build.md
+++ b/docs/build.md
@@ -259,8 +259,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
      && cmake --build build --config Release -- -j 16
  ```
-  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
-  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).

  To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.

@@ -296,6 +294,10 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.

+### Unified Memory
+
+On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
 ## Vulkan

 **Windows**
--- a/docs/development/HOWTO-add-model.md
+++ b/docs/development/HOWTO-add-model.md
@@ -9,10 +9,10 @@ Adding a model requires few steps:
 After following these steps, you can open PR.

 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
- [main](/examples/main/)
- [imatrix](/examples/imatrix/)
- [quantize](/examples/quantize/)
- [server](/examples/server/)
+- [main](/tools/main/)
+- [imatrix](/tools/imatrix/)
+- [quantize](/tools/quantize/)
+- [server](/tools/server/)

 ### 1. Convert the model to GGUF

--- a/examples/llava/MobileVLM-README.md
+++ b/examples/llava/MobileVLM-README.md
@@ -9,15 +9,15 @@ The implementation is based on llava, and is compatible with llava and mobileVLM
 Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.

 ## Usage
-Build with cmake or run `make llama-llava-cli` to build it.

-After building, run: `./llama-llava-cli` to see the usage. For example:
+Build the `llama-mtmd-cli` binary.
+
+After building, run: `./llama-mtmd-cli` to see the usage. For example:

 ```sh
-./llama-llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
+./llama-mtmd-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
    --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \
-    --image path/to/an/image.jpg \
-    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:"
+    --chat-template deepseek
 ```

 ## Model conversion
@@ -33,13 +33,13 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:

 ```sh
-python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
+python ./tools/mtmd/llava_surgery.py -m path/to/MobileVLM-1.7B
 ```

 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:

 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py \
+python ./tools/mtmd/convert_image_encoder_to_gguf.py \
    -m path/to/clip-vit-large-patch14-336 \
    --llava-projector path/to/MobileVLM-1.7B/llava.projector \
    --output-dir path/to/MobileVLM-1.7B \
@@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf.py \
 ```

 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py \
+python ./tools/mtmd/convert_image_encoder_to_gguf.py \
    -m path/to/clip-vit-large-patch14-336 \
    --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
    --output-dir path/to/MobileVLM-1.7B_V2 \
@@ -69,10 +69,10 @@ Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directo

 ## Android compile and run
 ### compile
-refer to `examples/llava/android/build_64.sh`
+refer to `tools/mtmd/android/build_64.sh`
 ```sh
-mkdir examples/llava/android/build_64
-cd examples/llava/android/build_64
+mkdir tools/mtmd/android/build_64
+cd tools/mtmd/android/build_64
 ../build_64.sh
 ```
 ### run on Android
@@ -82,7 +82,7 @@ refer to `android/adb_run.sh`, modify resources' `name` and `path`
 ### case 1
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
    -m /data/local/tmp/ggml-model-q4_k.gguf \
    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
    -t 4 \
@@ -102,7 +102,7 @@ llama_print_timings:       total time =   34731.93 ms
 ### case 2
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
    -m /data/local/tmp/ggml-model-q4_k.gguf \
    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
    -t 4 \
@@ -123,10 +123,10 @@ llama_print_timings:       total time =   34570.79 ms

 ## Some result on Android with `Snapdragon 778G` chip
 ### MobileVLM-1.7B case
-#### llava-cli release-b2005
+#### mtmd-cli release-b2005
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
    -m /data/local/tmp/ggml-model-q4_k.gguf \
    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
    -t 4 \
@@ -147,7 +147,7 @@ llama_print_timings: prompt eval time =    8119.49 ms /   191 tokens (   42.51 m
 llama_print_timings:        eval time =    1005.75 ms /    14 runs   (   71.84 ms per token,    13.92 tokens per second)
 llama_print_timings:       total time =   28038.34 ms /   205 tokens
 ```
-#### llava-cli latest-version
+#### mtmd-cli latest-version
 **input**

 Just the same as above.
@@ -169,7 +169,7 @@ llama_print_timings:        eval time =   43894.02 ms /    13 runs   ( 3376.46 m
 llama_print_timings:       total time =  865441.76 ms /   204 tokens
 ```
 ### MobileVLM_V2-1.7B case
-#### llava-cli release-2005b
+#### mtmd-cli release-2005b
 **input**

 Just the same as above.
@@ -200,7 +200,7 @@ make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
 ### case 1
 **input**
 ```sh
-./llama-llava-cli \
+./llama-mtmd-cli \
    -m /data/local/tmp/ggml-model-q4_k.gguf \
    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
    --image /data/local/tmp/demo.jpeg \
@@ -224,7 +224,7 @@ llama_print_timings:       total time =    1352.63 ms /   252 tokens
 ### case 2
 **input**
 ```sh
-./llama-llava-cli \
+./llama-mtmd-cli \
    -m /data/local/tmp/ggml-model-q4_k.gguf \
    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:" \
--- a/examples/llava/README-gemma3.md
+++ b/examples/llava/README-gemma3.md
@@ -11,26 +11,27 @@ You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli

 # alternatively, install from brew (MacOS)
 brew install llama.cpp

 # run it
-llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF

 # note: 1B model does not support vision
 ```

 ## How to get mmproj.gguf?

+Simply to add `--mmproj` in when converting model via `convert_hf_to_gguf.py`:
+
 ```bash
 cd gemma-3-4b-it
-python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py .
-
-# output file is mmproj.gguf
+python ../llama.cpp/convert_hf_to_gguf.py --outfile model.gguf --outtype f16 --mmproj .
+# output file: mmproj-model.gguf
 ```

 ## How to run it?
@@ -43,8 +44,8 @@ What you need:
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli

 # run it
-./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+./build/bin/llama-mtmd-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
 ```
--- a/examples/llava/README-glmedge.md
+++ b/examples/llava/README-glmedge.md
@@ -3,12 +3,12 @@
 Currently this implementation supports [glm-edge-v-2b](https://huggingface.co/THUDM/glm-edge-v-2b) and [glm-edge-v-5b](https://huggingface.co/THUDM/glm-edge-v-5b).

 ## Usage
-Build with cmake or run `make llama-llava-cli` to build it.
+Build the `llama-mtmd-cli` binary.

-After building, run: `./llama-llava-cli` to see the usage. For example:
+After building, run: `./llama-mtmd-cli` to see the usage. For example:

 ```sh
-./llama-llava-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf --image img_path/image.jpg -p "<|system|>\n system prompt <image><|user|>\n prompt <|assistant|>\n"
+./llama-mtmd-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf
 ```

 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
@@ -25,13 +25,13 @@ git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/T
 2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents:

 ```sh
-python ./examples/llava/glmedge-surgery.py -m ../model_path
+python ./tools/mtmd/glmedge-surgery.py -m ../model_path
 ```

 4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF:

 ```sh
-python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
+python ./tools/mtmd/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
 ```

 5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF:
--- a/examples/llava/README-granitevision.md
+++ b/examples/llava/README-granitevision.md
@@ -176,15 +176,11 @@ Note that currently you cannot quantize the visual encoder because granite visio


 ### 5. Running the Model in Llama cpp
-Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
+Build llama cpp normally; you should have a target binary named `llama-mtmd-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.

 ```bash
-$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
+$ ./build/bin/llama-mtmd-cli -m $LLM_GGUF_PATH \
    --mmproj $VISUAL_GGUF_PATH \
-    --image ./media/llama0-banner.png \
    -c 16384 \
-    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat does the text in this image say?\n<|assistant|>\n" \
    --temp 0
 ```
-
-Sample output: `The text in the image reads "LLAMA C++ Can it run DOOM Llama?"`
--- a/docs/multimodal/llava.md
+++ b/docs/multimodal/llava.md
@@ -11,12 +11,14 @@ For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](h
 After API is confirmed, more models will be supported / uploaded.

 ## Usage
-Build with cmake or run `make llama-llava-cli` to build it.
+Build the `llama-mtmd-cli` binary.

-After building, run: `./llama-llava-cli` to see the usage. For example:
+After building, run: `./llama-mtmd-cli` to see the usage. For example:

 ```sh
-./llama-llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+./llama-mtmd-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf \
+    --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf \
+    --chat-template vicuna
 ```

 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
@@ -35,19 +37,19 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Install the required Python packages:

 ```sh
-pip install -r examples/llava/requirements.txt
+pip install -r tools/mtmd/requirements.txt
 ```

 3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:

 ```sh
-python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
+python ./tools/mtmd/llava_surgery.py -m ../llava-v1.5-7b
 ```

 4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:

 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+python ./tools/mtmd/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```

 5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
@@ -67,12 +69,12 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 2) Install the required Python packages:

 ```sh
-pip install -r examples/llava/requirements.txt
+pip install -r tools/mtmd/requirements.txt
 ```

 3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
-python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
+python tools/mtmd/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory

@@ -86,7 +88,7 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso

 5) Create the visual gguf model:
 ```console
-python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+python ./tools/mtmd/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP

@@ -97,7 +99,7 @@ python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknow

 7) And finally we can run the llava cli using the 1.6 model version:
 ```console
-./llama-llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
+./llama-mtmd-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf
 ```

 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
@@ -122,17 +124,9 @@ model.language_model.save_pretrained(llm_export_path)

 Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.

-## llava-cli templating and llava-1.6 prompting
+## Chat template

-llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
-For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
-
-**For Mistral and using llava-cli binary:**
-Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
-The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
-
-**For the 34B this should work:**
-Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+For llava-1.5 and llava-1.6, you need to use `vicuna` chat template. Simply add `--chat-template vicuna` to activate this template.


 ## How to know if you are running in llava-1.5 or llava-1.6 mode
@@ -147,12 +141,3 @@ When running llava-cli you will see a visual information right before the prompt


 Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
-
-
-
-
-## TODO
-
- [x] Support non-CPU backend for the image encoding part.
- [ ] Support different sampling methods.
- [ ] Support more model variants.
--- a/examples/llava/README-minicpmo2.6.md
+++ b/examples/llava/README-minicpmo2.6.md
@@ -29,8 +29,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)

 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model

 # quantize int4 version
@@ -40,9 +40,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model

 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"

-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf
 ```
--- a/examples/llava/README-minicpmv2.5.md
+++ b/examples/llava/README-minicpmv2.5.md
@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)

 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model

 # quantize int4 version
@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model

 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"

-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf
 ```
--- a/examples/llava/README-minicpmv2.6.md
+++ b/examples/llava/README-minicpmv2.6.md
@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)

 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model

 # quantize int4 version
@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model

 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"

-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf
 ```
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -12,60 +12,30 @@ llama_add_compile_flags()

 # examples

-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(batched-bench)
    add_subdirectory(batched)
    add_subdirectory(embedding)
    add_subdirectory(eval-callback)

-    if (NOT WIN32)
-        # disabled on Windows because it uses internal functions not exported with LLAMA_API
-        add_subdirectory(gbnf-validator)
-    endif()
-
    add_subdirectory(gguf-hash)
-    add_subdirectory(gguf-split)
    add_subdirectory(gguf)
    add_subdirectory(gritlm)
-    add_subdirectory(imatrix)
    add_subdirectory(infill)
-    add_subdirectory(llama-bench)
    add_subdirectory(lookahead)
    add_subdirectory(lookup)
-    add_subdirectory(main)
    add_subdirectory(parallel)
    add_subdirectory(passkey)
-    add_subdirectory(perplexity)
-    add_subdirectory(quantize)
    add_subdirectory(retrieval)
-    if (LLAMA_BUILD_SERVER)
-        add_subdirectory(server)
-    endif()
    add_subdirectory(save-load-state)
-    add_subdirectory(run)
    add_subdirectory(simple)
    add_subdirectory(simple-chat)
    add_subdirectory(speculative)
    add_subdirectory(speculative-simple)
-    add_subdirectory(tokenize)
-    add_subdirectory(tts)
    add_subdirectory(gen-docs)
    if (NOT GGML_BACKEND_DL)
-        # these examples use the backends directly and cannot be built with dynamic loading
        add_subdirectory(convert-llama2c-to-ggml)
-        add_subdirectory(cvector-generator)
-        add_subdirectory(export-lora)
-        if (NOT WIN32)
-            # disabled on Windows because it uses internal functions not exported with LLAMA_API
-            add_subdirectory(quantize-stats)
-        endif()
-        add_subdirectory(llava)
-        if (GGML_RPC)
-            add_subdirectory(rpc)
-        endif()
+        # these examples use the backends directly and cannot be built with dynamic loading
        if (GGML_SYCL)
            add_subdirectory(sycl)
        endif()
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -89,6 +89,13 @@ int main(int argc, char ** argv) {
    common_init();

    params.embedding = true;
+
+    // utilize the full context
+    if (params.n_batch < params.n_ctx) {
+        LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
+        params.n_batch = params.n_ctx;
+    }
+
    // For non-causal models, batch size must be equal to ubatch size
    params.n_ubatch = params.n_batch;

@@ -134,7 +141,6 @@ int main(int argc, char ** argv) {

    // max batch size
    const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch >= params.n_ctx);

    // tokenize the prompts and trim
    std::vector<std::vector<int32_t>> inputs;
--- a/examples/gbnf-validator/CMakeLists.txt
+++ b/examples/gbnf-validator/CMakeLists.txt
@@ -1,5 +0,0 @@
-set(TARGET llama-gbnf-validator)
-add_executable(${TARGET} gbnf-validator.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
--- a/examples/json_schema_to_grammar.py
+++ b/examples/json_schema_to_grammar.py
@@ -10,6 +10,9 @@ from typing import Any, List, Optional, Set, Tuple, Union

 def _build_repetition(item_rule, min_items, max_items, separator_rule=None):

+    if max_items == 0:
+        return ""
+
    if min_items == 0 and max_items == 1:
        return f'{item_rule}?'

--- a/examples/llava/gemma3_convert_encoder_to_gguf.py
+++ b/examples/llava/gemma3_convert_encoder_to_gguf.py
@@ -1,307 +0,0 @@
-import gguf
-import argparse
-import logging
-import sys
-import torch
-import json
-import os
-import numpy as np
-from typing import cast, ContextManager, Any, Iterator
-from pathlib import Path
-from torch import Tensor
-
-logger = logging.getLogger("gemma3-mmproj")
-
-
-# (copied from convert_hf_to_gguf.py)
-# tree of lazy tensors
-class LazyTorchTensor(gguf.LazyBase):
-    _tensor_type = torch.Tensor
-    # to keep the type-checker happy
-    dtype: torch.dtype
-    shape: torch.Size
-
-    # only used when converting a torch.Tensor to a np.ndarray
-    _dtype_map: dict[torch.dtype, type] = {
-        torch.float16: np.float16,
-        torch.float32: np.float32,
-    }
-
-    # used for safetensors slices
-    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
-    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
-    _dtype_str_map: dict[str, torch.dtype] = {
-        "F64": torch.float64,
-        "F32": torch.float32,
-        "BF16": torch.bfloat16,
-        "F16": torch.float16,
-        # "U64": torch.uint64,
-        "I64": torch.int64,
-        # "U32": torch.uint32,
-        "I32": torch.int32,
-        # "U16": torch.uint16,
-        "I16": torch.int16,
-        "U8": torch.uint8,
-        "I8": torch.int8,
-        "BOOL": torch.bool,
-        "F8_E4M3": torch.float8_e4m3fn,
-        "F8_E5M2": torch.float8_e5m2,
-    }
-
-    def numpy(self) -> gguf.LazyNumpyTensor:
-        dtype = self._dtype_map[self.dtype]
-        return gguf.LazyNumpyTensor(
-            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
-            args=(self,),
-            func=(lambda s: s.numpy())
-        )
-
-    @classmethod
-    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
-        return torch.empty(size=shape, dtype=dtype, device="meta")
-
-    @classmethod
-    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
-        dtype = cls._dtype_str_map[st_slice.get_dtype()]
-        shape: tuple[int, ...] = tuple(st_slice.get_shape())
-        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
-        return cast(torch.Tensor, lazy)
-
-    @classmethod
-    def __torch_function__(cls, func, types, args=(), kwargs=None):
-        del types  # unused
-
-        if kwargs is None:
-            kwargs = {}
-
-        if func is torch.Tensor.numpy:
-            return args[0].numpy()
-
-        return cls._wrap_fn(func)(*args, **kwargs)
-
-
-class Gemma3VisionTower:
-    hparams: dict
-    gguf_writer: gguf.GGUFWriter
-    fname_out: Path
-    ftype: gguf.LlamaFileType
-
-    @staticmethod
-    def load_hparams(dir_model: Path):
-        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-            return json.load(f)
-
-    @staticmethod
-    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
-        part_names: list[str] = []
-        for filename in os.listdir(dir_model):
-            if filename.startswith(prefix) and filename.endswith(suffix):
-                part_names.append(filename)
-        part_names.sort()
-        return part_names
-
-    def __init__(self,
-                 dir_model: Path,
-                 fname_out: Path,
-                 ftype: gguf.LlamaFileType,
-                 is_big_endian: bool,):
-        hparams = Gemma3VisionTower.load_hparams(dir_model)
-        self.hparams = hparams
-        self.fname_out = fname_out
-        self.ftype = ftype
-        endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
-        self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess)
-
-        text_config = hparams["text_config"]
-        vision_config = hparams["vision_config"]
-
-        assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration"
-        assert text_config is not None
-        assert vision_config is not None
-
-        self.gguf_writer.add_string ("clip.projector_type",              "gemma3")
-        self.gguf_writer.add_bool   ("clip.has_text_encoder",            False)
-        self.gguf_writer.add_bool   ("clip.has_vision_encoder",          True)
-        self.gguf_writer.add_bool   ("clip.has_llava_projector",         False) # legacy
-        self.gguf_writer.add_uint32 ("clip.vision.image_size",           vision_config["image_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.patch_size",           vision_config["patch_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.embedding_length",     vision_config["hidden_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length",  vision_config["intermediate_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.projection_dim",       text_config["hidden_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.block_count",          vision_config["num_hidden_layers"])
-        self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"])
-        self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6))
-        # default values taken from HF tranformers code
-        self.gguf_writer.add_array  ("clip.vision.image_mean", [0.5, 0.5, 0.5])
-        self.gguf_writer.add_array  ("clip.vision.image_std",  [0.5, 0.5, 0.5])
-        self.gguf_writer.add_bool   ("clip.use_gelu", True)
-
-        # load tensors
-        for name, data_torch in self.get_tensors(dir_model):
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-            self.add_tensor(name, data_torch)
-
-    def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]:
-        part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors")
-        tensor_names_from_parts: set[str] = set()
-        for part_name in part_names:
-            logger.info(f"gguf: loading model part '{part_name}'")
-            from safetensors import safe_open
-            ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu"))
-            with ctx as model_part:
-                tensor_names_from_parts.update(model_part.keys())
-
-                for name in model_part.keys():
-                    data = model_part.get_slice(name)
-                    data = LazyTorchTensor.from_safetensors_slice(data)
-                    yield name, data
-
-    def add_tensor(self, name: str, data_torch: Tensor):
-        is_1d = len(data_torch.shape) == 1
-        is_embd = ".embeddings." in name
-        old_dtype = data_torch.dtype
-        can_quantize = not is_1d and not is_embd
-        data_qtype = gguf.GGMLQuantizationType.F32
-
-        # this is to support old checkpoint
-        # TODO: remove this when we have the final model
-        name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.")
-        name = name.replace("multimodal_projector.", "multi_modal_projector.")
-
-        # filter only vision tensors
-        if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."):
-            return
-        # prefix
-        name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.")
-        name = name.replace("vision_tower.vision_model.", "v.")
-        # projector and input embd
-        name = name.replace(".embeddings.patch_embedding.", ".patch_embd.")
-        name = name.replace(".embeddings.position_embedding.", ".position_embd.")
-        name = name.replace(
-            "multi_modal_projector.mm_input_projection_weight",
-            "mm.input_projection.weight"
-        )
-        name = name.replace(
-            "multi_modal_projector.mm_soft_emb_norm.weight",
-            "mm.soft_emb_norm.weight"
-        )
-        name = name.replace("post_layernorm.", "post_ln.")
-        # each block
-        name = name.replace(".self_attn.k_proj.", ".attn_k.")
-        name = name.replace(".self_attn.v_proj.", ".attn_v.")
-        name = name.replace(".self_attn.q_proj.", ".attn_q.")
-        name = name.replace(".self_attn.out_proj.", ".attn_out.")
-        name = name.replace(".layer_norm1.", ".ln1.")
-        name = name.replace(".layer_norm2.", ".ln2.")
-        name = name.replace(".mlp.fc1.", ".ffn_down.")
-        name = name.replace(".mlp.fc2.", ".ffn_up.")
-
-        if can_quantize:
-            if self.ftype == gguf.LlamaFileType.ALL_F32:
-                data_qtype = gguf.GGMLQuantizationType.F32
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
-                data_qtype = gguf.GGMLQuantizationType.F16
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                data_qtype = gguf.GGMLQuantizationType.BF16
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
-                data_qtype = gguf.GGMLQuantizationType.Q8_0
-            else:
-                raise ValueError(f"Unsupported file type: {self.ftype}")
-
-        # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
-        # the other norm values are part of SigLIP model, and they are already correct
-        # ref code: Gemma3RMSNorm
-        if "soft_emb_norm.weight" in name:
-            logger.info(f"Correcting norm value for '{name}'")
-            data_torch = data_torch + 1
-
-        data = data_torch.numpy()
-
-        try:
-            data = gguf.quants.quantize(data, data_qtype)
-        except Exception as e:
-            logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16")
-            data_qtype = gguf.GGMLQuantizationType.F16
-            data = gguf.quants.quantize(data, data_qtype)
-
-        # reverse shape to make it similar to the internal ggml dimension order
-        shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}"
-        logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
-
-        self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
-
-    def write(self):
-        self.gguf_writer.write_header_to_file(path=self.fname_out)
-        self.gguf_writer.write_kv_data_to_file()
-        self.gguf_writer.write_tensors_to_file(progress=True)
-        self.gguf_writer.close()
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Convert Gemma 3 vision tower safetensors to GGUF format",)
-    parser.add_argument(
-        "--outfile", type=Path, default="mmproj.gguf",
-        help="path to write to",
-    )
-    parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
-        help="output format",
-    )
-    parser.add_argument(
-        "--bigendian", action="store_true",
-        help="model is executed on big endian machine",
-    )
-    parser.add_argument(
-        "model", type=Path,
-        help="directory containing model file",
-        nargs="?",
-    )
-    parser.add_argument(
-        "--verbose", action="store_true",
-        help="increase output verbosity",
-    )
-
-    args = parser.parse_args()
-    if args.model is None:
-        parser.error("the following arguments are required: model")
-    return args
-
-
-def main() -> None:
-    args = parse_args()
-
-    if args.verbose:
-        logging.basicConfig(level=logging.DEBUG)
-    else:
-        logging.basicConfig(level=logging.INFO)
-
-    dir_model = args.model
-
-    if not dir_model.is_dir():
-        logger.error(f'Error: {args.model} is not a directory')
-        sys.exit(1)
-
-    ftype_map: dict[str, gguf.LlamaFileType] = {
-        "f32": gguf.LlamaFileType.ALL_F32,
-        "f16": gguf.LlamaFileType.MOSTLY_F16,
-        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
-        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
-    }
-
-    logger.info(f"Loading model: {dir_model.name}")
-
-    with torch.inference_mode():
-        gemma3_vision_tower = Gemma3VisionTower(
-            dir_model=dir_model,
-            fname_out=args.outfile,
-            ftype=ftype_map[args.outtype],
-            is_big_endian=args.bigendian,
-        )
-        gemma3_vision_tower.write()
-
-
-if __name__ == '__main__':
-    main()
-
--- a/examples/llava/llava-cli.cpp
+++ b/examples/llava/llava-cli.cpp
@@ -1,332 +0,0 @@
-#include "arg.h"
-#include "base64.hpp"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-    return true;
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
-static const char* IMG_BASE64_TAG_END = "\">";
-
-static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
-    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
-    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
-}
-
-static bool prompt_contains_image(const std::string& prompt) {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    return (begin != std::string::npos);
-}
-
-// replaces the base64 image tag in the prompt with `replacement`
-static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
-    size_t img_base64_str_start, img_base64_str_end;
-    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
-    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
-        return NULL;
-    }
-
-    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
-    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
-    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
-
-    auto required_bytes = base64::required_encode_size(base64_str.size());
-    auto img_bytes = std::vector<unsigned char>(required_bytes);
-    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
-
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
-    if (!embed) {
-        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
-        return NULL;
-    }
-
-    return embed;
-}
-
-static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    if (begin == std::string::npos || end == std::string::npos) {
-        return prompt;
-    }
-    auto pre = prompt.substr(0, begin);
-    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
-    return pre + replacement + post;
-}
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void print_usage(int, char ** argv) {
-    LOG("\n example usage:\n");
-    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
-
-    // load and preprocess the image
-    llava_image_embed * embed = NULL;
-    auto prompt = params->prompt;
-    if (prompt_contains_image(prompt)) {
-        if (!params->image.empty()) {
-            LOG_INF("using base64 encoded image instead of command line image path\n");
-        }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
-        if (!embed) {
-            LOG_ERR("%s: can't load image from prompt\n", __func__);
-            return NULL;
-        }
-        params->prompt = remove_image_from_prompt(prompt);
-    } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
-        if (!embed) {
-            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
-            return NULL;
-        }
-    }
-
-    return embed;
-}
-
-static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
-    int n_past = 0;
-
-    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
-
-    std::string system_prompt, user_prompt;
-    size_t image_pos = prompt.find("<image>");
-    if (image_pos != std::string::npos) {
-        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
-        system_prompt = prompt.substr(0, image_pos);
-        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    } else {
-        // llava-1.5 native mode
-        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
-        user_prompt = prompt + "\nASSISTANT:";
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
-    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-
-    // generate the response
-
-    LOG("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    if (!smpl) {
-        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
-        exit(1);
-    }
-
-    std::string response = "";
-    for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-        response += tmp;
-        if (strcmp(tmp, "</s>") == 0) break;
-        if (strstr(tmp, "###")) break; // Yi-VL behavior
-        LOG("%s", tmp);
-        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
-        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
-        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
-
-        fflush(stdout);
-    }
-
-    common_sampler_free(smpl);
-    LOG("\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->ctx_clip = ctx_clip;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        print_usage(argc, argv);
-        return 1;
-    }
-
-    auto * model = llava_init(&params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
-        return 1;
-    }
-
-    if (prompt_contains_image(params.prompt)) {
-        auto * ctx_llava = llava_init_context(&params, model);
-
-        auto * image_embed = load_image(ctx_llava, &params, "");
-
-        // process the prompt
-        process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-        llama_perf_context_print(ctx_llava->ctx_llama);
-        llava_image_embed_free(image_embed);
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    } else {
-        for (auto & image : params.image) {
-            auto * ctx_llava = llava_init_context(&params, model);
-
-            auto * image_embed = load_image(ctx_llava, &params, image);
-            if (!image_embed) {
-                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
-                return 1;
-            }
-
-            // process the prompt
-            process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-            llama_perf_context_print(ctx_llava->ctx_llama);
-            llava_image_embed_free(image_embed);
-            ctx_llava->model = NULL;
-            llava_free(ctx_llava);
-        }
-    }
-
-    llama_model_free(model);
-
-    return 0;
-}
--- a/examples/llava/minicpmv-cli.cpp
+++ b/examples/llava/minicpmv-cli.cpp
@@ -1,354 +0,0 @@
-#include "arg.h"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <algorithm>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-#include <iostream> // TODO: remove me
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    if (params->n_ctx < 2048) {
-        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
-        LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
-        ctx_params.n_ctx = 2048;
-    } else {
-        ctx_params.n_ctx = params->n_ctx;
-    }
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-    struct clip_context_params clip_params = {
-        /* use_gpu */   params->n_gpu_layers != 0,
-        /* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
-    };
-    auto * ctx_clip = clip_init(clip_path, clip_params);
-    return ctx_clip;
-}
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-}
-
-static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
-    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
-    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
-
-    auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
-    slice_embed->embed = image_embed;
-    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
-    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
-    llava_image_embed_free(slice_embed);
-}
-
-static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
-    std::string system_prompt;
-    int idx = 0;
-    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (has_minicpmv_projector == 2) {
-        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
-    }
-    else if (has_minicpmv_projector == 3) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    else if (has_minicpmv_projector == 4) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
-    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-    if (num_image_embeds > 1) {
-        if (has_minicpmv_projector == 2) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-            eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-        }
-        else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-        }
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
-    auto * ctx_clip = clip_init_context(params);
-    auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
-    if (!embeds) {
-        LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
-        return NULL;
-    }
-
-    // process the prompt
-    if (params->prompt.empty() && params->interactive == false) {
-        LOG_ERR("prompt should be given or interactive mode should be on");
-        return NULL;
-    }
-
-    auto * model = llava_init(params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
-        return NULL;
-    }
-    const int64_t t_llava_init_start_us = ggml_time_us();
-    auto * ctx_llava = llava_init_context(params, model);
-    ctx_llava->ctx_clip = ctx_clip;
-    const int64_t t_llava_init_end_us = ggml_time_us();
-    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
-    LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
-
-    const int64_t t_process_image_start_us = ggml_time_us();
-    process_image(ctx_llava, embeds, params, n_past);
-    const int64_t t_process_image_end_us = ggml_time_us();
-    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
-    LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
-
-    llava_image_embed_free(embeds);
-    return ctx_llava;
-}
-
-static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
-    std::string user_prompt = prompt;
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (!is_first) {
-        if (has_minicpmv_projector == 2) {
-            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 3) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 4) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-    if (has_minicpmv_projector == 2) {
-        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 3) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 4) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-
-    // generate the response
-
-    LOG_INF("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    return smpl;
-}
-
-static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
-
-    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-    return tmp;
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty())) {
-        show_additional_info(argc, argv);
-        return 1;
-    }
-
-    for (auto & image : params.image) {
-        int n_past = 0;
-        auto * ctx_llava = minicpmv_init(&params, image, n_past);
-
-        if (!params.prompt.empty()) {
-            LOG("<user>%s\n", params.prompt.c_str());
-            LOG("<assistant>");
-            auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
-            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-            std::string response;
-            bool have_tmp = false;
-            for (int i = 0; i < max_tgt_len; i++) {
-                const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                response += tmp;
-                if (strcmp(tmp, "</s>") == 0){
-                    if (!have_tmp) {
-                        continue;
-                    }
-                    break;
-                }
-                if (strstr(tmp, "###")) break; // Yi-VL behavior
-                have_tmp = true;
-                printf("%s", tmp);
-                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-
-                fflush(stdout);
-            }
-            common_sampler_free(smpl);
-        }else {
-            while (true) {
-                LOG("<user>");
-                std::string prompt;
-                std::getline(std::cin, prompt);
-                LOG("<assistant>");
-                auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
-                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-                std::string response;
-                for (int i = 0; i < max_tgt_len; i++) {
-                    const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                    response += tmp;
-                    if (strcmp(tmp, "</s>") == 0) break;
-                    printf("%s", tmp);// mistral llava-1.6
-                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-                    fflush(stdout);
-                }
-                common_sampler_free(smpl);
-            }
-        }
-        printf("\n");
-        llama_perf_context_print(ctx_llava->ctx_llama);
-
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    }
-
-    return 0;
-}
--- a/examples/llava/mtmd.cpp
+++ b/examples/llava/mtmd.cpp
@@ -1,341 +0,0 @@
-#include "clip.h"
-#include "clip-impl.h"
-#include "mtmd.h"
-
-#include "llama.h"
-
-#include <algorithm>
-#include <cerrno>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <limits>
-#include <vector>
-
-struct mtmd_context {
-    struct clip_ctx * ctx_clip;
-    const struct llama_model * text_model;
-    std::vector<float> image_embd_v; // image embedding vector
-    bool print_timings;
-    int n_threads;
-    std::string image_marker;
-
-    // TODO @ngxson : add timings
-
-    mtmd_context(const char * mmproj_fname,
-                   const llama_model * text_model,
-                   const mtmd_context_params & ctx_params) : print_timings(ctx_params.print_timings), n_threads(ctx_params.n_threads), image_marker(ctx_params.image_marker) {
-        clip_context_params ctx_clip_params;
-        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
-        ctx_clip_params.verbosity = ctx_params.verbosity;
-        ctx_clip = clip_init(mmproj_fname, ctx_clip_params);
-        if (!ctx_clip) {
-            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
-        }
-        this->text_model = text_model;
-    }
-
-    ~mtmd_context() {
-        clip_free(ctx_clip);
-    }
-};
-
-struct mtmd_image_tokens_data {
-    clip_image_f32_batch batch_f32; // preprocessed image patches
-};
-
-struct mtmd_image_tokens {
-    uint32_t nx; // number of tokens in x direction
-    uint32_t ny; // number of tokens in y direction
-    uint32_t n_tokens() const { return nx * ny; }
-    clip_image_f32_batch batch_f32; // preprocessed image patches
-};
-
-mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
-        const struct llama_model * text_model,
-        const struct mtmd_context_params ctx_params) {
-    try {
-        return new mtmd_context(mmproj_fname, text_model, ctx_params);
-    } catch (const std::exception & e) {
-        LOG_ERR("%s: error: %s\n", __func__, e.what());
-        return nullptr;
-    }
-}
-
-void mtmd_free(mtmd_context * ctx) {
-    if (ctx) {
-        delete ctx;
-    }
-}
-
-// copied from common_tokenize
-static std::vector<llama_token> mtmd_tokenize_text_internal(
-    const struct llama_vocab * vocab,
-           const std::string & text,
-                        bool   add_special,
-                        bool   parse_special) {
-    // upper limit for the number of tokens
-    int n_tokens = text.length() + 2 * add_special;
-    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-    return result;
-}
-
-mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
-                                const mtmd_input_text & text,
-                                const std::vector<mtmd_bitmap> & bitmaps) {
-    mtmd_input_chunks * output = new mtmd_input_chunks;
-    auto vocab = llama_model_get_vocab(ctx->text_model);
-
-    std::string prompt_modified(text.text);
-    std::string marker_modified(ctx->image_marker);
-    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
-    // a bit hacky here, but works for now
-    // for some models, we need to add prefix and suffix to the image embeddings
-    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
-        // <start_of_image> ... (image embeddings) ... <end_of_image>
-        marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-    }
-
-    std::vector<std::string> parts = string_split_str(text.text, ctx->image_marker);
-    output->clear();
-    output->reserve(parts.size());
-
-    size_t i_img = 0;
-
-    for (const auto & part : parts) {
-        //printf("tokenizing part: %s\n", part.c_str());
-        bool add_bos = &parts.front() == &part;
-        auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
-        if (tokens.empty()) {
-            continue;
-        }
-        mtmd_input_chunk chunk{
-            MTMD_INPUT_CHUNK_TYPE_TEXT,
-            std::move(tokens),
-            {},
-        };
-        output->emplace_back(std::move(chunk));
-
-        if (&parts.back() != &part) {
-            // add image token to middle of 2 parts
-
-            if (i_img >= bitmaps.size()) {
-                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
-                return nullptr;
-            }
-
-            // shim layer
-            clip_image_u8_ptr img_u8(clip_image_u8_init());
-            img_u8->nx = bitmaps[i_img].nx;
-            img_u8->ny = bitmaps[i_img].ny;
-            img_u8->buf.resize(bitmaps[i_img].data.size());
-            std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
-
-            // preprocess image
-            clip_image_f32_batch batch_f32;
-            bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32);
-            if (!ok) {
-                LOG_ERR("Unable to preprocess image\n");
-                return nullptr;
-            }
-
-            mtmd_image_tokens * image_tokens = new mtmd_image_tokens;
-            image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image
-            image_tokens->ny = 1; // TODO
-            image_tokens->batch_f32 = std::move(batch_f32);
-
-            mtmd_input_chunk chunk{
-                MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                {},
-                image_tokens,
-            };
-            output->emplace_back(std::move(chunk));
-            i_img++;
-        }
-    }
-
-    return output;
-}
-
-void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
-    for (auto & chunk : *chunks) {
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE && chunk.tokens_image) {
-            delete chunk.tokens_image;
-        }
-    }
-    delete chunks;
-}
-
-int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
-    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
-    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
-    bool ok = clip_image_batch_encode(
-        ctx->ctx_clip,
-        ctx->n_threads,
-        &image_tokens->batch_f32,
-        ctx->image_embd_v.data());
-    return ok ? 0 : 1;
-}
-
-float * mtmd_get_output_embd(mtmd_context * ctx) {
-    return ctx->image_embd_v.data();
-}
-
-size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks) {
-    size_t n_tokens = 0;
-    for (auto & chunk : *chunks) {
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            n_tokens += chunk.tokens_text.size();
-        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            n_tokens += chunk.tokens_image->n_tokens();
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
-    }
-    return n_tokens;
-}
-
-// helper struct to make working with embd batch easier
-// note: this will be removed after llama_batch_ext refactoring
-struct decode_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
-int32_t mtmd_helper_eval(mtmd_context * ctx,
-        llama_context * lctx,
-        mtmd_input_chunks * chunks,
-        llama_pos pos0,
-        llama_seq_id seq_id,
-        int32_t n_batch) {
-    int32_t ret;
-    llama_pos n_past = pos0;
-    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
-
-    for (auto & chunk : *chunks) {
-        bool is_last = &chunk == &chunks->back();
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            // TODO @ngxson : may need to split into smaller batches
-            text_batch.n_tokens = chunk.tokens_text.size();
-            for (size_t i = 0; i < chunk.tokens_text.size(); i++) {
-                text_batch.token   [i]    = chunk.tokens_text[i];
-                text_batch.pos     [i]    = n_past++;
-                text_batch.n_seq_id[i]    = 1;
-                text_batch.seq_id  [i][0] = seq_id;
-                text_batch.logits  [i]    = false;
-            }
-            if (is_last) {
-                // always get logits for last input chunk
-                text_batch.logits[text_batch.n_tokens - 1] = true;
-            }
-            ret = llama_decode(lctx, text_batch);
-            if (ret != 0) {
-                LOG_ERR("failed to decode text\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-
-        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
-            GGML_ASSERT(chunk.tokens_image != nullptr);
-            int64_t t0 = ggml_time_ms();
-            if (ctx->print_timings) {
-                LOG_INF("encoding image...\n");
-            }
-            ret = mtmd_encode(ctx, chunk.tokens_image);
-            if (ret != 0) {
-                LOG_ERR("failed to encode image\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-            if (ctx->print_timings) {
-                LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
-            }
-
-            int32_t n_tokens = chunk.tokens_image->n_tokens();
-            float * embd = mtmd_get_output_embd(ctx);
-            decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
-            int64_t t1 = ggml_time_ms();
-            ret = llama_decode(lctx, batch_img.batch);
-            if (ret != 0) {
-                LOG_ERR("failed to decode image\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-            if (ctx->print_timings) {
-                LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
-            }
-
-            n_past += n_tokens;
-
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
-    }
-
-    llama_batch_free(text_batch);
-    return 0;
-}
-
-int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) {
-    clip_image_u8_ptr img_u8(clip_image_u8_init());
-    bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
-    if (!ok) {
-        LOG_ERR("Unable to load image from buffer\n");
-        return 1;
-    }
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
-    output.data.resize(output.nx * output.ny * 3);
-    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
-    return 0;
-}
-
-int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) {
-    clip_image_u8_ptr img_u8(clip_image_u8_init());
-    bool ok = clip_image_load_from_file(fname, img_u8.get());
-    if (!ok) {
-        LOG_ERR("Unable to load image %s\n", fname);
-        return 1;
-    }
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
-    output.data.resize(output.nx * output.ny * 3);
-    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
-    return 0;
-}
--- a/examples/llava/mtmd.h
+++ b/examples/llava/mtmd.h
@@ -1,146 +0,0 @@
-#ifndef MTMD_H
-#define MTMD_H
-
-#include "ggml.h"
-#include "llama.h"
-#include "clip.h"
-
-#include <vector>
-#include <cinttypes>
-#include <memory>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define MTMD_API __declspec(dllexport)
-#        else
-#            define MTMD_API __declspec(dllimport)
-#        endif
-#    else
-#        define MTMD_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define MTMD_API
-#endif
-
-#ifdef __cplusplus
-
-enum mtmd_input_chunk_type {
-    MTMD_INPUT_CHUNK_TYPE_TEXT,
-    MTMD_INPUT_CHUNK_TYPE_IMAGE,
-};
-
-struct mtmd_context;
-struct mtmd_image_tokens;
-
-// represents raw image data, layout is RGBRGBRGB...
-// length of data must be nx * ny * 3
-struct mtmd_bitmap {
-    uint32_t nx;
-    uint32_t ny;
-    std::vector<unsigned char> data;
-};
-
-struct mtmd_input_chunk {
-    mtmd_input_chunk_type type;
-    std::vector<llama_token> tokens_text;
-    mtmd_image_tokens * tokens_image = nullptr;
-};
-
-using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
-
-struct mtmd_context_params {
-    bool use_gpu = true;
-    bool print_timings = true;
-    int n_threads = 4;
-    enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO;
-    const char * image_marker = "<__image__>";
-};
-
-struct mtmd_input_text {
-    std::string text;
-    bool add_special;
-    bool parse_special;
-};
-
-// initialize the mtmd context
-// return nullptr on failure
-MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
-                                                const llama_model * text_model,
-                                                const mtmd_context_params ctx_params);
-
-MTMD_API void mtmd_free(mtmd_context * ctx);
-
-// tokenize an input text prompt and an image
-// the prompt must have the input image marker (default: "<__image__>") in it
-// the marker will be replaced with the image tokens
-// for example:
-//   "here is an image: <__image__>\ndescribe it in detail."
-//   this will gives 3 chunks:
-//   1. "here is an image: <start_of_image>"
-//   2. (image tokens)
-//   3. "<end_of_image>\ndescribe it in detail."
-// number of bitmaps must be equal to the number of image markers in the prompt
-// this function is thread-safe (shared ctx)
-MTMD_API mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
-                                const mtmd_input_text & text,
-                                const std::vector<mtmd_bitmap> & bitmaps);
-
-// free image chunk data
-MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
-
-// returns 0 on success
-MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
-                            const mtmd_image_tokens * image_tokens);
-
-// get output embeddings from the last encode pass
-MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
-
-//
-// helper functions (can be implemented based on other functions)
-//
-
-// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
-MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks);
-
-// helper function that automatically:
-// 1. run llama_decode() on text chunks
-// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
-// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
-// otherwise, returns 0 on success
-MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
-                                llama_context * lctx,
-                                mtmd_input_chunks * chunks,
-                                llama_pos pos0,
-                                llama_seq_id seq_id,
-                                int32_t n_batch);
-
-// helper function to construct a mtmd_bitmap from a file
-// returns 0 on success
-// this function is thread-safe
-MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output);
-
-// helper function to construct a mtmd_bitmap from a buffer
-// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
-// returns 0 on success
-// this function is thread-safe
-MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output);
-
-// convenient unique_ptr wrappers
-struct mtmd_context_deleter {
-    void operator()(mtmd_context * val) { mtmd_free(val); }
-};
-using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
-
-struct mtmd_input_chunks_deleter {
-    void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
-};
-using mtmd_input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
-
-#else
-
-static_assert(false && "C header is not yet supported by this library");
-
-#endif
-
-#endif
--- a/examples/llava/qwen2_vl_surgery.py
+++ b/examples/llava/qwen2_vl_surgery.py
@@ -1,165 +0,0 @@
-import argparse
-from typing import Dict
-
-import torch
-import numpy as np
-from gguf import *
-from transformers import (
-    Qwen2VLForConditionalGeneration,
-    Qwen2VLProcessor,
-    AutoProcessor,
-    Qwen2VLConfig
-)
-
-
-VISION = "clip.vision"
-
-
-def k(raw_key: str, arch: str) -> str:
-    return raw_key.format(arch=arch)
-
-
-def to_gguf_name(name: str) -> str:
-    og = name
-    name = name.replace("text_model", "t").replace("vision_model", "v")
-    name = name.replace("blocks", "blk").replace("embeddings.", "")
-    name = name.replace("attn.", "attn_")
-    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-    name = name.replace("merger.mlp", 'mm')
-    print(f"[to_gguf_name] {og} --> {name}")
-    return name
-
-
-def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
-    vision_model = qwen2vl.visual
-    tensor_map = {}
-    for name, ten in vision_model.state_dict().items():
-        ten = ten.numpy()
-        if 'qkv' in name:
-            if ten.ndim == 2: # weight
-                c3, _ = ten.shape
-            else:             # bias
-                c3 = ten.shape[0]
-            assert c3 % 3 == 0
-            c = c3 // 3
-            wq = ten[:c]
-            wk = ten[c: c * 2]
-            wv = ten[c * 2:]
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-        elif 'merger' in name:
-            if name.endswith("ln_q.weight"):
-                tensor_map['v.post_ln.weight'] = ten
-            elif name.endswith("ln_q.bias"):
-                tensor_map['v.post_ln.bias'] = ten
-            else:
-                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                tensor_map[to_gguf_name(name)] = ten
-        elif 'patch_embed.proj.weight' in name:
-            # NOTE: split Conv3D into Conv2Ds
-            c1, c2, kt, kh, kw = ten.shape
-            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
-        else:
-            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
-
-    for new_name, ten in tensor_map.items():
-        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-            tensor_map[new_name] = ten.astype(np.float32)
-        else:
-            tensor_map[new_name] = ten.astype(dtype)
-    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-    return tensor_map
-
-
-def main(args):
-    if args.data_type == 'fp32':
-        dtype = torch.float32
-        np_dtype = np.float32
-        ftype = 0
-    elif args.data_type == 'fp16':
-        dtype = torch.float32
-        np_dtype = np.float16
-        ftype = 1
-    else:
-        raise ValueError()
-
-    local_model = False
-    model_path = ""
-    model_name = args.model_name
-    print("model_name: ", model_name)
-    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-        model_name, torch_dtype=dtype, device_map="cpu"
-    )
-    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-    vcfg = cfg.vision_config
-
-    if os.path.isdir(model_name):
-        local_model = True
-        if model_name.endswith(os.sep):
-            model_name = model_name[:-1]
-        model_path = model_name
-        model_name = os.path.basename(model_name)
-    fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
-
-    fout = GGUFWriter(path=fname_out, arch="clip")
-    fout.add_description("image encoder for Qwen2VL")
-
-    fout.add_file_type(ftype)
-    fout.add_bool("clip.has_text_encoder", False)
-    fout.add_bool("clip.has_vision_encoder", True)
-    fout.add_bool("clip.has_qwen2vl_merger", True)
-    fout.add_string("clip.projector_type", "qwen2vl_merger")
-
-    print(cfg.vision_config)
-    if 'silu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", True)
-        fout.add_bool("clip.use_gelu", False)
-    elif 'gelu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", False)
-        fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
-    else:
-        raise ValueError()
-
-    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
-    for name, data in tensor_map.items():
-        fout.add_tensor(name, data)
-
-    fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
-    fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
-    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
-    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
-    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
-    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
-    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0)  # not sure what this does, put 0 here as a placeholder
-    fout.add_name(model_name)
-    """
-    HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
-            it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
-    """
-
-    if local_model:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path)
-    else:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
-    fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
-    fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
-
-    fout.write_header_to_file()
-    fout.write_kv_data_to_file()
-    fout.write_tensors_to_file()
-    fout.close()
-    print("save model as: ", fname_out)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
-    parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
-    args = parser.parse_args()
-    main(args)
--- a/examples/llava/tests.sh
+++ b/examples/llava/tests.sh
@@ -1,81 +0,0 @@
-#!/bin/bash
-
-# make sure we are in the right directory
-SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
-cd $SCRIPT_DIR
-
-#export LLAMA_CACHE="$SCRIPT_DIR/tmp"
-
-set -eux
-
-mkdir -p $SCRIPT_DIR/output
-
-PROJ_ROOT="$SCRIPT_DIR/../.."
-cd $PROJ_ROOT
-
-###############
-
-arr_bin=()
-arr_hf=()
-
-add_test() {
-    local bin=$1
-    local hf=$2
-    arr_bin+=("$bin")
-    arr_hf+=("$hf")
-}
-
-add_test "llama-gemma3-cli"   "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
-add_test "llama-llava-cli"    "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
-add_test "llama-llava-cli"    "second-state/Llava-v1.5-7B-GGUF:Q2_K"
-add_test "llama-llava-cli"    "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"
-add_test "llama-llava-cli"    "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
-add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
-add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
-
-###############
-
-cmake --build build -j --target "${arr_bin[@]}"
-
-arr_res=()
-
-for i in "${!arr_bin[@]}"; do
-    bin="${arr_bin[$i]}"
-    hf="${arr_hf[$i]}"
-
-    echo "Running test with binary: $bin and HF model: $hf"
-    echo ""
-    echo ""
-
-    output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty)
-
-    echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
-
-    if echo "$output" | grep -iq "new york"; then
-        result="\033[32mOK\033[0m:   $bin $hf"
-    else
-        result="\033[31mFAIL\033[0m: $bin $hf"
-    fi
-    echo -e "$result"
-    arr_res+=("$result")
-
-    echo ""
-    echo ""
-    echo ""
-    echo "#################################################"
-    echo "#################################################"
-    echo ""
-    echo ""
-done
-
-set +x
-
-for i in "${!arr_res[@]}"; do
-    echo -e "${arr_res[$i]}"
-done
-echo ""
-echo "Output logs are saved in $SCRIPT_DIR/output"
--- a/examples/pydantic_models_to_grammar_examples.py
+++ b/examples/pydantic_models_to_grammar_examples.py
@@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar):
    """Calls the /completion API on llama-server.

    See
-    https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints
+    https://github.com/ggml-org/llama.cpp/tree/HEAD/tools/server#api-endpoints
    """
    print(f"  Request:\n    Grammar:\n{textwrap.indent(gbnf_grammar, '      ')}\n    Prompt:\n{textwrap.indent(prompt.rstrip(), '      ')}")
    headers = {"Content-Type": "application/json"}
--- a/examples/quantize-stats/CMakeLists.txt
+++ b/examples/quantize-stats/CMakeLists.txt
@@ -1,6 +0,0 @@
-set(TARGET llama-quantize-stats)
-add_executable(${TARGET} quantize-stats.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
--- a/examples/server/public/index.html.gz
+++ b/examples/server/public/index.html.gz
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -107,6 +107,7 @@ message(DEBUG "INS_ENB             : ${INS_ENB}")
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
+option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
@@ -170,7 +171,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -360,3 +360,27 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
              ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4005  # Macro redefinition
+        /wd4244  # Conversion from one type to another type, possible loss of data
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+    )
+    function(disable_msvc_warnings target_name)
+        if(TARGET ${target_name})
+            target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+        endif()
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+    disable_msvc_warnings(ggml-cpu-x64)
+    disable_msvc_warnings(ggml-cpu-sse42)
+    disable_msvc_warnings(ggml-cpu-sandybridge)
+    disable_msvc_warnings(ggml-cpu-haswell)
+    disable_msvc_warnings(ggml-cpu-skylakex)
+    disable_msvc_warnings(ggml-cpu-icelake)
+    disable_msvc_warnings(ggml-cpu-alderlake)
+endif()
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -38,7 +38,7 @@ extern "C" {
    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
    GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
    GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
-    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
    GGML_API bool                  ggml_backend_buft_is_host       (ggml_backend_buffer_type_t buft);
    GGML_API ggml_backend_dev_t    ggml_backend_buft_get_device    (ggml_backend_buffer_type_t buft);

@@ -59,7 +59,7 @@ extern "C" {
    GGML_API enum ggml_status               ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
    GGML_API size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
    GGML_API size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor);
    GGML_API void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
    GGML_API bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
    GGML_API void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
--- a/ggml/include/ggml-cpp.h
+++ b/ggml/include/ggml-cpp.h
@@ -24,7 +24,7 @@ typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;

 struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };

-typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+typedef std::unique_ptr<ggml_gallocr, ggml_gallocr_deleter> ggml_gallocr_ptr;

 // ggml-backend

--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -133,6 +133,11 @@ extern "C" {

    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);

+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t);
+
 #ifdef __cplusplus
 }
 #endif
--- a/ggml/include/ggml-rpc.h
+++ b/ggml/include/ggml-rpc.h
@@ -7,6 +7,9 @@
 extern "C" {
 #endif

+#define RPC_PROTO_MAJOR_VERSION    2
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16

 // backend API
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -393,8 +393,8 @@ extern "C" {

    // precision
    enum ggml_prec {
-        GGML_PREC_DEFAULT,
-        GGML_PREC_F32,
+        GGML_PREC_DEFAULT =  0, // stored as ggml_tensor.op_params, 0 by default
+        GGML_PREC_F32     = 10,
    };

    // model file types
@@ -481,6 +481,7 @@ extern "C" {
        GGML_OP_CONV_TRANSPOSE_1D,
        GGML_OP_IM2COL,
        GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D_DW,
        GGML_OP_CONV_TRANSPOSE_2D,
        GGML_OP_POOL_1D,
        GGML_OP_POOL_2D,
@@ -677,6 +678,9 @@ extern "C" {
    GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
    GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2

+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
+
    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);

@@ -1660,7 +1664,7 @@ extern "C" {
            struct ggml_tensor  * a,
            struct ggml_tensor  * b);

-    // depthwise
+    // depthwise (via im2col and mul_mat)
    GGML_API struct ggml_tensor * ggml_conv_2d_dw(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,  // convolution kernel
@@ -1672,6 +1676,22 @@ extern "C" {
            int                  d0,  // dilation dimension 0
            int                  d1); // dilation dimension 1

+    // Depthwise 2D convolution
+    // may be faster than ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
    GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -267,6 +267,7 @@ function(ggml_add_cpu_backend_variant tag_name)
    set(GGML_CPU_TAG_NAME ${tag_name})
    # other: OPENMP LLAMAFILE CPU_HBM
    foreach (feat NATIVE
+                  SSE42
                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
                  AMX_TILE AMX_INT8 AMX_BF16)
@@ -286,14 +287,16 @@ if (GGML_CPU_ALL_VARIANTS)
    if (NOT GGML_BACKEND_DL)
        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
    endif()
-    ggml_add_cpu_backend_variant(sandybridge    AVX)
-    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+    ggml_add_cpu_backend_variant(x64)
+    ggml_add_cpu_backend_variant(sse42        SSE42)
+    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
    if (NOT MSVC)
        # MSVC doesn't support AMX
-        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
    endif()
 elseif (GGML_CPU)
    ggml_add_cpu_backend_variant_impl("")
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
    size_t node_size = 0;
    if (!node->data && !node->view_src) {
-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
    }
    return talloc->size_max >= node_size;
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -56,7 +56,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
    return SIZE_MAX;
 }

-size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
    // get_alloc_size is optional, defaults to ggml_nbytes
    if (buft->iface.get_alloc_size) {
        size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -152,7 +152,7 @@ size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
 }

-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor) {
    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }

--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
--- a/ggml/src/ggml-cann/aclnn_ops.h
+++ b/ggml/src/ggml-cann/aclnn_ops.h
@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS

+#include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
 #include <aclnnop/aclnn_exp.h>
@@ -713,6 +714,270 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 */
 void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);

+/*
+ * @brief A generic wrapper for ACL resources with custom deleter support.
+ */
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+
+/**
+ * @brief Trait structure used to define how to destroy a given ACL resource type.
+ *
+ * @tparam T ACL resource type.
+ */
+template<typename T>
+struct acl_resource_traits;
+
+/**
+ * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
+ */
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
+ */
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
+ */
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
+ */
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
+};
+
+/**
+ * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ *
+ * @tparam T ACL resource type.
+ * @param ptr Raw pointer to ACL resource.
+ * @return any_acl_resource Smart pointer that handles destruction.
+ */
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
+}
+
+/**
+ * @brief Registers multiple ACL resources into a vector for lifetime management.
+ *
+ * @tparam Args Variadic list of ACL resource types.
+ * @param vec Target vector to hold ACL resources.
+ * @param args Raw pointers to ACL resources.
+ */
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+    (vec.emplace_back(make_acl_resource(args)), ...);
+}
+
+/**
+ * @brief Task class that wraps the execution of an aclnn function call.
+ */
+class aclnn_task : public cann_task {
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
+};
+
+/**
+ * @brief Task class that releases ACL resources after usage.
+ */
+class release_resource_task : public cann_task {
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
+
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
+    std::vector<any_acl_resource> resource_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory copy operations.
+ */
+class async_memcpy_task : public cann_task {
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
+    aclrtMemcpyKind kind_;
+    aclrtStream stream_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory set operations.
+ */
+class async_memset_task : public cann_task {
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
+};
+
+/**
+ * @brief Launches an asynchronous task using the memory allocator.
+ *
+ * This macro submit an asynchronous task on the specified stream.
+ * The task uses memory allocated by the allocator. It is guaranteed
+ * that the memory will not be accessed by other tasks until this task
+ * completes, due to the sequential execution order within the same stream.
+ *
+ * @param OP_NAME aclnn operator name.
+ * @param args Additional arguments required by the task.
+ *
+ * @note
+ * Memory from the allocator will be "freed" immediately and can be
+ * reallocated to other pointers. However, it won't be accessed by any
+ * other task before this asynchronous task ends, because all tasks in the
+ * same stream are executed in queue order.
+ */
+
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
+    } while (0)
+
+/**
+ * @brief Registers and releases multiple ACL resources, optionally deferring the release
+ *        using a task.
+ *
+ * @tparam Args Types of the ACL resources.
+ * @param ctx Backend context which manages task submission and async mode.
+ * @param args Pointers to ACL resources to be released.
+ */
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+    std::vector<any_acl_resource> resources;
+    register_acl_resources(resources, std::forward<Args>(args)...);
+    if(ctx.async_mode) {
+        auto task = std::make_unique<release_resource_task>(std::move(resources));
+        ctx.task_queue.submit_task(std::move(task));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param dst Destination memory address.
+ * @param src Source memory address.
+ * @param len Size of memory to copy (in bytes).
+ * @param kind Type of memory copy (host-to-device, device-to-host, etc).
+ */
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
+    }
+}
+
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx->async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
+        ctx->task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param buffer Memory buffer to be set.
+ * @param size Size of the memory buffer (in bytes).
+ * @param value Value to set in the buffer.
+ */
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
+    }
+}
+
 /**
 * @brief Applies a element-wise operation to two input tensors using the CANN
 * backend.
@@ -742,42 +1007,9 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
    binary_op(ctx, acl_src0, acl_src1, acl_dst);

-    ACL_CHECK(aclDestroyTensor(acl_src0));
-    ACL_CHECK(aclDestroyTensor(acl_src1));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }

-/**
- * @brief Launches an asynchronous task using the memory allocator.
- *
- * This macro submit an asynchronous task on the specified stream.
- * The task uses memory allocated by the allocator. It is guaranteed
- * that the memory will not be accessed by other tasks until this task
- * completes, due to the sequential execution order within the same stream.
- *
- * @param OP_NAME aclnn operator name.
- * @param args Additional arguments required by the task.
- *
- * @note
- * Memory from the allocator will be "freed" immediately and can be
- * reallocated to other pointers. However, it won't be accessed by any
- * other task before this asynchronous task ends, because all tasks in the
- * same stream are executed in queue order.
- */
-#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...)                                                \
-    do {                                                                                     \
-        uint64_t        workspaceSize = 0;                                                   \
-        aclOpExecutor * executor;                                                            \
-        void *          workspaceAddr = nullptr;                                             \
-                                                                                             \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
-                                                                                             \
-        if (workspaceSize > 0) {                                                             \
-            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);             \
-            workspaceAddr = workspace_allocator.get();                                       \
-        }                                                                                    \
-        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream()));     \
-    } while (0)

 /**
 * @brief Applies a unary operation to an input tensor using the CANN backend.
@@ -799,9 +1031,7 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
    aclTensor* acl_dst = ggml_cann_create_tensor(dst);

    unary_op(ctx, acl_src, acl_dst);
-
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }

 /**
@@ -832,7 +1062,7 @@ void ggml_cann_unary_op(
 *
 * Internally, the lambda will call:
 * @code
- * GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);
+ * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
 * @endcode
 *
 * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
@@ -840,14 +1070,14 @@ void ggml_cann_unary_op(
 * @see ggml_cann_unary_op
 * @see GGML_CANN_CALL_ACLNN_OP
 */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                         \
-    do {                                                         \
-        auto lambda = [](ggml_backend_cann_context& ctx,         \
-            aclTensor* acl_src,                                  \
-            aclTensor* acl_dst) {                                \
-            GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);  \
-        };                                                       \
-        ggml_cann_unary_op(lambda, ctx, dst);                    \
-    }                                                            \
+#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_unary_op(lambda, ctx, dst);                         \
+    }                                                                 \
    while (0)
 #endif  // CANN_ACLNN_OPS
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -31,9 +31,16 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <unistd.h>
+#include <functional>

 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
+#include "../ggml-impl.h"

 #define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
@@ -205,6 +212,127 @@ struct ggml_cann_pool_alloc {
    ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };

+/**
+ * @brief Function pointer type for ACLNN operator calls.
+ */
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+
+/**
+ * @brief Base class for all CANN tasks to be submitted to the task queue.
+ *
+ * Users should override the run_task() method with actual task logic.
+ */
+class cann_task {
+public:
+    virtual void run_task() {}
+};
+
+/**
+ * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
+ */
+class cann_task_queue {
+public:
+    /**
+     * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
+     *
+     * @param capacity Queue capacity. Must be a power of 2.
+     * @param device Target device ID (used for context setting).
+     */
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
+        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
+        mask_ = capacity_ - 1;
+    }
+
+    /**
+     * @brief Attempts to enqueue a task into the queue.
+     *
+     * @param item Unique pointer to the task.
+     * @return true if the task was successfully enqueued, false if the queue was full.
+     */
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
+        size_t next_tail = (tail_ + 1) & mask_;
+
+        if (next_tail == head_) {
+            return false;
+        }
+
+        buffer_[tail_] = std::move(item);
+        std::atomic_thread_fence(std::memory_order_release);
+        tail_ = next_tail;
+
+        return true;
+    }
+
+    /**
+     * @brief Submits a task to the queue, and starts the worker thread if not already running.
+     *
+     * @param task Task to be submitted.
+     */
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
+            std::this_thread::yield();
+            continue;
+        }
+
+        if (!running_) {
+            running_ = true;
+            thread_ = std::thread(&cann_task_queue::execute, this);
+        }
+
+    }
+
+    /**
+     * @brief Waits until the queue is completely empty and no tasks are being processed.
+     */
+    void wait() {
+        while (running_ && head_ != tail_) {
+            std::this_thread::yield();
+            continue;
+        }
+    }
+
+    /**
+     * @brief Stops the task queue and joins the worker thread.
+     */
+    void stop() {
+        running_ = false;
+        if (thread_.joinable()) {
+            thread_.join();
+        }
+    }
+
+private:
+    /**
+     * @brief Worker thread function that continuously dequeues and executes tasks.
+     */
+    void execute() {
+        ggml_cann_set_device(device_);
+
+        while (running_) {
+            if(head_ == tail_) {
+                std::this_thread::yield();
+                continue;
+            }
+
+            std::atomic_thread_fence(std::memory_order_acquire);
+            buffer_[head_]->run_task();
+            buffer_[head_].reset();
+            head_ = (head_ + 1) & mask_;
+        }
+    }
+
+    std::vector<std::unique_ptr<cann_task>> buffer_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
+};
+
 /**
 * @brief Context for managing CANN backend operations.
 */
@@ -213,6 +341,8 @@ struct ggml_backend_cann_context {
    std::string name;                /**< Name of the device. */
    std::string description;         /**< Description of the device. */
    aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    cann_task_queue task_queue;
+    bool async_mode;

    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */

@@ -221,9 +351,12 @@ struct ggml_backend_cann_context {
     * @param device Device ID.
     */
    explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
        ggml_cann_set_device(device);
        description = aclrtGetSocName();
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
    }

    /**
@@ -231,6 +364,7 @@ struct ggml_backend_cann_context {
     */
    ~ggml_backend_cann_context() {
        ggml_cann_set_device(device);
+        task_queue.stop();
        if (copy_event != nullptr) {
            ACL_CHECK(aclrtDestroyEvent(copy_event));
        }
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -156,195 +156,196 @@ const ggml_cann_device_info& ggml_cann_info() {
 * This class manages a pool of CANN buffers for a specific device.
 */
 struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
-        /**
-         * @brief The maximum reuse margin for a buffer.
-         */
-        static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB

-        /**
-         * @brief The minimum free margin for a buffer.
-         */
-        static const size_t min_free_margin = 1ull << 20;   // 1MB
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB

-        /**
-         * @brief The alignment for buffer allocation.
-         */
-        static const size_t alignment = 128;
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;

-        /**
-         * @brief The device ID associated with this buffer pool.
-         */
-        int device;
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;

-        /**
-         * @brief Whether to disable clean during buffer allocation.
-         */
-        bool disable_clean = false;
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;

-        /**
-         * @brief Structure representing a CANN buffer.
-         */
-        struct ggml_cann_buffer {
-            void* ptr = nullptr;  ///< Pointer to the buffer.
-            size_t size = 0;      ///< Size of the buffer.
-            std::chrono::steady_clock::time_point last_used;  ///< Last used time.
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer.
+        size_t size = 0;      ///< Size of the buffer.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.

-            bool operator>(const ggml_cann_buffer& other) const {
-                return size > other.size;
-            }
-        };
+        bool operator>(const ggml_cann_buffer& other) const {
+            return size > other.size;
+        }
+    };

-        /**
-         * @brief Array of CANN buffers in the pool.
-         */
-        std::unordered_map<void*, size_t> buffer_pool;
-        std::priority_queue<ggml_cann_buffer,
-                            std::vector<ggml_cann_buffer>,
-                            std::greater<>> free_buffers ;
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    std::unordered_map<void*, size_t> buffer_pool;
+    std::priority_queue<ggml_cann_buffer,
+                        std::vector<ggml_cann_buffer>,
+                        std::greater<>> free_buffers ;

-        /**
-         * @brief Total size of all buffers in the pool.
-         */
-        size_t pool_size = 0;
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;

-        /**
-         * @brief Constructor to initialize the buffer pool for a specific device.
-         *
-         * @param device The device ID to associate with this buffer pool.
-         */
-        explicit ggml_cann_pool_buf_prio(int device) : device(device) {
-            disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_buf_prio(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_buf_prio() {
+        ggml_cann_set_device(device);
+        for (auto& [b_ptr, b_size] : buffer_pool) {
+            aclrtFree(b_ptr);
+            pool_size -= b_size;
+        }
+        buffer_pool.clear();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
        }

-        /**
-         * @brief Destructor to free all buffers in the pool.
-         */
-        ~ggml_cann_pool_buf_prio() {
-            ggml_cann_set_device(device);
-            for (auto& [b_ptr, b_size] : buffer_pool) {
-                aclrtFree(b_ptr);
-               pool_size -= b_size;
-            }
-            buffer_pool.clear();
-            GGML_ASSERT(pool_size == 0);
-        }
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();

-        /**
-         * @brief Allocate a buffer of the given size.
-         *
-         * @param size The size of the buffer to allocate.
-         * @param actual_size A pointer to a variable to receive the actual size of
-         * the allocated buffer.
-         * @return A pointer to the allocated buffer.
-         */
-        void* alloc(size_t size, size_t* actual_size) override {
-            size = GGML_PAD(size, alignment);
-            if (size == 0) {
-                size = alignment;
-            }
+        std::vector<ggml_cann_buffer> free_buffers_rest;
+        free_buffers_rest.reserve(free_buffers.size());
+        while (!free_buffers.empty()) {
+            auto b = free_buffers.top();
+            free_buffers.pop();

-            void* ptr = nullptr;
-            auto now = std::chrono::steady_clock::now();
-
-            std::vector<ggml_cann_buffer> free_buffers_rest;
-            free_buffers_rest.reserve(free_buffers.size());
-            while (!free_buffers.empty()) {
-                auto b = free_buffers.top();
-                free_buffers.pop();
-
-                if (b.size >= size) {
-                    // reuse the buffer if the size is enough
-                    const size_t margin = b.size - size;
-                    if (margin <= max_reuse_margin) {
-                        *actual_size = b.size;
-                        ptr = b.ptr;
-    #ifdef DEBUG_CANN_MALLOC
-                        GGML_LOG_INFO(
-                            "cann pool[%d]: reused   %p, "
-                            "pool_size = %5u MB, "
-                            "size = %5u MB, "
-                            "margin = %5u MB\n",
-                            device, b.ptr,
-                            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
-                            (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
-                            (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
-    #endif
-                        break;
-                    }
-                }
-
-                bool should_clean = !disable_clean &&
-                                   b.size > min_free_margin &&
-                                   std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
-                if (should_clean) {
-                    // free the buffer if the size is needed to be freed
-                    ACL_CHECK(aclrtFree(b.ptr));
-                    pool_size -= b.size;
-                    buffer_pool.erase(b.ptr);
-    #ifdef DEBUG_CANN_MALLOC
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    ptr = b.ptr;
+#ifdef DEBUG_CANN_MALLOC
                    GGML_LOG_INFO(
-                        "cann pool[%d]: clean    %p, "
+                        "cann pool[%d]: reused   %p, "
                        "pool_size = %5u MB, "
-                        "size = %5u MB\n",
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
                        device, b.ptr,
                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
-                        (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
-    #endif
-                    continue;
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
+#endif
+                    break;
                }
-                free_buffers_rest.push_back(b);
-            }
-            for (ggml_cann_buffer &b : free_buffers_rest) {
-                free_buffers.push(std::move(b));
            }

-    #ifdef DEBUG_CANN_MALLOC
-            GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
-    #endif
-            if (ptr != nullptr) {
-                return ptr;
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+                buffer_pool.erase(b.ptr);
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                continue;
            }
+            free_buffers_rest.push_back(b);
+        }
+        for (ggml_cann_buffer &b : free_buffers_rest) {
+            free_buffers.push(std::move(b));
+        }

-            // allocate a new buffer if no buffer can be reused
-            ggml_cann_set_device(device);
-            ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-            *actual_size = size;
-            pool_size += size;
-    #ifdef DEBUG_CANN_MALLOC
-            GGML_LOG_INFO(
-                "cann pool[%d]: allocate %p, "
-                "pool_size = %5u MB, "
-                "size = %5u MB\n",
-                device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
-                (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
-    #endif
-            buffer_pool.emplace(ptr, size);
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+        if (ptr != nullptr) {
            return ptr;
        }

-        /**
-         * @brief Free a buffer and return it to the pool.
-         *
-         * @param ptr Pointer to the buffer to free.
-         * @param size Size of the buffer to free.
-         */
-        void free(void* ptr, size_t size) override {
-            auto it = buffer_pool.find(ptr);
-            if (it == buffer_pool.end()) {
-                GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
-            }
+        // allocate a new buffer if no buffer can be reused
+        ggml_cann_set_device(device);
+        ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: allocate %p, "
+            "pool_size = %5u MB, "
+            "size = %5u MB\n",
+            device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+            (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
+#endif
+        buffer_pool.emplace(ptr, size);
+        return ptr;
+    }

-            auto now = std::chrono::steady_clock::now();
-            free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
-    #ifdef DEBUG_CANN_MALLOC
-            GGML_LOG_INFO(
-                "cann pool[%d]: return   %p, "
-                "pool_size = %5u MB\n",
-                device, ptr,
-                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
-    #endif
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
+        auto it = buffer_pool.find(ptr);
+        if (it == buffer_pool.end()) {
+            GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
        }
-    };
+
+        auto now = std::chrono::steady_clock::now();
+        free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: return   %p, "
+            "pool_size = %5u MB\n",
+            device, ptr,
+            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+    }
+};

 /**
 * @brief A pool of CANN buffers(segment buffer).
@@ -531,6 +532,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
     * @param size Size of the buffer to free.
     */
    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
        for (int i = 0; i < MAX_BUFFERS; ++i) {
            ggml_cann_buffer& b = buffer_pool[i];
            if (b.ptr != ptr) {
@@ -1604,7 +1606,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                    auto lambda = [](ggml_backend_cann_context& ctx,
                        aclTensor* acl_src,
                        aclTensor* acl_dst) {
-                        GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst);
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
                    };
                    ggml_cann_unary_op(lambda, ctx, dst);
                } break;
@@ -1787,12 +1789,11 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
    delete backend;
 }

+
 /**
 * @brief Sets tensor data asynchronously in the CANN backend.
 *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
+ * This function asynchronously sets tensor data in the CANN backend.
 *
 * @param backend Pointer to the CANN backend structure.
 * @param tensor Pointer to the tensor structure to set data for.
@@ -1807,23 +1808,28 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
                                               size_t size) {
    ggml_backend_cann_context *cann_ctx =
        (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+        "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));

-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
+    ggml_cann_async_memcpy(cann_ctx, (char *)tensor->data + offset, data, size,
+        ACL_MEMCPY_HOST_TO_DEVICE);
 }

+/**
+ * @brief Gets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously gets tensor data in the CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to get data from.
+ * @param data Pointer to the host data to copy from the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
 static void ggml_backend_cann_get_tensor_async(
    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
    size_t offset, size_t size) {
@@ -1834,20 +1840,11 @@ static void ggml_backend_cann_get_tensor_async(

    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
                "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
+
+    ggml_cann_async_memcpy(cann_ctx, data, (char *)tensor->data + offset, size,
+        ACL_MEMCPY_DEVICE_TO_HOST);

-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
 }

 /**
@@ -1907,6 +1904,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
        ggml_cann_set_device(cann_ctx_src->device);
        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));

+        // wait for task_queue empty to keep task order.
+        cann_ctx_src->task_queue.wait();
        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
                                   cann_ctx_src->stream()));
@@ -1934,9 +1933,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
    ggml_backend_cann_context* cann_ctx =
        (ggml_backend_cann_context*)backend->context;
-
+    cann_ctx->task_queue.wait();
    ggml_cann_set_device(cann_ctx->device);
-
    ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }

@@ -2020,6 +2018,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                    return true;
                case GGML_TYPE_Q8_0:
                case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
                    // only support contiguous for quantized types.
                    return ggml_is_contiguous(op->src[0]) &&
                            ggml_is_contiguous(op->src[1]);
@@ -2087,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                return false;
            }

+            if(!ggml_is_contiguous(op->src[0])){
+                return false;
+            }
            return true;
        }
        case GGML_OP_UPSCALE: {
@@ -2102,6 +2107,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
        }
        case GGML_OP_POOL_2D: {
            const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
            const int       k0   = opts[1];
            const int       k1   = opts[2];
            const int       p0   = opts[5];
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -222,7 +222,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            elseif (GGML_AVX)
                list(APPEND ARCH_FLAGS /arch:AVX)
                list(APPEND ARCH_DEFINITIONS GGML_AVX)
-            else ()
+            elseif (GGML_SSE42)
                list(APPEND ARCH_FLAGS /arch:SSE4.2)
                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
            endif()
@@ -237,8 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            if (GGML_NATIVE)
                list(APPEND ARCH_FLAGS -march=native)
            else ()
-                list(APPEND ARCH_FLAGS -msse4.2)
-                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_SSE42)
+                    list(APPEND ARCH_FLAGS -msse4.2)
+                    list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                endif()
                if (GGML_F16C)
                    list(APPEND ARCH_FLAGS -mf16c)
                    list(APPEND ARCH_DEFINITIONS GGML_F16C)
@@ -350,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        # TODO: Separation to determine activation of VX/VXE/VXE2
        if (${S390X_M} MATCHES "8561|8562")
            message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+            list(APPEND ARCH_FLAGS -march=z15)
        elseif (${S390X_M} MATCHES "3931")
            message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=z17)
        else()
            message(STATUS "Unknown target")
            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
--- a/ggml/src/ggml-cpu/cpu-feats-x86.cpp
+++ b/ggml/src/ggml-cpu/cpu-feats-x86.cpp
@@ -263,7 +263,7 @@ void test_x86_is() {
 static int ggml_backend_cpu_x86_score() {
    // FIXME: this does not check for OS support

-    int score = 0;
+    int score = 1;
    cpuid_x86 is;

 #ifdef GGML_FMA
--- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
@@ -4044,7 +4044,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
    UNUSED(ncols_interleaved);
    UNUSED(blocklen);

-#if defined(__AVX2__)
+#if defined(__AVX2__) || defined(__AVX512F__)
    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx;
    const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
    int64_t b_nb = n / QK_K;
@@ -4054,8 +4054,748 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
    const __m256i m4b = _mm256_set1_epi8(0x0F);
    // Permute mask used for easier vector processing at later stages
    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
-
+    int64_t xstart = 0;
    int anr = nr - nr % 16;; // Used to align nr with boundary of 16
+#ifdef __AVX512F__
+    int anc = nc - nc % 16; // Used to align nc with boundary of 16
+    // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+    //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
+
+        const block_q8_Kx4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //4-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+
+                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
+                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
+                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
+                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
+                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
+                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
+                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
+                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
+                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
+                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+
+                    // Shuffle pattern two - right side input
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+
+                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
+
+                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
+                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
+                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
+                    utmp_00[2] = uaux_00;
+                    utmp_00[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
+                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
+                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
+                    utmp_01[2] = uaux_01;
+                    utmp_01[0] &= kmask1;
+
+                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
+                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
+                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
+                    utmp_10[2] = uaux_10;
+                    utmp_10[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
+                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
+                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
+                    utmp_11[2] = uaux_11;
+                    utmp_11[0] &= kmask1;
+
+                    // Scales of first sub block in the sb loop
+                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
+
+                    // Scales of second sub block in the sb loop
+                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
+
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    for (int rp = 0; rp < 4; rp++) {
+
+                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                        __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                        __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                        __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                        __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                        __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
+                        __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
+                        __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
+                        __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
+                        __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                        __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                        __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                        __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                        __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
+                        __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
+                        __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
+                        __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
+
+                        __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                        __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                        __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                        __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+                        __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
+                        __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
+                        __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
+                        __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
+
+                        __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                        __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                        __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                        __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+                        __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
+                        __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
+                        __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
+                        __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+
+                        // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
+                        __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb)));
+                        __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
+                        lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
+                        __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
+
+                        // Shuffle pattern one - left side input
+                        const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                        const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+                        const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                        const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+                        const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
+                        const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
+                        const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
+                        const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
+
+                        const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                        const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+                        const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                        const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+                        const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
+                        const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
+                        const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
+                        const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
+
+                        const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                        const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+                        const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                        const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+                        const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
+                        const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
+                        const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
+                        const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
+
+                        const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                        const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+                        const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                        const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+                        const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
+                        const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
+                        const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
+                        const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
+                        __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
+                        __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
+                        __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
+                        __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
+                        __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
+                        __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
+                        __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
+
+                        __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
+                        __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
+                        __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
+                        __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
+                        __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
+                        __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
+                        __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
+                        __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                        __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                        __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                        __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                        __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                        __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                        __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                        __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                        iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                        iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                        iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                        iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                        iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                        iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                        iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                        iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                        // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
+                        __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
+                        __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
+                        __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
+                        __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
+
+                        __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
+                        __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
+                        __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
+                        __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
+
+                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
+                        const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                        const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                        acc_rows[rp * 4  + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+
+                        __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
+                        __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
+                        __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
+                        __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
+
+                        acc_min_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]);
+                        acc_min_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]);
+                        acc_min_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]);
+                        acc_min_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[rp * 4 + 3]);
+                    }
+                }
+            }
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+
+    for (; y < nr / 4; y++) {
+
+        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //4-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+
+                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
+                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
+                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
+                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
+                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
+                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
+                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
+                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
+                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
+                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+
+                    // Shuffle pattern two - right side input
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+
+                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
+
+                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
+                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
+                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
+                    utmp_00[2] = uaux_00;
+                    utmp_00[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
+                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
+                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
+                    utmp_01[2] = uaux_01;
+                    utmp_01[0] &= kmask1;
+
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
+                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
+                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
+                    utmp_10[2] = uaux_10;
+                    utmp_10[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
+                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
+                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
+                    utmp_11[2] = uaux_11;
+                    utmp_11[0] &= kmask1;
+
+                    // Scales of first sub block in the sb loop
+                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
+
+                    // Scales of second sub block in the sb loop
+                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
+
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                    __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                    __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                    __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                    __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
+                    __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
+                    __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
+                    __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
+                    __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                    __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                    __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                    __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                    __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
+                    __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
+                    __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
+                    __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
+
+                    //Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into a 512 bit vector
+                    __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                    __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                    __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                    __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+                    __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
+                    __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
+                    __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
+                    __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
+
+                    __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                    __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                    __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                    __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+                    __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
+                    __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
+                    __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
+                    __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+
+                    // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
+                    __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb)));
+                    __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
+                    lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
+                    __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
+
+                    // Shuffle pattern one - left side input
+                    const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                    const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+                    const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                    const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+                    const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
+                    const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
+                    const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
+                    const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
+
+                    const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                    const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+                    const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                    const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+                    const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
+                    const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
+                    const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
+                    const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
+
+                    const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                    const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+                    const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                    const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+                    const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
+                    const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
+                    const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
+                    const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
+
+                    const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                    const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+                    const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                    const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+                    const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
+                    const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
+                    const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
+                    const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
+                    __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
+                    __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
+                    __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
+                    __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
+                    __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
+                    __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
+                    __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
+
+                    __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
+                    __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
+                    __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
+                    __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
+                    __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
+                    __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
+                    __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
+                    __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                    __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                    __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                    __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                    __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                    __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                    __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                    __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                    iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                    iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                    iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                    iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                    iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                    iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                    iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                    iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                    // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
+                    __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
+                    __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
+                    __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
+                    __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
+
+                    __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
+                    __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
+                    __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
+                    __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
+
+                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
+                    const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                    const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+
+                    __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
+                    __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
+                    __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
+                    __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
+
+                    acc_min_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]);
+                    acc_min_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]);
+                    acc_min_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]);
+                    acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
+                }
+            }
+            // Store accumlated values
+            for (int i = 0; i < 4; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+    if (anc != nc) {
+        xstart = anc/8;
+        y = 0;
+    }
+#endif //AVX512F
+
    // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
    for (; y < anr / 4; y += 4) {

@@ -4067,7 +4807,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
        }

        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
-        for (int64_t x = 0; x < nc / 8; x++) {
+        for (int64_t x = xstart; x < nc / 8; x++) {

            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);

@@ -4401,7 +5141,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c

        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);

-        for (int64_t x = 0; x < nc / 8; x++) {
+        for (int64_t x = xstart; x < nc / 8; x++) {

            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);

--- a/ggml/src/ggml-cpu/ggml-cpu-quants.c
+++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c
@@ -6596,7 +6596,118 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    }

    *s = hsum_float_8(acc);
+#elif defined(__VXE__) || defined(__VXE2__)
+    uint32_t aux[3];
+    uint32_t utmp[4];

+    const int32x4_t v_z = vec_splat_s32(0);
+    const uint8x16_t v_3m = vec_splat_u8(0x03);
+
+    const uint8x16_t v_0c = vec_splat_u8(1);
+    const uint8x16_t v_1c = vec_sl(v_0c, 1);
+    const uint8x16_t v_2c = vec_sl(v_0c, 2);
+    const uint8x16_t v_3c = vec_sl(v_0c, 3);
+
+    uint8x16_t q3h[4];
+    uint8x16_t q3b[2];
+    int8x16_t q3bytes[4];
+    int8x16_t q8bytes[4];
+    uint8x16_t qhbits[2];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict x0l = x[i].qs;
+        const uint8_t * restrict x0h = x[i].hmask;
+        const int8_t  * restrict y0  = y[i].qs;
+
+        qhbits[0] = vec_xl(0 , x0h);
+        qhbits[1] = vec_xl(16, x0h);
+
+        int32_t isum = 0;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            int32x4_t isum0, isum1, isum2, isum3;
+
+            q3b[0] = vec_xl(0 , x0l);
+            q3b[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            q8bytes[0] = vec_xl(0  , y0);
+            q8bytes[1] = vec_xl(16 , y0);
+            q8bytes[2] = vec_xl(32 , y0);
+            q8bytes[3] = vec_xl(48 , y0);
+            q8bytes[4] = vec_xl(64 , y0);
+            q8bytes[5] = vec_xl(80 , y0);
+            q8bytes[6] = vec_xl(96 , y0);
+            q8bytes[7] = vec_xl(112, y0);
+            y0 += 128;
+
+            q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2);
+            q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2);
+            q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1);
+            q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            q3h[0] = vec_andc(v_2c, qhbits[0]);
+            q3h[1] = vec_andc(v_2c, qhbits[1]);
+            q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1);
+            q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits[0] = vec_sr(qhbits[0], 4);
+                qhbits[1] = vec_sr(qhbits[1], 4);
+            }
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
 #else
    // scalar version
    // This function is written like this so the compiler can manage to vectorize most of it
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -215,7 +215,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
        .nrows                    = 1,
    },
    [GGML_TYPE_F16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp16,
        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
        .vec_dot_type             = GGML_TYPE_F16,
        .nrows                    = 1,
@@ -356,7 +356,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
        .from_float               = quantize_row_q8_K,
    },
    [GGML_TYPE_BF16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_bf16,
        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
        .vec_dot_type             = GGML_TYPE_BF16,
        .nrows                    = 1,
@@ -1932,6 +1932,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_im2col_back_f32(params, tensor);
            } break;
+        case GGML_OP_CONV_2D_DW:
+            {
+                ggml_compute_forward_conv_2d_dw(params, tensor);
+            } break;
        case GGML_OP_CONV_TRANSPOSE_2D:
            {
                ggml_compute_forward_conv_transpose_2d(params, tensor);
@@ -2268,6 +2272,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
            } break;
        case GGML_OP_IM2COL:
        case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D_DW:
        case GGML_OP_CONV_TRANSPOSE_1D:
        case GGML_OP_CONV_TRANSPOSE_2D:
            {
@@ -3161,6 +3166,93 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
    return ggml_graph_compute(cgraph, &cplan);
 }

+void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m512 x_vec = _mm512_loadu_ps(x + i);
+        __m256i y_vec = _mm512_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm256_storeu_si256((__m256i *)(y + i), y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m256 x_vec = _mm256_loadu_ps(x + i);
+        __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storeu_si128((__m128i *)(y + i), y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128 x_vec = _mm_loadu_ps(x + i);
+        __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storel_epi64((__m128i *)(y + i), y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(x[i]);
+    }
+}
+
+void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m256i x_vec = _mm256_loadu_si256((const __m256i *)(x + i));
+        __m512 y_vec = _mm512_cvtph_ps(x_vec);
+        _mm512_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m128i x_vec = _mm_loadu_si128((const __m128i *)(x + i));
+        __m256 y_vec = _mm256_cvtph_ps(x_vec);
+        _mm256_storeu_ps(y + i, y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128i x_vec = _mm_loadl_epi64((const __m128i *)(x + i));
+        __m128 y_vec = _mm_cvtph_ps(x_vec);
+        _mm_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+}
+
+void ggml_cpu_fp32_to_bf16(const float * x, ggml_bf16_t * y, int64_t n) {
+    int64_t i = 0;
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_BF16(x[i]);
+    }
+}
+
+void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__AVX2__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i,
+                        _mm512_castsi512_ps(
+                            _mm512_slli_epi32(
+                                _mm512_cvtepu16_epi32(
+                                    _mm256_loadu_si256(
+                                        (const __m256i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i,
+                        _mm256_castsi256_ps(
+                            _mm256_slli_epi32(
+                                _mm256_cvtepu16_epi32(
+                                    _mm_loadu_si128(
+                                        (const __m128i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_BF16_TO_FP32(x[i]);
+    }
+}

 int ggml_cpu_has_avx(void) {
 #if defined(__AVX__)
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -11,24 +11,26 @@
 #include <vector>

 #ifdef GGML_USE_CPU_HBM
-#include "ggml-cpu-hbm.h"
+#    include "ggml-cpu-hbm.h"
 #endif

 #ifdef GGML_USE_CPU_KLEIDIAI
-#include "kleidiai/kleidiai.h"
-#endif
-
-#if defined(__APPLE__)
-#include <sys/types.h>
-#include <sys/sysctl.h>
+#    include "kleidiai/kleidiai.h"
 #endif

 #if defined(_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-    #define NOMINMAX
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#else
+#    include <unistd.h>
 #endif
-#include <windows.h>
+
+#if defined(__APPLE__)
+#    include <sys/sysctl.h>
+#    include <sys/types.h>
 #endif

 // ggml-backend interface
@@ -70,8 +72,10 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_ty
 }

 static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
-        if (extra && extra == buft) return true;
+    for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra == buft) {
+            return true;
+        }
    }
    return false;
 }
@@ -330,9 +334,18 @@ static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t d
 }

 static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
+#ifdef _WIN32
+    MEMORYSTATUSEX status;
+    status.dwLength = sizeof(status);
+    GlobalMemoryStatusEx(&status);
+    *total = status.ullTotalPhys;
+    *free = status.ullAvailPhys;
+#else
+    long pages = sysconf(_SC_PHYS_PAGES);
+    long page_size = sysconf(_SC_PAGE_SIZE);
+    *total = pages * page_size;
+    *free = *total;
+#endif

    GGML_UNUSED(dev);
 }
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -1054,6 +1054,493 @@ class tinyBLAS_Q0_AVX {
      } \
   } \

+template <typename TA, typename TB, typename TC>
+class tinyBLAS_BF16_PPC {
+  public:
+    tinyBLAS_BF16_PPC(int64_t k,
+                const TA *A, int64_t lda,
+                const TB *B, int64_t ldb,
+                TC *C, int64_t ldc,
+                int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        if (numVec == 2) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[2], c[3], swiz1);
+            s[0] = vec_perm(t[0], t[1], swiz3);
+            s[1] = vec_perm(t[0], t[1], swiz4);
+            vec_xst(s[0], 0, (vec_t*)vecOffset);
+            vec_xst(s[1], 0, (vec_t*)(vecOffset + 16));
+        } else if (numVec == 4) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[0], c[1], swiz2);
+            t[2] = vec_perm(c[2], c[3], swiz1);
+            t[3] = vec_perm(c[2], c[3], swiz2);
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            for (int i = 0; i < 4; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        } else if (numVec == 8) {
+            for (int i = 0; i < 4; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            for (int i = 4; i < 8; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            s[4] = vec_perm(t[4], t[6], swiz3);
+            s[5] = vec_perm(t[4], t[6], swiz4);
+            s[6] = vec_perm(t[5], t[7], swiz3);
+            s[7] = vec_perm(t[5], t[7], swiz4);
+            for (int i = 0; i < 8; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        }
+    }
+
+    void packNormal(const TA* a, int64_t lda, int rows, int cols, unsigned char* vec) {
+        int64_t i, j;
+        TA *aoffset = NULL;
+        unsigned char *vecOffset = NULL;
+        TA * aoffsets[8];
+        vector unsigned char c_arr[8];
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                if (cols == 4) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 4; ++it)
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffset += 4 * lda;
+                    for (int i = 0; i < 4; ++i)
+                        c_arr[i] = vec_xl(0, (vector unsigned char*)aoffsets[i]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int i = 0; i<4; i++)
+                        aoffsets[i] = aoffsets[i]+lda;
+                    vecOffset +=64;
+                }
+                i = (cols >> 3);
+                if (i > 0) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 8; ++it) {
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    }
+                    aoffset += 8 * lda;
+                    do {
+                        for (int it = 0; it < 8; ++it)
+                            c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                        vector_permute_store(c_arr, 8, vecOffset);
+                        for (int it = 0; it < 8; ++it)
+                            aoffsets[it] = aoffsets[it] + 8*lda;
+                        vecOffset += 128;
+                        i--;
+                    } while(i > 0);
+                }
+                j--;
+            } while(j > 0);
+        }
+        if (rows & 4) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            aoffset += 4 * lda;
+            if (cols == 4) {
+                for (int it = 0; it < 4; ++it)
+                    c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                    aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    for (int it = 0; it < 4; ++it)
+                        c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it< 4; it++)
+                        aoffsets[it] = aoffsets[it] + 8*lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+        if (rows & 3) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            if (cols == 4) {
+                switch(rows) {
+                    case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                    case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                    case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                        break;
+                }
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                     aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    switch(rows) {
+                        case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                        case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                        case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                            break;
+                    }
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it <4; it++)
+                         aoffsets[it] = aoffsets[it] + 8* lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+    }
+
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        int m_rem = MIN(m - m0, 8);
+        int n_rem = MIN(n - n0, 8);
+
+        if (m_rem >= 8 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 8) {
+            mc = 4;
+            nc = 8;
+            gemm<4,8>(m0, m, n0, n);
+        } else if (m_rem >=8 && n_rem >=4){
+                mc = 8;
+                nc = 4;
+                gemm<8,4>(m0, m, n0, n);
+        } else if ((m_rem < 4) && (n_rem >= 8)) {
+            nc = 8;
+            switch(m_rem) {
+                case 1:
+                    mc = 1;
+                    gemm_Mx8<1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    mc = 2;
+                    gemm_Mx8<2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    mc = 3;
+                    gemm_Mx8<3>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        } else if (m_rem >= 4 && n_rem >= 4) {
+            mc = 4;
+            nc = 4;
+            gemm_small<4, 4>(m0, m, n0, n);
+        } else if ((m_rem > 4) && (n_rem < 4)) {
+            mc = 4;
+            switch(n_rem) {
+                case 1:
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+
+                default:
+                    return;
+            }
+        } else {
+            switch((m_rem << 4) | n_rem) {
+                case 0x43:
+                    mc = 4;
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+                case 0x42:
+                    mc = 4;
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 0x41:
+                    mc = 4;
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 0x34:
+                    mc = 3;
+                    nc = 4;
+                    gemm_small<3, 4>(m0, m, n0, n);
+                    break;
+                case 0x33:
+                    mc = 3;
+                    nc = 3;
+                    gemm_small<3, 3>(m0, m, n0, n);
+                    break;
+                case 0x32:
+                    mc = 3;
+                    nc = 2;
+                    gemm_small<3, 2>(m0, m, n0, n);
+                    break;
+                case 0x31:
+                    mc = 3;
+                    nc = 1;
+                    gemm_small<3, 1>(m0, m, n0, n);
+                    break;
+                case 0x24:
+                    mc = 2;
+                    nc = 4;
+                    gemm_small<2,4>(m0, m, n0, n);
+                    break;
+                case 0x23:
+                    mc = 2;
+                    nc = 3;
+                    gemm_small<2, 3>(m0, m, n0, n);
+                    break;
+                case 0x22:
+                    mc = 2;
+                    nc = 2;
+                    gemm_small<2, 2>(m0, m, n0, n);
+                    break;
+                case 0x21:
+                    mc = 2;
+                    nc = 1;
+                    gemm_small<2, 1>(m0, m, n0, n);
+                    break;
+                case 0x14:
+                    mc = 1;
+                    nc = 4;
+                    gemm_small<1, 4>(m0, m, n0, n);
+                    break;
+                case 0x13:
+                    mc = 1;
+                    nc = 3;
+                    gemm_small<1, 3>(m0, m, n0, n);
+                    break;
+                case 0x12:
+                    mc = 1;
+                    nc = 2;
+                    gemm_small<1, 2>(m0, m, n0, n);
+                    break;
+                case 0x11:
+                    mc = 1;
+                    nc = 1;
+                    gemm_small<1, 1>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[8] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+    }
+
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[4] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii+4, jj);
+    }
+
+
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[8], vec_C[4];
+        acc_t acc_0, acc_1, acc_2, acc_3;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        __builtin_mma_xxsetaccz(&acc_2);
+        __builtin_mma_xxsetaccz(&acc_3);
+        for (int l = 0; l < k; l+=8) {
+            packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
+            packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
+                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+            }
+        }
+
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+        SAVE_ACC(&acc_2, ii+4, jj);
+        SAVE_ACC(&acc_3, ii+4, jj+4);
+    }
+
+    template<int RM, int RN>
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0;
+            __builtin_mma_xxsetaccz(&acc_0);
+            vec_t vec_A[2], vec_B[2];
+            for (int l=0; l<k; l+=4) {
+                packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
+                for (int x = 0; x<2; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < RN; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM>
+    void gemm_Mx8(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int RN = 8;
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0, acc_1;
+            __builtin_mma_xxsetaccz(&acc_0);
+            __builtin_mma_xxsetaccz(&acc_1);
+            vec_t vec_A[4], vec_B[8];
+            for (int l=0; l<k; l+=8) {
+                packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
+                for (int x = 0; x<4; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_1);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+4+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+       if constexpr(RM == 4 && RN == 8) {
+          KERNEL_4x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 8) {
+          KERNEL_8x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 4) {
+          KERNEL_8x4(ii,jj);
+       } else {
+          static_assert(false, "RN/RM values not supported");
+       }
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            kernel<RM, RN>(ii, jj);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
 template <typename TA, typename TB, typename TC>
 class tinyBLAS_Q0_PPC {
  public:
@@ -2202,6 +2689,7 @@ class tinyBLAS_PPC {
        boffset = vec;
        j = (rows >> 3);
        if (j > 0) {
+
            do {
                aoffset1 = aoffset;
                aoffset2 = aoffset1 + lda;
@@ -2875,9 +3363,22 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                (float *)C, ldc};
            return tb.matmul(m, n);
        }
+#elif defined(__MMA__)
+        if ((k % 8))
+                return false;
+        if(Btype == GGML_TYPE_BF16) {
+           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+            (const ggml_bf16_t *)A, lda,
+            (const ggml_bf16_t *)B, ldb,
+            (float *)C, ldc,
+            params->ith, params->nth};
+        tb.matmul(m, n);
+        return true;
+        }
 #endif
        return false;
    }
+
    case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
        if (Btype == GGML_TYPE_F16) {
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -4222,7 +4222,7 @@ static void ggml_compute_forward_get_rows_f16(

        GGML_ASSERT(i01 >= 0 && i01 < ne01);

-        ggml_fp16_to_fp32_row(
+        ggml_cpu_fp16_to_fp32(
            (const ggml_fp16_t*) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                       (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
    }
@@ -4263,7 +4263,7 @@ static void ggml_compute_forward_get_rows_bf16(

        GGML_ASSERT(i01 >= 0 && i01 < ne01);

-        ggml_bf16_to_fp32_row(
+        ggml_cpu_bf16_to_fp32(
            (const ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
    }
@@ -6064,6 +6064,178 @@ void ggml_compute_forward_conv_transpose_2d(
    }
 }

+// ggml_compute_forward_conv_2d_dw
+
+struct ggml_conv_2d_dw_params {
+    int64_t channels;
+    int64_t batch;
+    int64_t src_w;
+    int64_t src_h;
+    int64_t dst_w;
+    int64_t dst_h;
+    int64_t knl_w;
+    int64_t knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+};
+
+static void ggml_compute_forward_conv_2d_dw_cwhn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t c = p.channels;
+    const float * knl_data = (const float *)kernel->data;
+
+    const int64_t rows_total = p.dst_h * p.batch;
+    const int64_t rows_per_thread = (rows_total + params->nth - 1) / params->nth;
+    const int64_t row_start = params->ith * rows_per_thread;
+    const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
+
+#ifdef GGML_SIMD
+    const int64_t pkg_size = GGML_F32_EPR;
+    const int64_t pkg_count = c / pkg_size;
+    const int64_t c_pkg_end = pkg_count * pkg_size;
+#else
+    const int64_t c_pkg_end = 0;
+#endif
+
+    for (int64_t row = row_start; row < row_end; ++row) {
+        const int64_t dst_y = row % p.dst_h;
+        const float * src_data = (const float *)src->data + (row / p.dst_h) * p.src_w * p.src_h * c;
+        for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+            float * dst_data = (float *)dst->data + (row * p.dst_w + dst_x) * c;
+            const int64_t src_y_base = dst_y * p.stride_y - p.pad_y;
+            const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;
+
+#ifdef GGML_SIMD
+            // Vectorized loop
+            for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {
+                GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(knl_data + (knl_y * p.knl_w + knl_x) * c + c_i);
+                        GGML_F32_VEC s = GGML_F32_VEC_LOAD(src_data + (src_y * p.src_w + src_x) * c + c_i);
+                        sum = GGML_F32_VEC_FMA(sum, k, s);
+                    }
+                }
+                GGML_F32_VEC_STORE(dst_data + c_i, sum);
+            }
+#endif
+            // Scalar loop
+            for (int64_t c_i = c_pkg_end; c_i < c; ++c_i) {
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[(knl_y * p.knl_w + knl_x) * c + c_i]
+                             * src_data[(src_y * p.src_w + src_x) * c + c_i];
+                    }
+                }
+                dst_data[c_i] = sum;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_2d_dw_whcn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t n = p.channels * p.batch;
+    const int64_t per_thread = (n + params->nth - 1) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end = MIN(start + per_thread, n);
+
+    for (int64_t i = start; i < end; ++i) {
+        const float * knl_data = (const float *)kernel->data + (i % p.channels) * p.knl_w * p.knl_h;
+        const float * src_data = (const float *)src->data + i * p.src_w * p.src_h;
+        float * dst_data = (float *)dst->data + i * p.dst_w * p.dst_h;
+
+        for (int64_t dst_y = 0; dst_y < p.dst_h; ++dst_y) {
+            for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[knl_y * p.knl_w + knl_x]
+                             * src_data[src_y * p.src_w + src_x];
+                    }
+                }
+                dst_data[dst_y * p.dst_w + dst_x] = sum;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_2d_dw(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * src = dst->src[1];
+    ggml_conv_2d_dw_params p;
+    p.channels = src->ne[2];
+    p.batch = src->ne[3];
+    p.src_w = src->ne[0];
+    p.src_h = src->ne[1];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.knl_w = kernel->ne[0];
+    p.knl_h = kernel->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(kernel->ne[3] == p.channels);
+    GGML_ASSERT(dst->ne[3] == p.batch);
+
+    if (ggml_is_contiguous(src)) {
+        ggml_compute_forward_conv_2d_dw_whcn(params, src, kernel, dst, p);
+    } else if (ggml_is_contiguous_channels(src)) {
+        // kernel should also have channels most contiguous in memory
+        GGML_ASSERT(kernel->nb[0] >= kernel->nb[2] && kernel->nb[1] >= kernel->nb[0]);
+        ggml_compute_forward_conv_2d_dw_cwhn(params, src, kernel, dst, p);
+    } else {
+        GGML_ABORT("non-contiguous memory layout not supported");
+    }
+}
+
 // ggml_compute_forward_pool_1d_sk_p0

 static void ggml_compute_forward_pool_1d_sk_p0(
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -65,6 +65,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
--- a/ggml/src/ggml-cpu/simd-mappings.h
+++ b/ggml/src/ggml-cpu/simd-mappings.h
@@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_EPR  4

 #define GGML_F32x4              vector float
-#define GGML_F32x4_ZERO         0.0f
+#define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
 #define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -133,6 +133,7 @@ if (CUDAToolkit_FOUND)
                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
                OUTPUT_VARIABLE CUDA_CCVER
                ERROR_QUIET
+                OUTPUT_STRIP_TRAILING_WHITESPACE
            )
        else()
            if (CUDA_CCFULLVER MATCHES Apple)
@@ -143,7 +144,7 @@ if (CUDAToolkit_FOUND)
            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
        endif()

-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+        message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")

        ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -78,13 +78,13 @@
 // Moore Threads
 #define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)

-#define GGML_CUDA_CC_QY1  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD

 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)

 #ifdef __CUDA_ARCH_LIST__
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -1,6 +1,8 @@
 #include "convert.cuh"
 #include "dequantize.cuh"

+#include <cstdint>
+
 #define CUDA_Q8_0_NE_ALIGN 2048

 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
 }

 template <typename src_t, typename dst_t>
-static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void convert_unary(
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;

-    if (i >= k) {
+    if (i00 >= ne00) {
        return;
    }

+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
    const src_t * x = (const src_t *) vx;

-    y[i] = float(x[i]);
+    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
+    y[iy] = float(x[ix]);
 }

 template <typename src_t, typename dst_t>
-static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void convert_unary_cuda(const void * vx, dst_t * y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    convert_unary_cuda<src_t>(vx, y, k, 1, 1, 1, k, k, k, stream);
 }

 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
    switch (type) {
        case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
        case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
        default:
            return nullptr;
    }
@@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
        case GGML_TYPE_IQ3_S:
            return dequantize_row_iq3_s_cuda;
        case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
        case GGML_TYPE_BF16:
-            return convert_unary_cuda<nv_bfloat16>;
+            return convert_unary_cont_cuda<nv_bfloat16>;
        default:
            return nullptr;
    }
@@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
        case GGML_TYPE_IQ3_S:
            return dequantize_row_iq3_s_cuda;
        case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cont_cuda<nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
        case GGML_TYPE_BF16:
            return convert_unary_cuda<nv_bfloat16>;
        default:
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -3,7 +3,7 @@
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256

 template<typename T>
-using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream);

 typedef to_t_cuda_t<float> to_fp32_cuda_t;
 typedef to_t_cuda_t<half> to_fp16_cuda_t;
@@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);

 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
+
+// TODO more general support for non-contiguous inputs
+
+template<typename T>
+using to_t_nc_cuda_t = void (*)(const void * x, T * y,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
+    int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
+
+typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -551,7 +551,7 @@ static void ggml_cpy_f16_f16_cuda(
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }

-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
    const int64_t ne = ggml_nelements(src0);
    GGML_ASSERT(ne == ggml_nelements(src1));

@@ -588,10 +588,12 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
    char ** dest_ptrs_d = nullptr;
    int graph_cpynode_index = -1;
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
    }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
        GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
@@ -636,16 +638,19 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
                ggml_type_name(src0->type), ggml_type_name(src1->type));
    }
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
    }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif

 }

 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
 }

 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
--- a/ggml/src/ggml-cuda/cpy.cuh
+++ b/ggml/src/ggml-cuda/cpy.cuh
@@ -2,7 +2,7 @@

 #define CUDA_CPY_BLOCK_SIZE 64

-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);

 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -33,8 +33,8 @@ static __global__ void k_get_rows(
    dfloat2 v;
    dequantize_kernel(src0_row, ib, iqs, v);

-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
 }

 template<typename src0_t, typename dst_t>
@@ -60,7 +60,7 @@ static __global__ void k_get_rows_float(
    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
    const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);

-    dst_row[i00] = src0_row[i00];
+    dst_row[i00] = float(src0_row[i00]);
 }

 template<typename grad_t, typename dst_t>
@@ -86,120 +86,159 @@ static __global__ void k_get_rows_back_float(
    dst[dst_row*ncols + col] = sum;
 }

-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
    const dim3 block_nums(block_num_x, ne10, ne11*ne12);

    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);

-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);

    GGML_ASSERT(ne00 % 2 == 0);

    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
        ne00, /*ne01, ne02, ne03,*/
        /*ne10, ne11,*/ ne12, /*ne13,*/
        /* s0,*/ s1, s2, s3,
        /* nb00,*/ nb01, nb02, nb03,
        s10, s11, s12/*, s13*/);
-
-    GGML_UNUSED(dst);
 }

-template<typename src0_t>
+template<typename src0_t, typename dst_t>
 static void get_rows_cuda_float(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne13 == 1);
-
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
    const dim3 block_nums(block_num_x, ne10, ne11*ne12);

    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);

-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);

    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
        ne00, /*ne01, ne02, ne03,*/
        /*ne10, ne11,*/ ne12, /*ne13,*/
        /* s0,*/ s1, s2, s3,
        /* nb00,*/ nb01, nb02, nb03,
        s10, s11, s12/*, s13*/);
+}

-    GGML_UNUSED(dst);
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
 }

 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];

-    const void    * src0_d = (const void    *) src0->data;
-    const int32_t * src1_d = (const int32_t *) src1->data;
-    float         * dst_d  = (float         *) dst->data;
-
    cudaStream_t stream = ctx.stream();

+    GGML_TENSOR_BINARY_OP_LOCALS
+
    GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ne13 == 1);

    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
    GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));

-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            break;
-    }
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
 }

 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
--- a/ggml/src/ggml-cuda/getrows.cuh
+++ b/ggml/src/ggml-cuda/getrows.cuh
@@ -3,6 +3,13 @@
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256

+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -96,31 +96,32 @@ int ggml_cuda_get_device() {

 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
    ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
-    auto res = hipMallocManaged(ptr, size);
-    if (res == hipSuccess) {
-        // if error we "need" to know why...
-        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
-    }
-    return res;
-#else
-
-#if !defined(GGML_USE_HIP)
    cudaError_t err;
    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
    {
        err = cudaMallocManaged(ptr, size);
+#if defined(GGML_USE_HIP)
+        if (err == hipSuccess) {
+            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+        }
+
+        // fall back to cudaMalloc if not supported (e.g. on Windows)
+        if (err == hipErrorNotSupported) {
+            static bool warned_unsupported = false;
+            if (!warned_unsupported) {
+                GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
+                warned_unsupported = true;
+            }
+
+            err = cudaMalloc(ptr, size);
+        }
+#endif // defined(GGML_USE_HIP)
    }
    else
    {
        err = cudaMalloc(ptr, size);
    }
    return err;
-#else
-    return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIP)
-
-#endif
 }

 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -554,8 +555,8 @@ static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer

    if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
        // initialize padding to 0 to avoid possible NaN values
-        size_t original_size = ggml_nbytes(tensor);
-        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+        const size_t original_size = ggml_nbytes(tensor);
+        const size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);

        if (padded_size > original_size) {
            ggml_cuda_set_device(ctx->device);
@@ -678,6 +679,7 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t

    if (ggml_is_quantized(tensor->type)) {
        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            GGML_ASSERT(tensor->nb[0] == ggml_element_size(tensor));
            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
        }
    }
@@ -799,6 +801,7 @@ static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buff

 static enum ggml_status ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
    GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");

    ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
@@ -850,6 +853,7 @@ static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buff
    // split tensors must always be set in their entirety at once
    GGML_ASSERT(offset == 0);
    GGML_ASSERT(size == ggml_nbytes(tensor));
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");

    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;

@@ -888,6 +892,7 @@ static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buff
    // split tensors must always be set in their entirety at once
    GGML_ASSERT(offset == 0);
    GGML_ASSERT(size == ggml_nbytes(tensor));
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");

    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;

@@ -969,6 +974,7 @@ static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buf

 static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
    ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
+    GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors");

    size_t total_size = 0;

@@ -1409,6 +1415,11 @@ static void ggml_cuda_op_mul_mat(
    const int64_t ne0 = dst->ne[0];
    const int64_t ne1 = dst->ne[1];

+    // const int64_t nb10 = src1->nb[0];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+
    const int64_t nb2 = dst->nb[2];
    const int64_t nb3 = dst->nb[3];

@@ -1544,7 +1555,10 @@ static void ggml_cuda_op_mul_mat(
            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);

            if (src1_on_device && src1_is_contiguous) {
-                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
+                quantize_src1(
+                    dev[id].src1_ddf, nullptr, dev[id].src1_ddq, src0->type, ne10,
+                    nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float),
+                    src1_padded_col_size, ne11, ne12, ne13, stream);
                CUDA_CHECK(cudaGetLastError());
            }
        }
@@ -1639,7 +1653,9 @@ static void ggml_cuda_op_mul_mat(
                }

                if (quantize_src1 && !src1_is_contiguous) {
-                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
+                    quantize_src1(
+                        src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_padded_col_size, src1_ncols, 1, 1, stream);
                    CUDA_CHECK(cudaGetLastError());
                }

@@ -1709,15 +1725,15 @@ static __global__ void k_compute_batched_ptrs(
        size_t  nb12, size_t  nb13,
        size_t  nbd2, size_t  nbd3,
        int64_t r2,   int64_t r3) {
-    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
+    const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;

    if (i13 >= ne13 || i12 >= ne12) {
        return;
    }

-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
+    const int64_t i03 = i13 / r3;
+    const int64_t i02 = i12 / r2;

    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
@@ -1731,6 +1747,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
    GGML_ASSERT(src0->type == GGML_TYPE_F16);

+    // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
+    // As long as dst is contiguous this does not matter though.
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t ne_dst = ggml_nelements(dst);
@@ -1739,21 +1759,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co

    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));

-    void * src0_ddq = src0->data;
-    half * src0_f16 = (half *) src0_ddq;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
+    const half * src0_f16 = (const half *) src0->data;
+    float * dst_ddf = (float *) dst->data;
+
+    const half * src1_f16 = (const half *) src1->data;
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    GGML_ASSERT(nb10 == ts_src1);
+    int64_t s11 = nb11 / ts_src1;
+    int64_t s12 = nb12 / ts_src1;
+    int64_t s13 = nb13 / ts_src1;
+    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());

    // convert src1 to fp16
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
    if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
        const int64_t ne_src1 = ggml_nelements(src1);
        src1_f16_alloc.alloc(ne_src1);
        GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+
+        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+
+        src1_f16 = src1_f16_alloc.get();
+        s11 = ne10;
+        s12 = ne11*s11;
+        s13 = ne12*s12;
    }
-    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();

    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
    char * dst_t;
@@ -1813,13 +1843,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                int i02 = i12 / r2;

                CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                    ne01, ne11, ne10,
+                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
+                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
+                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
+                    cu_compute_type,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
            }
        }
    }
@@ -1830,15 +1860,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
        CUBLAS_CHECK(
        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
-                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
-                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
-                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
+                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
+                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
                ne12*ne13,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
    } else {
        // use cublasGemmBatchedEx
-        const int ne23 = ne12*ne13;
+        const int64_t ne23 = ne12*ne13;

        ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
        ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
@@ -1850,8 +1880,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                ne12, ne13,
                ne23,
                nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
-                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
+                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
+                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
                nbd2, nbd3,
                r2, r3);
        CUDA_CHECK(cudaGetLastError());
@@ -1860,8 +1890,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
                ne23,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1877,7 +1907,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);

-    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
+    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
@@ -1918,12 +1948,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);

-    if (!split && use_mul_mat_vec && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
-        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
-               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_vec_q) {
+        ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_q) {
+        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
+            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
        // general KQ + KQV multi-batch without FlashAttention
        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
    } else if (use_mul_mat_vec) {
@@ -1937,196 +1971,146 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    }
 }

-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
    const ggml_tensor * ids  = dst->src[2];

+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+
    GGML_TENSOR_BINARY_OP_LOCALS

-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (ne2 == 1) {
+            if (ggml_is_quantized(src0->type)) {
+                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+            } else {
+                ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+            }
+            return;
+        }
+
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+            ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
+            return;
+        }
+    }

    cudaStream_t stream = ctx.stream();

-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
+        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
+    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
+    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<int32_t> ids_to_sorted_host;
+    ids_to_sorted_host.reserve(2*ne_get_rows);
+    std::vector<int32_t> ids_from_sorted_host(ne_get_rows);
+
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool(), 2*ne_get_rows);
+
+    std::vector<int32_t> tokens_per_expert(ne02);
+
+    ggml_cuda_pool_alloc<char> src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted);
+    ggml_cuda_pool_alloc<char>  dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted);

    std::vector<char> ids_host(ggml_nbytes(ids));
-    const char * ids_dev = (const char *) ids->data;
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
    CUDA_CHECK(cudaStreamSynchronize(stream));

-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
-
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
-
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    if (ne12 == 1) {
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-            for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-                const int64_t i11 = id % ne11;
-                const int64_t i12 = iid1;
-
-                const int64_t i1 = id;
-                const int64_t i2 = i12;
-
-                src0_row.data = src0_original + i02*nb02;
-                src1_row.data = src1_original + i11*nb11 + i12*nb12;
-                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
-
-                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-            }
-        }
-    } else {
-        ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
-        ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
-
-        src1_row.data = src1_contiguous.get();
-        dst_row.data  =  dst_contiguous.get();
-
-        for (int64_t i02 = 0; i02 < n_as; i02++) {
-            int64_t num_src1_rows = 0;
-
-            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-                for (int64_t id = 0; id < n_ids; id++) {
-                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                    if (row_id_i != i02) {
-                        continue;
-                    }
-
-                    num_src1_rows++;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size();
+                    ids_to_sorted_host.push_back(i12*ne11 + iex % ne11);
+                    tokens_per_expert[i02]++;
+                    break;
                }
            }
-
-            if (num_src1_rows == 0) {
-                continue;
-            }
-
-            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne10, 768u));
-                dim3 grid_dims(ids->ne[1], n_ids);
-                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        src1_original, src1_contiguous.get(),
-                        dev_cur_src1_row.get(), dev_row_mapping.get(),
-                        ids_dev, i02, ids->nb[1], ids->nb[0],
-                        ne11, ne10,
-                        nb11, nb12);
-                CUDA_CHECK(cudaGetLastError());
-            }
-
-            src0_row.data = src0_original + i02*nb02;
-
-            GGML_ASSERT(nb11 == sizeof(float)*ne10);
-            GGML_ASSERT(nb1 == sizeof(float)*ne0);
-
-            src1_row.ne[1] = num_src1_rows;
-            src1_row.nb[1] = nb11;
-            src1_row.nb[2] = num_src1_rows*nb11;
-            src1_row.nb[3] = num_src1_rows*nb11;
-
-            dst_row.ne[1] = num_src1_rows;
-            dst_row.nb[1] = nb1;
-            dst_row.nb[2] = num_src1_rows*nb1;
-            dst_row.nb[3] = num_src1_rows*nb1;
-
-            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne0, 768u));
-                dim3 grid_dims(num_src1_rows);
-                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        dst_original, dst_contiguous.get(),
-                        dev_row_mapping.get(),
-                        ne0,
-                        nb1, nb2);
-                CUDA_CHECK(cudaGetLastError());
-            }
        }
    }
+    GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows));
+
+    ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_to_sorted   = ids_buf_dev.ptr + 0*ne_get_rows;
+    const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows;
+
+    get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
+        ne10, nb11, nb12, nb13,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    char * src1_data_cur = (char *) src1_sorted.ptr;
+    char *  dst_data_cur = (char *)  dst_sorted.ptr;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) {
+        if (tokens_per_expert[i02] == 0) {
+            continue;
+        }
+
+        ggml_tensor src0_slice = *src0;
+        src0_slice.ne[2] = 1;
+        src0_slice.nb[3] = src0_slice.nb[2];
+        src0_slice.data  = (char *) src0->data + i02*nb02;
+        GGML_ASSERT(!ggml_cuda_should_use_mmq(src0->type, cc, ne11) || ne00 % MATRIX_ROW_PADDING == 0);
+
+        ggml_tensor src1_slice;
+        memset(&src1_slice, 0, sizeof(src1_slice));
+        src1_slice.buffer = src1->buffer;
+        src1_slice.type   = type_src1_sorted;
+        src1_slice.ne[0]  = ne10;
+        src1_slice.ne[1]  = tokens_per_expert[i02];
+        src1_slice.ne[2]  = 1;
+        src1_slice.ne[3]  = 1;
+        src1_slice.nb[0]  = ts_src1_sorted;
+        src1_slice.nb[1]  = src1_slice.ne[0] * src1_slice.nb[0];
+        src1_slice.nb[2]  = src1_slice.ne[1] * src1_slice.nb[1];
+        src1_slice.nb[3]  = src1_slice.ne[2] * src1_slice.nb[2];
+        src1_slice.data   = src1_data_cur;
+
+        ggml_tensor dst_slice;
+        memset(&dst_slice, 0, sizeof(dst_slice));
+        dst_slice.buffer = dst->buffer;
+        dst_slice.type   = type_dst_sorted;
+        dst_slice.ne[0]  = ne0;
+        dst_slice.ne[1]  = tokens_per_expert[i02];
+        dst_slice.ne[2]  = 1;
+        dst_slice.ne[3]  = 1;
+        dst_slice.nb[0]  = ts_dst_sorted;
+        dst_slice.nb[1]  = dst_slice.ne[0] * dst_slice.nb[0];
+        dst_slice.nb[2]  = dst_slice.ne[1] * dst_slice.nb[1];
+        dst_slice.nb[3]  = dst_slice.ne[2] * dst_slice.nb[2];
+        dst_slice.data   = dst_data_cur;
+
+        ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice);
+        CUDA_CHECK(cudaGetLastError());
+
+        src1_data_cur += src1_slice.nb[2];
+        dst_data_cur  +=  dst_slice.nb[2];
+    }
+
+    get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type,
+        ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        nb1, nb2, nb3, stream);
 }

 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
@@ -2488,7 +2472,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
        }

-        if (node->op == GGML_OP_MUL_MAT_ID || node->op == GGML_OP_CONT || node->op == GGML_OP_DUP) {
+        if (node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
            use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
@@ -3202,9 +3186,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
        }
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK: {
-            const size_t ts = ggml_type_size(op->src[0]->type);
-            const int64_t ne0_012 = op->src[0]->ne[0] * op->src[0]->ne[1] * op->src[0]->ne[2];
-            return op->src[0]->nb[0] == ts && op->src[0]->nb[3] == ne0_012*ts;
+            return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
        }
        case GGML_OP_IM2COL:
        case GGML_OP_POOL_2D:
@@ -3236,6 +3218,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            if (op->src[0]->ne[0] == 192) {
                return false;
            }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek MLA
+                return false;
+            }
            if (op->src[0]->ne[3] != 1) {
                return false;
            }
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -1,37 +1,10 @@
 #include "mmq.cuh"
+#include "quantize.cuh"

-void ggml_cuda_op_mul_mat_q(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
+#include <vector>

-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
-
-    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
-    // Also its fixup needs to allocate a temporary buffer in the memory pool.
-    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
-
-    switch (src0->type) {
+static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    switch (args.type_x) {
        case GGML_TYPE_Q4_0:
            mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
            break;
@@ -90,10 +63,205 @@ void ggml_cuda_op_mul_mat_q(
            GGML_ABORT("fatal error");
            break;
    }
+}
+
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    cudaStream_t stream = ctx.stream();
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+
+    const char  * src0_d = (const char  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    // If src0 is a temporary compute buffer, clear any potential padding.
+    if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+        const size_t size_data  = ggml_nbytes(src0);
+        const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
+        if (size_alloc > size_data) {
+            CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
+        }
+    }
+
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+
+    if (!ids) {
+        const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+            get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+        ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+        {
+            const int64_t s11 = src1->nb[1] / ts_src1;
+            const int64_t s12 = src1->nb[2] / ts_src1;
+            const int64_t s13 = src1->nb[3] / ts_src1;
+            quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
+                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        }
+
+        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        const int64_t s13 = ne12*s12;
+
+        const mmq_args args = {
+            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
+            ne00, ne01, ne1, s01, s1,
+            ne02, ne12, s02, s12, s2,
+            ne03, ne13, s03, s13, s3,
+            use_stream_k};
+        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+        return;
+    }
+
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    std::vector<int32_t> ids_src1_host;
+    ids_src1_host.reserve(ne_get_rows);
+    std::vector<int32_t> ids_dst_host;
+    ids_dst_host.reserve(ne_get_rows);
+    std::vector<int32_t> tokens_per_expert_host(ne02);
+    std::vector<int32_t> expert_bounds_host(ne02 + 1);
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
+                    ids_dst_host.push_back(i12*ne1 + iex);
+                    tokens_per_expert_host[i02]++;
+                    break;
+                }
+            }
+        }
+    }
+
+    int32_t cumsum = 0;
+    for (int64_t i = 0; i < ne02; ++i) {
+        expert_bounds_host[i] = cumsum;
+        cumsum += tokens_per_expert_host[i];
+    }
+    expert_bounds_host[ne02] = cumsum;
+
+    std::vector<int32_t> ids_buf_host;
+    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
+    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
+    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
+    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
+    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
+
+    const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+        get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+    const int64_t ne11_flat = ne12*n_expert_used;
+    const int64_t ne12_flat = 1;
+    const int64_t ne13_flat = 1;
+
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[2] / ts_src1;
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+    }
+
+    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s13 = ne12*s12;
+
+    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
+    const mmq_args args = {
+        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        ne00, ne01, ne_get_rows, s01, s1,
+        ne02, ne02, s02, s12, s2,
+        ne03, ne13, s03, s13, s3,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+}
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride01 = ne00 / ggml_blck_size(src0->type);
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const mmq_args args = {
+        src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
+        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        1, 1, 0, 0, 0,
+        1, 1, 0, 0, 0,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);

    GGML_UNUSED(src1);
    GGML_UNUSED(dst);
    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_padded_row_size);
 }

 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
--- a/ggml/src/ggml-cuda/mmv.cu
+++ b/ggml/src/ggml-cuda/mmv.cu
@@ -4,18 +4,23 @@

 template <typename T, typename type_acc, int block_size>
 static __global__ void mul_mat_vec(
-        const T * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
        const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) {
-    const int64_t row       = blockIdx.x;
-    const int64_t channel   = blockIdx.y;
-    const int64_t sample    = blockIdx.z;
-    const int     tid       = threadIdx.x;
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int64_t row         = blockIdx.x;
+    const int64_t channel_dst = blockIdx.y;
+    const int64_t channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int64_t channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
+    const int64_t sample_dst  = blockIdx.z;
+    const int64_t sample_x    = sample_dst / sample_ratio;
+    const int64_t sample_y    = sample_dst;
+    const int     tid         = threadIdx.x;
+    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();

-    x   +=  (sample/sample_ratio)*stride_sample_x   + (channel/channel_ratio)*stride_channel_x + row*stride_row;
-    y   +=   sample              *stride_sample_y   +  channel               *stride_channel_y;
-    dst +=   sample              *stride_sample_dst +  channel               *stride_channel_dst;
+    x   += sample_x  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    y   += sample_y  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;

    const float2 * y2 = (const float2 *) y;

@@ -31,12 +36,19 @@ static __global__ void mul_mat_vec(

    float sumf = 0.0f;

-    if constexpr (std::is_same<T, half>::value) {
+    if constexpr (std::is_same<T, float>::value) {
+        const float2 * x2 = (const float2 *) x;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = x2[col2];
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x*tmpy.x;
+            sumf += tmpx.y*tmpy.y;
+        }
+    } else if constexpr (std::is_same<T, half>::value) {
        const half2 * x2 = (const half2 *) x;

        if (std::is_same<type_acc, float>::value) {
-            sumf = 0.0f;
-
            for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
                const float2 tmpx = __half22float2(x2[col2]);
                const float2 tmpy = y2[col2];
@@ -59,8 +71,6 @@ static __global__ void mul_mat_vec(
        }
    } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
        const int * x2 = (const int *) x;
-        sumf = 0.0f;
-
        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
            const int    tmpx = x2[col2];
            const float2 tmpy = y2[col2];
@@ -92,17 +102,17 @@ static __global__ void mul_mat_vec(

 template <typename T, typename type_acc>
 static void launch_mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
        cudaStream_t stream) {
    GGML_ASSERT(ncols      % 2 == 0);
    GGML_ASSERT(stride_row % 2 == 0);
-    GGML_ASSERT(nchannels_y % nchannels_x == 0);
-    GGML_ASSERT(nsamples_y  % nsamples_x  == 0);
-    const int64_t channel_ratio = nchannels_y / nchannels_x;
-    const int64_t sample_ratio  = nsamples_y  / nsamples_x;
+    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
+    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
    int device;
    int warp_size;

@@ -124,48 +134,48 @@ static void launch_mul_mat_vec_cuda(
    }

    const int smem = warp_size*sizeof(float);
-    const dim3 block_nums(nrows, nchannels_y, nsamples_y);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
    const dim3 block_dims(block_size_best, 1, 1);
    switch (block_size_best) {
        case   32: {
            mul_mat_vec<T, type_acc,  32><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case   64: {
            mul_mat_vec<T, type_acc,  64><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case   96: {
            mul_mat_vec<T, type_acc,  96><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case  128: {
            mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case  160: {
            mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case  192: {
            mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case  224: {
            mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        case  256: {
            mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
        } break;
        default: {
            GGML_ABORT("fatal error");
@@ -175,28 +185,28 @@ static void launch_mul_mat_vec_cuda(

 template<typename T>
 static void mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
        enum ggml_prec prec, cudaStream_t stream) {
-    switch (prec) {
-        case GGML_PREC_DEFAULT: {
+    if constexpr(std::is_same<T, half>::value) {
+        if (prec == GGML_PREC_DEFAULT) {
            launch_mul_mat_vec_cuda<T, half>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
-        case GGML_PREC_F32: {
-            launch_mul_mat_vec_cuda<T, float>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
+                (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            return;
+        }
    }
+    launch_mul_mat_vec_cuda<T, float>
+        (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }

-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(         dst->type == GGML_TYPE_F32);

    GGML_TENSOR_BINARY_OP_LOCALS;

@@ -204,21 +214,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
    const size_t ts_src1 = ggml_type_size(src1->type);
    const size_t ts_dst  = ggml_type_size(dst->type);

-    GGML_ASSERT(ne11 == 1);
-    GGML_ASSERT(ne12 == ne2);
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
    GGML_ASSERT(ne13 == ne3);

-    GGML_ASSERT(nb00 == ts_src0);
-    GGML_ASSERT(nb10 == ts_src1);
-    GGML_ASSERT(nb0  == ts_dst);
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+    GGML_ASSERT(        nb0        == ts_dst);

    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;

-    const float * src1_d = (const float *) src1->data;
-    float       *  dst_d = (float       *)  dst->data;
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;

    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = src1->nb[1] / ts_src1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
    const int64_t s02 = src0->nb[2] / ts_src0;
    const int64_t s12 = src1->nb[2] / ts_src1;
    const int64_t s2  =  dst->nb[2] / ts_dst;
@@ -226,14 +239,33 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
    const int64_t s13 = src1->nb[3] / ts_src1;
    const int64_t s3  =  dst->nb[3] / ts_dst;

+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
+
+    GGML_ASSERT(ncols_dst == 1);
+
    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
        case GGML_TYPE_F16: {
            const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
        } break;
        case GGML_TYPE_BF16: {
            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
        } break;
        default:
            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -262,27 +294,34 @@ void ggml_cuda_op_mul_mat_vec(
    const int64_t stride_row         = ne00;
    const int64_t nchannels_x        = 1;
    const int64_t nchannels_y        = 1;
+    const int64_t nchannels_dst      = 1;
    const int64_t stride_channel_x   = 0;
    const int64_t stride_channel_y   = 0;
    const int64_t stride_channel_dst = 0;
    const int64_t nsamples_x         = 1;
-    const int64_t nsamples_y         = 1;
+    const int64_t nsamples_dst       = 1;
    const int64_t stride_sample_x    = 0;
    const int64_t stride_sample_y    = 0;
    const int64_t stride_sample_dst  = 0;

    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0_dd_i;
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
        case GGML_TYPE_F16: {
            const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
        } break;
        case GGML_TYPE_BF16: {
            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
        } break;
        default:
            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
--- a/ggml/src/ggml-cuda/mmv.cuh
+++ b/ggml/src/ggml-cuda/mmv.cuh
@@ -3,7 +3,7 @@
 // maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
 #define MMV_MAX_ROWS 512

-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);

 void ggml_cuda_op_mul_mat_vec(
    ggml_backend_cuda_context & ctx,
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -1,50 +1,57 @@
 #include "mmvq.cuh"
+#include "quantize.cuh"
 #include "vecdotq.cuh"

+#include <cstdint>
+
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);

 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? vec_dot_q4_0_q8_1 :
-        type == GGML_TYPE_Q4_1 ? vec_dot_q4_1_q8_1 :
-        type == GGML_TYPE_Q5_0 ? vec_dot_q5_0_q8_1 :
-        type == GGML_TYPE_Q5_1 ? vec_dot_q5_1_q8_1 :
-        type == GGML_TYPE_Q8_0 ? vec_dot_q8_0_q8_1 :
-        type == GGML_TYPE_Q2_K ? vec_dot_q2_K_q8_1 :
-        type == GGML_TYPE_Q3_K ? vec_dot_q3_K_q8_1 :
-        type == GGML_TYPE_Q4_K ? vec_dot_q4_K_q8_1 :
-        type == GGML_TYPE_Q5_K ? vec_dot_q5_K_q8_1 :
-        type == GGML_TYPE_Q6_K ? vec_dot_q6_K_q8_1 :
-        type == GGML_TYPE_IQ2_XXS ? vec_dot_iq2_xxs_q8_1 :
-        type == GGML_TYPE_IQ2_XS ? vec_dot_iq2_xs_q8_1 :
-        type == GGML_TYPE_IQ2_S ? vec_dot_iq2_s_q8_1 :
-        type == GGML_TYPE_IQ3_XXS ? vec_dot_iq3_xxs_q8_1 :
-        type == GGML_TYPE_IQ1_S ? vec_dot_iq1_s_q8_1 :
-        type == GGML_TYPE_IQ1_M ? vec_dot_iq1_m_q8_1 :
-        type == GGML_TYPE_IQ4_NL ? vec_dot_iq4_nl_q8_1 :
-        type == GGML_TYPE_IQ4_XS ? vec_dot_iq4_xs_q8_1 :
-        type == GGML_TYPE_IQ3_S ? vec_dot_iq3_s_q8_1 :
-        nullptr;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
+        case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
+        case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
+        case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
+        case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
+        case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
+        case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
+        case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
+        case GGML_TYPE_Q5_K:    return vec_dot_q5_K_q8_1;
+        case GGML_TYPE_Q6_K:    return vec_dot_q6_K_q8_1;
+        case GGML_TYPE_IQ2_XXS: return vec_dot_iq2_xxs_q8_1;
+        case GGML_TYPE_IQ2_XS:  return vec_dot_iq2_xs_q8_1;
+        case GGML_TYPE_IQ2_S:   return vec_dot_iq2_s_q8_1;
+        case GGML_TYPE_IQ3_XXS: return vec_dot_iq3_xxs_q8_1;
+        case GGML_TYPE_IQ1_S:   return vec_dot_iq1_s_q8_1;
+        case GGML_TYPE_IQ1_M:   return vec_dot_iq1_m_q8_1;
+        case GGML_TYPE_IQ4_NL:  return vec_dot_iq4_nl_q8_1;
+        case GGML_TYPE_IQ4_XS:  return vec_dot_iq4_xs_q8_1;
+        case GGML_TYPE_IQ3_S:   return vec_dot_iq3_s_q8_1;
+        default:                return nullptr;
+    }
 }

 static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_1    ? VDR_Q4_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_0    ? VDR_Q5_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_1    ? VDR_Q5_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q8_0    ? VDR_Q8_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q2_K    ? VDR_Q2_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q3_K    ? VDR_Q3_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_K    ? VDR_Q4_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_K    ? VDR_Q5_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q6_K    ? VDR_Q6_K_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XS  ? VDR_IQ2_XS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_S   ? VDR_IQ2_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_S   ? VDR_IQ3_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_NL  ? VDR_IQ4_NL_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_XS  ? VDR_IQ4_XS_Q8_1_MMVQ :
-        1;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_K:    return VDR_Q5_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q6_K:    return VDR_Q6_K_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XXS: return VDR_IQ2_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XS:  return VDR_IQ2_XS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_S:   return VDR_IQ2_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_XXS: return VDR_IQ3_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_S:   return VDR_IQ3_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_NL:  return VDR_IQ4_NL_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_XS:  return VDR_IQ4_XS_Q8_1_MMVQ;
+        default:                return 1;
+    }
 }

 enum mmvq_parameter_table_id {
@@ -73,9 +80,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
    return MMVQ_PARAMETERS_GENERIC;
 }

-static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
    if (table_id == MMVQ_PARAMETERS_GENERIC) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
            case 1:
            case 2:
            case 3:
@@ -90,7 +97,7 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
                return 1;
        }
    } else if (table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
            case 1:
            case 2:
            case 3:
@@ -107,9 +114,9 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
    return 1;
 }

-static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) {
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
            case 1:
                return 1;
            case 2:
@@ -127,19 +134,21 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int ta
    return 1;
 }

-template <ggml_type type, int ncols_y>
+template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
-__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols_x, const int nchannels_y, const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {

    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
    constexpr int vdr = get_vdr_mmvq(type);
    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(ncols_y, table_id);
-    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id);
+    constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
@@ -147,13 +156,21 @@ static __global__ void mul_mat_vec_q(
    const     int tid = warp_size*threadIdx.y + threadIdx.x;
    const     int row0 = rows_per_cuda_block*blockIdx.x;
    const     int blocks_per_row_x = ncols_x / qk;
-    const     int blocks_per_col_y = nrows_y / QK8_1;
    constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;

-    // partial sum for each thread
-    float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};
+    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;

-    const block_q8_1 * y = (const block_q8_1 *) vy;
+    // partial sum for each thread
+    float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;

    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
@@ -162,18 +179,19 @@ static __global__ void mul_mat_vec_q(
        const int kqs = vdr * (tid % (qi/vdr));

 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
            for (int i = 0; i < rows_per_cuda_block; ++i) {
-                tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
+                tmp[j][i] += vec_dot_q_cuda(
+                    vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
            }
        }
    }

-    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
    if (threadIdx.y > 0) {
 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
            for (int i = 0; i < rows_per_cuda_block; ++i) {
                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
@@ -185,9 +203,11 @@ static __global__ void mul_mat_vec_q(
        return;
    }

+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
+
    // sum up partial sums and write back result
 #pragma unroll
-    for (int j = 0; j < ncols_y; ++j) {
+    for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
        for (int i = 0; i < rows_per_cuda_block; ++i) {
 #pragma unroll
@@ -197,88 +217,121 @@ static __global__ void mul_mat_vec_q(
            tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
        }

-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
-            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + int(threadIdx.x) < stride_col_dst)) {
+            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
        }
    }
-
-    GGML_UNUSED(nrows_x);
 }

-static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
-    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id);
-    const dim3 block_nums(nblocks, 1, 1);
-    const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1);
+static std::pair<dim3, dim3> calc_launch_params(
+        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int warp_size, const mmvq_parameter_table_id table_id) {
+    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
+    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
    return {block_nums, block_dims};
 }

 template <ggml_type type>
-static void mul_mat_vec_q_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+static void mul_mat_vec_q_switch_ncols_dst(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {

    GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
-    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+    GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
+
+    const int channel_ratio = nchannels_dst / nchannels_x;
+    const int sample_ratio  = nsamples_dst  / nsamples_x;

    const int device = ggml_cuda_get_device();
    const int warp_size = ggml_cuda_info().devices[device].warp_size;
    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);

-    switch (ncols_y) {
+    GGML_ASSERT(!ids || ncols_dst == 1);
+    switch (ncols_dst) {
        case 1:
        {
-            constexpr int c_ncols_y = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 1;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 2:
        {
-            constexpr int c_ncols_y = 2;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 2;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 3:
        {
-            constexpr int c_ncols_y = 3;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 3;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 4:
        {
-            constexpr int c_ncols_y = 4;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 4;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 5:
        {
-            constexpr int c_ncols_y = 5;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 5;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 6:
        {
-            constexpr int c_ncols_y = 6;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 6;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 7:
        {
-            constexpr int c_ncols_y = 7;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 7;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        case 8:
        {
-            constexpr int c_ncols_y = 8;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 8;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
            break;
        }
        default:
@@ -287,137 +340,223 @@ static void mul_mat_vec_q_cuda(
    }
 }

-static void mul_mat_vec_q4_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+static void mul_mat_vec_q_switch_type(
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {
+    switch (type_x) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
 }

-static void mul_mat_vec_q4_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+void ggml_cuda_mul_mat_vec_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.

-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_TENSOR_BINARY_OP_LOCALS;

-static void mul_mat_vec_q5_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    cudaStream_t stream = ctx.stream();

-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);

-static void mul_mat_vec_q5_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));

-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.

-static void mul_mat_vec_q8_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;

-    mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    // If src0 is a temporary compute buffer, clear any potential padding.
+    if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+        const size_t size_data  = ggml_nbytes(src0);
+        const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
+        if (size_alloc > size_data) {
+            CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
+        }
+    }

-static void mul_mat_vec_q2_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1);
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[3] / ts_src1;
+        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+    }

-    mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = ne10_padded / QK8_1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;

-static void mul_mat_vec_q3_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const int64_t s12 = ne11*s11;
+    const int64_t s13 = ne12*s12;

-    mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;

-static void mul_mat_vec_q4_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q5_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q6_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_m_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_nl_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+    mul_mat_vec_q_switch_type(
+        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
+        ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+        ne03,              ne3,           s03, s13,              s3,                 stream);
 }

 void ggml_cuda_op_mul_mat_vec_q(
@@ -440,68 +579,12 @@ void ggml_cuda_op_mul_mat_vec_q(
    // nrows_dst == nrows of the matrix that the kernel writes into
    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;

-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XXS:
-            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XS:
-            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_S:
-            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_XXS:
-            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_S:
-            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_M:
-            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_NL:
-            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_XS:
-            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_S:
-            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
+    const int stride_row_x = ne00 / ggml_blck_size(src0->type);
+    const int stride_col_y = src1_padded_row_size / QK8_1;
+
+    mul_mat_vec_q_switch_type(
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);

    GGML_UNUSED(src1);
    GGML_UNUSED(dst);
--- a/ggml/src/ggml-cuda/mmvq.cuh
+++ b/ggml/src/ggml-cuda/mmvq.cuh
@@ -2,6 +2,9 @@

 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.

+void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
 void ggml_cuda_op_mul_mat_vec_q(
    ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
--- a/ggml/src/ggml-cuda/quantize.cu
+++ b/ggml/src/ggml-cuda/quantize.cu
@@ -1,30 +1,40 @@
 #include "quantize.cuh"
 #include <cstdint>

-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx0_padded) {
-    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void quantize_q8_1(
+        const float * __restrict__ x, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
+    const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;

-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
        return;
    }

-    const int64_t ix1 = blockIdx.y;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;

-    const int64_t i_padded = ix1*kx0_padded + ix0;
+    const int64_t & i00 = i0;
+    const int64_t & i01 = i1;
+    const int64_t & i02 = i2;
+    const int64_t & i03 = i3;
+
+    const int64_t i_cont = ((i3*ne2 + i2) * ne1 + i1) * ne0 + i0;

    block_q8_1 * y = (block_q8_1 *) vy;

-    const int64_t ib = i_padded / QK8_1; // block index
-    const int64_t iqs = i_padded % QK8_1; // quant index
+    const int64_t ib  = i_cont / QK8_1; // block index
+    const int64_t iqs = i_cont % QK8_1; // quant index

-    const float xi = ix0 < kx ? x[ix1*kx + ix0] : 0.0f;
+    const float xi = i0 < ne00 ? x[i03*s03 + i02*s02 + i01*s01 + i00] : 0.0f;
    float amax = fabsf(xi);
    float sum = xi;

    amax = warp_reduce_max(amax);
-    sum = warp_reduce_sum(sum);
+    sum  = warp_reduce_sum(sum);

-    const float d = amax / 127;
+    const float  d = amax / 127;
    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);

    y[ib].qs[iqs] = q;
@@ -39,29 +49,38 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest

 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+        const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {

    constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
    constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;

-    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;

-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
        return;
    }

-    const float4 * x4 = (const float4 *) x;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;

-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+    const int64_t i00 = i0;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    const float4 * x4 = (const float4 *) x;

    block_q8_1_mmq * y = (block_q8_1_mmq *) vy;

    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block

    // Load 4 floats per thread and calculate max. abs. value between them:
-    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
    float amax = fabsf(xi.x);
    amax = fmaxf(amax, fabsf(xi.y));
    amax = fmaxf(amax, fabsf(xi.z));
@@ -77,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
    if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
        sum = xi.x + xi.y + xi.z + xi.w;

-        // Exchange calculate sum across vals_per_sum/4 threads.
+        // Calculate sums across vals_per_sum/4 threads.
 #pragma unroll
        for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
            sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -127,40 +146,41 @@ static __global__ void quantize_mmq_q8_1(
 }

 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(!ids);
+    GGML_ASSERT(ne0 % QK8_1 == 0);

-    GGML_ASSERT(kx0_padded % QK8_1 == 0);
-
-    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, kx1*channels, 1);
+    const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx0_padded);
-
-    GGML_UNUSED(type_x);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    GGML_UNUSED(type_src0);
 }

 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(ne00 % 4 == 0);
+    GGML_ASSERT(ne0 % (4*QK8_1) == 0);

-    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
-
-    const int64_t block_num_x = (kx0_padded + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
-    const dim3 num_blocks(block_num_x, kx1, channels);
+    const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
-    switch (mmq_get_q8_1_ds_layout(type_x)) {
+    switch (mmq_get_q8_1_ds_layout(type_src0)) {
        case MMQ_Q8_1_DS_LAYOUT_D4:
            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
            break;
        case MMQ_Q8_1_DS_LAYOUT_DS4:
            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
            break;
        case MMQ_Q8_1_DS_LAYOUT_D2S6:
            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
            break;
        default:
            GGML_ABORT("fatal error");
--- a/ggml/src/ggml-cuda/quantize.cuh
+++ b/ggml/src/ggml-cuda/quantize.cuh
@@ -12,13 +12,16 @@ static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");

 typedef void (*quantize_cuda_t)(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
--- a/ggml/src/ggml-cuda/vecdotq.cuh
+++ b/ggml/src/ggml-cuda/vecdotq.cuh
@@ -1,3 +1,5 @@
+#pragma once
+
 #include "common.cuh"
 #include <cstdint>

--- a/Show More
+++ b/Show More