vulkan: support CPY from any type to itself (#13695 )

Reuse the f16/f32 copy shaders, and just scale the number of elements according to the type size.
vulkan: Disable coopmat/coopmat2/bfloat extensions if glslc doesn't support it (#13696 )
2026-06-29 00:57:39 +02:00 · 2025-05-23 06:45:02 +02:00 · 2025-05-23 06:33:45 +02:00 · 2025-05-23 06:33:08 +02:00 · 2025-05-23 00:21:37 +02:00 · 2025-05-22 22:21:07 +03:00
96 changed files with 99344 additions and 2520 deletions
@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04

 ## Build Image

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc3.1.1
+ARG MUSA_VERSION=rc4.0.1
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}

-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}

 FROM ${BASE_MUSA_DEV_CONTAINER} AS build

@@ -21,21 +21,14 @@ RUN apt-get update && \
    libcurl4-openssl-dev \
    libgomp1

-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
-
-RUN pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt
-
 WORKDIR /app

 COPY . .

-# Use the default MUSA archs if not specified
 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_BUILD_TESTS=OFF -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 -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 && \
@@ -48,3 +48,7 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[tools/mtmd/miniaudio.h]
+trim_trailing_whitespace = unset
+insert_final_newline = unset
@@ -351,7 +351,7 @@ jobs:

  ubuntu-22-cmake-musa:
    runs-on: ubuntu-22.04
-    container: mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04

    steps:
      - name: Clone
@@ -899,7 +899,7 @@ jobs:
        shell: bash

    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
    steps:
@@ -1,4 +1,4 @@
-name: Create Release
+name: Release

 on:
  workflow_dispatch: # allows manual triggering
@@ -227,6 +227,66 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
          name: llama-bin-ubuntu-vulkan-x64.zip

+  windows-cpu:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        include:
+          - arch: 'x64'
+          - arch: 'arm64'
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-cpu-${{ matrix.arch }}
+          variant: ccache
+          evict-old-files: 1d
+
+      - name: Install Ninja
+        run: |
+          choco install ninja
+
+      - name: libCURL
+        id: get_libcurl
+        uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
+
+      - name: Build
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake `
+            -DGGML_NATIVE=OFF `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_OPENMP=OFF `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
+          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
+          name: llama-bin-win-cpu-${{ matrix.arch }}.zip
+
  windows:
    runs-on: windows-latest

@@ -237,52 +297,30 @@ jobs:
    strategy:
      matrix:
        include:
-          - build: 'cpu-x64'
+          - backend: 'vulkan'
            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
-          #- build: 'openblas-x64'
-          #  arch: 'x64'
-          #  defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -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'
-            arch: 'x64'
-            defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
-          - build: 'cpu-arm64'
-            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
-          - build: 'opencl-adreno-arm64'
+            defines: '-DGGML_VULKAN=ON'
+            target: 'ggml-vulkan'
+          - backend: 'opencl-adreno'
            arch: 'arm64'
            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'
+            target: 'ggml-opencl'

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0

      - name: ccache
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-${{ matrix.build }}
+          key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
          variant: ccache
          evict-old-files: 1d

-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
-          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/openblas
-          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
-          $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()))
-          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
-          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
-
      - name: Install Vulkan SDK
        id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.backend == 'vulkan' }}
        run: |
          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
@@ -296,7 +334,7 @@ jobs:

      - name: Install OpenCL Headers and Libs
        id: install_opencl
-        if: ${{ matrix.build == 'opencl-adreno-arm64' }}
+        if: ${{ matrix.backend == 'opencl-adreno' && matrix.arch == 'arm64' }}
        run: |
          git clone https://github.com/KhronosGroup/OpenCL-Headers
          cd OpenCL-Headers
@@ -314,46 +352,22 @@ jobs:
            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
          cmake --build build-arm64-release --target install --config release

-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
-
      - name: Build
        id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
-          cmake -S . -B build ${{ matrix.defines }} `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Add libopenblas.dll
-        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          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: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake -S . -B build ${{ matrix.defines }} -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_BACKEND_DL=ON -DLLAMA_CURL=OFF
+          cmake --build build --config Release --target ${{ matrix.target }}

      - name: Pack artifacts
        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
+          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll

      - name: Upload artifacts
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
+          path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip

  windows-cuda:
    runs-on: windows-2019
@@ -366,8 +380,6 @@ jobs:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0

      - name: Install ccache
        uses: hendrikmuhs/ccache-action@v1.2.16
@@ -386,45 +398,30 @@ jobs:
        run: |
          choco install ninja

-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
      - name: Build
        id: cmake_build
        shell: cmd
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
          cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
-            ${{ env.CMAKE_ARGS }}
+            -DLLAMA_CURL=OFF
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

      - name: Pack artifacts
        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll

      - name: Upload artifacts
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip

      - name: Copy and pack Cuda runtime
        run: |
@@ -432,13 +429,13 @@ jobs:
          $dst='.\build\bin\cudart\'
          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*

      - name: Upload Cuda runtime
        uses: actions/upload-artifact@v4
        with:
-          path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip

  windows-sycl:
    runs-on: windows-latest
@@ -448,15 +445,14 @@ jobs:
        shell: bash

    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+
    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0

      - name: ccache
        uses: hendrikmuhs/ccache-action@v1.2.16
@@ -469,15 +465,18 @@ jobs:
        run:  |
          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL

-      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
      - name: Build
        id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+        shell: cmd
+        run: |
+          call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+          cmake -G "Ninja" -B build ^
+            -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx ^
+            -DCMAKE_BUILD_TYPE=Release ^
+            -DGGML_BACKEND_DL=ON -DBUILD_SHARED_LIBS=ON ^
+            -DGGML_CPU=OFF -DGGML_SYCL=ON ^
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-sycl -j

      - name: Build the release package
        id: pack_artifacts
@@ -502,12 +501,12 @@ jobs:
          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin

          echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+          7z a llama-bin-win-sycl-x64.zip ./build/bin/*

      - name: Upload the release package
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          path: llama-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

  windows-hip:
@@ -515,14 +514,14 @@ jobs:

    strategy:
      matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0

      - name: Clone rocWMMA repository
        id: clone_rocwmma
@@ -532,7 +531,7 @@ jobs:
      - name: ccache
        uses: hendrikmuhs/ccache-action@v1.2.16
        with:
-          key: windows-latest-cmake-hip-release
+          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
          evict-old-files: 1d

      - name: Install
@@ -550,50 +549,39 @@ jobs:
        run: |
          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version

-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
      - name: Build
        id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
          cmake -G "Unix Makefiles" -B build -S . `
            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
            -DCMAKE_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
            -DGGML_HIP_ROCWMMA_FATTN=ON `
            -DGGML_HIP=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
          md "build\bin\rocblas\library\"
          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"

-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
      - name: Pack artifacts
        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*

      - name: Upload artifacts
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip

  ios-xcode-build:
    runs-on: macos-latest
@@ -655,14 +643,16 @@ jobs:
    runs-on: ubuntu-latest

    needs:
-      - ubuntu-22-cpu
-      - ubuntu-22-vulkan
      - windows
+      - windows-cpu
      - windows-cuda
      - windows-sycl
      - windows-hip
+      - ubuntu-22-cpu
+      - ubuntu-22-vulkan
      - macOS-arm64
      - macOS-x64
+      - ios-xcode-build

    steps:
      - name: Clone
@@ -680,10 +670,43 @@ jobs:
        uses: actions/download-artifact@v4
        with:
          path: ./artifact
+          merge-multiple: true

      - name: Move artifacts
        id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+        run: |
+          mkdir -p release
+
+          echo "Adding CPU backend files to existing zips..."
+          for arch in x64 arm64; do
+            cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
+            temp_dir=$(mktemp -d)
+            echo "Extracting CPU backend for $arch..."
+            unzip "$cpu_zip" -d "$temp_dir"
+
+            echo "Adding CPU files to $arch zips..."
+            for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
+              if [[ "$target_zip" == "$cpu_zip" ]]; then
+                continue
+              fi
+              echo "Adding CPU backend to $(basename "$target_zip")"
+              realpath_target_zip=$(realpath "$target_zip")
+              (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
+            done
+
+            rm -rf "$temp_dir"
+          done
+
+          echo "Renaming and moving zips to release..."
+          for zip_file in artifact/llama-bin-win-*.zip; do
+            base_name=$(basename "$zip_file" .zip)
+            zip_name="llama-${{ steps.tag.outputs.name }}-${base_name#llama-}.zip"
+            echo "Moving $zip_file to release/$zip_name"
+            mv "$zip_file" "release/$zip_name"
+          done
+
+          echo "Moving other artifacts..."
+          mv -v artifact/*.zip release

      - name: Create release
        id: create_release
@@ -702,7 +725,7 @@ jobs:
            const path = require('path');
            const fs = require('fs');
            const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
+            for (let file of await fs.readdirSync('./release')) {
              if (path.extname(file) === '.zip') {
                console.log('uploadReleaseAsset', file);
                await github.repos.uploadReleaseAsset({
@@ -710,7 +733,7 @@ jobs:
                  repo: context.repo.repo,
                  release_id: release_id,
                  name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                  data: await fs.readFileSync(`./release/${file}`)
                });
              }
            }
@@ -37,7 +37,7 @@ range of hardware - locally and in the cloud.
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity

@@ -237,7 +237,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
 | [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](docs/build.md#musa) | Moore Threads MTT GPU |
+| [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
@@ -580,3 +580,4 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
 - [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
 - [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
+- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
@@ -54,7 +54,7 @@ docker run --privileged -it \
    -v $HOME/llama.cpp/ci-cache:/ci-cache \
    -v $HOME/llama.cpp/ci-results:/ci-results \
    -v $PWD:/ws -w /ws \
-    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 ```

 Inside the container, execute the following commands:
@@ -39,7 +39,7 @@
 using json = nlohmann::ordered_json;

 std::initializer_list<enum llama_example> mmproj_examples = {
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
    LLAMA_EXAMPLE_SERVER,
 };

@@ -1445,6 +1445,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.n_keep = value;
        }
    ));
+    add_opt(common_arg(
+        {"--swa-full"},
+        string_format("use full-size SWA cache (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)", params.swa_full ? "true" : "false"),
+        [](common_params & params) {
+            params.swa_full = true;
+        }
+    ).set_env("LLAMA_ARG_SWA_FULL"));
    add_opt(common_arg(
        {"--no-context-shift"},
        string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1670,7 +1678,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.warmup = false;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
    add_opt(common_arg(
        {"--spm-infill"},
        string_format(
@@ -2057,13 +2065,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.grp_attn_w = value;
        }
    ).set_env("LLAMA_ARG_GRP_ATTN_W").set_examples({LLAMA_EXAMPLE_MAIN}));
-    add_opt(common_arg(
-        {"-dkvc", "--dump-kv-cache"},
-        "verbose print of the KV cache",
-        [](common_params & params) {
-            params.dump_kv_cache = true;
-        }
-    ));
    add_opt(common_arg(
        {"-nkvo", "--no-kv-offload"},
        "disable KV offload",
@@ -2232,12 +2233,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
    add_opt(common_arg(
-        {"--image"}, "FILE",
-        "path to an image file. use with multimodal models. Specify multiple times for batching",
+        {"--image", "--audio"}, "FILE",
+        "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",
        [](common_params & params, const std::string & value) {
            params.image.emplace_back(value);
        }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples({LLAMA_EXAMPLE_MTMD}));
    if (llama_supports_rpc()) {
        add_opt(common_arg(
            {"--rpc"}, "SERVERS",
@@ -2585,7 +2586,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, int value) {
            params.n_junk = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_PASSKEY}));
+    ).set_examples({LLAMA_EXAMPLE_PASSKEY, LLAMA_EXAMPLE_PARALLEL}));
    add_opt(common_arg(
        {"--pos"}, "N",
        string_format("position of the passkey in the junk text (default: %d)", params.i_pos),
@@ -2648,7 +2649,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.is_pp_shared = true;
        }
-    ).set_examples({LLAMA_EXAMPLE_BENCH}));
+    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
    add_opt(common_arg(
        {"-npp"}, "n0,n1,...",
        "number of prompt tokens",
@@ -2867,7 +2868,7 @@ 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, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
    add_opt(common_arg(
        {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
        string_format(
@@ -2880,6 +2881,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            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(
+        {"--no-prefill-assistant"},
+        string_format(
+            "whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)\n"
+            "when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled\n"
+        ),
+        [](common_params & params) {
+            params.prefill_assistant = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_PREFILL_ASSISTANT"));
    add_opt(common_arg(
        {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
        string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),
@@ -1102,6 +1102,9 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
    }

+    mparams.progress_callback           = params.load_progress_callback;
+    mparams.progress_callback_user_data = params.load_progress_callback_user_data;
+
    return mparams;
 }

@@ -1133,6 +1136,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.flash_attn        = params.flash_attn;
    cparams.no_perf           = params.no_perf;
    cparams.op_offload        = !params.no_op_offload;
+    cparams.swa_full          = params.swa_full;

    if (params.reranking) {
        cparams.embeddings    = true;
@@ -1325,81 +1329,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
    return text;
 }

-//
-// KV cache utils
-//
-
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        int seq_count = 0;
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) { seq_count++; }
-        }
-        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    std::unordered_map<llama_seq_id, size_t> seqs;
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] < 0) { continue; }
-            if (seqs.find(cs_curr[j]) == seqs.end()) {
-                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                const size_t sz = seqs.size();
-                seqs[cs_curr[j]] = sz;
-            }
-        }
-        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-    }
-
-    printf("=== Sequence legend: ");
-    for (const auto & it : seqs) {
-        printf("%zu=%d, ", it.second, it.first);
-    }
-    printf("'+'=other sequence ids");
-
-    c_curr = view.cells;
-    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) {
-                const auto & it = seqs.find(cs_curr[j]);
-                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
-            } else {
-                putchar('.');
-            }
-        }
-        putchar(' ');
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
 //
 // Embedding utils
 //
@@ -76,7 +76,7 @@ enum llama_example {
    LLAMA_EXAMPLE_SERVER,
    LLAMA_EXAMPLE_CVECTOR_GENERATOR,
    LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
    LLAMA_EXAMPLE_LOOKUP,
    LLAMA_EXAMPLE_PARALLEL,
    LLAMA_EXAMPLE_TTS,
@@ -323,13 +323,13 @@ struct common_params {
    bool flash_attn        = false; // flash attention
    bool no_perf           = false; // disable performance metrics
    bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)

    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
    bool use_mmap          = true;  // use mmap for faster loads
    bool use_mlock         = false; // use mlock to keep model in memory
    bool verbose_prompt    = false; // print prompt tokens before generation
    bool display_prompt    = true;  // print prompt before generation
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
    bool no_kv_offload     = false; // disable KV offloading
    bool warmup            = true;  // warmup run
    bool check_tensors     = false; // validate tensor data
@@ -368,6 +368,7 @@ struct common_params {
    bool use_jinja = false;                                                                                 // NOLINT
    bool enable_chat_template = true;
    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response

    std::vector<std::string> api_keys;

@@ -427,6 +428,11 @@ struct common_params {

    // common params
    std::string out_file; // output filename for all example programs
+    // optional callback for model loading progress and cancellation:
+    // called with a progress value between 0.0 and 1.0.
+    // return false from callback to abort model loading or true to continue
+    llama_progress_callback load_progress_callback = NULL;
+    void *                  load_progress_callback_user_data = NULL;
 };

 // call once at the start of a program if it uses libcommon
@@ -615,16 +621,6 @@ std::string common_detokenize(
        const std::vector<llama_token> & tokens,
                                  bool   special = true);

-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// Dump the KV cache view showing individual sequences in each cell (long output).
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
 //
 // Embedding utils
 //
@@ -45,7 +45,7 @@ class SentencePieceTokenTypes(IntEnum):

 class ModelType(IntEnum):
    TEXT = 1
-    VISION = 2
+    MMPROJ = 2


 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
@@ -54,7 +54,7 @@ AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
 class ModelBase:
    _model_classes: dict[ModelType, dict[str, type[ModelBase]]] = {
        ModelType.TEXT: {},
-        ModelType.VISION: {},
+        ModelType.MMPROJ: {},
    }

    dir_model: Path
@@ -88,7 +88,7 @@ class ModelBase:
                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
        if type(self) is ModelBase or \
                type(self) is TextModel or \
-                type(self) is VisionModel:
+                type(self) is MmprojModel:
            raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")

        self.dir_model = dir_model
@@ -308,6 +308,8 @@ class ModelBase:
                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
                            gguf.MODEL_TENSOR.POSNET_NORM1,
                            gguf.MODEL_TENSOR.POSNET_NORM2,
+                            gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
+                            gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
                        )
                    )
                    or not new_name.endswith(".weight")
@@ -437,7 +439,7 @@ class ModelBase:
        assert names

        def func(modelcls: AnyModel) -> AnyModel:
-            model_type = ModelType.VISION if modelcls.model_arch == gguf.MODEL_ARCH.CLIP_VISION else ModelType.TEXT
+            model_type = ModelType.MMPROJ if modelcls.model_arch == gguf.MODEL_ARCH.MMPROJ else ModelType.TEXT
            for name in names:
                cls._model_classes[model_type][name] = modelcls
            return modelcls
@@ -1113,60 +1115,87 @@ class TextModel(ModelBase):
            self.gguf_writer.add_pooling_type(pooling_type)


-class VisionModel(ModelBase):
-    model_type = ModelType.VISION
-    model_arch = gguf.MODEL_ARCH.CLIP_VISION
+class MmprojModel(ModelBase):
+    model_type = ModelType.MMPROJ
+    model_arch = gguf.MODEL_ARCH.MMPROJ
    preprocessor_config: dict[str, Any]
    global_config: dict[str, Any]

+    has_vision_encoder: bool = True # by default
+    has_audio_encoder: bool = False
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

-        if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
-            raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
+        if self.model_arch != gguf.MODEL_ARCH.MMPROJ:
+            raise TypeError("MmprojModel must be subclassed with model_arch = gguf.MODEL_ARCH.MMPROJ")
+
+        if self.has_vision_encoder and self.has_audio_encoder:
+            raise NotImplementedError("both vision + audio not supported yet")

        # get n_embd of the text model
        if "text_config" not in self.hparams:
            self.hparams["text_config"] = {}
+        if "audio_config" not in self.hparams:
+            self.hparams["audio_config"] = {}
        text_config = {**self.hparams, **self.hparams["text_config"]}
        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
        assert self.n_embd_text > 0, "n_embd not found in hparams"

-        if "vision_config" not in self.hparams:
-            raise ValueError("vision_config not found in hparams")
        # move vision config to the top level, while preserving the original hparams in global_config
        self.global_config = self.hparams
-        self.hparams = self.hparams["vision_config"]
+
+        if "vision_config" in self.hparams:
+            self.hparams = self.hparams["vision_config"]
+        elif "audio_config" in self.hparams:
+            self.hparams = self.hparams["audio_config"]
+        else:
+            raise ValueError("vision_config / audio_config not found in hparams")

        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)

        # load preprocessor config
        with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
            self.preprocessor_config = json.load(f)

    def set_type(self):
-        self.gguf_writer.add_type(gguf.GGUFType.CLIP_VISION)
+        self.gguf_writer.add_type(gguf.GGUFType.MMPROJ)

    def set_gguf_parameters(self):
        self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
-        self.gguf_writer.add_vision_has_vision_encoder(True)

-        # vision config
-        self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
-        self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
-        self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
-        self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_vision_block_count(self.block_count)
-        self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+        if self.has_vision_encoder:
+            self.gguf_writer.add_clip_has_vision_encoder(True)
+            self.gguf_writer.add_vision_projection_dim(self.n_embd_text)

-        # preprocessor config
-        self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
-        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
+            # vision config
+            self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
+            self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
+            self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
+            self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
+            self.gguf_writer.add_vision_block_count(self.block_count)
+            self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+
+            # preprocessor config
+            self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
+            self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
+
+        elif self.has_audio_encoder:
+            self.gguf_writer.add_clip_has_audio_encoder(True)
+            self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+
+            # audio config
+            self.gguf_writer.add_audio_embedding_length(self.find_hparam(["hidden_size"]))
+            self.gguf_writer.add_audio_feed_forward_length(self.find_hparam(["intermediate_size"]))
+            self.gguf_writer.add_audio_block_count(self.block_count)
+            self.gguf_writer.add_audio_head_count(self.find_hparam(["num_attention_heads"]))
+
+        else:
+            raise ValueError("MmprojModel must have either vision or audio encoder")

    def write_vocab(self):
-        raise ValueError("VisionModel does not support vocab writing")
+        raise ValueError("MmprojModel does not support vocab writing")


@ModelBase.register("GPTNeoXForCausalLM")
@@ -1950,7 +1979,7 @@ class LlamaModel(TextModel):
    "LlavaForConditionalGeneration", # pixtral
    "Mistral3ForConditionalGeneration", # mistral small 3.1
 )
-class LlavaVisionModel(VisionModel):
+class LlavaVisionModel(MmprojModel):
    img_break_tok_id = -1

    def __init__(self, *args, **kwargs):
@@ -1976,7 +2005,7 @@ class LlavaVisionModel(VisionModel):
        super().set_gguf_parameters()
        hparams = self.hparams
        if hparams["model_type"] == "pixtral":
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PIXTRAL)
            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])

            # hidden_act
@@ -2015,7 +2044,7 @@ class LlavaVisionModel(VisionModel):


@ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
-class SmolVLMModel(VisionModel):
+class SmolVLMModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        if self.hparams["model_type"] == "smolvlm_vision":
@@ -2027,7 +2056,7 @@ class SmolVLMModel(VisionModel):

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.IDEFICS3)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.IDEFICS3)
        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
        self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
        self.gguf_writer.add_vision_use_gelu(True)
@@ -2092,6 +2121,26 @@ class Llama4Model(LlamaModel):
        return super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("Llama4ForConditionalGeneration")
+class Llama4VisionModel(MmprojModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LLAMA4)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams["norm_eps"])
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / self.hparams["pixel_shuffle_ratio"]))
+        assert self.hparams["hidden_act"] == "gelu"
+        self.gguf_writer.add_vision_use_gelu(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid # unused
+        if "multi_modal_projector" in name or "vision_model" in name:
+            # process vision tensors
+            if "positional_embedding_vlm" in name and ".weight" not in name:
+                name += ".weight"
+            return [(self.map_tensor_name(name), data_torch)]
+        return []
+
+
@ModelBase.register("Mistral3ForConditionalGeneration")
 class Mistral3Model(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA
@@ -2624,7 +2673,7 @@ class Qwen2Model(TextModel):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+@ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(TextModel):
    model_arch = gguf.MODEL_ARCH.QWEN2VL

@@ -2648,8 +2697,8 @@ class Qwen2VLModel(TextModel):
        return [(self.map_tensor_name(name), data_torch)]


-@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
-class Qwen2VLVisionModel(VisionModel):
+@ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+class Qwen2VLVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.hparams["image_size"] = self.hparams.get("image_size", 560)
@@ -2664,9 +2713,9 @@ class Qwen2VLVisionModel(VisionModel):
        super().set_gguf_parameters()
        hparams = self.hparams
        if self.global_config['model_type'] == 'qwen2_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN2VL)
        elif self.global_config['model_type'] == 'qwen2_5_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN25VL)
            self.gguf_writer.add_vision_use_silu(True)
            # find n_wa_pattern (window attention pattern)
            fullatt_block_indexes = hparams.get("fullatt_block_indexes")
@@ -2725,11 +2774,11 @@ class Qwen2VLVisionModel(VisionModel):


@ModelBase.register("InternVisionModel")
-class InternVisionModel(VisionModel):
+class InternVisionModel(MmprojModel):
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hparams = self.hparams
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.INTERNVL)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.INTERNVL)
        self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
        # hidden_act
        if hparams["hidden_act"] == "silu":
@@ -3987,11 +4036,11 @@ class Gemma3Model(TextModel):


@ModelBase.register("Gemma3ForConditionalGeneration")
-class Gemma3VisionModel(VisionModel):
+class Gemma3VisionModel(MmprojModel):
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hparams = self.hparams
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.GEMMA3)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GEMMA3)
        # default values below are taken from HF tranformers code
        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
        self.gguf_writer.add_vision_use_gelu(True)
@@ -5938,6 +5987,52 @@ class ChameleonModel(TextModel):
        return data_torch


+@ModelBase.register("UltravoxModel")
+class UltravoxModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA # dummy
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        raise NotImplementedError("Ultravox does not have text decoder. Please use --mmproj argument")
+
+
+@ModelBase.register("UltravoxModel")
+class UltravoxAudioModel(MmprojModel):
+    has_vision_encoder = False # no vision encoder
+    has_audio_encoder = True
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["hidden_size"] = self.hparams["d_model"]
+        self.hparams["intermediate_size"] = self.hparams["encoder_ffn_dim"]
+        self.hparams["num_attention_heads"] = self.hparams["encoder_attention_heads"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.ULTRAVOX)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams["num_mel_bins"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
+        self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"])
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, new_name, n_dims  # unused
+        if ".conv" in name and ".weight" in name:
+            return gguf.GGMLQuantizationType.F16
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # prevent clash naming with vision tensors
+        if name.startswith("multi_modal_projector"):
+            name = "audio." + name
+
+        if "conv1.bias" in name or "conv2.bias" in name:
+            # transpose conv1 and conv2 bias
+            data_torch = data_torch.unsqueeze(-1)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
 ###### CONVERSION LOGIC ######


@@ -6113,13 +6208,15 @@ def split_str_to_n_bytes(split_str: str) -> int:


 def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str:
+    # TODO @ngxson : this won't work correctly if the model has both audio & vision encoders
+    # maybe we should fallback to text model's arch in that case, since not many models have both
    text_config = hparams.get("text_config", {})
    vision_config = hparams.get("vision_config", {})
    arch = hparams["architectures"][0]
    # if "architectures" is found in the sub-config, use that instead
    if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
        arch = text_config["architectures"][0]
-    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+    elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
        arch = vision_config["architectures"][0]
    return arch

@@ -6182,7 +6279,7 @@ def main() -> None:

    with torch.inference_mode():
        output_type = ftype_map[args.outtype]
-        model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        model_type = ModelType.MMPROJ if args.mmproj else ModelType.TEXT
        hparams = ModelBase.load_hparams(dir_model)
        model_architecture = get_model_architecture(hparams, model_type)
        logger.info(f"Model architecture: {model_architecture}")
@@ -56,60 +56,82 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi

 ## Model Supports

-| Model Name                  | FP16  | Q8_0 | Q4_0 |
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
 |:----------------------------|:-----:|:----:|:----:|
-| AquilaChat2-7B              |   √   |   √  |   √  |
-| Baichuan-7b                 |   √   |   √  |   √  |
-| Baichuan2-7B-Chat           |   √   |   √  |   √  |
-| bitnet_b1_58-large          |   √   |   √  |   √  |
-| bloom-560m                  |   √   |   x  |   √  |
-| bloomz-alpaca-560m          |   √   |   x  |   √  |
-| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
-| chatglm3-6B                 |   x   |   x  |   x  |
-| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
-| CodeShell-7B                |   √   |   √  |   √  |
-| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
-| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
-| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
-| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
-| falcon-7b-instruct          |   √   |   √  |   √  |
-| flan-t5-large               |   √   |   √  |   √  |
-| gemma-2-9b-it               |   √   |   √  |   √  |
-| glm-4-9B                    |   x   |   x  |   x  |
-| gpt2                        |   √   |   √  |   √  |
-| Gpt2-163M                   |   √   |   √  |   √  |
-| granite-3B-code-instruct    |   √   |   √  |   √  |
+| Llama-2                     |   √   |   √  |   √  |
+| Llama-3                     |   √   |   √  |   √  |
+| Mistral-7B                  |   √   |   √  |   √  |
+| Mistral MOE                 |   √   |   √  |   √  |
+| DBRX                        |   -   |   -  |   -  |
+| Falcon                      |   √   |   √  |   √  |
+| Chinese LLaMA/Alpaca        |   √   |   √  |   √  |
+| Vigogne(French)             |   √   |   √  |   √  |
+| BERT                        |   x   |   x  |   x  |
+| Koala                       |   √   |   √  |   √  |
+| Baichuan                    |   √   |   √  |   √  |
+| Aquila 1 & 2                |   √   |   √  |   √  |
+| Starcoder models            |   √   |   √  |   √  |
+| Refact                      |   √   |   √  |   √  |
+| MPT                         |   √   |   √  |   √  |
+| Bloom                       |   √   |   √  |   √  |
+| Yi models                   |   √   |   √  |   √  |
+| stablelm models             |   √   |   √  |   √  |
+| DeepSeek models             |   x   |   x  |   x  |
+| Qwen models                 |   √   |   √  |   √  |
+| PLaMo-13B                   |   √   |   √  |   √  |
+| Phi models                  |   √   |   √  |   √  |
+| PhiMoE                      |   √   |   √  |   √  |
+| GPT-2                       |   √   |   √  |   √  |
+| Orion                       |   √   |   √  |   √  |
+| InternlLM2                  |   √   |   √  |   √  |
+| CodeShell                   |   √   |   √  |   √  |
+| Gemma                       |   √   |   √  |   √  |
+| Mamba                       |   √   |   √  |   √  |
+| Xverse                      |   √   |   √  |   √  |
+| command-r models            |   √   |   √  |   √  |
+| Grok-1                      |   -   |   -  |   -  |
+| SEA-LION                    |   √   |   √  |   √  |
 | GritLM-7B                   |   √   |   √  |   √  |
-| internlm2_5-7b-chat         |   √   |   √  |   √  |
-| koala-7B-HF                 |   √   |   √  |   √  |
-| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
-| Llama-3-Smaug-8B            |   √   |   √  |   √  |
-| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
-| Llama3-8B                   |   √   |   √  |   √  |
-| Llama3-8b-chinese           |   √   |   √  |   √  |
-| mamba-130m-hf               |   √   |   √  |   √  |
-| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
-| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
-| mpt-7B                      |   √   |   √  |   √  |
-| OLMo-1B-hf                  |   √   |   √  |   √  |
-| OpenELM-3B-Instruct         |   √   |   √  |   √  |
-| Orion-14b-base              |   √   |   √  |   √  |
-| phi1                        |   x   |   x  |   x  |
-| phi2                        |   x   |   x  |   x  |
-| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
-| plamo-13b                   |   √   |   √  |   √  |
-| pythia-70M                  |   x   |   x  |   x  |
-| Qwen-7B                     |   √   |   √  |   √  |
-| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
-| Refact-1_6B-fim             |   √   |   √  |   √  |
-| SmolLM-135M                 |   √   |   √  |   √  |
-| stablelm-zephyr             |   x   |   x  |   x  |
-| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
-| starcoderbase-1b            |   √   |   √  |   √  |
-| starcoder2-3b               |   √   |   √  |   √  |
-| vigogne-7b-chat             |   √   |   √  |   √  |
-| xverse-7b-chat              |   √   |   √  |   √  |
-| Yi-6b-Chat                  |   √   |   √  |   √  |
+| OLMo                        |   √   |   √  |   √  |
+| OLMo 2                      |   √   |   √  |   √  |
+| OLMoE                       |   √   |   √  |   √  |
+| Granite models              |   √   |   √  |   √  |
+| GPT-NeoX                    |   √   |   √  |   √  |
+| Pythia                      |   √   |   √  |   √  |
+| Snowflake-Arctic MoE        |   -   |   -  |   -  |
+| Smaug                       |   √   |   √  |   √  |
+| Poro 34B                    |   √   |   √  |   √  |
+| Bitnet b1.58 models         |   √   |   x  |   x  |
+| Flan-T5                     |   √   |   √  |   √  |
+| Open Elm models             |   x   |   √  |   √  |
+| chatGLM3-6B + ChatGLM4-9b +  GLMEdge-1.5b + GLMEdge-4b    |   √   |   √  |   √  |
+| GLM-4-0414                  |   √   |   √  |   √  |
+| SmolLM                      |   √   |   √  |   √  |
+| EXAONE-3.0-7.8B-Instruct    |   √   |   √  |   √  |
+| FalconMamba Models          |   √   |   √  |   √  |
+| Jais Models                 |   -   |   x  |   x  |
+| Bielik-11B-v2.3             |   √   |   √  |   √  |
+| RWKV-6                      |   -   |   √  |   √  |
+| QRWKV-6                     |   √   |   √  |   √  |
+| GigaChat-20B-A3B            |   x   |   x  |   x  |
+| Trillion-7B-preview         |   √   |   √  |   √  |
+| Ling models                 |   √   |   √  |   √  |
+
+
+**Multimodal**
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
+|:----------------------------|:-----:|:----:|:----:|
+| LLaVA 1.5 models, LLaVA 1.6 models      |   x   |   x  |   x  |
+|  BakLLaVA                   |   √   |   √  |   √  |
+|  Obsidian                   |   √   |   -  |   -  |
+|  ShareGPT4V                 |   x   |   -  |   -  |
+|  MobileVLM 1.7B/3B models   |   -   |   -  |   -  |
+|  Yi-VL                      |   -   |   -  |   -  |
+|  Mini CPM                   |   √   |   √  |   √  |
+|  Moondream                  |   √   |   √  |   √  |
+|  Bunny                      |   √   |   -  |   -  |
+|  GLM-EDGE                   |   √   |   √  |   √  |
+|  Qwen2-VL                   |   √   |   √  |   √  |



@@ -17,25 +17,25 @@

 **SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.

-**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to Intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:

 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.

 ### Llama.cpp + SYCL

-The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
+The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
+SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.

 ## Recommended Release

-The SYCL backend would be broken by some PRs due to no online CI.
-
-The following release is verified with good quality:
+The following releases are verified and recommended:

 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
+|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |ArcB580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||

@@ -106,15 +106,14 @@ SYCL backend supports Intel GPU Family:
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
-| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake    |
-| Intel iGPU                    | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750, B580         |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
+| Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |

 *Notes:*

 - **Memory**
  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
-
  - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.

 - **Execution Unit (EU)**
@@ -138,9 +137,11 @@ Note: AMD GPU support is highly experimental and is incompatible with F16.
 Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.

 ## Docker
-The docker build option is currently limited to *intel GPU* targets.
+
+The docker build option is currently limited to *Intel GPU* targets.

 ### Build image
+
 ```sh
 # Using FP16
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
@@ -148,9 +149,10 @@ docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f

 *Notes*:

-To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.
+To build in default FP32 *(Slower than FP16 alternative)*, set `--build-arg="GGML_SYCL_F16=OFF"` in the previous command.

 You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
+Check the [documentation for Docker](../docker.md) to see the available images.

 ### Run container

@@ -250,7 +252,7 @@ sycl-ls

 - **Intel GPU**

-When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
+When targeting an intel GPU, the user should expect one or more devices among the available SYCL devices. Please make sure that at least one GPU is present via `sycl-ls`, for instance `[level_zero:gpu]` in the sample output below:

 ```
 [opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
@@ -282,7 +284,7 @@ For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:

 #### Intel GPU

-```
+```sh
 ./examples/sycl/build.sh
 ```

@@ -351,7 +353,7 @@ cmake --build build --config Release -j -v

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).

 ##### Check device

@@ -398,11 +400,15 @@ Choose one of following methods to run.

 ```sh
 ./examples/sycl/run-llama2.sh 0
+# OR
+./examples/sycl/run-llama3.sh 0
 ```
 - Use multiple devices:

 ```sh
 ./examples/sycl/run-llama2.sh
+# OR
+./examples/sycl/run-llama3.sh
 ```

 2. Command line
@@ -425,13 +431,13 @@ Examples:
 - Use device 0:

 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm none -mg 0
 ```

 - Use multiple devices:

 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer
 ```

 *Notes:*
@@ -452,7 +458,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 1. Install GPU driver

-Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).

 2. Install Visual Studio

@@ -629,7 +635,7 @@ Once it is completed, final results will be in **build/Release/bin**

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).

 ##### Check device

@@ -648,7 +654,7 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow
 build\bin\llama-ls-sycl-device.exe
 ```

-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *Intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -658,13 +664,14 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|

 ```
+
 #### Choose level-zero devices

 |Chosen Device ID|Setting|
 |-|-|
-|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|0|Default option. You may also want to `set ONEAPI_DEVICE_SELECTOR="level_zero:0"`|
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|

 #### Execute

@@ -673,7 +680,13 @@ Choose one of following methods to run.
 1. Script

 ```
-examples\sycl\win-run-llama2.bat
+examples\sycl\win-run-llama-2.bat
+```
+
+or
+
+```
+examples\sycl\win-run-llama-3.bat
 ```

 2. Command line
@@ -697,13 +710,13 @@ Examples:
 - Use device 0:

 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm none -mg 0
 ```

 - Use multiple devices:

 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer
 ```


@@ -714,7 +727,9 @@ Note:
 ```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
+
 Or
+
 ```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
@@ -726,15 +741,17 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
 | GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
 | GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |

+1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+
 #### Runtime

 | Name              | Value            | Function                                                                                                                  |
@@ -752,7 +769,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 ## Q&A

- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+- Error:  `error while loading shared libraries: libsycl.so: cannot open shared object file: No such file or directory`.

  - Potential cause: Unavailable oneAPI installation or not set ENV variables.
  - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
@@ -781,18 +798,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512

  It's same for other projects including llama.cpp SYCL backend.

- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+- `Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)`, `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 3503030272 Bytes of memory on device`, or `failed to allocate SYCL0 buffer`

-  Device Memory is not enough.
+  You are running out of Device Memory.

  |Reason|Solution|
  |-|-|
-  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
-  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
+  | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
+  | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|

 ### **GitHub contribution**:
-Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
+Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.

 ## TODO

- NA
+- Review ZES_ENABLE_SYSMAN: https://github.com/intel/compute-runtime/blob/master/programmers-guide/SYSMAN.md#support-and-limitations
@@ -22,6 +22,9 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)

 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).

@@ -104,7 +107,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment

 The defaults are:

- `MUSA_VERSION` set to `rc3.1.1`
+- `MUSA_VERSION` set to `rc4.0.1`

 The resulting images, are essentially the same as the non-MUSA images:

@@ -4,7 +4,9 @@ llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools
 - [llama-mtmd-cli](../tools/mtmd/README.md)
 - [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API

-To enable it, can use use one of the 2 methods below:
+Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
+
+To enable it, you can use one of the 2 methods below:

 - Use `-hf` option with a supported model (see a list of pre-quantized model below)
    - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
@@ -37,6 +39,8 @@ Replaces the `(tool_name)` with the name of binary you want to use. For example,

 NOTE: some models may require large context window, for example: `-c 8192`

+**Vision models**:
+
 ```sh
 # Gemma 3
 (tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
@@ -74,4 +78,15 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
 (tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
 (tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
+
+# Llama 4 Scout
+(tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
+```
+
+**Audio models**:
+
+```sh
+# Ultravox 0.5
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF
 ```
@@ -50,8 +50,6 @@ int main(int argc, char ** argv) {
    const int N = 5;  // n-gram size
    const int G = 15; // max verification n-grams

-    const bool dump_kv_cache = params.dump_kv_cache;
-
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -152,9 +150,6 @@ int main(int argc, char ** argv) {
    // here we keep adding new n-grams as we go
    ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);

-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
-
    const auto t_dec_start = ggml_time_us();

    // sample first token
@@ -172,12 +167,6 @@ int main(int argc, char ** argv) {
    }

    while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
        // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
        //
        // Example for W = 5, N = 4, G = 2:
@@ -473,8 +462,6 @@ int main(int argc, char ** argv) {

    common_sampler_free(smpl);

-    llama_kv_cache_view_free(&kvc_view);
-
    llama_batch_free(batch);

    llama_backend_free();
@@ -24,8 +24,6 @@ int main(int argc, char ** argv){
    // max. number of additional tokens to draft if match is found
    const int n_draft = params.speculative.n_max;

-    const bool dump_kv_cache = params.dump_kv_cache;
-
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -110,18 +108,9 @@ int main(int argc, char ** argv){

    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);

-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
-
    const auto t_dec_start = ggml_time_us();

    while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
        // print current draft sequence
        LOG_DBG("drafted %s\n", string_from(ctx, draft).c_str());

@@ -1,3 +1,14 @@
 # llama.cpp/example/parallel

 Simplified simulation of serving incoming requests in parallel
+
+## Example
+
+Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of 10 junk questions (`-j 10`) followed by the actual question.
+
+```bash
+llama-parallel -m model.gguf -np 8 -ns 128 --top-k 1 -pps --junk 10 -c 16384
+```
+
+> [!NOTE]
+> It's recommended to use base models with this example. Instruction tuned models might not be able to properly follow the custom chat template specified here, so the results might not be as expected.
@@ -34,11 +34,61 @@ static std::string k_system =
 R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
 The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.

-User: Recommend a nice restaurant in the area.
-Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
-User: Who is Richard Feynman?
-Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
-User:)";
+User:
+Recommend a nice restaurant in the area.
+Assistant:
+I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User:
+Who is Richard Feynman?
+Assistant:
+Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+)";
+
+static std::vector<std::string> k_questions = {
+    "What is the tallest mountain in the world?",
+    "Who was the first person to win two Nobel Prizes?",
+    "Which country invented paper?",
+    "What organ is primarily responsible for pumping blood throughout the body?",
+    "Which planet is known for its prominent ring system?",
+    "Who directed the movie 'Inception'?",
+    "What is the freezing point of water in Fahrenheit?",
+    "Which animal is known to have the longest lifespan?",
+    "What language has the most native speakers worldwide?",
+    "What is the capital city of Canada?",
+    "Who is credited with inventing the World Wide Web?",
+    "Which metal is liquid at room temperature?",
+    "What is the term for an animal that eats both plants and meat?",
+    "Who painted 'The Starry Night'?",
+    "What gas do humans exhale that plants use for photosynthesis?",
+    "What year did World War II end?",
+    "Which continent has the most countries?",
+    "Who wrote the novel 'Frankenstein'?",
+    "What does DNA stand for?",
+    "What is the main ingredient in traditional Japanese miso soup?"
+};
+
+static std::vector<std::string> k_answers = {
+    "The tallest mountain in the world is Mount Everest.",
+    "Marie Curie was the first person to win two Nobel Prizes.",
+    "Paper was invented in China.",
+    "The heart is the organ responsible for pumping blood.",
+    "Saturn is known for its prominent ring system.",
+    "Christopher Nolan directed the movie 'Inception'.",
+    "The freezing point of water in Fahrenheit is 32°F.",
+    "The bowhead whale is known to have the longest lifespan among mammals.",
+    "Mandarin Chinese has the most native speakers in the world.",
+    "The capital city of Canada is Ottawa.",
+    "Tim Berners-Lee is credited with inventing the World Wide Web.",
+    "Mercury is the metal that is liquid at room temperature.",
+    "An animal that eats both plants and meat is called an omnivore.",
+    "'The Starry Night' was painted by Vincent van Gogh.",
+    "Humans exhale carbon dioxide, which plants use in photosynthesis.",
+    "World War II ended in 1945.",
+    "Africa is the continent with the most countries.",
+    "The novel 'Frankenstein' was written by Mary Shelley.",
+    "DNA stands for Deoxyribonucleic Acid.",
+    "The main ingredient in traditional Japanese miso soup is fermented soybean paste."
+};

 static std::vector<std::string> k_prompts = {
    "What is the meaning of life?",
@@ -49,7 +99,7 @@ static std::vector<std::string> k_prompts = {
    "What is the best way to learn a new language?",
    "How to get a job at Google?",
    "If you could have any superpower, what would it be?",
-    "I want to learn how to play the piano.",
+    "I want to learn how to play the piano. What would be the best way to do it?",
 };

 struct client {
@@ -68,6 +118,7 @@ struct client {
    int64_t t_start_prompt;
    int64_t t_start_gen;

+    int32_t n_past    = 0;
    int32_t n_prompt  = 0;
    int32_t n_decoded = 0;
    int32_t i_batch   = -1;
@@ -107,6 +158,7 @@ int main(int argc, char ** argv) {
    common_params params;

    params.n_predict = 128;
+    params.n_junk = 0;

    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
        return 1;
@@ -126,7 +178,11 @@ int main(int argc, char ** argv) {
    // insert new requests as soon as the previous one is done
    const bool cont_batching = params.cont_batching;

-    const bool dump_kv_cache = params.dump_kv_cache;
+    // is the system prompt shared in the cache
+    const bool is_sp_shared = params.is_pp_shared;
+
+    // extra text to insert in each client's prompt in order to make it larger
+    const int32_t n_junk = params.n_junk;

    // init llama.cpp
    llama_backend_init();
@@ -169,6 +225,7 @@ int main(int argc, char ** argv) {
    }

    std::vector<llama_token> tokens_system;
+
    tokens_system = common_tokenize(ctx, k_system, true);
    const int32_t n_tokens_system = tokens_system.size();

@@ -182,15 +239,13 @@ int main(int argc, char ** argv) {
    int32_t n_total_gen    = 0;
    int32_t n_cache_miss   = 0;

-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
-
    const auto t_main_start = ggml_time_us();

    LOG_INF("%s: Simulating parallel requests from clients:\n", __func__);
    LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
    LOG_INF("\n");

-    {
+    if (is_sp_shared) {
        LOG_INF("%s: Evaluating the system prompt ...\n", __func__);

        for (int32_t i = 0; i < n_tokens_system; ++i) {
@@ -213,11 +268,6 @@ int main(int argc, char ** argv) {
    LOG_INF("Processing requests ...\n\n");

    while (true) {
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
        common_batch_clear(batch);

        // decode any currently ongoing sequences
@@ -228,7 +278,7 @@ int main(int argc, char ** argv) {

            client.i_batch = batch.n_tokens;

-            common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            common_batch_add(batch, client.sampled, client.n_past++, { client.id + 1 }, true);

            client.n_decoded += 1;
        }
@@ -254,9 +304,23 @@ int main(int argc, char ** argv) {
                    client.t_start_gen    = 0;

                    client.input    = k_prompts[rand() % k_prompts.size()];
-                    client.prompt   = client.input + "\nAssistant:";
                    client.response = "";

+                    // construct the prompt:
+                    // [system prompt] + [junk] + [user prompt]
+                    client.n_past = 0;
+                    client.prompt = "";
+                    if (is_sp_shared) {
+                        client.n_past = n_tokens_system;
+                    } else {
+                        client.prompt += k_system;
+                    }
+                    for (int i = 0; i < n_junk; ++i) {
+                        const int r = rand() % k_questions.size();
+                        client.prompt += "User:\n" + k_questions[r] + "\nAssistant:\n " + k_answers[r] + "\n";
+                    }
+                    client.prompt += "User:\n" + client.input + "\nAssistant:\n";
+
                    common_sampler_reset(client.smpl);

                    // do not prepend BOS because we have a system prompt!
@@ -264,7 +328,7 @@ int main(int argc, char ** argv) {
                    tokens_prompt = common_tokenize(ctx, client.prompt, false);

                    for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        common_batch_add(batch, tokens_prompt[i], client.n_past++, { client.id + 1 }, false);
                    }

                    // extract the logits only for the last token
@@ -363,10 +427,9 @@ int main(int argc, char ** argv) {
                //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());

                if (client.n_decoded > 2 &&
-                        (llama_vocab_is_eog(vocab, id) ||
-                         (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
-                         client.response.find("User:") != std::string::npos ||
-                         client.response.find('\n') != std::string::npos)) {
+                    (llama_vocab_is_eog(vocab, id) ||
+                     (params.n_predict > 0 && client.n_decoded >= params.n_predict) ||
+                     client.response.find("User:") != std::string::npos)) {
                    // basic reverse prompt
                    const size_t pos = client.response.find("User:");
                    if (pos != std::string::npos) {
@@ -81,14 +81,14 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
    }
 }

-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+static void batch_encode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
    // clear previous kv_cache values (irrelevant for embeddings)
    llama_kv_self_clear(ctx);

    // run model
    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        LOG_ERR("%s : failed to decode\n", __func__);
+    if (llama_encode(ctx, batch) < 0) {
+        LOG_ERR("%s : failed to encode\n", __func__);
    }

    for (int i = 0; i < batch.n_tokens; i++) {
@@ -233,7 +233,7 @@ int main(int argc, char ** argv) {
        // encode if at capacity
        if (batch.n_tokens + n_toks > n_batch) {
            float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd);
+            batch_encode(ctx, batch, out, s, n_embd);
            common_batch_clear(batch);
            p += s;
            s = 0;
@@ -246,7 +246,7 @@ int main(int argc, char ** argv) {

    // final batch
    float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd);
+    batch_encode(ctx, batch, out, s, n_embd);

    // save embeddings to chunks
    for (int i = 0; i < n_chunks; i++) {
@@ -267,7 +267,7 @@ int main(int argc, char ** argv) {
        batch_add_seq(query_batch, query_tokens, 0);

        std::vector<float> query_emb(n_embd, 0);
-        batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
+        batch_encode(ctx, query_batch, query_emb.data(), 1, n_embd);

        common_batch_clear(query_batch);

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
    auto generate = [&](const std::string & prompt) {
        std::string response;

-        const bool is_first = llama_kv_self_used_cells(ctx) == 0;
+        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;

        // tokenize the prompt
        const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
        while (true) {
            // check if we have enough space in the context to evaluate this batch
            int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_kv_self_used_cells(ctx);
+            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
            if (n_ctx_used + batch.n_tokens > n_ctx) {
                printf("\033[0m\n");
                fprintf(stderr, "context size exceeded\n");
@@ -84,13 +84,13 @@ int main(int argc, char ** argv) {
    model_params.n_gpu_layers = ngl;

    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
-    const llama_vocab * vocab = llama_model_get_vocab(model);

    if (model == NULL) {
        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
        return 1;
    }

+    const llama_vocab * vocab = llama_model_get_vocab(model);
    // tokenize the prompt

    // find the number of tokens in the prompt
@@ -12,16 +12,16 @@ source /opt/intel/oneapi/setvars.sh

 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
 MODEL_FILE=models/llama-2-7b.Q4_0.gguf
-NGL=33
-CONEXT=4096
+NGL=99
+CONTEXT=4096

 if [ $# -gt 0 ]; then
    GGML_SYCL_DEVICE=$1
    echo "use $GGML_SYCL_DEVICE as main GPU"
    #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none

 else
    #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi
@@ -0,0 +1,28 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2025 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+# If you want more control, DPC++ Allows selecting a specific device through the
+# following environment variable
+#export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
+NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
+CONTEXT=4096
+
+if [ $# -gt 0 ]; then
+    GGML_SYCL_DEVICE=$1
+    echo "Using $GGML_SYCL_DEVICE as the main GPU"
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+else
+    #use multiple GPUs with same max compute units
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT}
+fi
@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force


-.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0
@@ -0,0 +1,9 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+
+.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -e -ngl 99
@@ -128,6 +128,8 @@ extern "C" {
    // set gradients to zero, initilize loss, and optionally reset the optimizer
    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);

+    GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically
+
    // get underlying tensors that store data
    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
@@ -528,14 +528,15 @@ extern "C" {
        GGML_UNARY_OP_STEP,
        GGML_UNARY_OP_TANH,
        GGML_UNARY_OP_ELU,
-        GGML_UNARY_OP_RELU,
        GGML_UNARY_OP_SIGMOID,
        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_ERF,
        GGML_UNARY_OP_GELU_QUICK,
        GGML_UNARY_OP_SILU,
        GGML_UNARY_OP_HARDSWISH,
        GGML_UNARY_OP_HARDSIGMOID,
        GGML_UNARY_OP_EXP,
+        GGML_UNARY_OP_RELU,

        GGML_UNARY_OP_COUNT,
    };
@@ -1024,6 +1025,16 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

+    // GELU using erf (error function) when possible
+    // some backends may fallback to approximation based on Abramowitz and Stegun formula
+    GGML_API struct ggml_tensor * ggml_gelu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_erf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
    GGML_API struct ggml_tensor * ggml_gelu_quick(
            struct ggml_context * ctx,
            struct ggml_tensor  * a);
@@ -65,6 +65,7 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
+#include <aclnnop/aclnn_grouped_matmul_v2.h>
 #include <float.h>

 #include <cmath>
@@ -2587,3 +2588,149 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst){

    ggml_cann_release_resources(ctx, acl_src, acl_dst, alpha);
 }
+
+/**
+ * @brief Performs expert-specific matrix multiplication (MoE) with
+ * floating-point precision using the CANN backend.
+ *
+ * This function executes a matrix multiplication operation tailored for
+ * Mixture of Experts (MoE) models, where the input tensor is multiplied
+ * with expert-specific weight matrices. It uses the CANN backend for
+ * efficient computation and stores the result in the destination tensor `dst`.
+ * The operation may leverage identity-based optimizations or routing masks
+ * as part of sparse expert selection.
+ *
+ * @param ctx The context for executing CANN backend operations.
+ * @param dst The destination tensor where the MoE multiplication result
+ * will be stored.
+ *
+ * @note This function assumes floating-point data types and is designed for
+ * MoE architectures, possibly involving sparse expert routing.
+ */
+static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    //dst   [M, K, N, 1]
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // copy index from npu to cpu
+    int64_t n_as = ne02; // A
+    int64_t n_ids = ids->ne[0]; // K
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
+        ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+    size_t ori_src0_nb[4] = {nb00, nb01, nb02, nb03};
+
+    // src0 is F16, src1 is F32, dst is F32
+    ggml_cann_pool_alloc src0_cast_allocator;
+    if (src0->type == GGML_TYPE_F16) {
+        src0_cast_allocator.alloc(ctx.pool(), sizeof(float) * ggml_nelements(src0));
+        void* src0_cast_buf = src0_cast_allocator.get();
+
+        size_t cast_nb[GGML_MAX_DIMS];
+        cast_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            cast_nb[i] = cast_nb[i - 1] * src0->ne[i - 1];
+        }
+
+        aclTensor* acl_src0_f16 = ggml_cann_create_tensor(src0);
+        aclTensor* acl_cast = ggml_cann_create_tensor(src0_cast_buf,
+            ACL_FLOAT, sizeof(float), src0->ne, cast_nb, 4);
+        GGML_CANN_CALL_ACLNN_OP(ctx, Cast, acl_src0_f16, ACL_FLOAT, acl_cast);
+        ggml_cann_release_resources(ctx, acl_cast, acl_src0_f16);
+
+        src0_original = (char *) src0_cast_buf;
+        memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
+    }
+
+    std::vector<aclTensor*> src0_tensor_vec;
+    std::vector<aclTensor*> src1_tensor_vec;
+    std::vector<aclTensor*> dst_tensor_vec;
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // src0_row [M, D] -> weight && permute
+            int64_t src0_ne[2] = {ne01, ne00};
+            size_t src0_nb[2] = {ori_src0_nb[1], ori_src0_nb[0]};
+            // src1_row [D, 1] -> input
+            int64_t src1_ne[2] = {ne10, 1};
+            size_t src1_nb[2] = {nb10, nb11};
+            // dst_row [M, 1] -> out
+            int64_t dst_ne[2] = {ne0, 1};
+            size_t dst_nb[2] = {nb0, nb1};
+
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            aclTensor* acl_src0 = ggml_cann_create_tensor(src0_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                src0_ne, src0_nb, 2);
+            aclTensor* acl_src1 = ggml_cann_create_tensor(src1_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                src1_ne, src1_nb, 2);
+            aclTensor* acl_dst = ggml_cann_create_tensor(dst_tmp_ptr,
+                ACL_FLOAT, sizeof(float),
+                dst_ne, dst_nb, 2);
+
+            src0_tensor_vec.push_back(acl_src0);
+            src1_tensor_vec.push_back(acl_src1);
+            dst_tensor_vec.push_back(acl_dst);
+        }
+    }
+
+    // GroupedMatmulV2 required tensor_list.size < 128
+    size_t GROUP_SIZE = 128;
+    std::vector<std::vector<aclTensor*>> src0_tensor_vec_vec;
+    std::vector<std::vector<aclTensor*>> src1_tensor_vec_vec;
+    std::vector<std::vector<aclTensor*>> dst_tensor_vec_vec;
+
+    // split and call GroupedMatmulV2
+    for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
+        size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
+        std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
+        std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
+        std::vector<aclTensor*> dst_tensor_vec_split(dst_tensor_vec.begin() + i, dst_tensor_vec.begin() + end);
+
+        aclTensorList* src0_tensor_list = aclCreateTensorList(src0_tensor_vec_split.data(), src0_tensor_vec_split.size());
+        aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
+        aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+            nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
+
+        ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
+    }
+    return;
+}
+
+void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    const enum ggml_type type = dst->src[0]->type;
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            ggml_cann_mul_mat_id_fp(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupported type for mul_mat_id");
+            break;
+    }
+}
@@ -978,6 +978,33 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
    }
 }

+/**
+ * @brief   Performs sparse expert-based matrix multiplication using the CANN backend.
+ *
+ * @details This function implements a MoE-style batched matrix multiplication, where each input token
+ *          is routed to one or more experts, and each expert corresponds to a specific [D, M] weight matrix
+ *          in the source tensor `src0`. The routing indices are provided via the `ids` tensor.
+ *
+ *          For each token (from `src1`), the function selects the corresponding expert(s) as specified by `ids`,
+ *          performs the matrix multiplication with the selected expert's weight submatrix (from `src0`),
+ *          and stores the results in `dst`. This operation is optimized and executed on the CANN backend.
+ *
+ *          Dimensions:
+ *              - src0: [D, M, A, 1], where A is the number of experts
+ *              - src1: [D, B, N, 1], where N is batch size and B is the slot count per sample
+ *              - ids : [K, N],       where K is the number of experts each token is routed to
+ *              - dst : [M, K, N, 1], output tensor storing the result of expert × token multiplication
+ *
+ *          The function handles two main modes:
+ *              - If `ne12 == 1`, a simpler per-token loop is used.
+ *              - TODO: If `ne12 > 1`, grouped multiplication and memory copying is used for efficiency.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the expert-weighted token outputs are stored.
+ *            Expected to be of shape [M, K, N, 1].
+ */
+void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
 * @brief Applies a element-wise operation to two input tensors using the CANN
 * backend.
@@ -1672,7 +1672,8 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
            ggml_cann_mul_mat(ctx, dst);
            break;
        case GGML_OP_MUL_MAT_ID:
-            return false;
+            ggml_cann_mul_mat_id(ctx, dst);
+            break;
        case GGML_OP_SCALE:
            ggml_cann_scale(ctx, dst);
            break;
@@ -2030,7 +2031,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
            }
        }
        case GGML_OP_MUL_MAT_ID:
-            return false;
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
+                    return true;
+                default:
+                    return false;
+            }
        // embedding
        case GGML_OP_GET_ROWS: {
            switch (op->src[0]->type) {
@@ -2202,6 +2202,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                    } break;

                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_ERF:
                case GGML_UNARY_OP_GELU_QUICK:
                case GGML_UNARY_OP_SILU:
                    {
@@ -2691,6 +2691,109 @@ static void ggml_compute_forward_gelu(
    }
 }

+// ggml_compute_forward_gelu_erf
+
+static void ggml_compute_forward_gelu_erf_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_erf_f32(nc,
+                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_erf_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_erf_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_erf(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_gelu_erf_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_gelu_erf_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_gelu_quick

 static void ggml_compute_forward_gelu_quick_f32(
@@ -7749,6 +7852,10 @@ void ggml_compute_forward_unary(
            {
                ggml_compute_forward_gelu(params, dst);
            } break;
+        case GGML_UNARY_OP_GELU_ERF:
+            {
+                ggml_compute_forward_gelu_erf(params, dst);
+            } break;
        case GGML_UNARY_OP_GELU_QUICK:
            {
                ggml_compute_forward_gelu_quick(params, dst);
@@ -428,6 +428,7 @@ inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
 static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+static const float SQRT_2_INV      = 0.70710678118654752440084436210484f;

 inline static float ggml_gelu_f32(float x) {
    return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
@@ -440,6 +441,14 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
    }
 }

+inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_FP16_TO_FP32(x[i]);
+        float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+        y[i] = GGML_FP32_TO_FP16(res);
+    }
+}
+
 #ifdef GGML_GELU_FP16
 inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
    uint16_t t;
@@ -463,6 +472,13 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
 }
 #endif

+inline static void ggml_vec_gelu_erf_f32(const int n, float * y, const float * x) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+    }
+}
+
 inline static float ggml_gelu_quick_f32(float x) {
    return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
 }
@@ -1,5 +1,8 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
+#ifdef GGML_USE_MUSA
+#include "ggml-musa/mudnn.cuh"
+#endif // GGML_USE_MUSA

 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);

@@ -597,7 +600,14 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 #endif
    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
        GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
-        CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
+#ifdef GGML_USE_MUSA
+        if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
+            CUDA_CHECK(mudnnMemcpyAsync(ctx, src1, src0));
+        } else
+#endif // GGML_USE_MUSA
+        {
+            CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
+        }
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
@@ -772,7 +772,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
-    GGML_UNUSED(kb0);
+    GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"

 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // GGML_USE_HIP
 static __global__ void flash_attn_vec_ext_f16(
        const char * __restrict__ Q,
        const char * __restrict__ K,
@@ -48,6 +48,12 @@ static __global__ void flash_attn_vec_ext_f16(
        NO_DEVICE_CODE;
        return;
    }
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    if (ncols > 1) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)

    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.

@@ -91,6 +97,13 @@ static __global__ void flash_attn_vec_ext_f16(
            kqsum_shared[j][threadIdx.x] = 0.0f;
        }
    }
+
+    __shared__ half maskh_shared[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        maskh_shared[j*D + tid] = 0.0f;
+    }
+
    __syncthreads();

    // Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
@@ -175,6 +188,35 @@ static __global__ void flash_attn_vec_ext_f16(
    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
        // Calculate KQ tile and keep track of new maximum KQ values:

+        if (mask) {
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + k_VKQ_0 + tid];
+            }
+
+            __syncthreads();
+
+            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
+            // In such cases, skip the KV slice.
+            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
+#ifndef GGML_USE_HIP
+            bool skip = true;
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    const float2 tmp = __half22float2(((const half2 *) maskh_shared)[j*(D/2) + i]);
+                    skip = skip && isinf(tmp.x) && isinf(tmp.y);
+                }
+            }
+            if (__all_sync(0xFFFFFFFF, skip)) {
+                continue;
+            }
+#endif // GGML_USE_HIP
+        }
+
        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
@@ -202,7 +244,7 @@ static __global__ void flash_attn_vec_ext_f16(
                    sum = logit_softcap*tanhf(sum);
                }

-                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+                sum += maskh_shared[j*D + i_KQ];

                if (ncols == 1) {
                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
@@ -335,7 +377,9 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml
    float logit_softcap;
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

-    if (Q->ne[1] == 1) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
        constexpr int cols_per_block = 1;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"

 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // GGML_USE_HIP
 static __global__ void flash_attn_vec_ext_f32(
        const char * __restrict__ Q,
        const char * __restrict__ K,
@@ -60,6 +60,12 @@ static __global__ void flash_attn_vec_ext_f32(
        NO_DEVICE_CODE;
        return;
    }
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    if (ncols > 1) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)

    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.

@@ -104,6 +110,13 @@ static __global__ void flash_attn_vec_ext_f32(
            kqsum_shared[j][threadIdx.x] = 0.0f;
        }
    }
+
+    __shared__ float maskf_shared[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        maskf_shared[j*D + tid] = 0.0f;
+    }
+
    __syncthreads();

    // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
@@ -181,6 +194,34 @@ static __global__ void flash_attn_vec_ext_f32(
    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
        // Calculate KQ tile and keep track of new maximum KQ values:

+        if (mask) {
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + k_VKQ_0 + tid]);
+            }
+
+            __syncthreads();
+
+            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
+            // In such cases, skip the KV slice.
+            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
+#ifndef GGML_USE_HIP
+            bool skip = true;
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+                for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    skip = skip && isinf(maskf_shared[j*D + i]);
+                }
+            }
+            if (__all_sync(0xFFFFFFFF, skip)) {
+                continue;
+            }
+#endif // GGML_USE_HIP
+        }
+
        float kqmax_new_arr[ncols];
 #pragma unroll
        for (int j = 0; j < ncols; ++j) {
@@ -204,7 +245,7 @@ static __global__ void flash_attn_vec_ext_f32(
                    sum = logit_softcap*tanhf(sum);
                }

-                sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+                sum += maskf_shared[j*D + i_KQ];

                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);

@@ -326,7 +367,9 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml
    float logit_softcap;
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

-    if (Q->ne[1] == 1) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
        constexpr int cols_per_block = 1;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
@@ -149,6 +149,8 @@ enum ggml_metal_kernel_type {
    GGML_METAL_KERNEL_TYPE_SIGMOID,
    GGML_METAL_KERNEL_TYPE_GELU,
    GGML_METAL_KERNEL_TYPE_GELU_4,
+    GGML_METAL_KERNEL_TYPE_GELU_ERF,
+    GGML_METAL_KERNEL_TYPE_GELU_ERF_4,
    GGML_METAL_KERNEL_TYPE_GELU_QUICK,
    GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
    GGML_METAL_KERNEL_TYPE_SILU,
@@ -1103,6 +1105,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID,                         sigmoid,                         true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                            gelu,                            true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                          gelu_4,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_ERF,                        gelu_erf,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_ERF_4,                      gelu_erf_4,                      true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                      gelu_quick,                      true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                    gelu_quick_4,                    true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                            silu,                            true);
@@ -1613,6 +1617,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                case GGML_UNARY_OP_RELU:
                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_ERF:
                case GGML_UNARY_OP_GELU_QUICK:
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_ELU:
@@ -2251,6 +2256,25 @@ static bool ggml_metal_encode_node(

                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                } break;
+                case GGML_UNARY_OP_GELU_ERF:
+                {
+                    int64_t n = ggml_nelements(dst);
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    if (n % 4 == 0) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_ERF_4].pipeline;
+                        n /= 4;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_ERF].pipeline;
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                case GGML_UNARY_OP_GELU_QUICK:
                {
                    int64_t n = ggml_nelements(dst);
@@ -856,6 +856,7 @@ kernel void kernel_tanh(
 constant float GELU_COEF_A     = 0.044715f;
 constant float GELU_QUICK_COEF = -1.702f;
 constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;

 kernel void kernel_gelu(
    device const float * src0,
@@ -897,6 +898,42 @@ kernel void kernel_gelu_quick_4(
    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
 }

+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+kernel void kernel_gelu_erf(
+    device const float * src0,
+    device       float * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+
+    dst[tpig] = 0.5f*x*(1.0f+erf_approx<float>(x*SQRT_2_INV));
+}
+
+kernel void kernel_gelu_erf_4(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    dst[tpig] = 0.5f*x*(1.0f+erf_approx<float4>(x*SQRT_2_INV));
+}
+
 kernel void kernel_silu(
        device const float * src0,
        device       float * dst,
@@ -3255,7 +3292,7 @@ template<
    typename kd4x4_t, // key type in device memory
    short nl_k,
    void (*deq_k)(device const kd4x4_t *, short, thread k4x4_t &),
-    typename vd4x4_t, // key type in device memory
+    typename vd4x4_t, // value type in device memory
    short nl_v,
    void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
    short DK,        // K head size
@@ -3776,7 +3813,7 @@ template<
    typename kd4_t, // key type in device memory
    short nl_k,
    void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &),
-    typename vd4_t, // key type in device memory
+    typename vd4_t, // value type in device memory
    short nl_v,
    void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
    short DK,       // K head size
@@ -27,12 +27,15 @@ if (MUSAToolkit_FOUND)

    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
+    list(APPEND GGML_HEADERS_MUSA "../ggml-musa/mudnn.cuh")

    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
    list(APPEND GGML_SOURCES_MUSA ${SRCS})
    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-musa/*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})

    if (GGML_CUDA_FA_ALL_QUANTS)
        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
@@ -62,7 +65,9 @@ if (MUSAToolkit_FOUND)
                            )

    # TODO: do not use CUDA definitions for MUSA
-    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+    if (NOT GGML_BACKEND_DL)
+        target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+    endif()

    add_compile_definitions(GGML_USE_MUSA)
    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
@@ -92,9 +97,10 @@ if (MUSAToolkit_FOUND)
    endif()

    if (GGML_STATIC)
+        # TODO: mudnn has not provided static libraries yet
        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
    else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas mudnn)
    endif()

    if (GGML_CUDA_NO_VMM)
@@ -0,0 +1,112 @@
+#include <mutex>
+#include <mudnn.h>
+
+#include "mudnn.cuh"
+
+namespace mudnn = musa::dnn;
+
+// Returns a human-readable error string for mudnn::Status
+const char* mudnnGetErrorString(mudnn::Status err) {
+    switch (err) {
+        case mudnn::Status::SUCCESS:
+            return "Success";
+        case mudnn::Status::INVALID_PARAMETER:
+            return "Invalid parameter";
+        case mudnn::Status::NOT_INITIALIZED:
+            return "Not initialized";
+        case mudnn::Status::ALLOC_FAILED:
+            return "Allocation failed";
+        case mudnn::Status::NOT_SUPPORTED:
+            return "Not supported";
+        case mudnn::Status::INTERNAL_ERROR:
+            return "Internal error";
+        case mudnn::Status::ARCH_MISMATCH:
+            return "Architecture mismatch";
+        case mudnn::Status::EXECUTION_FAILED:
+            return "Execution failed";
+        default:
+            return "Unknown mudnn status";
+    }
+}
+
+// Error checking macro for MUDNN calls
+#define MUDNN_CHECK(err) CUDA_CHECK_GEN(err, mudnn::Status::SUCCESS, mudnnGetErrorString)
+
+namespace {
+    // Thread-safe cache for mudnn::Handle objects per device
+    std::unordered_map<int, std::unique_ptr<mudnn::Handle>> handle_cache;
+    std::mutex handle_cache_mutex;
+
+    mudnn::Handle* get_cached_handle(int device_id) {
+        std::lock_guard<std::mutex> lock(handle_cache_mutex);
+        auto it = handle_cache.find(device_id);
+        if (it != handle_cache.end()) {
+            return it->second.get();
+        }
+        auto handle = std::make_unique<mudnn::Handle>(device_id);
+        mudnn::Handle* handle_ptr = handle.get();
+        handle_cache[device_id] = std::move(handle);
+        return handle_ptr;
+    }
+}
+
+// Extracts dimensions and strides from a ggml_tensor
+int get_ggml_dims_and_strides(const ggml_tensor* tensor,
+                              std::vector<int64_t>& dims,
+                              std::vector<int64_t>& strides) {
+    const int ndims = ggml_n_dims(tensor);
+    const size_t element_size = ggml_element_size(tensor);
+
+    dims.resize(ndims);
+    strides.resize(ndims);
+
+    for (int i = 0; i < ndims; ++i) {
+        dims[i] = tensor->ne[i];
+        strides[i] = tensor->nb[i] / static_cast<int64_t>(element_size);
+    }
+    return ndims;
+}
+
+// Converts ggml_type to mudnn::Tensor::Type
+mudnn::Tensor::Type ggml_type_to_mudnn_type(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return mudnn::Tensor::Type::FLOAT;
+        case GGML_TYPE_F16:
+            return mudnn::Tensor::Type::HALF;
+
+        // TODO: Add support for other types
+
+        default:
+            MUDNN_CHECK(mudnn::Status::NOT_SUPPORTED);
+    }
+
+    return mudnn::Tensor::Type::FLOAT; // Default fallback
+}
+
+// Asynchronous memory copy using mudnn::Unary::IDENTITY
+musaError_t mudnnMemcpyAsync(ggml_backend_cuda_context& ctx, const ggml_tensor* dst, const ggml_tensor* src) {
+    mudnn::Tensor tensor_dst, tensor_src;
+
+    MUDNN_CHECK(tensor_dst.SetType(ggml_type_to_mudnn_type(dst->type)));
+    MUDNN_CHECK(tensor_src.SetType(ggml_type_to_mudnn_type(src->type)));
+
+    std::vector<int64_t> dims, strides;
+    const int ndims = get_ggml_dims_and_strides(src, dims, strides);
+
+    MUDNN_CHECK(tensor_dst.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_src.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_dst.SetAddr(dst->data));
+    MUDNN_CHECK(tensor_src.SetAddr(src->data));
+
+    mudnn::Unary op;
+    MUDNN_CHECK(op.SetMode(mudnn::Unary::Mode::IDENTITY));
+    MUDNN_CHECK(op.SetAlpha(0.0f));
+    MUDNN_CHECK(op.SetBeta(0.0f));
+
+    mudnn::Handle* handle = get_cached_handle(ctx.device);
+    MUDNN_CHECK(handle->SetStream(ctx.stream()));
+    MUDNN_CHECK(op.Run(*handle, tensor_dst, tensor_src));
+
+    return musaSuccess;
+}
@@ -0,0 +1,12 @@
+#pragma once
+
+#include "../include/ggml.h"
+#include "../ggml-cuda/common.cuh"
+
+// Asynchronously copies data from src tensor to dst tensor using the provided context.
+// Returns a musaError_t indicating success or failure.
+musaError_t mudnnMemcpyAsync(
+    ggml_backend_cuda_context &ctx,
+    const ggml_tensor *dst,
+    const ggml_tensor *src
+);
@@ -27,6 +27,7 @@
 #include <cmath>
 #include <memory>
 #include <charconv>
+#include <mutex>

 #undef MIN
 #undef MAX
@@ -74,6 +75,7 @@ struct ggml_cl_version {
    cl_uint minor = 0;
 };

+
 struct ggml_cl_compiler_version {
    ADRENO_CL_COMPILER_TYPE type;
    int major = -1;
@@ -91,6 +93,14 @@ struct ggml_cl_compiler_version {
    }
 };

+static size_t align_to(size_t value, size_t to_alignment) {
+    GGML_ASSERT(to_alignment && "Invalid alignment (must be non-zero)");
+    GGML_ASSERT((to_alignment & (to_alignment - 1)) == 0 && "to_alignment must be power-of-two");
+
+    return ((value + to_alignment - 1) / to_alignment) * to_alignment;
+}
+
+
 // Parses a version string of form "XX.YY ". On an error returns ggml_cl_version with all zeroes.
 static ggml_cl_version parse_cl_version(std::string_view str) {
    size_t major_str_begin = 0;
@@ -221,13 +231,25 @@ static ggml_cl_compiler_version get_adreno_cl_compiler_version(const char *drive
    return { type, major, minor, patch };
 }

+struct ggml_backend_opencl_context;
+
 // backend device context
 struct ggml_backend_opencl_device_context {
    cl_platform_id platform;
    std::string platform_name;

-    cl_device_id device;
-    std::string device_name;
+    cl_device_id   device;
+    std::string    device_name;
+    cl_device_type device_type;
+    std::string    device_version;
+
+    // Initialized by ggml_cl2_init().
+    ggml_backend_opencl_context * backend_ctx = nullptr;
+
+    // Initialized by ggml_backend_opencl_device_get_buffer_type()
+    ggml_backend_buffer_type buffer_type;
+
+    cl_context context = nullptr;
 };

 // backend context
@@ -248,6 +270,8 @@ struct ggml_backend_opencl_context {

    int adreno_wave_size;

+    cl_bool non_uniform_workgroups;
+
    cl_context context;
    cl_command_queue queue;

@@ -344,15 +368,8 @@ struct ggml_backend_opencl_context {
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 };

-static ggml_backend_device                 g_ggml_backend_opencl_device;
-static ggml_backend_opencl_device_context  g_ggml_ctx_dev_main {
-    /*.platform         =*/ nullptr,
-    /*.platform_nane    =*/ "",
-    /*.device           =*/ nullptr,
-    /*.device_name      =*/ "",
-};
-
-static int ggml_backend_opencl_n_devices = 0;
+// All registered devices with a default device in the front.
+static std::vector<ggml_backend_device> g_ggml_backend_opencl_devices;

 // Profiling
 #ifdef GGML_OPENCL_PROFILING
@@ -1107,25 +1124,19 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
    GGML_LOG_CONT("\n");
 }

-static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
-    static bool initialized = false;
-    static ggml_backend_opencl_context *backend_ctx = nullptr;
+// XXX static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
+// XXX    static bool initialized = false;
+// XXX    static ggml_backend_opencl_context *backend_ctx = nullptr;

-    if (initialized) {
-        return backend_ctx;
-    }
+static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev);

-    ggml_backend_opencl_device_context *dev_ctx = (ggml_backend_opencl_device_context *)dev->context;
-    GGML_ASSERT(dev_ctx);
-    GGML_ASSERT(dev_ctx->platform == nullptr);
-    GGML_ASSERT(dev_ctx->device == nullptr);
-    GGML_ASSERT(backend_ctx == nullptr);
+namespace /* anonymous */ {
+extern struct ggml_backend_device_i ggml_backend_opencl_device_i;
+}

-    initialized = true;
-    backend_ctx = new ggml_backend_opencl_context();
-    backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
-
-    cl_int err;
+// Look for available and suitable devices.
+static std::vector<ggml_backend_device> ggml_opencl_probe_devices(ggml_backend_reg * reg) {
+    std::vector<ggml_backend_device> found_devices;

 #ifdef GGML_OPENCL_PROFILING
    GGML_LOG_INFO("ggml_opencl: OpenCL profiling enabled\n");
@@ -1158,11 +1169,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    struct cl_device devices[NDEV];
    unsigned n_devices = 0;
    struct cl_device * default_device = NULL;
+    unsigned           default_platform_number = 0;

    cl_platform_id platform_ids[NPLAT];
    if (clGetPlatformIDs(NPLAT, platform_ids, &n_platforms) != CL_SUCCESS) {
        GGML_LOG_ERROR("ggml_opencl: plaform IDs not available.\n");
-        return backend_ctx;
+        return found_devices;
    }

    for (unsigned i = 0; i < n_platforms; i++) {
@@ -1197,19 +1209,22 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
        }

        if (default_device == NULL && p->default_device != NULL) {
-            default_device = p->default_device;
+            default_device          = p->default_device;
+            default_platform_number = i;
        }
    }

    if (n_devices == 0) {
        GGML_LOG_ERROR("ggml_opencl: could find any OpenCL devices.\n");
-        return backend_ctx;
+        return found_devices;
    }

-    char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
-    char * user_device_string = getenv("GGML_OPENCL_DEVICE");
-    int user_platform_number = -1;
-    int user_device_number = -1;
+    char *      user_platform_string = getenv("GGML_OPENCL_PLATFORM");
+    char *      user_device_string   = getenv("GGML_OPENCL_DEVICE");
+    int         user_platform_number = -1;
+    int         user_device_number   = -1;
+    cl_device * candidate_devices    = nullptr;
+    unsigned    n_candidate_devices  = 0;

    unsigned n;
    if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
@@ -1224,12 +1239,11 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
            GGML_LOG_ERROR("ggml_opencl: invalid device number %d\n", user_device_number);
            exit(1);
        }
-        default_device = &platform->devices[user_device_number];
+        default_device      = &platform->devices[user_device_number];
+        candidate_devices   = platform->devices;
+        n_candidate_devices = platform->n_devices;
    } else {
-
-        struct cl_device * selected_devices = devices;
-        unsigned n_selected_devices = n_devices;
-
+        // Choose a platform by matching a substring.
        if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
            for (unsigned i = 0; i < n_platforms; i++) {
                struct cl_platform * p = &platforms[i];
@@ -1244,20 +1258,20 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
                exit(1);
            }
        }
-        if (user_platform_number != -1) {
-            struct cl_platform * p = &platforms[user_platform_number];
-            selected_devices = p->devices;
-            n_selected_devices = p->n_devices;
-            default_device = p->default_device;
-            if (n_selected_devices == 0) {
-                GGML_LOG_ERROR("ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
-                exit(1);
-            }
+
+        int                  platform_idx = user_platform_number != -1 ? user_platform_number : default_platform_number;
+        struct cl_platform * p            = &platforms[platform_idx];
+        candidate_devices                 = p->devices;
+        n_candidate_devices               = p->n_devices;
+        default_device                    = p->default_device;
+        if (n_candidate_devices == 0) {
+            GGML_LOG_ERROR("ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
+            exit(1);
        }

        if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
-            for (unsigned i = 0; i < n_selected_devices; i++) {
-                struct cl_device * d = &selected_devices[i];
+            for (unsigned i = 0; i < n_candidate_devices; i++) {
+                struct cl_device * d = &candidate_devices[i];
                if (strstr(d->name, user_device_string) != NULL) {
                    user_device_number = d->number;
                    break;
@@ -1269,71 +1283,145 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
            }
        }
        if (user_device_number != -1) {
-            selected_devices = &devices[user_device_number];
-            n_selected_devices = 1;
-            default_device = &selected_devices[0];
+            candidate_devices   = &devices[user_device_number];
+            n_candidate_devices = 1;
+            default_device      = &candidate_devices[0];
        }

-        GGML_ASSERT(n_selected_devices > 0);
+        GGML_ASSERT(n_candidate_devices > 0);

        if (default_device == NULL) {
-            default_device = &selected_devices[0];
+            default_device = &candidate_devices[0];
        }
    }

-    GGML_LOG_INFO("ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
-    GGML_LOG_INFO("ggml_opencl: selecting device: '%s (%s)'\n", default_device->name, default_device->version);
-    if (default_device->type != CL_DEVICE_TYPE_GPU) {
-        GGML_LOG_WARN("ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
+    GGML_ASSERT(n_candidate_devices != 0 && candidate_devices);
+
+    // Put the default device in front.
+    for (unsigned i = 1; i < n_candidate_devices; i++) {
+        if (&candidate_devices[i] == default_device) {
+            std::swap(candidate_devices[0], candidate_devices[i]);
+            default_device = &candidate_devices[0];
+            break;
+        }
    }

-    dev_ctx->platform = default_device->platform->id;
-    dev_ctx->device = default_device->id;
-    backend_ctx->device = default_device->id;
+    GGML_LOG_INFO("ggml_opencl: selected platform: '%s'\n", default_device->platform->name);

-    if (strstr(default_device->name, "Adreno") ||
-        strstr(default_device->name, "Qualcomm") ||
-        strstr(default_device->version, "Adreno")) {
+    std::vector<cl_device_id> device_ids;
+    for (auto dev = candidate_devices, dev_end = candidate_devices + n_candidate_devices; dev != dev_end; dev++) {
+        device_ids.push_back(dev->id);
+    }
+
+    cl_int                err;
+    cl_context            shared_context;
+    cl_context_properties properties[] = { (intptr_t) CL_CONTEXT_PLATFORM, (intptr_t) default_device->platform->id, 0 };
+
+    CL_CHECK(
+        (shared_context = clCreateContext(properties, device_ids.size(), device_ids.data(), NULL, NULL, &err), err));
+
+    for (auto dev = candidate_devices, dev_end = candidate_devices + n_candidate_devices; dev != dev_end; dev++) {
+        GGML_LOG_INFO("\nggml_opencl: device: '%s (%s)'\n", dev->name, dev->version);
+
+        auto dev_ctx = std::unique_ptr<ggml_backend_opencl_device_context>(new ggml_backend_opencl_device_context{
+            /*.platform         =*/dev->platform->id,
+            /*.platform_nane    =*/dev->platform->name,
+            /*.device           =*/dev->id,
+            /*.device_name      =*/dev->name,
+            /*.device_type      =*/dev->type,
+            /*.device_version   =*/dev->version,
+            /*.backend_ctx      =*/nullptr,
+            /*.buffer_type      =*/{},
+            /*.context          =*/shared_context,
+        });
+
+        found_devices.push_back(ggml_backend_device{
+            /* .iface   = */ ggml_backend_opencl_device_i,
+            /* .reg     = */ reg,
+            /* .context = */ dev_ctx.get(),
+        });
+
+        if (!ggml_cl2_init(&found_devices.back())) {
+            found_devices.pop_back();
+            GGML_LOG_INFO("ggml_opencl: drop unsupported device.\n");
+            continue;
+        }
+
+        dev_ctx.release();
+    }
+
+    if (found_devices.size()) {
+        auto * dev_ctx = static_cast<ggml_backend_opencl_device_context *>(found_devices.front().context);
+        GGML_LOG_INFO("ggml_opencl: default device: '%s (%s)'\n", dev_ctx->device_name.c_str(),
+                      dev_ctx->device_version.c_str());
+
+        if (dev_ctx->device_type != CL_DEVICE_TYPE_GPU) {
+            GGML_LOG_WARN("ggml_opencl: warning, the default device is not a GPU: '%s'.\n",
+                          dev_ctx->device_name.c_str());
+        }
+    }
+
+    return found_devices;
+}
+
+// Initialize device if it is supported (returns nullptr if it is not).
+static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
+    GGML_ASSERT(dev);
+    GGML_ASSERT(dev->context);
+
+    ggml_backend_opencl_device_context * dev_ctx = (ggml_backend_opencl_device_context *) dev->context;
+    GGML_ASSERT(dev_ctx->platform);
+    GGML_ASSERT(dev_ctx->device);
+
+    if (dev_ctx->backend_ctx) {
+        return dev_ctx->backend_ctx;
+    }
+
+    auto backend_ctx        = std::make_unique<ggml_backend_opencl_context>();
+    backend_ctx->device     = dev_ctx->device;
+    backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
+
+    if (strstr(dev_ctx->device_name.c_str(), "Adreno") ||
+        strstr(dev_ctx->device_name.c_str(), "Qualcomm") ||
+        strstr(dev_ctx->device_version.c_str(), "Adreno")) {
        backend_ctx->gpu_family = GPU_FAMILY::ADRENO;
        // Usually device version contains the detailed device name
-        backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->version);
+        backend_ctx->adreno_gen = get_adreno_gpu_gen(dev_ctx->device_version.c_str());
        if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::ADRENO_UNKNOWN) {
-            backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name);
+            backend_ctx->adreno_gen = get_adreno_gpu_gen(dev_ctx->device_name.c_str());
        }

        // Use wave size of 64 for all Adreno GPUs.
        backend_ctx->adreno_wave_size = 64;
-    } else if (strstr(default_device->name, "Intel")) {
+    } else if (strstr(dev_ctx->device_name.c_str(), "Intel")) {
        backend_ctx->gpu_family = GPU_FAMILY::INTEL;
    } else {
-        GGML_LOG_ERROR("Unsupported GPU: %s\n", default_device->name);
+        GGML_LOG_ERROR("Unsupported GPU: %s\n", dev_ctx->device_name.c_str());
        backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
-        return backend_ctx;
+        return nullptr;
    }

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    if (backend_ctx->gpu_family != GPU_FAMILY::ADRENO) {
        GGML_LOG_ERROR("ggml_opencl: Adreno-specific kernels should not be enabled for non-Adreno GPUs; "
            "run on an Adreno GPU or recompile with CMake option `-DGGML_OPENCL_USE_ADRENO_KERNELS=OFF`\n");
-        return backend_ctx;
+        return nullptr;
    }
 #endif

    // Populate backend device name
-    dev_ctx->platform_name = default_device->platform->name;
-    dev_ctx->device_name = default_device->name;
-    backend_ctx->device_name = default_device->name;
+    backend_ctx->device_name = dev_ctx->device_name;

    // A local ref of cl_device_id for convenience
    cl_device_id device = backend_ctx->device;

-    ggml_cl_version platform_version = get_opencl_platform_version(default_device->platform->id);
+    ggml_cl_version platform_version = get_opencl_platform_version(dev_ctx->platform);

    // Check device OpenCL version, OpenCL 2.0 or above is required
    ggml_cl_version opencl_c_version = get_opencl_c_version(platform_version, device);
    if (opencl_c_version.major < 2) {
        GGML_LOG_ERROR("ggml_opencl: OpenCL 2.0 or above is required\n");
-        return backend_ctx;
+        return nullptr;
    }

    // Check driver version
@@ -1364,7 +1452,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    // fp16 is required
    if (!backend_ctx->fp16_support) {
        GGML_LOG_ERROR("ggml_opencl: device does not support FP16\n");
-        return backend_ctx;
+        return nullptr;
    }

    // If OpenCL 3.0 is supported, then check for cl_khr_subgroups, which becomes
@@ -1373,7 +1461,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
        strstr(ext_buffer, "cl_intel_subgroups") == NULL) {
        GGML_LOG_ERROR("ggml_opencl: device does not support subgroups (cl_khr_subgroups or cl_intel_subgroups) "
            "(note that subgroups is an optional feature in OpenCL 3.0)\n");
-        return backend_ctx;
+        return nullptr;
    }

    cl_uint base_align_in_bits;
@@ -1397,6 +1485,15 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    GGML_LOG_INFO("ggml_opencl: SVM atomics support: %s\n",
        svm_caps & CL_DEVICE_SVM_ATOMICS ? "true" : "false");

+    if (opencl_c_version.major >= 3) {
+        CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_NON_UNIFORM_WORK_GROUP_SUPPORT, sizeof(cl_bool),
+                                 &backend_ctx->non_uniform_workgroups, 0));
+    } else {
+        GGML_ASSERT(opencl_c_version.major == 2);
+        // Non-uniform workgroup sizes is mandatory feature in v2.x.
+        backend_ctx->non_uniform_workgroups = true;
+    }
+
    // Print out configurations
 #ifdef GGML_OPENCL_SOA_Q
    GGML_LOG_INFO("ggml_opencl: flattening quantized weights representation as struct of arrays (GGML_OPENCL_SOA_Q)\n");
@@ -1406,14 +1503,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    GGML_LOG_INFO("ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)\n");
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS

-    cl_context_properties properties[] = {
-        (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)dev_ctx->platform, 0
-    };
-
-    CL_CHECK((backend_ctx->context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
+    cl_int err;

    // A local ref of cl_context for convenience
-    cl_context context = backend_ctx->context;
+    cl_context context = backend_ctx->context = dev_ctx->context;

    //CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
    //    (err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
@@ -1426,7 +1519,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    CL_CHECK((backend_ctx->queue = clCreateCommandQueue(context, device, command_queue_props, &err), err));

    // Load kernels
-    load_cl_kernels(backend_ctx, opencl_c_version);
+    load_cl_kernels(backend_ctx.get(), opencl_c_version);

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    // Allocate intermediate buffers and images
@@ -1456,10 +1549,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    CL_CHECK((backend_ctx->B_d_max   = clCreateBuffer(context, 0, max_B_d_bytes,   NULL, &err), err));
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS

-    // For now we support a single devices
-    ggml_backend_opencl_n_devices = 1;
-
-    return backend_ctx;
+    dev_ctx->backend_ctx = backend_ctx.release();
+    return dev_ctx->backend_ctx;
 }

 static void ggml_cl2_free(void) {
@@ -1664,10 +1755,46 @@ static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
    GGML_UNUSED(backend);
 }

+// Syncronizes the 'backend_ctx's device with others so that commands
+// enqueued to it won't start until commands in the other devices have
+// completed.
+static void sync_with_other_backends(ggml_backend_opencl_context * backend_ctx) {
+    if (g_ggml_backend_opencl_devices.size() < 2)
+      return; // No other devices to synchronize with.
+
+    std::vector<cl_event> events;
+    events.reserve(g_ggml_backend_opencl_devices.size());
+
+    for (ggml_backend_device & backend_dev : g_ggml_backend_opencl_devices) {
+        auto * other_backend_ctx = ggml_cl2_init(&backend_dev);
+        if (backend_ctx != other_backend_ctx) {
+            cl_event ev;
+            CL_CHECK(clEnqueueMarkerWithWaitList(other_backend_ctx->queue, 0, nullptr, &ev));
+            CL_CHECK(clFlush(other_backend_ctx->queue));
+            events.push_back(ev);
+        }
+    }
+
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, events.size(), events.data(), nullptr));
+    for (auto ev : events) {
+        CL_CHECK(clReleaseEvent(ev));
+    }
+}
+
+static void sync_with_other_backends(ggml_backend_t backend) {
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+    sync_with_other_backends(backend_ctx);
+}
+
 static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
    for (int i = 0; i < cgraph->n_nodes; i++) {
        ggml_tensor * node = cgraph->nodes[i];

+        // NOTE: this may oversynchronize by synchronizing with
+        //       backends/devices which don't compute 'cgraph's
+        //       dependencies.
+        sync_with_other_backends(backend);
+
        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
            continue;
        }
@@ -2058,15 +2185,16 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        // The original tensor memory is divided into scales and quants, i.e.,
        // we first store scales, then quants.
        // Create subbuffer for scales.
-        region.origin = extra_orig->offset + tensor->view_offs + offset;
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
        region.size = size_d;
        extra->d = clCreateSubBuffer(
            extra_orig->data_device, CL_MEM_READ_WRITE,
            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
        CL_CHECK(err);
+        auto previous_origin = region.origin;

        // Create subbuffer for quants.
-        region.origin = extra_orig->offset + tensor->view_offs + offset + size_d;
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
        region.size = size_q;
        extra->q = clCreateSubBuffer(
            extra_orig->data_device, CL_MEM_READ_WRITE,
@@ -2271,8 +2399,8 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
    cl_context context = backend_ctx->context;
    cl_command_queue queue = backend_ctx->queue;

-    // Make sure all previously submitted commands are finished.
-    CL_CHECK(clFinish(queue));
+    // Make sure all previously submitted commands in other devices are finished.
+    sync_with_other_backends(backend_ctx);

 #ifdef GGML_OPENCL_SOA_Q
    // In end-to-end runs, get_tensor is usually used to get back the logits,
@@ -2376,13 +2504,8 @@ static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_b
 }

 static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
-    // FIXME: not thread safe, device may not be initialized yet
-    static cl_uint alignment = -1;
-    if (alignment == (cl_uint)-1) {
-        ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer_type->device);
-        alignment = backend_ctx->alignment;
-    }
-    return alignment;
+    ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer_type->device);
+    return backend_ctx->alignment;
 }

 static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
@@ -2409,16 +2532,6 @@ static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
    /* .is_host          = */ NULL,
 };

-ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
-    static ggml_backend_buffer_type buffer_type = {
-        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
-        /* .device  = */ &g_ggml_backend_opencl_device,
-        /* .context = */ nullptr,
-    };
-
-    return &buffer_type;
-}
-
 //
 // backend device
 //
@@ -2476,9 +2589,15 @@ static ggml_backend_t ggml_backend_opencl_device_init(ggml_backend_dev_t dev, co
 }

 static ggml_backend_buffer_type_t ggml_backend_opencl_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_opencl_buffer_type();
+    auto * dev_ctx = static_cast<ggml_backend_opencl_device_context *>(dev->context);

-    GGML_UNUSED(dev);
+    dev_ctx->buffer_type = ggml_backend_buffer_type{
+        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
+        /* .device  = */ dev,
+        /* .context = */ nullptr,
+    };
+
+    return &dev_ctx->buffer_type;
 }

 static ggml_backend_buffer_t ggml_backend_opencl_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
@@ -2494,12 +2613,21 @@ static bool ggml_backend_opencl_device_supports_op(ggml_backend_dev_t dev, const
 }

 static bool ggml_backend_opencl_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_opencl_buffer_type_get_name;
+    // Check 'dev' and 'buffer_type' are not objects belonging to this backend.
+    if (dev->iface.get_name != ggml_backend_opencl_device_get_name ||
+        buft->iface.get_name != ggml_backend_opencl_buffer_type_get_name) {
+        return false;
+    }

-    GGML_UNUSED(dev);
+    // Check cl_context is the same. clEnqueue* commands may not use
+    // buffers from another cl_context.
+    ggml_backend_opencl_context * backend_ctx0 = ggml_cl2_init(dev);
+    ggml_backend_opencl_context * backend_ctx1 = ggml_cl2_init(buft->device);
+    return backend_ctx0->context == backend_ctx1->context;
 }

-static struct ggml_backend_device_i ggml_backend_opencl_device_i = {
+namespace /* anonymous */ {
+struct ggml_backend_device_i ggml_backend_opencl_device_i = {
    /* .get_name             = */ ggml_backend_opencl_device_get_name,
    /* .get_description      = */ ggml_backend_opencl_device_get_description,
    /* .get_memory           = */ ggml_backend_opencl_device_get_memory,
@@ -2516,6 +2644,7 @@ static struct ggml_backend_device_i ggml_backend_opencl_device_i = {
    /* .event_free           = */ NULL,
    /* .event_synchronize    = */ NULL,
 };
+}

 // Backend registry

@@ -2526,15 +2655,15 @@ static const char * ggml_backend_opencl_reg_get_name(ggml_backend_reg_t reg) {
 }

 static size_t ggml_backend_opencl_reg_device_count(ggml_backend_reg_t reg) {
-    return ggml_backend_opencl_n_devices;
+    return g_ggml_backend_opencl_devices.size();

    GGML_UNUSED(reg);
 }

 static ggml_backend_dev_t ggml_backend_opencl_reg_device_get(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
+    GGML_ASSERT(index < ggml_backend_opencl_reg_device_count(reg));

-    return &g_ggml_backend_opencl_device;
+    return &g_ggml_backend_opencl_devices[index];

    GGML_UNUSED(reg);
    GGML_UNUSED(index);
@@ -2548,27 +2677,23 @@ static struct ggml_backend_reg_i ggml_backend_opencl_reg_i = {
 };

 ggml_backend_reg_t ggml_backend_opencl_reg(void) {
-    // TODO: make this thread-safe somehow?
+    static std::mutex mutex;
    static ggml_backend_reg reg;
    static bool initialized = false;
+    std::lock_guard<std::mutex> lock(mutex);

-    if (!initialized) {
-        reg = ggml_backend_reg {
-            /* .api_version = */ GGML_BACKEND_API_VERSION,
-            /* .iface   = */ ggml_backend_opencl_reg_i,
-            /* .context = */ NULL,
-        };
-
-        g_ggml_backend_opencl_device = ggml_backend_device {
-            /* .iface   = */ ggml_backend_opencl_device_i,
-            /* .reg     = */ &reg,
-            /* .context = */ &g_ggml_ctx_dev_main,
-        };
-
-        ggml_cl2_init(&g_ggml_backend_opencl_device);
-
-        initialized = true;
+    if (initialized) {
+        return &reg;
    }
+    initialized = true;
+
+    g_ggml_backend_opencl_devices = ggml_opencl_probe_devices(&reg);
+
+    reg = ggml_backend_reg{
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_opencl_reg_i,
+        /* .context     = */ NULL,
+    };

    return &reg;
 }
@@ -2942,14 +3067,19 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
        size_t global_work_size[] = {(size_t)n, 1, 1};
        size_t local_work_size[] = {64, 1, 1};

+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+        }
+
 #ifdef GGML_OPENCL_PROFILING
        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
    } else {
        unsigned int nth = MIN(64, ne0);
@@ -3077,14 +3207,19 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
        size_t global_work_size[] = {(size_t)n, 1, 1};
        size_t local_work_size[] = {64, 1, 1};

+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+        }
+
 #ifdef GGML_OPENCL_PROFILING
        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
    } else {
        unsigned int nth = MIN(64, ne0);
@@ -3233,14 +3368,19 @@ static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const
    size_t global_work_size[] = {(size_t)n, 1, 1};
    size_t local_work_size[] = {64, 1, 1};

+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
 #ifdef GGML_OPENCL_PROFILING
    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
 }

@@ -3273,14 +3413,19 @@ static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const
    size_t global_work_size[] = {(size_t)n, 1, 1};
    size_t local_work_size[] = {64, 1, 1};

+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
 #ifdef GGML_OPENCL_PROFILING
    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
 }

@@ -3320,14 +3465,19 @@ static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, cons
    size_t global_work_size[] = {(size_t)n, 1, 1};
    size_t local_work_size[] = {64, 1, 1};

+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
 #ifdef GGML_OPENCL_PROFILING
    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
 }

@@ -4230,14 +4380,19 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
    size_t global_work_size[] = {(size_t)n, 1, 1};
    size_t local_work_size[] = {64, 1, 1};

+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
 #ifdef GGML_OPENCL_PROFILING
    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
 }

@@ -4418,14 +4573,19 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr
        size_t global_work_size[] = {(size_t)ne00, (size_t)ne01, (size_t)ne02};
        size_t local_work_size[] = {64, 1, 1};

+        size_t * local_work_size_ptr = local_work_size;
+        if (ne00 % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+        }
+
 #ifdef GGML_OPENCL_PROFILING
        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));

        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
 #else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
 #endif
    }
 }
@@ -576,6 +576,10 @@ void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
    }
 }

+bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->static_graphs;
+}
+
 struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
    return opt_ctx->inputs;
 }
@@ -842,6 +846,7 @@ void ggml_opt_epoch(
        int64_t                 idata_split,
        ggml_opt_epoch_callback callback_train,
        ggml_opt_epoch_callback callback_eval) {
+    GGML_ASSERT(ggml_opt_static_graphs(opt_ctx) && "ggml_opt_epoch requires static graphs");
    struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
    struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
    struct ggml_tensor * data   = ggml_opt_dataset_data(dataset);
@@ -385,16 +385,17 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
    ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
    ggml_sycl_set_device(ctx->device);
    auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
-    SYCL_CHECK(
-        CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
+    SYCL_CHECK(CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
+#ifndef _WIN32
    // Note: Use host buffer to save the data from mmap(), then copy to device. It's workaround for mmap() issue on PVC GPU.
    // This function will be called during load model from disk. Use memory buffer replace dynamic won't save more time and brings potential memory leak risk here.
-    char* host_buf = (char*)malloc(size);
+    char * host_buf = (char *) malloc(size);
    memcpy(host_buf, data, size);
-    SYCL_CHECK(
-        CHECK_TRY_ERROR((*stream).memcpy((char *)tensor->data + offset, host_buf, size)
-                             .wait()));
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy((char *) tensor->data + offset, host_buf, size).wait()));
    free(host_buf);
+#else
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy((char *) tensor->data + offset, data, size).wait()));
+#endif
 }
 catch (sycl::exception const &exc) {
  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -3027,7 +3028,7 @@ static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_ten
    return !g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
            ctx.opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
            dst->op == GGML_OP_MUL_MAT &&   //limit to some supported cases of Q4_0, to do for more cases.
-            dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
+            dst->src[1]->ne[1]==1 && dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
 }

 static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * /* src1 */,
@@ -3150,8 +3151,6 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, convert_src1_to_q8_1);
    } else {
        constexpr bool convert_src1_to_q8_1 = false;
-        // MUL_MAT_SYCL supports reorder
-        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MUL_MAT_SYCL);
        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1);
    }
    GGML_SYCL_DEBUG("call %s done\n", __func__);
@@ -3741,7 +3740,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
    GGML_ASSERT(buf->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
    const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
    SYCL_CHECK(CHECK_TRY_ERROR((stream)->memcpy(
-        data, (const char *)tensor->data + offset, size).wait()));
+        data, (const char *)tensor->data + offset, size)));
 }
 catch (sycl::exception const &exc) {
  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -3761,7 +3760,7 @@ static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
        */
        const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
        SYCL_CHECK(CHECK_TRY_ERROR((stream)->memcpy(
-            dst->data, src->data, ggml_nbytes(dst)).wait()));
+            dst->data, src->data, ggml_nbytes(dst))));
        return true;
    }

@@ -3810,11 +3809,43 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
    }
 }

+#ifdef GGML_SYCL_GRAPH
+static bool check_graph_compatibility(ggml_cgraph * cgraph) {
+    if (ggml_sycl_info().device_count > 1) {
+        // A sycl_ex::command_graph object can only be created for a single device
+        GGML_LOG_INFO("%s: disabling SYCL graphs due to multiple devices\n", __func__);
+        return false;
+    }
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        const ggml_op node_op = cgraph->nodes[i]->op;
+        switch (node_op) {
+            default:
+                break;
+            case GGML_OP_CONCAT:
+                // ggml_sycl_op_concat() does a blocking host wait after memcpy operations,
+                // but wait() can't be called on the events returned by a queue recording
+                // to a graph.
+                [[fallthrough]];
+            case GGML_OP_MUL_MAT_ID:
+                // ggml_sycl_mul_mat_id() does a blocking host wait on the sycl queue after
+                // submitting a memcpy operation, but wait() can't be called on a queue that
+                // is recording to a graph.
+                GGML_LOG_INFO("%s: disabling SYCL graphs due to unsupported node type %s\n", __func__,
+                              ggml_op_name(node_op));
+                return false;
+        }
+    }
+    return true;
+}
+#endif
+
 static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
    auto * sycl_ctx = static_cast<ggml_backend_sycl_context *>(backend->context);

 #ifdef GGML_SYCL_GRAPH
-    if (!g_ggml_sycl_disable_graph) {
+    bool use_sycl_graph = !g_ggml_sycl_disable_graph && check_graph_compatibility(cgraph);
+    if (use_sycl_graph) {
        const bool graph_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_limited_graph);
        if (!graph_support) {
            GGML_SYCL_DEBUG("[SYCL-GRAPH] can not use graphs on device:%d\n", sycl_ctx->device);
@@ -54,6 +54,11 @@ if (Vulkan_FOUND)
        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
    )

+    set(VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS "")
+    if (CMAKE_BUILD_TYPE AND CMAKE_BUILD_TYPE MATCHES "Debug|Release|MinSizeRel|RelWithDebInfo")
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS --config=${CMAKE_BUILD_TYPE})
+    endif()
+
    # Test all shader extensions
    test_shader_extension_support(
        "GL_KHR_cooperative_matrix"
@@ -149,7 +154,7 @@ if (Vulkan_FOUND)
        vulkan-shaders-gen
        SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-        BUILD_COMMAND ${CMAKE_COMMAND} --build .
+        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
        INSTALL_DIR ${CMAKE_BINARY_DIR}
    )
@@ -2031,25 +2031,25 @@ static void ggml_vk_load_shaders(vk_device& device) {
            CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
        }
 #endif
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_S].f16acc,   matmul_iq1_s_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_M].f16acc,   matmul_iq1_m_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_XS].f16acc,  matmul_iq4_xs_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0], matmul_q4_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1], matmul_q4_1_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0], matmul_q5_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_1], matmul_q5_1_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0], matmul_q8_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K], matmul_q2_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K], matmul_q3_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K], matmul_q4_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K], matmul_q5_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K], matmul_q6_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_S],   matmul_iq1_s_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_M],   matmul_iq1_m_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XS],  matmul_iq2_xs_f16,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_S],   matmul_iq2_s_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_S],   matmul_iq3_s_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f16,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f16,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)

        CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
 #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
@@ -2117,47 +2117,47 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #endif

        if (device->coopmat_acc_f16_support) {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0], matmul_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1], matmul_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0], matmul_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );

-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc,   matmul_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc,   matmul_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f16acc,  matmul_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K], matmul_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S],   matmul_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M],   matmul_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS],  matmul_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S],   matmul_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S],   matmul_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
        } else {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );

-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc,   matmul_iq1_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc,   matmul_iq1_m_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f16acc,  matmul_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f32acc,   matmul_iq1_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f32acc,   matmul_iq1_m_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f32acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc,  matmul_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc,   matmul_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f32acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc,  matmul_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
        }

        CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
@@ -2232,13 +2232,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
        if (device->mul_mat ## ID ## _s[TYPE]) \
            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \

-#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
-        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
-        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        if (device->mul_mat ## ID ## _l[TYPE]) { \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->l, #NAMELC "_f16acc_l", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->l, #NAMELC        "_l", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        } \
+        if (device->mul_mat ## ID ## _m[TYPE]) { \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->m, #NAMELC "_f16acc_m", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->m, #NAMELC        "_m", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        } \
+        if (device->mul_mat ## ID ## _s[TYPE]) { \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->s, #NAMELC "_f16acc_s", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->s, #NAMELC        "_s", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        } \

        // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -2252,34 +2258,34 @@ static void ggml_vk_load_shaders(vk_device& device) {

        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );

-        CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0], matmul_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1], matmul_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0], matmul_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );

-        CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc,   matmul_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc,   matmul_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f16acc,  matmul_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K], matmul_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S],   matmul_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M],   matmul_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS],  matmul_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S],   matmul_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S],   matmul_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );

 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
        if (device->integer_dot_product) {
-            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1], matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0], matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1], matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
        }
 #endif

@@ -2328,13 +2334,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
        if (device->mul_mat ## ID ## _s[TYPE]) \
            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \

-#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC "_l", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC "_m", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC "_s", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \

        CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
        CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
@@ -2366,11 +2372,11 @@ static void ggml_vk_load_shaders(vk_device& device) {

 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
        if (device->integer_dot_product) {
-            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
+            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
        }
 #endif

@@ -2798,23 +2804,29 @@ static vk_device ggml_vk_get_device(size_t idx) {
                pipeline_robustness = true;
            } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
                device->subgroup_size_control = true;
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
            } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
                       !getenv("GGML_VK_DISABLE_COOPMAT")) {
                device->coopmat_support = true;
                device->coopmat_m = 0;
                device->coopmat_n = 0;
                device->coopmat_k = 0;
+#endif
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
            } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
                       !getenv("GGML_VK_DISABLE_COOPMAT2")) {
                coopmat2_support = true;
+#endif
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
            } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 &&
                       !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
                device->integer_dot_product = true;
 #endif
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
            } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
                       !getenv("GGML_VK_DISABLE_BFLOAT16")) {
                bfloat16_support = true;
+#endif
            }
        }

@@ -3711,7 +3723,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
 }

 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
-    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
+    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ", " << prec << ")");
    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
        return ctx->device->pipeline_matmul_f32;
    }
@@ -3739,7 +3751,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte

    // MMQ
    if (src1_type == GGML_TYPE_Q8_1) {
-        vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f16acc;
+        vk_matmul_pipeline pipelines = (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;

        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
            return nullptr;
@@ -3779,9 +3791,12 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte

    if (ctx->device->coopmat2) {
        assert(src1_type == GGML_TYPE_F16);
-        return ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc;
+        return prec == GGML_PREC_DEFAULT ? ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f32acc;
    }
-    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
+    if (ctx->device->coopmat_support) {
+        return (ctx->device->fp16 && ctx->device->coopmat_acc_f16_support && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
+    }
+    return (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }

 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
@@ -4504,6 +4519,8 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
        return aligned ? mmp->a_m : mmp->m;
    }
    return aligned ? mmp->a_l : mmp->l;
+
+    GGML_UNUSED(src1_type);
 }

 static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type src0_type, ggml_type src1_type) {
@@ -4659,6 +4676,19 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
        }
    }

+    if (src->type == to) {
+        // Copy two or four bytes at a time, depending on block size.
+        // For quantized types, we scale by block size/type size. But
+        // this path is also used for bf16->bf16 for example, where the
+        // type size must be exactly 2 or 4.
+        GGML_ASSERT(ggml_is_quantized(to) || ggml_type_size(src->type) == 2 || ggml_type_size(src->type) == 4);
+        if ((ggml_type_size(src->type) % 4) == 0) {
+            return ctx->device->pipeline_contig_cpy_f32_f32;
+        } else {
+            return ctx->device->pipeline_contig_cpy_f16_f16;
+        }
+    }
+
    std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
    GGML_ABORT("fatal error");
 }
@@ -5872,10 +5902,17 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
    vk_pipeline *pipelines;
    bool small_rows = N <= get_fa_num_small_rows(path);

+    // coopmat1 does not actually support "small rows" (it needs 16 rows).
+    // So use scalar instead.
    if (small_rows && path == FA_COOPMAT1) {
        path = FA_SCALAR;
    }

+    // scalar is faster than coopmat2 when N==1
+    if (N == 1 && path == FA_COOPMAT2) {
+        path = FA_SCALAR;
+    }
+
    bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;

    switch (path) {
@@ -6713,7 +6750,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
    case GGML_OP_UNARY:
    case GGML_OP_CONV_2D_DW:
        {
-            const uint32_t ne = ggml_nelements(dst);
+            uint32_t ne = ggml_nelements(dst);
+            if (op == GGML_OP_CPY && ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+                // Convert from number of logical elements to 2- or 4-byte units.
+                ne /= ggml_blck_size(src0->type);
+                if ((ggml_type_size(src0->type) % 4) == 0) {
+                    ne *= ggml_type_size(src0->type) / 4;
+                } else {
+                    ne *= ggml_type_size(src0->type) / 2;
+                }
+            }
            if (ne > 262144) {
                elements = { 512, 512, CEIL_DIV(ne, 262144) };
            } else if (ne > 512) {
@@ -7263,8 +7309,19 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
    const uint32_t src0_type_size = ggml_type_size(src0->type);
    const uint32_t dst_type_size = ggml_type_size(dst->type);

+    uint32_t ne = (uint32_t)ggml_nelements(src0);
+    if (ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+        // Convert from number of logical elements to 2- or 4-byte units.
+        ne /= ggml_blck_size(src0->type);
+        if ((ggml_type_size(src0->type) % 4) == 0) {
+            ne *= ggml_type_size(src0->type) / 4;
+        } else {
+            ne *= ggml_type_size(src0->type) / 2;
+        }
+    }
+
    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
-        (uint32_t)ggml_nelements(src0),
+        ne,
        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
        0,
@@ -9246,8 +9303,7 @@ static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_
    try {
        ptr = ggml_vk_host_malloc(vk_instance.devices[0], size);
    } catch (vk::SystemError& e) {
-        std::cerr << "ggml_vulkan: Failed to allocate pinned memory." << std::endl;
-        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        GGML_LOG_WARN("ggml_vulkan: Failed to allocate pinned memory (%s)\n", e.what());
        // fallback to cpu buffer
        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
    }
@@ -9849,6 +9905,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
                    return true;
                }
+
+                // We can handle copying from a type to the same type if it's
+                // contiguous (memcpy). We use f16 or f32 shaders to do the copy,
+                // so the type/block size must be a multiple of 4.
+                if (src0_type == src1_type &&
+                    ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op) &&
+                    (ggml_type_size(src0_type) % 2) == 0) {
+                    return true;
+                }
                return false;
            } break;
        case GGML_OP_REPEAT:
@@ -10254,7 +10319,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
    } else if (tensor->op == GGML_OP_CONCAT) {
        tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params);
    } else if (tensor->op == GGML_OP_UPSCALE) {
-        tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->op_params[0], tensor->op_params[1], (ggml_scale_mode) tensor->op_params[0]);
+        tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
    } else if (tensor->op == GGML_OP_SCALE) {
        const float * params = (const float *)tensor->op_params;
        tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]);
@@ -10543,7 +10608,8 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
                        ggml_vk_print_graph_origin(tensor, done);
                        GGML_ABORT("fatal error");
                    }
-                    if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) {
+                    const double denom = std::fabs(correct) > 1.0f ? (std::fabs(correct) > 1e-8 ? std::fabs(correct) : 1e-8) : 1.0f;
+                    if (first_error[0] == -1 && std::fabs(correct - result) / denom > 0.5) {
                        first_error[0] = i0;
                        first_error[1] = i1;
                        first_error[2] = i2;
@@ -10555,7 +10621,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
                    // Special case, value is infinite, avoid NaN result in avg_err
                    // NaN also appears in results, if both are nan error is 0
                    if (!std::isinf(correct) && !std::isinf(result) && !std::isnan(correct) && !std::isnan(result)) {
-                        avg_err += std::fabs(correct - result);
+                        avg_err += std::fabs(correct - result) / denom;
                    }
                    counter++;
                }
@@ -10590,7 +10656,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
        ggml_vk_print_graph_origin(tensor, done);
    }

-    if (avg_err > 0.05 || std::isnan(avg_err)) {
+    if (avg_err > 0.5 || std::isnan(avg_err)) {
        std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
        if (src0 != nullptr) {
@@ -1,6 +1,6 @@
 #version 450

-#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require

 #include "dequant_head.comp"

@@ -9,60 +9,13 @@
 #extension GL_KHR_shader_subgroup_shuffle : enable

 #include "types.comp"
+#include "flash_attn_base.comp"

-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
-layout (constant_id = 1) const uint32_t Br = 1;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-
-layout (constant_id = 5) const uint32_t D_split = 16;
 const uint32_t D_per_thread = D / D_split;

 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;

-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;

 layout (binding = 0) readonly buffer Q {float data_q[];};
 layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
@@ -71,39 +24,6 @@ layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
 layout (binding = 2) readonly buffer V {float16_t data_v[];};
 layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
 layout (binding = 3) readonly buffer M {float16_t data_m[];};
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#if defined(A_TYPE_PACKED16)
-#define BINDING_IDX_K 0
-#define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
-#endif
-
-#if defined(DATA_A_Q4_0)
-#define BLOCK_BYTE_SIZE 18
-
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-#define BLOCK_BYTE_SIZE 34
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
-}
-#endif
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))

 // Store the output when doing grouped query attention.
 // Rows index by Q's dimension 2, and the first N rows are valid.
@@ -114,27 +34,6 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
    return elem;
 }

-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
 shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];

@@ -146,58 +45,12 @@ void main() {
    init_iq_shmem(gl_WorkGroupSize);
 #endif

-    const uint32_t tid = gl_LocalInvocationIndex;
-    const uint32_t N = p.N;
-    const uint32_t KV = p.KV;
+    init_indices();

+    const uint32_t tid = gl_LocalInvocationIndex;
    const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
    const uint32_t col_tid = gl_LocalInvocationIndex / D_split;

-    uint32_t i = gl_WorkGroupID.x;
-    uint32_t split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    const uint32_t Tr = CEIL_DIV(N, Br);
-
-    const uint32_t start_j = split_k_index * p.split_kv / Bc;
-    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    const uint32_t iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    const uint32_t rk2 = p.neq2/p.nek2;
-    const uint32_t rk3 = p.neq3/p.nek3;
-
-    const uint32_t rv2 = p.neq2/p.nev2;
-    const uint32_t rv3 = p.neq3/p.nev3;
-
-    // k indices
-    const uint32_t ik3 = iq3 / rk3;
-    const uint32_t ik2 = iq2 / rk2;
-
-    // v indices
-    const uint32_t iv3 = iq3 / rv3;
-    const uint32_t iv2 = iq2 / rv2;
-
-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    uint32_t k_stride = p.nb11;
-    uint32_t v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
-
    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;

    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
@@ -0,0 +1,162 @@
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
+layout (constant_id = 1) const uint32_t Br = 1;
+layout (constant_id = 2) const uint32_t Bc = 32;
+layout (constant_id = 3) const uint32_t D = 32;
+layout (constant_id = 4) const uint32_t Clamp = 0;
+layout (constant_id = 5) const uint32_t D_split = 16;
+
+
+layout (push_constant) uniform parameter {
+    uint32_t N;
+    uint32_t KV;
+
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+
+    uint32_t neq2;
+    uint32_t neq3;
+    uint32_t nek2;
+    uint32_t nek3;
+    uint32_t nev2;
+    uint32_t nev3;
+    uint32_t nem1;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t nb21;
+    uint32_t nb22;
+    uint32_t nb23;
+    uint32_t nb31;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t mask;
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+
+    uint32_t gqa_ratio;
+    uint32_t split_kv;
+    uint32_t k_num;
+} p;
+
+layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+
+#if defined(A_TYPE_PACKED16)
+#define BINDING_IDX_K 0
+#define BINDING_IDX_V 1
+layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
+#endif
+
+#if defined(DATA_A_Q4_0)
+#define BLOCK_BYTE_SIZE 18
+
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+    uint shift = (iqs & 0x10) >> 2;
+    vui_lo >>= shift;
+    vui_hi >>= shift;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+#define BLOCK_BYTE_SIZE 34
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+
+// Store column zero. This is used to save per-row m and L values for split_k.
+ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    if (r < N && c == 0) {
+        uint32_t offset = iq2 + r;
+        data_o[o_offset + offset] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Load the slope matrix, indexed by Q's dimension 2.
+ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+
+    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
+}
+
+uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
+         iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
+         q_stride, k_stride, v_stride, m_stride;
+
+void init_indices()
+{
+    N = p.N;
+    KV = p.KV;
+
+    i = gl_WorkGroupID.x;
+    split_k_index = 0;
+
+    if (p.k_num > 1) {
+        i = 0;
+        split_k_index = gl_WorkGroupID.x;
+    }
+
+    Tr = CEIL_DIV(N, Br);
+
+    start_j = split_k_index * p.split_kv / Bc;
+    end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
+
+    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
+    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
+    iq2 = gl_WorkGroupID.y * p.gqa_ratio;
+    iq3 = gl_WorkGroupID.z;
+
+    // broadcast factors
+    rk2 = p.neq2/p.nek2;
+    rk3 = p.neq3/p.nek3;
+
+    rv2 = p.neq2/p.nev2;
+    rv3 = p.neq3/p.nev3;
+
+    // k indices
+    ik3 = iq3 / rk3;
+    ik2 = iq2 / rk2;
+
+    // v indices
+    iv3 = iq3 / rv3;
+    iv2 = iq2 / rv2;
+
+    // nb?1 are already divided by the type size and are in units of elements.
+    // When using grouped query attention, Q is indexed by iq2, so the stride
+    // should be nb02 (which is in bytes).
+    q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
+    k_stride = p.nb11;
+    v_stride = p.nb21;
+    // When using grouped query attention, all rows use the same mask (stride 0).
+    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
+    // that prevents the compiler from folding the "&" through the select
+    // and breaking the alignment detection.
+    m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
+}
@@ -11,14 +11,7 @@
 #extension GL_KHR_cooperative_matrix : enable

 #include "types.comp"
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 1) const uint32_t Br = 1;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-
-layout (constant_id = 5) const uint32_t D_split = 16;
+#include "flash_attn_base.comp"

 const uint32_t D_per_thread = D / D_split;
 const uint32_t row_split = 4;
@@ -26,46 +19,6 @@ const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;

-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;

 layout (binding = 0) readonly buffer Q {float data_q[];};
 layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
@@ -74,39 +27,6 @@ layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
 layout (binding = 2) readonly buffer V {float16_t data_v[];};
 layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
 layout (binding = 3) readonly buffer M {float16_t data_m[];};
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#if defined(A_TYPE_PACKED16)
-#define BINDING_IDX_K 0
-#define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
-#endif
-
-#if defined(DATA_A_Q4_0)
-#define BLOCK_BYTE_SIZE 18
-
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-#define BLOCK_BYTE_SIZE 34
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
-}
-#endif
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))

 // Store the output when doing grouped query attention.
 // Rows index by Q's dimension 2, and the first N rows are valid.
@@ -117,27 +37,6 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
    return elem;
 }

-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
 // These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd
 const uint32_t MatBr = 16;
 const uint32_t MatBc = 16;
@@ -162,9 +61,9 @@ void main() {
    init_iq_shmem(gl_WorkGroupSize);
 #endif

+    init_indices();
+
    const uint32_t tid = gl_LocalInvocationIndex;
-    const uint32_t N = p.N;
-    const uint32_t KV = p.KV;

    const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
    const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
@@ -173,51 +72,6 @@ void main() {

 #define tile_row(r) (row_tid * rows_per_thread + (r))

-    uint32_t i = gl_WorkGroupID.x;
-    uint32_t split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    const uint32_t Tr = CEIL_DIV(N, Br);
-
-    const uint32_t start_j = split_k_index * p.split_kv / Bc;
-    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    const uint32_t iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    const uint32_t rk2 = p.neq2/p.nek2;
-    const uint32_t rk3 = p.neq3/p.nek3;
-
-    const uint32_t rv2 = p.neq2/p.nev2;
-    const uint32_t rv3 = p.neq3/p.nev3;
-
-    // k indices
-    const uint32_t ik3 = iq3 / rk3;
-    const uint32_t ik2 = iq2 / rk2;
-
-    // v indices
-    const uint32_t iv3 = iq3 / rv3;
-    const uint32_t iv2 = iq2 / rv2;
-
-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    uint32_t k_stride = p.nb11;
-    uint32_t v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
-
    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;

    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
@@ -18,62 +18,12 @@

 #include "types.comp"
 #include "dequant_funcs_cm2.comp"
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 1) const uint32_t Br = 32;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = gl_CooperativeMatrixClampModeConstantNV;
-
-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;
+#include "flash_attn_base.comp"

 layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
 layout (binding = 1) readonly buffer K {uint8_t data_k[];};
 layout (binding = 2) readonly buffer V {uint8_t data_v[];};
 layout (binding = 3) readonly buffer M {uint8_t data_m[];};
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))

 ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
    return max(x, y);
@@ -118,67 +68,12 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
    return elem;
 }

-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
    init_iq_shmem(gl_WorkGroupSize);
 #endif

-    const uint32_t N = p.N;
-    const uint32_t KV = p.KV;
-
-    uint32_t i = gl_WorkGroupID.x;
-    uint32_t split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    const uint32_t Tr = CEIL_DIV(N, Br);
-
-    const uint32_t start_j = split_k_index * p.split_kv / Bc;
-    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    const uint32_t iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    const uint32_t rk2 = p.neq2/p.nek2;
-    const uint32_t rk3 = p.neq3/p.nek3;
-
-    const uint32_t rv2 = p.neq2/p.nev2;
-    const uint32_t rv3 = p.neq3/p.nev3;
-
-    // k indices
-    const uint32_t ik3 = iq3 / rk3;
-    const uint32_t ik2 = iq2 / rk2;
-
-    // v indices
-    const uint32_t iv3 = iq3 / rv3;
-    const uint32_t iv2 = iq2 / rv2;
+    init_indices();

    tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
    tensorLayoutNV<2, Clamp> tensorLayoutK = createTensorLayoutNV(2, Clamp);
@@ -195,17 +90,6 @@ void main() {
    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);

-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    uint32_t k_stride = p.nb11;
-    uint32_t v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
    // hint to the compiler that strides are aligned for the aligned variant of the shader
    if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
    {
@@ -7,7 +7,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
 #if defined(DATA_A_IQ1_M)
-#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 #endif

 #if defined(DATA_A_BF16) && defined(COOPMAT)
@@ -64,12 +64,17 @@
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];

-#if (defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)) && \
-    (!defined(TARGET_OS_TV) && !defined(TARGET_OS_WATCH))
+#if defined(__linux__) || \
+    defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
+    (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH)
+
 #include <unistd.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
+#if defined(__linux__)
+#include <sys/prctl.h>
+#endif

 #if defined(__ANDROID__)
 #include <unwind.h>
@@ -133,10 +138,36 @@ static void ggml_print_backtrace(void) {
    if (GGML_NO_BACKTRACE) {
        return;
    }
-    char attach[32];
-    snprintf(attach, sizeof(attach), "attach %d", getpid());
-    int pid = fork();
-    if (pid == 0) {
+#if defined(__linux__)
+    FILE * f = fopen("/proc/self/status", "r");
+    size_t size = 0;
+    char * line = NULL;
+    ssize_t length = 0;
+    while ((length = getline(&line, &size, f)) > 0) {
+        if (!strncmp(line, "TracerPid:", sizeof("TracerPid:") - 1) &&
+            (length != sizeof("TracerPid:\t0\n") - 1 || line[length - 2] != '0')) {
+            // Already being debugged, and the breakpoint is the later abort()
+            free(line);
+            fclose(f);
+            return;
+        }
+    }
+    free(line);
+    fclose(f);
+    int lock[2] = { -1, -1 };
+    (void) !pipe(lock); // Don't start gdb until after PR_SET_PTRACER
+#endif
+    const int parent_pid = getpid();
+    const int child_pid = fork();
+    if (child_pid < 0) { // error
+        return;
+    } else if (child_pid == 0) { // child
+        char attach[32];
+        snprintf(attach, sizeof(attach), "attach %d", parent_pid);
+#if defined(__linux__)
+        close(lock[1]);
+        (void) !read(lock[0], lock, 1);
+#endif
        // try gdb
        execlp("gdb", "gdb", "--batch",
            "-ex", "set style enabled on",
@@ -149,18 +180,18 @@ static void ggml_print_backtrace(void) {
        execlp("lldb", "lldb", "--batch",
            "-o", "bt",
            "-o", "quit",
-            "-p", attach,
+            "-p", &attach[sizeof("attach ") - 1],
            (char *) NULL);
-        exit(EXIT_FAILURE);
-    } else {
-        int wstatus;
-        waitpid(pid, &wstatus, 0);
-        if (WIFEXITED(wstatus)) {
-            if (WEXITSTATUS(wstatus) == EXIT_FAILURE) {
-                // gdb failed, fallback to backtrace_symbols
-                ggml_print_backtrace_symbols();
-            }
-        }
+        // gdb failed, fallback to backtrace_symbols
+        ggml_print_backtrace_symbols();
+        _Exit(0);
+    } else { // parent
+#if defined(__linux__)
+        prctl(PR_SET_PTRACER, child_pid);
+        close(lock[1]);
+        close(lock[0]);
+#endif
+        waitpid(child_pid, NULL, 0);
    }
 }
 #else
@@ -1068,9 +1099,10 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
    "HARDSWISH",
    "HARDSIGMOID",
    "EXP",
+    "GELU_ERF",
 };

-static_assert(GGML_UNARY_OP_COUNT == 14, "GGML_UNARY_OP_COUNT != 14");
+static_assert(GGML_UNARY_OP_COUNT == 15, "GGML_UNARY_OP_COUNT != 15");


 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -2470,6 +2502,20 @@ struct ggml_tensor * ggml_gelu_inplace(
    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU);
 }

+// ggml_gelu_erf
+
+struct ggml_tensor * ggml_gelu_erf(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_GELU_ERF);
+}
+
+struct ggml_tensor * ggml_gelu_erf_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU_ERF);
+}
+
 // ggml_gelu_quick

 struct ggml_tensor * ggml_gelu_quick(
@@ -219,10 +219,13 @@ class Keys:
        TYPE       = "adapter.type"
        LORA_ALPHA = "adapter.lora.alpha"

-    class ClipVision:
+    class Clip:
        PROJECTOR_TYPE      = "clip.projector_type"
        HAS_VISION_ENCODER  = "clip.has_vision_encoder"
+        HAS_AUDIO_ENCODER   = "clip.has_audio_encoder"
        HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
+
+    class ClipVision:
        IMAGE_SIZE          = "clip.vision.image_size"
        PATCH_SIZE          = "clip.vision.patch_size"
        EMBEDDING_LENGTH    = "clip.vision.embedding_length"
@@ -243,19 +246,33 @@ class Keys:
        class Projector:
            SCALE_FACTOR    = "clip.vision.projector.scale_factor"

+    class ClipAudio:
+        NUM_MEL_BINS        = "clip.audio.num_mel_bins"
+        EMBEDDING_LENGTH    = "clip.audio.embedding_length"
+        FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
+        PROJECTION_DIM      = "clip.audio.projection_dim"
+        BLOCK_COUNT         = "clip.audio.block_count"
+
+        class Attention:
+            HEAD_COUNT      = "clip.audio.attention.head_count"
+            LAYERNORM_EPS   = "clip.audio.attention.layer_norm_epsilon"
+
+        class Projector:
+            STACK_FACTOR    = "clip.audio.projector.stack_factor"
+
 #
 # recommended mapping of model tensor names for storage in gguf
 #


 class GGUFType:
-    MODEL       = "model"
-    ADAPTER     = "adapter"
-    CLIP_VISION = "clip-vision"
+    MODEL   = "model"
+    ADAPTER = "adapter"
+    MMPROJ  = "mmproj" # dummy, unused for now


 class MODEL_ARCH(IntEnum):
-    CLIP_VISION      = auto() # dummy arch for clip.cpp
+    MMPROJ           = auto() # dummy arch for clip.cpp
    LLAMA            = auto()
    LLAMA4           = auto()
    DECI             = auto()
@@ -482,14 +499,15 @@ class MODEL_TENSOR(IntEnum):
    V_ENC_EMBD_CLS       = auto()
    V_ENC_EMBD_PATCH     = auto()
    V_ENC_EMBD_POS       = auto()
+    V_ENC_INPUT_NORM     = auto()
    V_ENC_ATTN_Q         = auto()
    V_ENC_ATTN_Q_NORM    = auto()
    V_ENC_ATTN_K         = auto()
    V_ENC_ATTN_K_NORM    = auto()
    V_ENC_ATTN_V         = auto()
-    V_ENC_INPUT_NORM     = auto()
-    V_ENC_OUTPUT         = auto()
-    V_ENC_OUTPUT_NORM    = auto()
+    V_ENC_ATTN_O         = auto()
+    V_ENC_ATTN_O_NORM    = auto()
+    V_ENC_POST_ATTN_NORM = auto()
    V_ENC_FFN_UP         = auto()
    V_ENC_FFN_GATE       = auto()
    V_ENC_FFN_DOWN       = auto()
@@ -513,10 +531,27 @@ class MODEL_TENSOR(IntEnum):
    V_RESMPL_QUERY       = auto() # minicpmv
    V_TOK_EMBD_IMG_BREAK = auto() # pixtral
    V_MM_PATCH_MERGER    = auto() # mistral small 3.1
+    # audio (mtmd)
+    A_ENC_EMBD_POS       = auto()
+    A_ENC_CONV1D         = auto()
+    A_PRE_NORM           = auto()
+    A_POST_NORM          = auto()
+    A_ENC_ATTN_Q         = auto()
+    A_ENC_ATTN_K         = auto()
+    A_ENC_ATTN_V         = auto()
+    A_ENC_INPUT_NORM     = auto()
+    A_ENC_OUTPUT         = auto()
+    A_ENC_OUTPUT_NORM    = auto()
+    A_ENC_FFN_UP         = auto()
+    A_ENC_FFN_GATE       = auto()
+    A_ENC_FFN_DOWN       = auto()
+    A_MMPROJ             = auto()
+    A_MM_NORM_PRE        = auto()
+    A_MM_NORM_MID        = auto()


 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
-    MODEL_ARCH.CLIP_VISION:      "clip", # dummy arch for clip.cpp
+    MODEL_ARCH.MMPROJ:           "clip", # dummy arch for clip.cpp
    MODEL_ARCH.LLAMA:            "llama",
    MODEL_ARCH.LLAMA4:           "llama4",
    MODEL_ARCH.DECI:             "deci",
@@ -749,8 +784,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.V_ENC_ATTN_K_NORM:         "v.blk.{bid}.attn_k_norm",
    MODEL_TENSOR.V_ENC_ATTN_V:              "v.blk.{bid}.attn_v",
    MODEL_TENSOR.V_ENC_INPUT_NORM:          "v.blk.{bid}.ln1",
-    MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
-    MODEL_TENSOR.V_ENC_OUTPUT_NORM:         "v.blk.{bid}.ln2",
+    MODEL_TENSOR.V_ENC_ATTN_O:              "v.blk.{bid}.attn_out",
+    MODEL_TENSOR.V_ENC_ATTN_O_NORM:         "v.blk.{bid}.attn_out_norm",
+    MODEL_TENSOR.V_ENC_POST_ATTN_NORM:      "v.blk.{bid}.ln2",
    MODEL_TENSOR.V_ENC_FFN_UP:              "v.blk.{bid}.ffn_up",
    MODEL_TENSOR.V_ENC_FFN_GATE:            "v.blk.{bid}.ffn_gate",
    MODEL_TENSOR.V_ENC_FFN_DOWN:            "v.blk.{bid}.ffn_down",
@@ -774,10 +810,27 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
    MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
    MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
+    # audio (mtmd)
+    MODEL_TENSOR.A_ENC_EMBD_POS:            "a.position_embd",
+    MODEL_TENSOR.A_ENC_CONV1D:              "a.conv1d.{bid}",
+    MODEL_TENSOR.A_PRE_NORM:                "a.pre_ln",
+    MODEL_TENSOR.A_POST_NORM:               "a.post_ln",
+    MODEL_TENSOR.A_ENC_ATTN_Q:              "a.blk.{bid}.attn_q",
+    MODEL_TENSOR.A_ENC_ATTN_K:              "a.blk.{bid}.attn_k",
+    MODEL_TENSOR.A_ENC_ATTN_V:              "a.blk.{bid}.attn_v",
+    MODEL_TENSOR.A_ENC_INPUT_NORM:          "a.blk.{bid}.ln1",
+    MODEL_TENSOR.A_ENC_OUTPUT:              "a.blk.{bid}.attn_out",
+    MODEL_TENSOR.A_ENC_OUTPUT_NORM:         "a.blk.{bid}.ln2",
+    MODEL_TENSOR.A_ENC_FFN_UP:              "a.blk.{bid}.ffn_up",
+    MODEL_TENSOR.A_ENC_FFN_GATE:            "a.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.A_ENC_FFN_DOWN:            "a.blk.{bid}.ffn_down",
+    MODEL_TENSOR.A_MMPROJ:                  "mm.a.mlp.{bid}",
+    MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
+    MODEL_TENSOR.A_MM_NORM_MID:             "mm.a.norm_mid",
 }

 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
-    MODEL_ARCH.CLIP_VISION: [
+    MODEL_ARCH.MMPROJ: [
        MODEL_TENSOR.V_MMPROJ,
        MODEL_TENSOR.V_MMPROJ_FC,
        MODEL_TENSOR.V_MMPROJ_MLP,
@@ -785,14 +838,15 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.V_ENC_EMBD_CLS,
        MODEL_TENSOR.V_ENC_EMBD_PATCH,
        MODEL_TENSOR.V_ENC_EMBD_POS,
+        MODEL_TENSOR.V_ENC_INPUT_NORM,
        MODEL_TENSOR.V_ENC_ATTN_Q,
        MODEL_TENSOR.V_ENC_ATTN_Q_NORM,
        MODEL_TENSOR.V_ENC_ATTN_K,
        MODEL_TENSOR.V_ENC_ATTN_K_NORM,
        MODEL_TENSOR.V_ENC_ATTN_V,
-        MODEL_TENSOR.V_ENC_INPUT_NORM,
-        MODEL_TENSOR.V_ENC_OUTPUT,
-        MODEL_TENSOR.V_ENC_OUTPUT_NORM,
+        MODEL_TENSOR.V_ENC_ATTN_O,
+        MODEL_TENSOR.V_ENC_ATTN_O_NORM,
+        MODEL_TENSOR.V_ENC_POST_ATTN_NORM,
        MODEL_TENSOR.V_ENC_FFN_UP,
        MODEL_TENSOR.V_ENC_FFN_GATE,
        MODEL_TENSOR.V_ENC_FFN_DOWN,
@@ -816,6 +870,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.V_RESMPL_QUERY,
        MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
        MODEL_TENSOR.V_MM_PATCH_MERGER,
+        # audio
+        MODEL_TENSOR.A_ENC_EMBD_POS,
+        MODEL_TENSOR.A_ENC_CONV1D,
+        MODEL_TENSOR.A_PRE_NORM,
+        MODEL_TENSOR.A_POST_NORM,
+        MODEL_TENSOR.A_ENC_ATTN_Q,
+        MODEL_TENSOR.A_ENC_ATTN_K,
+        MODEL_TENSOR.A_ENC_ATTN_V,
+        MODEL_TENSOR.A_ENC_INPUT_NORM,
+        MODEL_TENSOR.A_ENC_OUTPUT,
+        MODEL_TENSOR.A_ENC_OUTPUT_NORM,
+        MODEL_TENSOR.A_ENC_FFN_UP,
+        MODEL_TENSOR.A_ENC_FFN_GATE,
+        MODEL_TENSOR.A_ENC_FFN_DOWN,
+        MODEL_TENSOR.A_MMPROJ,
+        MODEL_TENSOR.A_MM_NORM_PRE,
+        MODEL_TENSOR.A_MM_NORM_MID,
    ],
    MODEL_ARCH.LLAMA: [
        MODEL_TENSOR.TOKEN_EMBD,
@@ -2180,8 +2251,10 @@ class VisionProjectorType:
    GEMMA3 = "gemma3"
    IDEFICS3 = "idefics3"
    PIXTRAL = "pixtral"
+    LLAMA4 = "llama4"
    QWEN2VL = "qwen2vl_merger"
    QWEN25VL = "qwen2.5vl_merger"
+    ULTRAVOX = "ultravox"
    INTERNVL = "internvl"


@@ -251,7 +251,7 @@ class GGUFReader:
                offs += curr_size
            return offs - orig_offs, aparts, data_idxs, types
        # We can't deal with this one.
-        raise ValueError('Unknown/unhandled field type {gtype}')
+        raise ValueError(f'Unknown/unhandled field type {gtype}')

    def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
        offs = orig_offs
@@ -896,7 +896,7 @@ class GGUFWriter:
    def add_remove_extra_whitespaces(self, value: bool) -> None:
        self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value)

-    def add_precompiled_charsmap(self, charsmap: Sequence[bytes]) -> None:
+    def add_precompiled_charsmap(self, charsmap: bytes) -> None:
        self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)

    def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
@@ -936,12 +936,18 @@ class GGUFWriter:

    # for vision models

+    def add_clip_has_vision_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.Clip.HAS_VISION_ENCODER, value)
+
+    def add_clip_has_audio_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.Clip.HAS_AUDIO_ENCODER, value)
+
+    def add_clip_projector_type(self, value: str) -> None:
+        self.add_string(Keys.Clip.PROJECTOR_TYPE, value)
+
    def add_vision_projection_dim(self, value: int) -> None:
        self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)

-    def add_vision_has_vision_encoder(self, value: bool) -> None:
-        self.add_bool(Keys.ClipVision.HAS_VISION_ENCODER, value)
-
    def add_vision_patch_size(self, value: int) -> None:
        self.add_uint32(Keys.ClipVision.PATCH_SIZE, value)

@@ -957,9 +963,6 @@ class GGUFWriter:
    def add_vision_head_count(self, value: int) -> None:
        self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value)

-    def add_vision_projector_type(self, value: str) -> None:
-        self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
-
    def add_vision_attention_layernorm_eps(self, value: float) -> None:
        self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)

@@ -987,6 +990,32 @@ class GGUFWriter:
    def add_vision_n_wa_pattern(self, value: int) -> None:
        self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)

+    # audio models
+
+    def add_audio_projection_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.PROJECTION_DIM, value)
+
+    def add_audio_embedding_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.EMBEDDING_LENGTH, value)
+
+    def add_audio_feed_forward_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.FEED_FORWARD_LENGTH, value)
+
+    def add_audio_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.BLOCK_COUNT, value)
+
+    def add_audio_head_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.Attention.HEAD_COUNT, value)
+
+    def add_audio_attention_layernorm_eps(self, value: float) -> None:
+        self.add_float32(Keys.ClipAudio.Attention.LAYERNORM_EPS, value)
+
+    def add_audio_num_mel_bins(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)
+
+    def add_audio_stack_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
+
    def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
        pack_prefix = ''
        if not skip_pack_prefix:
@@ -902,10 +902,12 @@ class TensorNameMap:

        MODEL_TENSOR.V_MMPROJ_FC: (
            "model.connector.modality_projection.proj", # SmolVLM
+            "multi_modal_projector.linear_1", # llama 4
        ),

        MODEL_TENSOR.V_MMPROJ_MLP: (
            "model.mm_projector.mlp.mlp.{bid}",
+            "vision_model.vision_adapter.mlp.fc{bid}", # llama 4
            "mlp1.{bid}", # InternVL
        ),

@@ -915,6 +917,7 @@ class TensorNameMap:

        MODEL_TENSOR.V_ENC_EMBD_CLS: (
            "vision_tower.vision_model.embeddings.class_embedding",
+            "vision_model.class_embedding", # llama 4
        ),

        MODEL_TENSOR.V_ENC_EMBD_PATCH: (
@@ -922,6 +925,7 @@ class TensorNameMap:
            "vpm.embeddings.patch_embedding",
            "model.vision_model.embeddings.patch_embedding", # SmolVLM
            "vision_tower.patch_conv", # pixtral
+            "vision_model.patch_embedding.linear", # llama 4
            "visual.patch_embed.proj", # qwen2vl
        ),

@@ -929,12 +933,14 @@ class TensorNameMap:
            "vision_tower.vision_model.embeddings.position_embedding",
            "vpm.embeddings.position_embedding",
            "model.vision_model.embeddings.position_embedding", # SmolVLM
+            "vision_model.positional_embedding_vlm", # llama 4
        ),

        MODEL_TENSOR.V_ENC_ATTN_Q: (
            "vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj",
            "vpm.encoder.layers.{bid}.self_attn.q_proj",
            "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
+            "vision_model.model.layers.{bid}.self_attn.q_proj", # llama4
            "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
            "visual.blocks.{bid}.attn.q", # qwen2vl, generated
        ),
@@ -947,6 +953,7 @@ class TensorNameMap:
            "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
            "vpm.encoder.layers.{bid}.self_attn.k_proj",
            "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
+            "vision_model.model.layers.{bid}.self_attn.k_proj", # llama4
            "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
            "visual.blocks.{bid}.attn.k", # qwen2vl, generated
        ),
@@ -959,6 +966,7 @@ class TensorNameMap:
            "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
            "vpm.encoder.layers.{bid}.self_attn.v_proj",
            "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
+            "vision_model.model.layers.{bid}.self_attn.v_proj", # llama4
            "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
            "visual.blocks.{bid}.attn.v", # qwen2vl, generated
        ),
@@ -969,23 +977,26 @@ class TensorNameMap:
            "vpm.encoder.layers.{bid}.layer_norm1",
            "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
            "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
+            "vision_model.model.layers.{bid}.input_layernorm", # llama4
            "visual.blocks.{bid}.norm1", # qwen2vl
        ),

-        MODEL_TENSOR.V_ENC_OUTPUT: (
+        MODEL_TENSOR.V_ENC_ATTN_O: (
            "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
            "vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL
            "vpm.encoder.layers.{bid}.self_attn.out_proj",
            "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
+            "vision_model.model.layers.{bid}.self_attn.o_proj", # llama4
            "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
            "visual.blocks.{bid}.attn.proj", # qwen2vl
        ),

-        MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
+        MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
            "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
            "vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL
            "vpm.encoder.layers.{bid}.layer_norm2",
            "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
+            "vision_model.model.layers.{bid}.post_attention_layernorm", # llama4
            "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
            "visual.blocks.{bid}.norm2", # qwen2vl
        ),
@@ -995,6 +1006,7 @@ class TensorNameMap:
            "vpm.encoder.layers.{bid}.mlp.fc1",
            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3
            "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+            "vision_model.model.layers.{bid}.mlp.fc1", # llama4
            "visual.blocks.{bid}.mlp.fc1", # qwen2vl
            "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
        ),
@@ -1009,6 +1021,7 @@ class TensorNameMap:
            "vpm.encoder.layers.{bid}.mlp.fc2",
            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3
            "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
+            "vision_model.model.layers.{bid}.mlp.fc2", # llama4
            "visual.blocks.{bid}.mlp.fc2", # qwen2vl
            "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
        ),
@@ -1024,11 +1037,13 @@ class TensorNameMap:
        MODEL_TENSOR.V_PRE_NORM: (
            "vision_tower.vision_model.pre_layrnorm",
            "vision_tower.ln_pre", # pixtral
+            "vision_model.layernorm_pre", # llama4
        ),

        MODEL_TENSOR.V_POST_NORM: (
            "vision_tower.vision_model.post_layernorm",
            "model.vision_model.post_layernorm", # SmolVLM
+            "vision_model.layernorm_post", # llama4
            "visual.merger.ln_q", # qwen2vl
        ),

@@ -1095,6 +1110,68 @@ class TensorNameMap:
        MODEL_TENSOR.V_MM_PATCH_MERGER: (
            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
        ),
+
+        # audio (mtmd)
+
+        MODEL_TENSOR.A_ENC_EMBD_POS: (
+            "audio_tower.embed_positions", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV1D: (
+            "audio_tower.conv{bid}", # ultravox
+        ),
+
+        MODEL_TENSOR.A_PRE_NORM: (),
+
+        MODEL_TENSOR.A_POST_NORM: (
+            "audio_tower.layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_Q: (
+            "audio_tower.layers.{bid}.self_attn.q_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_K: (
+            "audio_tower.layers.{bid}.self_attn.k_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_V: (
+            "audio_tower.layers.{bid}.self_attn.v_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_INPUT_NORM: (
+            "audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_OUTPUT: (
+            "audio_tower.layers.{bid}.self_attn.out_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_OUTPUT_NORM: (
+            "audio_tower.layers.{bid}.final_layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_FFN_UP: (
+            "audio_tower.layers.{bid}.fc1", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_FFN_GATE: (),
+
+        MODEL_TENSOR.A_ENC_FFN_DOWN: (
+            "audio_tower.layers.{bid}.fc2", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MMPROJ: (
+            "audio.multi_modal_projector.linear_{bid}", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MM_NORM_PRE: (
+            "audio.multi_modal_projector.ln_pre", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MM_NORM_MID: (
+            "audio.multi_modal_projector.ln_mid", # ultravox
+        ),
    }

    # architecture-specific block mappings
@@ -361,10 +361,11 @@ extern "C" {

        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
        bool embeddings;  // if true, extract embeddings (together with logits)
-        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-        bool no_perf;     // whether to measure performance timings
-        bool op_offload;  // whether to offload host tensor operations to device
+        bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
+        bool flash_attn;  // use flash attention [EXPERIMENTAL]
+        bool no_perf;     // measure performance timings
+        bool op_offload;  // offload host tensor operations to device
+        bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
    };

    // model quantization parameters
@@ -607,71 +608,14 @@ extern "C" {
    // KV cache
    //

-    // TODO: start using struct llama_kv_cache
-
-    // Information associated with an individual cell in the KV cache view.
-    struct llama_kv_cache_view_cell {
-        // The position for this cell. Takes KV cache shifts into account.
-        // May be negative if the cell is not populated.
-        llama_pos pos;
-    };
-
-    // An updateable view of the KV cache.
-    struct llama_kv_cache_view {
-        // Number of KV cache cells. This will be the same as the context size.
-        int32_t n_cells;
-
-        // Maximum number of sequences that can exist in a cell. It's not an error
-        // if there are more sequences in a cell than this value, however they will
-        // not be visible in the view cells_sequences.
-        int32_t n_seq_max;
-
-        // Number of tokens in the cache. For example, if there are two populated
-        // cells, the first with 1 sequence id in it and the second with 2 sequence
-        // ids then you'll have 3 tokens.
-        int32_t token_count;
-
-        // Number of populated cache cells.
-        int32_t used_cells;
-
-        // Maximum contiguous empty slots in the cache.
-        int32_t max_contiguous;
-
-        // Index to the start of the max_contiguous slot range. Can be negative
-        // when cache is full.
-        int32_t max_contiguous_idx;
-
-        // Information for an individual cell.
-        struct llama_kv_cache_view_cell * cells;
-
-        // The sequences for each cell. There will be n_seq_max items per cell.
-        llama_seq_id * cells_sequences;
-    };
-
-    // Create an empty KV cache view. (use only for debugging purposes)
-    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);
-
-    // Free a KV cache view. (use only for debugging purposes)
-    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
-
-    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
-    // TODO: change signature to llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_context * ctx)
-    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
-
-    ///
-
    // Returns the number of tokens in the KV cache (slow, use only for debug)
    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx);
-
-    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx),
-            "use llama_kv_self_n_tokens instead");
+    DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx),
+               "Use llama_kv_self_seq_pos_max() instead");

    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx);
-
-    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx),
-            "use llama_kv_self_used_cells instead");
+    DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx),
+               "Use llama_kv_self_seq_pos_max() instead");

    // Clear the KV cache - both cell info is erased and KV data is zeroed
    LLAMA_API void llama_kv_self_clear(
@@ -730,10 +674,18 @@ extern "C" {
                       llama_pos   p1,
                             int   d);

+    // Returns the smallest position present in the KV cache for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
    // Returns the largest position present in the KV cache for the specified sequence
+    // Return -1 if the sequence is empty
    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
            struct llama_context * ctx,
-                     llama_seq_id   seq_id);
+                    llama_seq_id   seq_id);

    // Defragment the KV cache
    // This will be applied:
@@ -747,61 +699,6 @@ extern "C" {
    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);

-    DEPRECATED(LLAMA_API void llama_kv_cache_clear(
-            struct llama_context * ctx),
-            "use llama_kv_self_clear instead");
-
-    DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "use llama_kv_self_seq_rm instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id_src,
-                    llama_seq_id   seq_id_dst,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "use llama_kv_self_seq_cp instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "use llama_kv_self_seq_keep instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_seq_add(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                       llama_pos   delta),
-            "use llama_kv_self_seq_add instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_seq_div(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                             int   d),
-            "use llama_kv_self_seq_div instead");
-
-    DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "use llama_kv_self_seq_pos_max instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx),
-            "use llama_kv_self_defrag instead");
-
-    DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx),
-            "use llama_kv_self_can_shift instead");
-
-    DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx),
-            "use llama_kv_self_update instead");
-
-
    //
    // State / sessions
    //
@@ -943,9 +840,12 @@ extern "C" {
    // Requires KV cache.
    // For encode-decoder contexts, processes the batch using the decoder.
    // Positive return values does not mean a fatal error, but rather a warning.
-    //   0 - success
-    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    //    0 - success
+    //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    //    2 - aborted
+    //   -1 - invalid input batch
+    // < -1 - error
    LLAMA_API int32_t llama_decode(
            struct llama_context * ctx,
              struct llama_batch   batch);
@@ -1,3 +1,7 @@
 -r ./requirements-convert_legacy_llama.txt
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch~=2.2.1
+torch~=2.2.1; platform_machine != "s390x"
+
+# torch s390x packages can only be found from nightly builds
+--extra-index-url https://download.pytorch.org/whl/nightly
+torch>=0.0.0.dev0; platform_machine == "s390x"
@@ -1,3 +1,7 @@
 -r ./requirements-convert_legacy_llama.txt
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch~=2.2.1
+torch~=2.2.1; platform_machine != "s390x"
+
+# torch s390x packages can only be found from nightly builds
+--extra-index-url https://download.pytorch.org/whl/nightly
+torch>=0.0.0.dev0; platform_machine == "s390x"
@@ -1,2 +1,4 @@
 -r ./requirements-convert_hf_to_gguf.txt
 --extra-index-url https://download.pytorch.org/whl/cpu
+# torch s390x packages can only be found from nightly builds
+--extra-index-url https://download.pytorch.org/whl/nightly
@@ -1 +1 @@
-9b048bb72b811f50b0c30d9e5c84d6ff9f4bf005
+7c06c10c532a6cda913c17fc56341e8880ae341d
@@ -1,5 +1,6 @@
 #include "llama-batch.h"

+#include <cassert>
 #include <cstring>
 #include <algorithm>

@@ -281,9 +282,10 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
    batch = in_batch;
    GGML_ASSERT(batch.n_tokens > 0);
    if (!batch.pos) {
+        assert(p0 >= 0);
        pos.resize(batch.n_tokens);
        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = i + p0;
+            pos[i] = p0 + i;
        }
        batch.pos = pos.data();
    }
@@ -93,6 +93,7 @@ llama_context::llama_context(
    }

    cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
+
    cparams.op_offload = params.op_offload;

    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
@@ -176,8 +177,9 @@ llama_context::llama_context(
    // init the memory module
    if (!hparams.vocab_only) {
        llama_memory_params params_mem = {
-            /*.type_k =*/ params.type_k,
-            /*.type_v =*/ params.type_v,
+            /*.type_k   =*/ params.type_k,
+            /*.type_v   =*/ params.type_v,
+            /*.swa_full =*/ params.swa_full,
        };

        memory.reset(model.create_memory(params_mem, cparams));
@@ -855,11 +857,17 @@ int llama_context::decode(llama_batch & inp_batch) {
        return -1;
    }

+    if (!inp_batch.pos) {
+        if (inp_batch.seq_id) {
+            LLAMA_LOG_ERROR("%s: pos == NULL, but seq_id != NULL\n", __func__);
+            return -1;
+        }
+    }
+
    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());

    // temporary allocate memory for the input batch if needed
-    // TODO: this is incorrect for multiple sequences because get_pos_max() is the maximum across all sequences
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->get_pos_max() + 1);
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->seq_pos_max(0) + 1);

    const llama_batch & batch = batch_allocr.batch;

@@ -947,8 +955,6 @@ int llama_context::decode(llama_batch & inp_batch) {

        // find KV slot
        if (!kv_self->find_slot(ubatch)) {
-            LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-
            return 1;
        }

@@ -2093,6 +2099,7 @@ llama_context_params llama_context_default_params() {
        /*.flash_attn                  =*/ false,
        /*.no_perf                     =*/ true,
        /*.op_offload                  =*/ true,
+        /*.swa_full                    =*/ true,
    };

    return result;
@@ -2287,65 +2294,51 @@ int32_t llama_apply_adapter_cvec(
    return res ? 0 : -1;
 }

-//
-// kv cache view
-//
-
-llama_kv_cache_view llama_kv_cache_view_init(const llama_context * ctx, int32_t n_seq_max) {
-    const auto * kv = ctx->get_kv_self();
-    if (kv == nullptr) {
-        LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__);
-        return {};
-    }
-
-    return llama_kv_cache_view_init(*kv, n_seq_max);
-}
-
-void llama_kv_cache_view_update(const llama_context * ctx, llama_kv_cache_view * view) {
-    const auto * kv = ctx->get_kv_self();
-    if (kv == nullptr) {
-        LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__);
-        return;
-    }
-
-    llama_kv_cache_view_update(view, kv);
-}
-
 //
 // kv cache
 //

 // deprecated
-int32_t llama_get_kv_cache_token_count(const llama_context * ctx) {
-    return llama_kv_self_n_tokens(ctx);
-}
-
 int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
    const auto * kv = ctx->get_kv_self();
    if (!kv) {
        return 0;
    }

-    return kv->get_n_tokens();
+    int32_t res = 0;
+
+    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
+        const llama_pos p0 = kv->seq_pos_min(s);
+        const llama_pos p1 = kv->seq_pos_max(s);
+
+        if (p0 >= 0) {
+            res += (p1 - p0) + 1;
+        }
+    }
+
+    return res;
 }

 // deprecated
-int32_t llama_get_kv_cache_used_cells(const llama_context * ctx) {
-    return llama_kv_self_used_cells(ctx);
-}
-
+// note: this is the same as above - will be removed anyway, so it's ok
 int32_t llama_kv_self_used_cells(const llama_context * ctx) {
    const auto * kv = ctx->get_kv_self();
    if (!kv) {
        return 0;
    }

-    return kv->get_used_cells();
-}
+    int32_t res = 0;

-// deprecated
-void llama_kv_cache_clear(llama_context * ctx) {
-    llama_kv_self_clear(ctx);
+    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
+        const llama_pos p0 = kv->seq_pos_min(s);
+        const llama_pos p1 = kv->seq_pos_max(s);
+
+        if (p0 >= 0) {
+            res += (p1 - p0) + 1;
+        }
+    }
+
+    return res;
 }

 void llama_kv_self_clear(llama_context * ctx) {
@@ -2357,15 +2350,6 @@ void llama_kv_self_clear(llama_context * ctx) {
    kv->clear();
 }

-// deprecated
-bool llama_kv_cache_seq_rm(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1) {
-    return llama_kv_self_seq_rm(ctx, seq_id, p0, p1);
-}
-
 bool llama_kv_self_seq_rm(
        llama_context * ctx,
         llama_seq_id   seq_id,
@@ -2379,16 +2363,6 @@ bool llama_kv_self_seq_rm(
    return kv->seq_rm(seq_id, p0, p1);
 }

-// deprecated
-void llama_kv_cache_seq_cp(
-        llama_context * ctx,
-         llama_seq_id   seq_id_src,
-         llama_seq_id   seq_id_dst,
-            llama_pos   p0,
-            llama_pos   p1) {
-    llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1);
-}
-
 void llama_kv_self_seq_cp(
        llama_context * ctx,
         llama_seq_id   seq_id_src,
@@ -2403,13 +2377,6 @@ void llama_kv_self_seq_cp(
    kv->seq_cp(seq_id_src, seq_id_dst, p0, p1);
 }

-// deprecated
-void llama_kv_cache_seq_keep(
-        llama_context * ctx,
-         llama_seq_id   seq_id) {
-    llama_kv_self_seq_keep(ctx, seq_id);
-}
-
 void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
    auto * kv = ctx->get_kv_self();
    if (!kv) {
@@ -2419,16 +2386,6 @@ void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
    kv->seq_keep(seq_id);
 }

-// deprecated
-void llama_kv_cache_seq_add(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-            llama_pos   delta) {
-    llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta);
-}
-
 void llama_kv_self_seq_add(
        llama_context * ctx,
         llama_seq_id   seq_id,
@@ -2443,16 +2400,6 @@ void llama_kv_self_seq_add(
    kv->seq_add(seq_id, p0, p1, delta);
 }

-// deprecated
-void llama_kv_cache_seq_div(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-                  int   d) {
-    llama_kv_self_seq_div(ctx, seq_id, p0, p1, d);
-}
-
 void llama_kv_self_seq_div(
        llama_context * ctx,
         llama_seq_id   seq_id,
@@ -2467,25 +2414,24 @@ void llama_kv_self_seq_div(
    kv->seq_div(seq_id, p0, p1, d);
 }

-// deprecated
-llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    return llama_kv_self_seq_pos_max(ctx, seq_id);
+llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
+    const auto * kv = ctx->get_kv_self();
+    if (!kv) {
+        return -1;
+    }
+
+    return kv->seq_pos_min(seq_id);
 }

 llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
    const auto * kv = ctx->get_kv_self();
    if (!kv) {
-        return 0;
+        return -1;
    }

    return kv->seq_pos_max(seq_id);
 }

-// deprecated
-void llama_kv_cache_defrag(llama_context * ctx) {
-    llama_kv_self_defrag(ctx);
-}
-
 void llama_kv_self_defrag(llama_context * ctx) {
    auto * kv = ctx->get_kv_self();
    if (!kv) {
@@ -2496,11 +2442,6 @@ void llama_kv_self_defrag(llama_context * ctx) {
    kv->defrag_sched(-1.0f);
 }

-// deprecated
-bool llama_kv_cache_can_shift(const llama_context * ctx) {
-    return llama_kv_self_can_shift(ctx);
-}
-
 bool llama_kv_self_can_shift(const llama_context * ctx) {
    const auto * kv = ctx->get_kv_self();
    if (!kv) {
@@ -2510,11 +2451,6 @@ bool llama_kv_self_can_shift(const llama_context * ctx) {
    return kv->get_can_shift();
 }

-// deprecated
-void llama_kv_cache_update(llama_context * ctx) {
-    llama_kv_self_update(ctx);
-}
-
 // llama state API

 // deprecated
@@ -2637,7 +2573,21 @@ int32_t llama_encode(
 int32_t llama_decode(
        llama_context * ctx,
          llama_batch   batch) {
-    const int ret = ctx->decode(batch);
+    int ret = ctx->decode(batch);
+
+    // defrag and try again
+    // TODO: distinguish return code when we are sure that even after defrag there is no space available
+    if (ret == 1) {
+        llama_kv_self_defrag(ctx);
+        ret = ctx->decode(batch);
+
+        if (ret == 1) {
+            LLAMA_LOG_WARN("%s: failed to find KV cache slot for batch of size %d\n", __func__, batch.n_tokens);
+
+            return ret;
+        }
+    }
+
    if (ret != 0) {
        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
    }
@@ -9,33 +9,6 @@
 #include <cmath>
 #include <cstring>

-static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
-    // TODO move to hparams if a T5 variant appears that uses a different value
-    const int64_t max_distance = 128;
-
-    if (bidirectional) {
-        n_buckets >>= 1;
-    }
-
-    const int64_t max_exact = n_buckets >> 1;
-
-    int32_t relative_position = x - y;
-    int32_t relative_bucket = 0;
-
-    if (bidirectional) {
-        relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
-    } else {
-        relative_position = -std::min<int32_t>(relative_position, 0);
-    }
-
-    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
-    relative_position_if_large = std::min<int32_t>(relative_position_if_large, n_buckets - 1);
-    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);
-
-    return relative_bucket;
-}
-
 void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
    if (ubatch->token) {
        const int64_t n_tokens = ubatch->n_tokens;
@@ -110,22 +83,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {

 void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
    if (pos_bucket) {
-        const int64_t n_tokens = ubatch->n_tokens;
-
-        GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer));
-        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
-
-        int32_t * data = (int32_t *) pos_bucket->data;
-
-        const int64_t n_kv = kv_self->n;
-
-        for (int h = 0; h < 1; ++h) {
-            for (int j = 0; j < n_tokens; ++j) {
-                for (int i = 0; i < n_kv; ++i) {
-                    data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(kv_self->cells[i].pos, ubatch->pos[j], hparams.n_rel_attn_bkts, false);
-                }
-            }
-        }
+        kv_self->set_input_pos_bucket(pos_bucket, ubatch);
    }
 }

@@ -403,99 +361,18 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 }

 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask || self_kq_mask_swa) {
-        const int64_t n_kv         = kv_self->n;
-        const int64_t n_tokens     = ubatch->n_tokens;
-        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
-        const int64_t n_seqs       = ubatch->n_seqs;
+    if (self_kq_mask) {
+        kv_self->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    }
+}

-        float * data     = nullptr;
-        float * data_swa = nullptr;
+void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
+    if (self_kq_mask) {
+        kv_self->get_kv_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    }

-        if (self_kq_mask) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
-            data = (float *) self_kq_mask->data;
-        }
-
-        if (self_kq_mask_swa) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
-            data_swa = (float *) self_kq_mask_swa->data;
-        }
-
-        // Use only the previous KV cells of the correct sequence for each token of the ubatch.
-        // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
-        // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
-        //   Causal mask:
-        //      xxx-------
-        //      xxxx------
-        //      xxxxx-----
-        //   Non-causal mask:
-        //      xxxxx-----
-        //      xxxxx-----
-        //      xxxxx-----
-        // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
-        for (int h = 0; h < 1; ++h) {
-            for (int s = 0; s < n_seqs; ++s) {
-                const llama_seq_id seq_id = ubatch->seq_id[s][0];
-
-                for (int j = 0; j < n_seq_tokens; ++j) {
-                    const llama_pos pos = ubatch->pos[s*n_seq_tokens + j];
-                    for (int i = 0; i < n_kv; ++i) {
-                        float f;
-                        // mask the token if:
-                        if (!kv_self->cells[i].has_seq_id(seq_id) // not the correct sequence
-                            || (cparams.causal_attn && kv_self->cells[i].pos > pos) // for causal, mask future tokens
-                        ) {
-                            f = -INFINITY;
-                        } else {
-                            if (hparams.use_alibi) {
-                                f = -std::abs(kv_self->cells[i].pos - pos);
-                            } else {
-                                f = 0.0f;
-                            }
-                        }
-
-                        if (data) {
-                            data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
-                        }
-
-                        // may need to cut off old tokens for sliding window
-                        // TODO @ngxson : we are currently re-using the swa logic to store the chunked mask, we should rename SWA to something more generic like "aux mask"
-                        if (data_swa) {
-                            if (hparams.n_attn_chunk) {
-                                llama_pos pos_chunk_start = (pos / hparams.n_attn_chunk) * hparams.n_attn_chunk;
-                                if (kv_self->cells[i].pos < pos_chunk_start || pos < pos_chunk_start) {
-                                    f = -INFINITY;
-                                }
-                            } else {
-                                if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) {
-                                    f = -INFINITY;
-                                }
-                            }
-                            data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
-                        }
-                    }
-                }
-            }
-
-            // mask padded tokens
-            if (data) {
-                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                    for (int j = 0; j < n_kv; ++j) {
-                        data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
-                    }
-                }
-            }
-
-            // mask padded tokens
-            if (data_swa) {
-                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                    for (int j = 0; j < n_kv; ++j) {
-                        data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
-                    }
-                }
-            }
-        }
+    if (self_kq_mask_swa) {
+        kv_self->get_kv_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
    }
 }

@@ -545,7 +422,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
    n_layer          (hparams.n_layer),
    n_rot            (hparams.n_rot),
    n_ctx            (cparams.n_ctx),
-    n_ctx_per_seq    (cparams.n_ctx / cparams.n_seq_max),
    n_head           (hparams.n_head()),
    n_head_kv        (hparams.n_head_kv()),
    n_embd_head_k    (hparams.n_embd_head_k),
@@ -1153,7 +1029,7 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {

    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_self);

-    const auto n_kv = kv_self->n;
+    const auto n_kv = kv_self->get_n();

    auto & cur = inp->pos_bucket;

@@ -1188,16 +1064,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_tensor * kq_b,
         ggml_tensor * kq_mask,
         ggml_tensor * v_mla,
-             bool      v_trans,
             float     kq_scale) const {
-  //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-  //const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+    const bool v_trans = v->nb[1] > v->nb[2];

-  //const int64_t n_head    = hparams.n_head(il);
-  //const int64_t n_head_kv = hparams.n_head_kv(il);
-
-  //const auto & n_embd_head_k = hparams.n_embd_head_k;
-  //const auto & n_embd_head_v = hparams.n_embd_head_v;
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3);
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3);
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3);

    const auto n_tokens = q->ne[1];
    const auto n_head   = q->ne[2];
@@ -1336,17 +1208,11 @@ ggml_tensor * llm_graph_context::build_attn(

    const auto & kq_mask = inp->get_kq_mask();

-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
-    //cb(k, "v", il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = k_cur;
+    ggml_tensor * v = v_cur;

+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1369,22 +1235,16 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()

    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_self);

-    const auto n_kv = kv_self->n;
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");

-    inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-    //cb(inp->self_kq_mask, "KQ_mask", -1);
-    ggml_set_input(inp->self_kq_mask);
+        const auto n_kv = kv_self->get_n();

-    inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        ggml_set_input(inp->self_kq_mask);

-    if (hparams.n_swa_pattern > 1) {
-        GGML_ASSERT(hparams.n_swa > 0);
-
-        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
-        ggml_set_input(inp->self_kq_mask_swa);
-
-        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
    }

    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
@@ -1409,87 +1269,110 @@ ggml_tensor * llm_graph_context::build_attn(
    ggml_build_forward_expand(gf, v_cur);

    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
-    const auto & n_ctx = cparams.n_ctx;
-
-    const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-    const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-    const auto n_tokens = q_cur->ne[2];
-
-    const bool v_trans = !cparams.flash_attn;

    // store to KV cache
    {
-        const auto kv_head = kv_self->head;
-
-        GGML_ASSERT(kv_self->size == n_ctx);
-
-        ggml_tensor * k_cache_view = ggml_view_1d(ctx0, kv_self->k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa)*kv_head);
-        //cb(k_cache_view, "k_cache_view", il);
-
-        // note: storing RoPE-ed version of K in the KV cache
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_cur, k_cache_view));
-
-        v_cur = ggml_reshape_2d(ctx0, v_cur, n_embd_v_gqa, n_tokens);
-
-        ggml_tensor * v_cache_view = nullptr;
-
-        if (!v_trans) {
-            v_cache_view = ggml_view_1d(ctx0, kv_self->v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa)*kv_head);
-        } else {
-            // note: the V cache is transposed when not using flash attention
-            v_cache_view = ggml_view_2d(ctx0, kv_self->v_l[il], n_tokens, n_embd_v_gqa,
-                    (  n_ctx)*ggml_element_size(kv_self->v_l[il]),
-                    (kv_head)*ggml_element_size(kv_self->v_l[il]));
-
-            v_cur = ggml_transpose(ctx0, v_cur);
-        }
-        //cb(v_cache_view, "v_cache_view", il);
-
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_cur, v_cache_view));
+        ggml_build_forward_expand(gf, kv_self->cpy_k(ctx0, k_cur, il));
+        ggml_build_forward_expand(gf, kv_self->cpy_v(ctx0, v_cur, il));
    }

-    const bool is_swa = hparams.is_swa(il);
+    const auto & kq_mask = inp->get_kq_mask();

-    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv_self->get_k(ctx0, il);
+    ggml_tensor * v = kv_self->get_v(ctx0, il);

-    const auto n_kv = kv_self->n;
-
-    const int64_t n_head_kv = hparams.n_head_kv(il);
-
-    const auto & n_embd_head_k = hparams.n_embd_head_k;
-    const auto & n_embd_head_v = hparams.n_embd_head_v;
-
-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k =
-        ggml_view_3d(ctx0, kv_self->k_l[il],
-                n_embd_head_k, n_kv, n_head_kv,
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k),
-                0);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = !v_trans ?
-        ggml_view_3d(ctx0, kv_self->v_l[il],
-                n_embd_head_v, n_kv, n_head_kv,
-                ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
-                ggml_row_size(kv_self->v_l[il]->type, n_embd_head_v),
-                0) :
-        ggml_view_3d(ctx0, kv_self->v_l[il],
-                n_kv, n_embd_head_v, n_head_kv,
-                ggml_element_size(kv_self->v_l[il])*n_ctx,
-                ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
-                0);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
    cb(cur, "kqv_out", il);

    if (wo) {
        cur = build_lora_mm(wo, cur);
    }

+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
+llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
+    const llama_kv_cache_unified_iswa * kv_self = static_cast<const llama_kv_cache_unified_iswa *>(memory);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_self);
+
+    {
+        const auto n_kv = kv_self->get_kv_base()->get_n();
+
+        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        ggml_set_input(inp->self_kq_mask);
+
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+    }
+
+    {
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+
+        const auto n_kv = kv_self->get_kv_swa()->get_n();
+
+        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    }
+
+    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_kv_unified_iswa * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
+            float     kq_scale,
+            int       il) const {
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    ggml_build_forward_expand(gf, q_cur);
+    ggml_build_forward_expand(gf, k_cur);
+    ggml_build_forward_expand(gf, v_cur);
+
+    const bool is_swa = hparams.is_swa(il);
+
+    const llama_kv_cache_unified_iswa * kv_self = static_cast<const llama_kv_cache_unified_iswa *>(memory);
+
+    const auto * kv = is_swa ? kv_self->get_kv_swa() : kv_self->get_kv_base();
+
+    // store to KV cache
+    {
+        ggml_build_forward_expand(gf, kv->cpy_k(ctx0, k_cur, il));
+        ggml_build_forward_expand(gf, kv->cpy_v(ctx0, v_cur, il));
+    }
+
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv->get_k(ctx0, il);
+    ggml_tensor * v = kv->get_v(ctx0, il);
+
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
    if (wo_b) {
        //cb(cur, "kqv_wo", il);
    }
@@ -1534,17 +1417,11 @@ ggml_tensor * llm_graph_context::build_attn(

    const auto & kq_mask = inp->get_kq_mask_cross();

-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
-    //cb(k, "v", il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = k_cur;
+    ggml_tensor * v = v_cur;

+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
    cb(cur, "kqv_out", il);

    if (wo) {
@@ -1712,3 +1589,30 @@ void llm_graph_context::build_pooling(

    ggml_build_forward_expand(gf, cur);
 }
+
+int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
+    // TODO move to hparams if a T5 variant appears that uses a different value
+    const int64_t max_distance = 128;
+
+    if (bidirectional) {
+        n_buckets >>= 1;
+    }
+
+    const int64_t max_exact = n_buckets >> 1;
+
+    int32_t relative_position = x - y;
+    int32_t relative_bucket = 0;
+
+    if (bidirectional) {
+        relative_bucket += (relative_position > 0) * n_buckets;
+        relative_position = abs(relative_position);
+    } else {
+        relative_position = -std::min<int32_t>(relative_position, 0);
+    }
+
+    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
+    relative_position_if_large = std::min<int32_t>(relative_position_if_large, n_buckets - 1);
+    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);
+
+    return relative_bucket;
+}
@@ -19,6 +19,7 @@ struct llama_cparams;

 class llama_memory_i;
 class llama_kv_cache_unified;
+class llama_kv_cache_unified_iswa;
 class llama_kv_cache_recurrent;

 // certain models (typically multi-modal) can produce different types of graphs
@@ -255,6 +256,31 @@ public:

    void set_input(const llama_ubatch * ubatch) override;

+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_kv_unified_iswa(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_unified_iswa * kv_self) :
+        hparams(hparams),
+        cparams(cparams),
+        kv_self(kv_self) {
+    }
+    ~llm_graph_input_attn_kv_unified_iswa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }

@@ -266,7 +292,7 @@ public:
    const llama_hparams & hparams;
    const llama_cparams & cparams;

-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_unified_iswa * kv_self;
 };

 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -378,7 +404,6 @@ struct llm_graph_context {
    const int64_t n_layer;
    const int64_t n_rot;
    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
-    const int64_t n_ctx_per_seq;
    const int64_t n_head;
    const int64_t n_head_kv;
    const int64_t n_embd_head_k;
@@ -507,13 +532,12 @@ struct llm_graph_context {

    ggml_tensor * build_attn_mha(
             ggml_cgraph * gf,
-             ggml_tensor * q,     // [n_embd_head_q, n_tokens, n_head_q]
-             ggml_tensor * k,     // [n_embd_head_k, n_tokens, n_head_k]
-             ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
+             ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
+             ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
+             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
             ggml_tensor * kq_b,
             ggml_tensor * kq_mask,
-             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                    bool   v_trans,
+             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale) const;

    llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
@@ -546,6 +570,21 @@ struct llm_graph_context {
                  float   kq_scale,
                    int   il) const;

+    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_kv_unified_iswa * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
    llm_graph_input_attn_cross * build_attn_inp_cross() const;

    ggml_tensor * build_attn(
@@ -596,3 +635,6 @@ struct llm_graph_context {
            ggml_tensor * cls_out,
            ggml_tensor * cls_out_b) const;
 };
+
+// TODO: better name
+int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional);
@@ -72,7 +72,7 @@ uint32_t llama_hparams::n_embd_v_s() const {

 bool llama_hparams::is_swa(uint32_t il) const {
    if (il < n_layer) {
-        return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+        return n_swa_pattern == 0 || (il % n_swa_pattern < (n_swa_pattern - 1));
    }

    GGML_ABORT("fatal error");
@@ -14,6 +14,12 @@ enum llama_expert_gating_func_type {
    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
 };

+enum llama_swa_type {
+    LLAMA_SWA_TYPE_NONE     = 0,
+    LLAMA_SWA_TYPE_STANDARD = 1,
+    LLAMA_SWA_TYPE_CHUNKED  = 2,
+};
+
 struct llama_hparams_posnet {
    uint32_t n_embd;
    uint32_t n_layer;
@@ -35,8 +41,6 @@ struct llama_hparams {
    uint32_t n_embd_features = 0;
    uint32_t n_layer;
    uint32_t n_rot;
-    uint32_t n_swa = 0; // sliding window attention (SWA)
-    uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
    uint32_t n_expert = 0;
@@ -96,6 +100,23 @@ struct llama_hparams {

    std::array<int, 4> rope_sections;

+    // Sliding Window Attention (SWA)
+    llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+
+    uint32_t n_swa = 0;         // the size of the sliding window (0 - no SWA)
+    uint32_t n_swa_pattern = 1; // this value n means that every nth layer is dense (i.e. non-SWA)
+                                // by default n == 1, all layers are dense
+                                // note that if n_swa_pattern == 0, all layers are SWA
+                                // example: n_swa_pattern = 3
+                                //   il == 0: swa
+                                //   il == 1: swa
+                                //   il == 2: dense
+                                //   il == 3: swa
+                                //   il == 4: swa
+                                //   il == 5: dense
+                                //   il == 6: swa
+                                //   etc ...
+
    // for State Space Models
    uint32_t ssm_d_conv  = 0;
    uint32_t ssm_d_inner = 0;
@@ -116,11 +137,10 @@ struct llama_hparams {
    bool causal_attn   = true;
    bool use_alibi     = false;
    bool attn_soft_cap = false;
+    bool use_kq_norm   = true;

+    // llama4
    uint32_t n_moe_layer_step        = 0;
-    bool     use_kq_norm             = true;
-    uint32_t n_attn_chunk            = 0;
-    // values below seems to be fixed on llama4
    uint32_t n_no_rope_layer_step    = 4;
    uint32_t n_attn_temp_floor_scale = 8192;
    float    f_attn_temp_scale       = 0.1;
@@ -8,6 +8,7 @@
 #include "ggml-cpp.h"

 #include <set>
+#include <unordered_map>
 #include <vector>

 struct llama_cparams;
@@ -40,6 +41,9 @@ struct llama_kv_cache : public llama_memory_i {
    // batch processing
    //

+    // =============================================================================================================
+    // TODO: refactor  and simplify this
+
    virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;

    // different KV caches require different batch splitting strategies
@@ -48,11 +52,10 @@ struct llama_kv_cache : public llama_memory_i {
    // find an empty slot of size "n_tokens" in the cache
    virtual bool find_slot(const llama_ubatch & batch) = 0;

+    // =============================================================================================================
+
    // getters
-    virtual int32_t   get_n_tokens()   const = 0;
-    virtual int32_t   get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
-    virtual llama_pos get_pos_max()    const = 0;
-    virtual bool      get_can_shift()  const = 0;
+    virtual bool get_can_shift() const = 0;

    bool get_can_edit() const override { return get_can_shift(); }

@@ -87,38 +90,25 @@ private:
 // llama_kv_cache_unified
 //

-// TODO: add notion of max sequences
 class llama_kv_cache_unified : public llama_kv_cache {
 public:
-    struct kv_cell {
-        llama_pos pos   = -1;
-        llama_pos delta =  0;
-
-        std::set<llama_seq_id> seq_id;
-
-        bool has_seq_id(const llama_seq_id & id) const {
-            return seq_id.find(id) != seq_id.end();
-        }
-
-        bool is_empty() const {
-            return seq_id.empty();
-        }
-
-        bool is_same_seq(const kv_cell & other) const {
-            return seq_id == other.seq_id;
-        }
-    };
-
    static uint32_t get_padding(const llama_cparams & cparams);

+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
    llama_kv_cache_unified(
-            const llama_model & model,
-                    ggml_type   type_k,
-                    ggml_type   type_v,
-                         bool   v_trans,
-                         bool   offload,
-                     uint32_t   kv_size,
-                     uint32_t   padding);
+            const llama_model &  model,
+              layer_filter_cb && filter,
+                    ggml_type    type_k,
+                    ggml_type    type_v,
+                         bool    v_trans,
+                         bool    offload,
+                     uint32_t    kv_size,
+                     uint32_t    n_seq_max,
+                     uint32_t    n_pad,
+                     uint32_t    n_swa,
+               llama_swa_type    swa_type);

    ~llama_kv_cache_unified() = default;

@@ -130,10 +120,11 @@ public:

    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
-    void seq_keep(llama_seq_id seq_id) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;

+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
    llama_pos seq_pos_max(llama_seq_id seq_id) const override;

    //
@@ -150,7 +141,6 @@ public:
    void set_full() override;

    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
-
    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;

    // updates the cache head
@@ -158,50 +148,106 @@ public:
    // to the first cell of the slot.
    bool find_slot(const llama_ubatch & batch) override;

-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos get_pos_max() const override;
-
    bool get_can_shift() const override;

    // state write/load

    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;

-    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
-    uint32_t size = 0; // total number of cells, shared across all sequences
-    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+    //
+    // llama_kv_cache_unified specific API
+    //

-    // computed before each graph build
-    uint32_t n = 0;
+    uint32_t get_n() const;
+    uint32_t get_size() const;

-    std::vector<kv_cell> cells;
+    // get views of the current state of the cache
+    ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
+    ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;

-    std::vector<ggml_tensor *> k_l; // per layer
-    std::vector<ggml_tensor *> v_l;
+    // store k_cur and v_cur in the cache based on the current head location
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+
+    void prune_swa(llama_seq_id seq_id, llama_pos pmin, llama_pos pmax);
+
+    void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
+    void set_input_k_shift   (ggml_tensor * dst) const;
+    void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;

 private:
    const llama_model & model;
    const llama_hparams & hparams;

+    struct kv_cell {
+        llama_pos pos   = -1;
+        llama_pos delta =  0;
+
+        // TODO: replace with bitset uint64_t
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    struct kv_layer {
+        // layer index in the model
+        // note: can be different from the layer index in the KV cache
+        uint32_t il;
+
+        ggml_tensor * k;
+        ggml_tensor * v;
+    };
+
    bool has_shift = false;
    bool do_defrag = false;
-
    bool v_trans   = true;  // the value tensor is transposed
-    bool can_shift = false;
+
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
+    uint32_t used = 0; // used cells (i.e. at least one seq_id) (TODO: add `struct kv_cells` and keep track automaticallt)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    const uint32_t n_seq_max = 1;

    // required padding
-    uint32_t padding = 1;
+    const uint32_t n_pad = 1;

-    ggml_type type_k = GGML_TYPE_F16;
-    ggml_type type_v = GGML_TYPE_F16;
+    // SWA
+    const uint32_t n_swa = 0;
+
+    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;

    std::vector<ggml_context_ptr>        ctxs;
    std::vector<ggml_backend_buffer_ptr> bufs;

+    std::vector<kv_cell>  cells;  // TODO: replace with `struct kv_cells`
+    std::vector<kv_layer> layers;
+
+    // model layer id -> KV cache layer id
+    std::unordered_map<int32_t, int32_t> map_layer_ids;
+
+    // recovery information used to restore the KV cells to their original state in case of a failure
+    struct {
+        void clear() {
+            cells.clear();
+        }
+
+        std::unordered_map<uint32_t, kv_cell> cells;
+    } recovery;
+
    // defrag
    struct {
        std::vector<uint32_t> ids;
@@ -210,17 +256,6 @@ private:
    // return true if cells have been moved
    bool defrag_prepare(int32_t n_max_nodes);

-    // commit/restore cache
-    struct slot_range {
-        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
-        uint32_t c1 = 0;
-    };
-
-    // pending cell updates that are not yet committed
-    struct {
-        std::vector<slot_range> ranges;
-    } pending;
-
    // find how many cells are currently in use
    uint32_t cell_max() const;

@@ -229,6 +264,8 @@ private:
    size_t size_k_bytes() const;
    size_t size_v_bytes() const;

+    bool is_masked_swa(llama_pos p0, llama_pos p1) const;
+
    ggml_tensor * build_rope_shift(
            const llama_cparams & cparams,
                   ggml_context * ctx,
@@ -255,6 +292,100 @@ private:
    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };

+//
+// llama_kv_cache_unified_iswa
+//
+
+// utilizes two instances of llama_kv_cache_unified
+//   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
+//   upon successful commit, the SWA cache removes old tokens outside the n_swa window
+
+class llama_kv_cache_unified_iswa : public llama_kv_cache {
+public:
+    llama_kv_cache_unified_iswa(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   swa_full,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_batch,
+                     uint32_t   n_pad);
+
+    ~llama_kv_cache_unified_iswa() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    void clear() override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & ctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
+
+    bool find_slot(const llama_ubatch & batch) override;
+
+    bool get_can_shift() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+
+    //
+    // llama_kv_cache_unified_iswa specific API
+    //
+
+    llama_kv_cache_unified * get_kv_base() const;
+    llama_kv_cache_unified * get_kv_swa () const;
+
+private:
+    const llama_hparams & hparams;
+
+    bool do_prune = true;
+
+    struct {
+        struct entry {
+            llama_pos pmin;
+            llama_pos pmax;
+        };
+
+        void clear() {
+            pos.clear();
+        }
+
+        // used to perform SWA pruning of old tokens
+        std::unordered_map<llama_seq_id, entry> pos;
+    } pending;
+
+    std::unique_ptr<llama_kv_cache_unified> kv_base;
+    std::unique_ptr<llama_kv_cache_unified> kv_swa;
+};
+
 //
 // llama_kv_cache_recurrent
 //
@@ -286,7 +417,8 @@ public:
                    ggml_type   type_k,
                    ggml_type   type_v,
                         bool   offload,
-                     uint32_t   kv_size);
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max);

    ~llama_kv_cache_recurrent() = default;

@@ -298,10 +430,11 @@ public:

    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
-    void seq_keep(llama_seq_id seq_id) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;

+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
    llama_pos seq_pos_max(llama_seq_id seq_id) const override;

    //
@@ -311,24 +444,17 @@ public:
    void restore() override;
    void commit()  override;

-    bool update(llama_context & lctx) override;
+    bool update(llama_context & ctx) override;

    void defrag_sched(float thold) override;

    void set_full() override;

    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
-
    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;

    bool find_slot(const llama_ubatch & batch) override;

-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos get_pos_max() const override;
-
    bool get_can_shift() const override;

    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
@@ -368,8 +494,7 @@ private:
        std::vector<slot_range> ranges;
    } pending;

-    ggml_type type_k = GGML_TYPE_F16;
-    ggml_type type_v = GGML_TYPE_F16;
+    const uint32_t n_seq_max = 1;

    std::vector<ggml_context_ptr>        ctxs;
    std::vector<ggml_backend_buffer_ptr> bufs;
@@ -388,12 +513,3 @@ private:
    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
-
-
-//
-// kv cache view
-//
-
-llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max);
-
-void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv);
@@ -7,8 +7,8 @@ struct llama_memory_params {
    ggml_type type_k;
    ggml_type type_v;

-    // parameters for other types of memory
-    // ...
+    // use full-size SWA cache
+    bool swa_full;
 };

 // general concept of LLM memory
@@ -25,6 +25,7 @@ public:
    virtual void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) = 0;
    virtual void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) = 0;

+    virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;

    virtual bool get_can_edit() const = 0;
@@ -571,9 +571,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
                ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);
+
+                hparams.swa_type      = LLAMA_SWA_TYPE_CHUNKED;
+                hparams.n_swa         = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
                hparams.n_swa_pattern = 4;    // pattern: 3 chunked - 1 full
-                hparams.n_attn_chunk  = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
-                hparams.n_swa = 1; // TODO @ngxson : this is added to trigger the SWA branch (we store the chunked attn mask in the SWA tensor), will need to clean this up later

                switch (hparams.n_expert) {
                    case 16:  type = LLM_TYPE_17B_16E; break;
@@ -852,22 +853,17 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                    default: type = LLM_TYPE_UNKNOWN;
                }

-                // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
-                if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
-                    // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
-                    hparams.n_swa = 2047;
-                } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
-                    // default value for Phi-3-mini-128k-instruct
-                    // note: this seems incorrect because the window is bigger than the train context?
-                    hparams.n_swa = 262144;
-                } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
-                    // default value for Phi-3-medium-128k-instruct
-                    // note: this seems incorrect because the window is equal to the train context?
-                    hparams.n_swa = 131072;
-                }
-                bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
-                if (!found_swa && hparams.n_swa == 0) {
-                    throw std::runtime_error("invalid value for sliding_window");
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+
+                if (found_swa && hparams.n_swa > 0) {
+                    LLAMA_LOG_WARN("%s: Phi SWA is currently disabled - results might be suboptimal for some models (see %s)\n",
+                            __func__, "https://github.com/ggml-org/llama.cpp/pull/13676");
+
+                    // TODO: fix conversion scripts to correctly populate `n_swa` and `n_swa_pattern`
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+
+                    hparams.n_swa         = 0;
+                    hparams.n_swa_pattern = 1;
                }
            } break;
        case LLM_ARCH_PHIMOE:
@@ -937,6 +933,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            } break;
        case LLM_ARCH_GEMMA2:
            {
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                hparams.n_swa = 4096; // default value of gemma 2
                hparams.n_swa_pattern = 2;
                hparams.attn_soft_cap = true;
@@ -955,6 +952,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            } break;
        case LLM_ARCH_GEMMA3:
            {
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                hparams.n_swa_pattern = 6;

                hparams.rope_freq_base_train_swa  = 10000.0f;
@@ -1039,6 +1037,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            } break;
        case LLM_ARCH_COHERE2:
            {
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                hparams.n_swa_pattern = 4;

                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
@@ -4489,7 +4488,17 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const {
    return it->second;
 }

-ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
+float llama_model::get_rope_freq_base (const llama_cparams & cparams, int il) const {
+    return hparams.is_swa(il) ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base;
+}
+
+float llama_model::get_rope_freq_scale(const llama_cparams & cparams, int il) const {
+    return hparams.is_swa(il) ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
+}
+
+ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const {
+    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+
    // choose long/short freq factors based on the context size
    if (layers[il].rope_freqs != nullptr) {
        return layers[il].rope_freqs;
@@ -4517,22 +4526,13 @@ struct llm_build_llama : public llm_graph_context {
        // inp_pos - contains the positions
        ggml_tensor * inp_pos = build_inp_pos();

-        // temperature tuning
-        ggml_tensor * inp_attn_scale = nullptr;
-        if (arch == LLM_ARCH_LLAMA4) {
-            inp_attn_scale = build_inp_attn_scale();
-        }
-
        auto * inp_attn = build_attn_inp_kv_unified();

        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
        for (int il = 0; il < n_layer; ++il) {
            ggml_tensor * inpSA = inpL;

-            bool use_rope = arch == LLM_ARCH_LLAMA4
-                ? (il + 1) % hparams.n_no_rope_layer_step != 0
-                : true;
-
            // norm
            cur = build_norm(inpL,
                    model.layers[il].attn_norm, NULL,
@@ -4542,7 +4542,169 @@ struct llm_build_llama : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network (non-MoE)
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+                cb(cur, "ffn_moe_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_llama_iswa : public llm_graph_context {
+    llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        // temperature tuning
+        ggml_tensor * inp_attn_scale = nullptr;
+        inp_attn_scale = build_inp_attn_scale();
+
+        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -4590,7 +4752,7 @@ struct llm_build_llama : public llm_graph_context {
                cb(Kcur, "Kcur", il);
                cb(Vcur, "Vcur", il);

-                if (arch == LLM_ARCH_LLAMA4 && use_rope && hparams.use_kq_norm) {
+                if (use_rope && hparams.use_kq_norm) {
                    // Llama4TextL2Norm
                    Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
                    Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
@@ -4616,7 +4778,6 @@ struct llm_build_llama : public llm_graph_context {

            // feed-forward network (non-MoE)
            if (model.layers[il].ffn_gate_inp == nullptr) {
-
                cur = build_norm(ffn_inp,
                        model.layers[il].ffn_norm, NULL,
                        LLM_NORM_RMS, il);
@@ -4629,9 +4790,7 @@ struct llm_build_llama : public llm_graph_context {
                        NULL,
                        LLM_FFN_SILU, LLM_FFN_PAR, il);
                cb(cur, "ffn_out", il);
-
-            } else if (arch == LLM_ARCH_LLAMA4) {
-                // llama4 MoE
+            } else {
                ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
                        model.layers[il].ffn_norm, NULL,
                        LLM_NORM_RMS, il);
@@ -4660,26 +4819,6 @@ struct llm_build_llama : public llm_graph_context {

                cur = ggml_add(ctx0, moe_out, shexp_out);
                cb(cur, "ffn_moe_out_merged", il);
-
-            } else {
-                // MoE branch
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(cur, "ffn_moe_out", il);
            }

            cur = ggml_add(ctx0, cur, ffn_inp);
@@ -4753,7 +4892,7 @@ struct llm_build_deci : public llm_graph_context {
            } else if (n_head > 0) {
                // self-attention
                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -7202,6 +7341,7 @@ struct llm_build_phi2 : public llm_graph_context {
    }
 };

+template<bool iswa>
 struct llm_build_phi3 : public llm_graph_context {
    llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
        const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -7217,7 +7357,14 @@ struct llm_build_phi3 : public llm_graph_context {
        // inp_pos - contains the positions
        ggml_tensor * inp_pos = build_inp_pos();

-        auto * inp_attn = build_attn_inp_kv_unified();
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        inp_attn_type * inp_attn = nullptr;
+
+        if constexpr (iswa) {
+            inp_attn = build_attn_inp_kv_unified_iswa();
+        } else {
+            inp_attn = build_attn_inp_kv_unified();
+        }

        for (int il = 0; il < n_layer; ++il) {
            auto * residual = inpL;
@@ -7225,7 +7372,7 @@ struct llm_build_phi3 : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                ggml_tensor* attn_norm_output = build_norm(inpL,
                        model.layers[il].attn_norm,
@@ -7977,7 +8124,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
        for (int il = 0; il < n_layer; ++il) {
            ggml_tensor * inpSA = inpL;

-            ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

            // norm
            cur = build_norm(inpL,
@@ -8277,8 +8424,8 @@ struct llm_build_gemma : public llm_graph_context {
    }
 };

-struct llm_build_gemma2 : public llm_graph_context {
-    llm_build_gemma2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+struct llm_build_gemma2_iswa : public llm_graph_context {
+    llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
        const int64_t n_embd_head = hparams.n_embd_head_k;

        ggml_tensor * cur;
@@ -8292,7 +8439,7 @@ struct llm_build_gemma2 : public llm_graph_context {
        // inp_pos - contains the positions
        ggml_tensor * inp_pos = build_inp_pos();

-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv_unified_iswa();

        for (int il = 0; il < n_layer; ++il) {
            // norm
@@ -8414,8 +8561,8 @@ struct llm_build_gemma2 : public llm_graph_context {
    }
 };

-struct llm_build_gemma3 : public llm_graph_context {
-    llm_build_gemma3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+struct llm_build_gemma3_iswa : public llm_graph_context {
+    llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
        const int64_t n_embd_head = hparams.n_embd_head_k;

        ggml_tensor * cur;
@@ -8433,13 +8580,11 @@ struct llm_build_gemma3 : public llm_graph_context {
        ggml_tensor * inp_pos = build_inp_pos();

        // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv_unified_iswa();

        for (int il = 0; il < n_layer; ++il) {
-            const bool is_swa = hparams.is_swa(il);
-
-            const float freq_base_l  = is_swa ? hparams.rope_freq_base_train_swa  : cparams.rope_freq_base;
-            const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
+            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);

            // norm
            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
@@ -9016,8 +9161,8 @@ struct llm_build_command_r : public llm_graph_context {
    }
 };

-struct llm_build_cohere2 : public llm_graph_context {
-    llm_build_cohere2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+struct llm_build_cohere2_iswa : public llm_graph_context {
+    llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
        const int64_t n_embd_head = hparams.n_embd_head_v;

        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -9032,7 +9177,7 @@ struct llm_build_cohere2 : public llm_graph_context {
        // inp_pos - contains the positions
        ggml_tensor * inp_pos = build_inp_pos();

-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv_unified_iswa();

        for (int il = 0; il < n_layer; ++il) {
            const bool is_swa = hparams.is_swa(il);
@@ -9045,7 +9190,7 @@ struct llm_build_cohere2 : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -9983,7 +10128,7 @@ struct llm_build_deepseek : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -11347,7 +11492,7 @@ struct llm_build_exaone : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -12263,7 +12408,7 @@ struct llm_build_granite : public llm_graph_context {
                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);

                if (use_rope) {
-                    ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                    ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
                    Qcur = ggml_rope_ext(
                            ctx0, Qcur, inp_pos, rope_factors,
                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -12916,7 +13061,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

                // compute Q and K and RoPE them
                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -13044,6 +13189,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
        case LLM_ARCH_JINA_BERT_V2:
        case LLM_ARCH_NOMIC_BERT:
        case LLM_ARCH_NOMIC_BERT_MOE:
+        case LLM_ARCH_WAVTOKENIZER_DEC:
            {
                res = nullptr;
            } break;
@@ -13058,7 +13204,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                        GGML_TYPE_F32,
                        GGML_TYPE_F32,
                        cparams.offload_kqv,
-                        std::max((uint32_t) 1, cparams.n_seq_max));
+                        std::max((uint32_t) 1, cparams.n_seq_max),
+                        cparams.n_seq_max);
            } break;
        default:
            {
@@ -13068,14 +13215,36 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,

                LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);

-                res = new llama_kv_cache_unified(
-                        *this,
-                        params.type_k,
-                        params.type_v,
-                        !cparams.flash_attn,
-                        cparams.offload_kqv,
-                        cparams.n_ctx,
-                        padding);
+                if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                    GGML_ASSERT(hparams.n_swa_pattern != 1);
+
+                    res = new llama_kv_cache_unified_iswa(
+                            *this,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            params.swa_full,
+                            cparams.n_ctx,
+                            cparams.n_seq_max,
+                            cparams.n_batch,
+                            padding);
+                } else {
+                    GGML_ASSERT(hparams.n_swa_pattern == 1);
+
+                    res = new llama_kv_cache_unified(
+                            *this,
+                            nullptr,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            cparams.n_ctx,
+                            cparams.n_seq_max,
+                            padding,
+                            hparams.n_swa,
+                            hparams.swa_type);
+                }
            }
    }

@@ -13090,11 +13259,14 @@ llm_graph_result_ptr llama_model::build_graph(

    switch (arch) {
        case LLM_ARCH_LLAMA:
-        case LLM_ARCH_LLAMA4:
        case LLM_ARCH_MINICPM:
            {
                llm = std::make_unique<llm_build_llama>(*this, params, gf);
            } break;
+        case LLM_ARCH_LLAMA4:
+            {
+                llm = std::make_unique<llm_build_llama_iswa>(*this, params, gf);
+            } break;
        case LLM_ARCH_DECI:
            {
                llm = std::make_unique<llm_build_deci>(*this, params, gf);
@@ -13169,7 +13341,11 @@ llm_graph_result_ptr llama_model::build_graph(
        case LLM_ARCH_PHI3:
        case LLM_ARCH_PHIMOE:
            {
-                llm = std::make_unique<llm_build_phi3>(*this, params, gf);
+                if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                    llm = std::make_unique<llm_build_phi3<true>> (*this, params, gf);
+                } else {
+                    llm = std::make_unique<llm_build_phi3<false>>(*this, params, gf);
+                }
            } break;
        case LLM_ARCH_PLAMO:
            {
@@ -13201,11 +13377,11 @@ llm_graph_result_ptr llama_model::build_graph(
            } break;
        case LLM_ARCH_GEMMA2:
            {
-                llm = std::make_unique<llm_build_gemma2>(*this, params, gf);
+                llm = std::make_unique<llm_build_gemma2_iswa>(*this, params, gf);
            } break;
        case LLM_ARCH_GEMMA3:
            {
-                llm = std::make_unique<llm_build_gemma3>(*this, params, gf);
+                llm = std::make_unique<llm_build_gemma3_iswa>(*this, params, gf);
            } break;
        case LLM_ARCH_STARCODER2:
            {
@@ -13225,7 +13401,7 @@ llm_graph_result_ptr llama_model::build_graph(
            } break;
        case LLM_ARCH_COHERE2:
            {
-                llm = std::make_unique<llm_build_cohere2>(*this, params, gf);
+                llm = std::make_unique<llm_build_cohere2_iswa>(*this, params, gf);
            } break;
        case LLM_ARCH_DBRX:
            {
@@ -398,7 +398,10 @@ struct llama_model {

    const struct ggml_tensor * get_tensor(const char * name) const;

-    ggml_tensor * get_rope_factors(uint32_t n_ctx_per_seq, int il) const;
+    float get_rope_freq_base (const llama_cparams & cparams, int il) const;
+    float get_rope_freq_scale(const llama_cparams & cparams, int il) const;
+
+    ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;

    // note: can mutate `cparams`
    // TODO: move this to new llm_arch_model_i interface
@@ -128,7 +128,7 @@ int main(void) {

    if (common_has_curl()) {
        printf("test-arg-parser: test curl-related functions\n\n");
-        const char * GOOD_URL = "https://raw.githubusercontent.com/ggml-org/llama.cpp/refs/heads/master/README.md";
+        const char * GOOD_URL = "https://ggml.ai/";
        const char * BAD_URL  = "https://www.google.com/404";
        const char * BIG_FILE = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v1.bin";

@@ -80,10 +80,6 @@ Using the `-d <n>` option, each test can be run at a specified context depth, pr

 For a description of the other options, see the [main example](../main/README.md).

-Note:
-
- When using SYCL backend, there would be hang issue in some cases. Please set `--mmp 0`.
-
 ## Examples

 ### Text generation with different models
@@ -991,6 +991,7 @@ struct cmd_params_instance {
        cparams.flash_attn   = flash_attn;
        cparams.embeddings   = embeddings;
        cparams.op_offload   = !no_op_offload;
+        cparams.swa_full     = false;

        return cparams;
    }
@@ -1,5 +1,15 @@
 # mtmd

+# compile mtmd-audio separately to avoid long compile times with miniaudio.h
+# TODO @ngxson : move miniaudio.h and stb_image.h to mtmd-helper.cpp, then compile the helper as a separate library
+add_library(mtmd_audio STATIC mtmd-audio.cpp mtmd-audio.h)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(mtmd_audio PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+target_link_libraries(mtmd_audio PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(mtmd_audio PRIVATE cxx_std_17)
+target_include_directories(mtmd_audio PRIVATE .)
+
 add_library(mtmd OBJECT
            mtmd.cpp
            mtmd-helper.cpp
@@ -9,7 +19,7 @@ add_library(mtmd OBJECT
            clip-impl.h
            )

-target_link_libraries(mtmd PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(mtmd PRIVATE ggml llama mtmd_audio ${CMAKE_THREAD_LIBS_INIT})

 target_include_directories(mtmd PUBLIC .)
 target_include_directories(mtmd PRIVATE ../..)
@@ -22,12 +32,13 @@ if (BUILD_SHARED_LIBS)
    set_target_properties(mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
    target_compile_definitions(mtmd PRIVATE LLAMA_SHARED LLAMA_BUILD)
    add_library(mtmd_shared SHARED $<TARGET_OBJECTS:mtmd>)
-    target_link_libraries(mtmd_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(mtmd_shared PRIVATE ggml llama mtmd_audio ${CMAKE_THREAD_LIBS_INIT})
    install(TARGETS mtmd_shared LIBRARY)
 endif()

 if (NOT MSVC)
    target_compile_options(mtmd PRIVATE -Wno-cast-qual) # stb_image.h
+    target_compile_options(mtmd_audio PRIVATE -Wno-cast-qual) # miniaudio.h
 endif()

 if(TARGET BUILD_INFO)
@@ -4,6 +4,7 @@

 #include <climits>
 #include <cstdarg>
+#include <cinttypes>
 #include <string>
 #include <map>
 #include <sstream>
@@ -15,22 +16,26 @@
 #define KEY_FTYPE               "general.file_type"
 #define KEY_NAME                "general.name"
 #define KEY_DESCRIPTION         "general.description"
-#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_HAS_AUDIO_ENC       "clip.has_audio_encoder"
+#define KEY_HAS_VISION_ENC      "clip.has_vision_encoder"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_USE_SILU            "clip.use_silu"
-#define KEY_N_EMBD              "clip.vision.embedding_length"
-#define KEY_N_FF                "clip.vision.feed_forward_length"
-#define KEY_N_BLOCK             "clip.vision.block_count"
-#define KEY_N_HEAD              "clip.vision.attention.head_count"
-#define KEY_LAYER_NORM_EPS      "clip.vision.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM            "clip.vision.projection_dim"
+
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+
+// vision-specific
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
-#define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"

 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
@@ -38,6 +43,11 @@
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
 #define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
 #define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
+#define KEY_MINICPMV_VERSION      "clip.minicpmv_version"
+
+// audio-specific
+#define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"
+#define KEY_A_PROJ_STACK_FACTOR "clip.audio.projector.stack_factor"


 //
@@ -94,6 +104,12 @@
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"

+// ultravox
+#define TN_CONV1D       "a.conv1d.%d.%s"
+#define TN_MM_AUDIO_MLP "mm.a.mlp.%d.%s"
+#define TN_MM_NORM_PRE  "mm.a.norm_pre.%s"
+#define TN_MM_NORM_MID  "mm.a.norm_mid.%s"
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))

@@ -109,7 +125,9 @@ enum projector_type {
    PROJECTOR_TYPE_IDEFICS3,
    PROJECTOR_TYPE_PIXTRAL,
    PROJECTOR_TYPE_QWEN25VL,
+    PROJECTOR_TYPE_ULTRAVOX,
    PROJECTOR_TYPE_INTERNVL,
+    PROJECTOR_TYPE_LLAMA4,
    PROJECTOR_TYPE_UNKNOWN,
 };

@@ -124,7 +142,9 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
+    { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},
    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
+    { PROJECTOR_TYPE_LLAMA4,    "llama4"},
 };

 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -144,8 +164,10 @@ struct clip_image_u8 {
    std::vector<uint8_t> buf;
 };

-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
+// For images, buf.size() == nx*ny*3
+//     Memory layout: RGBRGBRGB...
+// For audio, only one channel is used, buf.size() == nx*ny
+//     nx will be n_frames and ny will be n_mel
 struct clip_image_f32 {
    int nx;
    int ny;
@@ -239,9 +261,20 @@ struct clip_image_u8_batch {

 struct clip_image_f32_batch {
    std::vector<clip_image_f32_ptr> entries;
+    bool is_audio = false;
+
+    // for llava-uhd style models, we need to know the grid size
+    // note: entries.size() == grid_x * grid_y + 1 (one overview image)
+    int grid_x = 0;
+    int grid_y = 0;

    clip_image_f32_batch clone() const {
-        clip_image_f32_batch new_batch;
+        clip_image_f32_batch new_batch{
+            /* entries  */ {},
+            /* is_audio */ is_audio,
+            /* grid_x   */ grid_x,
+            /* grid_y   */ grid_y,
+        };
        new_batch.entries.reserve(entries.size());
        for (const auto & entry : entries) {
            new_batch.entries.emplace_back(new clip_image_f32(*entry));
@@ -358,6 +391,70 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
    }
 }

+//
+// debugging
+//
+
+static void print_tensor_shape(ggml_tensor * t) {
+    printf("%s.shape = [", t->name);
+    for (int i = 0; i < ggml_n_dims(t); ++i) {
+        printf("%" PRId64, t->ne[i]);
+        if (i < ggml_n_dims(t) - 1) {
+            printf(", ");
+        }
+    }
+    printf("]\n");
+}
+
+static void print_tensor_data(ggml_tensor * t, uint8_t * data, int64_t n) {
+    ggml_type type = t->type;
+    int64_t * ne = t->ne;
+    size_t * nb = t->nb;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        printf("%s.data: [\n", t->name);
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2*n) {
+                printf("     ..., \n");
+                i2 = ne[2] - n;
+            }
+            printf("     [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2*n) {
+                    printf("      ..., \n");
+                    i1 = ne[1] - n;
+                }
+                printf("      [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2*n) {
+                        printf("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+                    float v;
+                    if (type == GGML_TYPE_F16) {
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
+                    } else if (type == GGML_TYPE_F32) {
+                        v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I32) {
+                        v = (float) *(int32_t *) &data[i];
+                    } else if (type == GGML_TYPE_I16) {
+                        v = (float) *(int16_t *) &data[i];
+                    } else if (type == GGML_TYPE_I8) {
+                        v = (float) *(int8_t *) &data[i];
+                    } else {
+                        GGML_ABORT("fatal error");
+                    }
+                    printf("%8.4f", v);
+                    if (i0 < ne[0] - 1) printf(", ");
+                }
+                printf("],\n");
+            }
+            printf("     ],\n");
+        }
+        printf("    ]\n");
+    }
+}
+
 //
 // API used internally with mtmd
 //
@@ -35,6 +35,7 @@ struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callbac

 enum ffn_op_type {
    FFN_GELU,
+    FFN_GELU_ERF,
    FFN_SILU,
    FFN_GELU_QUICK,
 };
@@ -165,6 +166,9 @@ enum patch_merge_type {
 };

 struct clip_hparams {
+    bool has_vision = false;
+    bool has_audio = false;
+
    int32_t image_size;
    int32_t patch_size;
    int32_t n_embd;
@@ -191,6 +195,10 @@ struct clip_hparams {
    int32_t attn_window_size = 0;
    int32_t n_wa_pattern = 0;
    int32_t spatial_merge_size = 0;
+
+    // audio
+    int32_t n_mel_bins = 0; // whisper preprocessor
+    int32_t proj_stack_factor = 0; // ultravox
 };

 struct clip_layer {
@@ -332,6 +340,14 @@ struct clip_vision_model {
    // pixtral
    ggml_tensor * token_embd_img_break = nullptr;
    ggml_tensor * mm_patch_merger_w = nullptr;
+
+    // ultravox / whisper encoder
+    ggml_tensor * conv1d_1_w = nullptr;
+    ggml_tensor * conv1d_1_b = nullptr;
+    ggml_tensor * conv1d_2_w = nullptr;
+    ggml_tensor * conv1d_2_b = nullptr;
+    ggml_tensor * mm_norm_pre_w = nullptr;
+    ggml_tensor * mm_norm_mid_w = nullptr;
 };

 struct clip_ctx {
@@ -359,9 +375,12 @@ struct clip_ctx {
    int max_nodes = 8192;
    ggml_backend_sched_ptr sched;

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

    clip_ctx(clip_context_params & ctx_params) {
+        debug_graph = std::getenv("MTMD_DEBUG_GRAPH") != nullptr;
        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
        if (!backend_cpu) {
            throw std::runtime_error("failed to initialize CPU backend");
@@ -440,7 +459,7 @@ struct clip_graph {
        };
        ctx0_ptr.reset(ggml_init(params));
        ctx0 = ctx0_ptr.get();
-        gf = ggml_new_graph(ctx0);
+        gf = ggml_new_graph_custom(ctx0, ctx->max_nodes, false);
    }

    ggml_cgraph * build_siglip() {
@@ -522,7 +541,7 @@ struct clip_graph {
        ggml_set_input(pos_w);

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

        ggml_tensor * inp = build_inp();
@@ -936,6 +955,101 @@ struct clip_graph {
        return gf;
    }

+    ggml_cgraph * build_llama4() {
+        GGML_ASSERT(model.class_embedding != nullptr);
+        GGML_ASSERT(model.position_embeddings != nullptr);
+
+        const int n_pos = n_patches + 1; // +1 for [CLS]
+
+        // 2D input positions
+        ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+        ggml_set_name(pos_h, "pos_h");
+        ggml_set_input(pos_h);
+
+        ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+        ggml_set_name(pos_w, "pos_w");
+        ggml_set_input(pos_w);
+
+        ggml_tensor * inp = build_inp_raw();
+
+        // Llama4UnfoldConvolution
+        {
+            ggml_tensor * kernel = ggml_reshape_4d(ctx0, model.patch_embeddings_0,
+                                                    patch_size, patch_size, 3, n_embd);
+            inp = ggml_im2col(ctx0, kernel, inp, patch_size, patch_size, 0, 0, 1, 1, true, inp->type);
+            inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+            inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+            cb(inp, "patch_conv", -1);
+        }
+
+        // add CLS token
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+        // build ViT with 2D position embeddings
+        auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+            // first half is X axis and second half is Y axis
+            // ref: https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/llama4/modeling_llama4.py#L1312
+            // ref: https://github.com/Blaizzy/mlx-vlm/blob/a57156aa87b33cca6e5ee6cfc14dd4ef8f611be6/mlx_vlm/models/llama4/vision.py#L441
+            return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+        };
+        ggml_tensor * cur = build_vit(
+                                inp, n_pos,
+                                NORM_TYPE_NORMAL,
+                                hparams.ffn_op,
+                                model.position_embeddings,
+                                add_pos);
+
+        // remove CLS token
+        cur = ggml_view_2d(ctx0, cur,
+            n_embd, n_patches,
+            ggml_row_size(cur->type, n_embd), 0);
+
+        // pixel shuffle
+        // based on Llama4VisionPixelShuffleMLP
+        // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
+        {
+            const int scale_factor = model.hparams.proj_scale_factor;
+            const int bsz = 1; // batch size, always 1 for now since we don't support batching
+            GGML_ASSERT(scale_factor > 0);
+            GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
+            cur = ggml_reshape_4d(ctx0, cur,
+                n_embd * scale_factor,
+                n_patches_x / scale_factor,
+                n_patches_y,
+                bsz);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+                n_embd * scale_factor * scale_factor,
+                n_patches_x / scale_factor,
+                n_patches_y / scale_factor,
+                bsz);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            // flatten to 2D
+            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+                n_embd * scale_factor * scale_factor,
+                n_patches / scale_factor / scale_factor);
+            cb(cur, "pixel_shuffle", -1);
+        }
+
+        // based on Llama4VisionMLP2 (always uses GELU activation, no bias)
+        {
+            cur = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, cur);
+            cur = ggml_gelu(ctx0, cur);
+            cur = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, cur);
+            cur = ggml_gelu(ctx0, cur);
+            cb(cur, "adapter_mlp", -1);
+        }
+
+        // Llama4MultiModalProjector
+        cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+        cb(cur, "projected", -1);
+
+        // build the graph
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
    // this graph is used by llava, granite and glm
    // due to having embedding_stack (used by granite), we cannot reuse build_vit
    ggml_cgraph * build_llava() {
@@ -1310,16 +1424,118 @@ struct clip_graph {
        return gf;
    }

+    // whisper encoder with custom projector
+    ggml_cgraph * build_whisper_enc() {
+        const int n_frames = img.nx;
+        const int n_pos    = n_frames / 2;
+        GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);
+
+        ggml_tensor * inp = build_inp_raw(1);
+
+        // conv1d block
+        {
+            // convolution + gelu
+            ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
+            cur = ggml_add(ctx0, cur, model.conv1d_1_b);
+
+            cur = ggml_gelu_erf(ctx0, cur);
+
+            cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
+            cur = ggml_add(ctx0, cur, model.conv1d_2_b);
+
+            cur = ggml_gelu_erf(ctx0, cur);
+            // transpose
+            inp = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+            cb(inp, "after_conv1d", -1);
+        }
+
+        // sanity check (only check one layer, but it should be the same for all)
+        GGML_ASSERT(model.layers[0].ln_1_w && model.layers[0].ln_1_b);
+        GGML_ASSERT(model.layers[0].ln_2_w && model.layers[0].ln_2_b);
+        GGML_ASSERT(model.layers[0].q_b);
+        GGML_ASSERT(model.layers[0].v_b);
+        GGML_ASSERT(!model.layers[0].k_b); // no bias for k
+        GGML_ASSERT(model.post_ln_w && model.post_ln_b);
+
+        ggml_tensor * pos_embd_selected = ggml_view_2d(
+            ctx0, model.position_embeddings,
+            model.position_embeddings->ne[0], n_pos,
+            model.position_embeddings->nb[1], 0
+        );
+        ggml_tensor * cur = build_vit(
+                                inp, n_pos,
+                                NORM_TYPE_NORMAL,
+                                hparams.ffn_op,
+                                pos_embd_selected,
+                                nullptr);
+
+        cb(cur, "after_transformer", -1);
+
+        // StackAudioFrames
+        // https://huggingface.co/fixie-ai/ultravox-v0_5-llama-3_2-1b/blob/main/ultravox_model.py
+        {
+            int64_t stride = n_embd * hparams.proj_stack_factor;
+            int64_t padded_len = GGML_PAD(ggml_nelements(cur), stride);
+            int64_t pad = padded_len - ggml_nelements(cur);
+            if (pad > 0) {
+                cur = ggml_view_1d(ctx0, cur, ggml_nelements(cur), 0);
+                cur = ggml_pad(ctx0, cur, pad, 0, 0, 0);
+            }
+            cur = ggml_view_2d(ctx0, cur, stride, padded_len / stride,
+                                ggml_row_size(cur->type, stride), 0);
+        }
+
+        cb(cur, "after_stacked", -1);
+
+        // UltravoxProjector
+        {
+            // pre-norm
+            cur = ggml_rms_norm(ctx0, cur, 1e-6);
+            cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
+
+            // ffn in
+            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+
+            // swiglu
+            {
+                int64_t split_point = cur->ne[0] / 2;
+                ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
+                ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
+
+                // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
+                x1 = ggml_silu(ctx0, x1);
+                cur = ggml_mul(ctx0, x0, x1);
+            }
+
+            // mid-norm
+            cur = ggml_rms_norm(ctx0, cur, 1e-6);
+            cur = ggml_mul(ctx0, cur, model.mm_norm_mid_w);
+
+            // ffn out
+            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+        }
+
+        cb(cur, "projected", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
 private:
    //
    // utility functions
    //

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

    // build vision transformer (ViT) cgraph
@@ -1460,8 +1676,8 @@ private:
        return inp;
    }

-    ggml_tensor * build_inp_raw() {
-        ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, img.nx, img.ny, 3);
+    ggml_tensor * build_inp_raw(int channels = 3) {
+        ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, img.nx, img.ny, channels);
        ggml_set_name(inp_raw, "inp_raw");
        ggml_set_input(inp_raw);
        return inp_raw;
@@ -1539,6 +1755,11 @@ private:
                    cur = ggml_gelu(ctx0, cur);
                    cb(cur, "ffn_gelu", il);
                } break;
+            case FFN_GELU_ERF:
+                {
+                    cur = ggml_gelu_erf(ctx0, cur);
+                    cb(cur, "ggml_gelu_erf", il);
+                } break;
            case FFN_GELU_QUICK:
                {
                    cur = ggml_gelu_quick(ctx0, cur);
@@ -1630,9 +1851,10 @@ private:
    static ggml_tensor * build_rope_2d(
        ggml_context * ctx0,
        ggml_tensor * cur,
-        ggml_tensor * pos_h,
-        ggml_tensor * pos_w,
-        const float freq_base
+        ggml_tensor * pos_a, // first half
+        ggml_tensor * pos_b, // second half
+        const float freq_base,
+        const bool interleave_freq
    ) {
        const int64_t n_dim  = cur->ne[0];
        const int64_t n_head = cur->ne[1];
@@ -1646,7 +1868,9 @@ private:
        //  ^ don't ask me why, it's math! -2(2i) / n_dim == -2i / (n_dim/2)
        // then for the second half, we use freq_scale to shift the inv_freq
        //  ^ why? replace (2i) with (2i+1) in the above equation
-        const float freq_scale_odd = std::pow(freq_base, (float)-2/n_dim);
+        const float freq_scale_odd = interleave_freq
+                                    ? std::pow(freq_base, (float)-2/n_dim)
+                                    : 1.0;

        // first half
        ggml_tensor * first;
@@ -1659,7 +1883,7 @@ private:
            first = ggml_rope_ext(
                ctx0,
                first,
-                pos_h,      // positions
+                pos_a,      // positions
                nullptr,    // freq factors
                n_dim/2,    // n_dims
                0, 0, freq_base,
@@ -1679,7 +1903,7 @@ private:
            second = ggml_rope_ext(
                ctx0,
                second,
-                pos_w,      // positions
+                pos_b,      // positions
                nullptr,    // freq factors
                n_dim/2,    // n_dims
                0, 0, freq_base,
@@ -1723,6 +1947,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
            {
                res = graph.build_internvl();
            } break;
+        case PROJECTOR_TYPE_LLAMA4:
+            {
+                res = graph.build_llama4();
+            } break;
+        case PROJECTOR_TYPE_ULTRAVOX:
+            {
+                res = graph.build_whisper_enc();
+            } break;
        default:
            {
                res = graph.build_llava();
@@ -1806,18 +2038,30 @@ struct clip_model_loader {

        // other hparams
        {
-            get_i32(KEY_MINICPMV_VERSION, ctx_clip.minicpmv_version, false); // legacy
+            get_bool(KEY_HAS_AUDIO_ENC,  hparams.has_audio, false);
+            get_bool(KEY_HAS_VISION_ENC, hparams.has_vision, false);

-            get_u32(KEY_N_EMBD,         hparams.n_embd);
-            get_u32(KEY_N_HEAD,         hparams.n_head);
-            get_u32(KEY_N_FF,           hparams.n_ff);
-            get_u32(KEY_N_BLOCK,        hparams.n_layer);
-            get_u32(KEY_PROJ_DIM,       hparams.projection_dim);
-            get_f32(KEY_LAYER_NORM_EPS, hparams.eps);
-            get_u32(KEY_IMAGE_SIZE,     hparams.image_size);
-            get_u32(KEY_PATCH_SIZE,     hparams.patch_size);
-            get_u32(KEY_IMAGE_CROP_RESOLUTION,    hparams.image_crop_resolution, false);
-            get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false);
+            const char * prefix = hparams.has_vision ? "vision" : "audio";
+            get_u32(string_format(KEY_N_EMBD,         prefix), hparams.n_embd);
+            get_u32(string_format(KEY_N_HEAD,         prefix), hparams.n_head);
+            get_u32(string_format(KEY_N_FF,           prefix), hparams.n_ff);
+            get_u32(string_format(KEY_N_BLOCK,        prefix), hparams.n_layer);
+            get_u32(string_format(KEY_PROJ_DIM,       prefix), hparams.projection_dim);
+            get_f32(string_format(KEY_LAYER_NORM_EPS, prefix), hparams.eps);
+
+            if (hparams.has_vision) {
+                get_u32(KEY_IMAGE_SIZE, hparams.image_size);
+                get_u32(KEY_PATCH_SIZE, hparams.patch_size);
+                get_u32(KEY_IMAGE_CROP_RESOLUTION,    hparams.image_crop_resolution, false);
+                get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false);
+                get_i32(KEY_MINICPMV_VERSION, ctx_clip.minicpmv_version, false); // legacy
+
+            } else if (hparams.has_audio) {
+                get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
+
+            } else {
+                throw std::runtime_error(string_format("%s: neither vision nor audio encoder is present\n", __func__));
+            }

            // default warmup value
            hparams.warmup_image_size = hparams.image_size;
@@ -1855,7 +2099,7 @@ struct clip_model_loader {
                }
            }

-            {
+            if (hparams.has_vision) {
                int idx_mean = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_MEAN);
                int idx_std  = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_STD);
                GGML_ASSERT(idx_mean >= 0 && "image_mean not found");
@@ -1926,24 +2170,56 @@ struct clip_model_loader {
                        hparams.warmup_image_size = hparams.patch_size * 8;
                        get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern);
                    } break;
+                case PROJECTOR_TYPE_LLAMA4:
+                    {
+                        hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor);
+
+                        // borrowed from llava-1.6
+                        const int isize = hparams.image_size;
+                        hparams.image_grid_pinpoints = {
+                            isize,   isize*2, // 336, 672
+                            isize*2, isize,   // 672, 336
+                            isize*2, isize*2, // 672, 672
+                            isize*3, isize,   // 1008, 336
+                            isize,   isize*3, // 336, 1008
+                        };
+                    } break;
+                case PROJECTOR_TYPE_ULTRAVOX:
+                    {
+                        get_u32(KEY_A_PROJ_STACK_FACTOR, hparams.proj_stack_factor);
+                        if (hparams.n_mel_bins != 128) {
+                            throw std::runtime_error(string_format("%s: only 128 mel bins are supported for ultravox\n", __func__));
+                        }
+                        hparams.ffn_op = FFN_GELU_ERF;
+                        log_ffn_op = "gelu_erf"; // temporary solution for logging
+                    } break;
                default:
                    break;
            }

            LOG_INF("%s: projector:          %s\n", __func__, proj_type.c_str());
+            LOG_INF("%s: has_vision_encoder: %d\n", __func__, hparams.has_vision);
+            LOG_INF("%s: has_audio_encoder:  %d\n", __func__, hparams.has_audio);
            LOG_INF("%s: n_embd:             %d\n", __func__, hparams.n_embd);
            LOG_INF("%s: n_head:             %d\n", __func__, hparams.n_head);
            LOG_INF("%s: n_ff:               %d\n", __func__, hparams.n_ff);
            LOG_INF("%s: n_layer:            %d\n", __func__, hparams.n_layer);
-            LOG_INF("%s: projection_dim:     %d\n", __func__, hparams.projection_dim);
-            LOG_INF("%s: image_size:         %d\n", __func__, hparams.image_size);
-            LOG_INF("%s: patch_size:         %d\n", __func__, hparams.patch_size);
-            LOG_INF("\n");
-            LOG_INF("%s: has_llava_proj:     %d\n", __func__, ctx_clip.has_llava_projector);
-            LOG_INF("%s: minicpmv_version:   %d\n", __func__, ctx_clip.minicpmv_version);
-            LOG_INF("%s: proj_scale_factor:  %d\n", __func__, hparams.proj_scale_factor);
-            LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
            LOG_INF("%s: ffn_op:             %s\n", __func__, log_ffn_op.c_str());
+            LOG_INF("%s: projection_dim:     %d\n", __func__, hparams.projection_dim);
+            LOG_INF("\n");
+            if (hparams.has_vision) {
+                LOG_INF("%s: image_size:         %d\n", __func__, hparams.image_size);
+                LOG_INF("%s: patch_size:         %d\n", __func__, hparams.patch_size);
+                LOG_INF("%s: has_llava_proj:     %d\n", __func__, ctx_clip.has_llava_projector);
+                LOG_INF("%s: minicpmv_version:   %d\n", __func__, ctx_clip.minicpmv_version);
+                LOG_INF("%s: proj_scale_factor:  %d\n", __func__, hparams.proj_scale_factor);
+                LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
+            } else if (hparams.has_audio) {
+                LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
+                LOG_INF("%s: proj_stack_factor:  %d\n", __func__, hparams.proj_stack_factor);
+            }
+            LOG_INF("\n");
            LOG_INF("%s: model size:         %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0);
            LOG_INF("%s: metadata size:      %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0);
        }
@@ -1954,6 +2230,9 @@ struct clip_model_loader {
        std::map<std::string, size_t> tensor_offset;
        std::vector<ggml_tensor *> tensors_to_load;

+        // TODO @ngxson : support both audio and video in the future
+        const char * prefix = hparams.has_audio ? "a" : "v";
+
        // get offsets
        for (int64_t i = 0; i < gguf_get_n_tensors(ctx_gguf.get()); ++i) {
            const char * name = gguf_get_tensor_name(ctx_gguf.get(), i);
@@ -1991,47 +2270,47 @@ struct clip_model_loader {

        vision_model.class_embedding = get_tensor(TN_CLASS_EMBD, false);

-        vision_model.pre_ln_w = get_tensor(string_format(TN_LN_PRE, "v", "weight"), false);
-        vision_model.pre_ln_b = get_tensor(string_format(TN_LN_PRE, "v", "bias"),   false);
+        vision_model.pre_ln_w = get_tensor(string_format(TN_LN_PRE, prefix, "weight"), false);
+        vision_model.pre_ln_b = get_tensor(string_format(TN_LN_PRE, prefix, "bias"),   false);

-        vision_model.post_ln_w = get_tensor(string_format(TN_LN_POST, "v", "weight"), false);
-        vision_model.post_ln_b = get_tensor(string_format(TN_LN_POST, "v", "bias"),   false);
+        vision_model.post_ln_w = get_tensor(string_format(TN_LN_POST, prefix, "weight"), false);
+        vision_model.post_ln_b = get_tensor(string_format(TN_LN_POST, prefix, "bias"),   false);

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

-        vision_model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, "v"), false);
+        vision_model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, prefix), false);

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

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

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

            // some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here
            // note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check!
@@ -2173,6 +2452,17 @@ struct clip_model_loader {
                    vision_model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
                    vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                } break;
+            case PROJECTOR_TYPE_ULTRAVOX:
+                {
+                    vision_model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
+                    vision_model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
+                    vision_model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
+                    vision_model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
+                    vision_model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
+                    vision_model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
+                    vision_model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"));
+                    vision_model.mm_norm_mid_w = get_tensor(string_format(TN_MM_NORM_MID, "weight"));
+                } break;
            case PROJECTOR_TYPE_INTERNVL:
                {
                    vision_model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
@@ -2182,6 +2472,12 @@ struct clip_model_loader {
                    vision_model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
                    vision_model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
                } break;
+            case PROJECTOR_TYPE_LLAMA4:
+                {
+                    vision_model.mm_model_proj    = get_tensor(TN_MM_PROJECTOR);
+                    vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "weight"));
+                } break;
            default:
                GGML_ASSERT(false && "unknown projector type");
        }
@@ -2224,13 +2520,19 @@ struct clip_model_loader {
    }

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

        // create a fake batch
        clip_image_f32_batch batch;
        clip_image_f32_ptr img(clip_image_f32_init());
-        img->nx = ctx_clip.vision_model.hparams.warmup_image_size;
-        img->ny = ctx_clip.vision_model.hparams.warmup_image_size;
+        if (hparams.has_vision) {
+            img->nx = hparams.warmup_image_size;
+            img->ny = hparams.warmup_image_size;
+        } else {
+            img->nx = 1024; // TODO @ngxson : use a better default
+            img->ny = hparams.n_mel_bins;
+        }
        img->buf.resize(img->nx * img->ny * 3);
        batch.entries.push_back(std::move(img));

@@ -2328,14 +2630,6 @@ struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_p
    return ctx_clip;
 }

-void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
-    ctx_clip->load_image_size = *load_image_size; // copy
-}
-
-struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) {
-    return &ctx_clip->load_image_size;
-}
-
 struct clip_image_size * clip_image_size_init() {
    struct clip_image_size * load_image_size = new struct clip_image_size();
    load_image_size->width = 448;
@@ -2849,7 +3143,7 @@ private:

    // used by llava 1.6 with custom list of pinpoints
    static clip_image_size select_best_resolution(const std::vector<int32_t> & pinpoints, const clip_image_size & original_size) {
-        std::vector<clip_image_size> possible_resolutions;
+        std::vector<clip_image_size> possible_resolutions; // TODO @ngxson : construct this inside hparams, not here
        for (size_t i = 0; i < pinpoints.size(); i += 2) {
            possible_resolutions.push_back(clip_image_size{pinpoints[i], pinpoints[i+1]});
        }
@@ -2916,12 +3210,6 @@ private:
    }
 };

-// TODO @ngxson : decprecate the load_image_size singleton pattern
-int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
-    const auto inst = llava_uhd::get_slice_instructions(ctx_clip, ctx_clip->load_image_size);
-    return inst.grid_size.width;
-}
-
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) {
@@ -2943,9 +3231,12 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
            normalize_image_u8_to_f32(*imgs[i], *res, ctx->image_mean, ctx->image_std);
            res_imgs->entries.push_back(std::move(res));
        }
+
+        res_imgs->grid_x = inst.grid_size.width;
+        res_imgs->grid_y = inst.grid_size.height;
        return true;
-    }
-    else if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
+
+    } else if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
        clip_image_u8 resized;
        auto patch_size = params.patch_size * 2;
        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, patch_size, params.image_size);
@@ -2971,8 +3262,8 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
        normalize_image_u8_to_f32(resized_image, *img_f32, ctx->image_mean, ctx->image_std);
        res_imgs->entries.push_back(std::move(img_f32));
        return true;
-    }
-    else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
+
+    } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
        clip_image_u8 resized_image;
        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, params.patch_size, params.image_size);
        image_manipulation::bilinear_resize(*img, resized_image, new_size.width, new_size.height);
@@ -2980,6 +3271,22 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
        normalize_image_u8_to_f32(resized_image, *img_f32, ctx->image_mean, ctx->image_std);
        res_imgs->entries.push_back(std::move(img_f32));
        return true;
+
+    } else if (ctx->proj_type == PROJECTOR_TYPE_LLAMA4) {
+        GGML_ASSERT(!params.image_grid_pinpoints.empty());
+        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
+        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
+
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            clip_image_f32_ptr res(clip_image_f32_init());
+            normalize_image_u8_to_f32(*imgs[i], *res, ctx->image_mean, ctx->image_std);
+            res_imgs->entries.push_back(std::move(res));
+        }
+
+        res_imgs->grid_x = inst.grid_size.width;
+        res_imgs->grid_y = inst.grid_size.height;
+        return true;
+
    }

    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
@@ -3098,6 +3405,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
    const auto & params = ctx->vision_model.hparams;

    int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+    int scale_factor = ctx->vision_model.hparams.proj_scale_factor;

    if (ctx->proj_type == PROJECTOR_TYPE_LDP
            || ctx->proj_type == PROJECTOR_TYPE_LDPV2
@@ -3136,6 +3444,12 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
        int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
        int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
        n_patches = n_patches_y*n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
+    } else if (ctx->proj_type == PROJECTOR_TYPE_LLAMA4) {
+        n_patches /= (scale_factor * scale_factor);
+    } else if (ctx->proj_type == PROJECTOR_TYPE_ULTRAVOX) {
+        const int proj_stack_factor = ctx->vision_model.hparams.proj_stack_factor;
+        const int n_len = CLIP_ALIGN(img->nx, proj_stack_factor);
+        n_patches = n_len / proj_stack_factor / 2;
    }

    return n_patches;
@@ -3247,6 +3561,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    }

    // build the inference graph
+    ctx->debug_print_tensors.clear();
    ggml_backend_sched_reset(ctx->sched.get());
    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
    ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);
@@ -3261,8 +3576,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    const int patch_size    = hparams.patch_size;
    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
    const int n_pos = num_patches + (model.class_embedding ? 1 : 0);
-    const int pos_w = ctx->load_image_size.width  / patch_size;
-    const int pos_h = ctx->load_image_size.height / patch_size;
+    const int pos_w = image_size_width  / patch_size;
+    const int pos_h = image_size_height / patch_size;

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

@@ -3292,7 +3607,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    };

    // set input pixel values
-    {
+    if (!imgs.is_audio) {
        size_t nelem = 0;
        for (const auto & img : imgs.entries) {
            nelem += img->nx * img->ny * 3;
@@ -3329,6 +3644,16 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
            }
        }
        set_input_f32("inp_raw", inp_raw);
+
+    } else {
+        // audio input
+        GGML_ASSERT(imgs.entries.size() == 1);
+        const auto & mel_inp = imgs.entries[0];
+        const int n_step = mel_inp->nx;
+        const int n_mel  = mel_inp->ny;
+        std::vector<float> inp_raw(n_step * n_mel);
+        std::memcpy(inp_raw.data(), mel_inp->buf.data(), n_step * n_mel * sizeof(float));
+        set_input_f32("inp_raw", inp_raw);
    }

    // set input per projector
@@ -3525,9 +3850,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
        case PROJECTOR_TYPE_GEMMA3:
        case PROJECTOR_TYPE_IDEFICS3:
        case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_ULTRAVOX:
            {
                // do nothing
            } break;
+        case PROJECTOR_TYPE_LLAMA4:
+            {
+                // set the 2D positions
+                int n_patches_per_col = image_size_width / patch_size;
+                std::vector<int> pos_data(num_patches + 1, 0); // +1 for the [CLS] token
+                // last pos is always kept 0, it's for CLS
+                // dimension H
+                for (int i = 0; i < num_patches; i++) {
+                    pos_data[i] = (i / n_patches_per_col) + 1;
+                }
+                set_input_i32("pos_h", pos_data);
+                // dimension W
+                for (int i = 0; i < num_patches; i++) {
+                    pos_data[i] = (i % n_patches_per_col) + 1;
+                }
+                set_input_i32("pos_w", pos_data);
+            } break;
        default:
            GGML_ABORT("Unknown projector type");
    }
@@ -3548,6 +3891,18 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
        return false;
    }

+    // print debug nodes
+    if (ctx->debug_graph) {
+        LOG_INF("\n\n---\n\n");
+        LOG_INF("\n\nDebug graph:\n\n");
+        for (ggml_tensor * t : ctx->debug_print_tensors) {
+            std::vector<uint8_t> data(ggml_nbytes(t));
+            ggml_backend_tensor_get(t, data.data(), 0, ggml_nbytes(t));
+            print_tensor_shape(t);
+            print_tensor_data(t, data.data(), 3);
+        }
+    }
+
    // the last node is the embedding tensor
    ggml_tensor * embeddings = ggml_graph_node(gf, -1);

@@ -3594,8 +3949,12 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
            return ctx->vision_model.mm_input_proj_w->ne[0];
        case PROJECTOR_TYPE_IDEFICS3:
            return ctx->vision_model.projection->ne[1];
+        case PROJECTOR_TYPE_ULTRAVOX:
+            return ctx->vision_model.mm_2_w->ne[1];
        case PROJECTOR_TYPE_INTERNVL:
            return ctx->vision_model.mm_3_w->ne[1];
+        case PROJECTOR_TYPE_LLAMA4:
+            return ctx->vision_model.mm_model_proj->ne[1];
        default:
            GGML_ABORT("Unknown projector type");
    }
@@ -3624,6 +3983,14 @@ bool clip_is_gemma3(const struct clip_ctx * ctx) {
    return ctx->proj_type == PROJECTOR_TYPE_GEMMA3;
 }

+bool clip_has_vision_encoder(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.has_vision;
+}
+
+bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.has_audio;
+}
+
 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
    clip_image_f32 clip_img;
    clip_img.buf.resize(h * w * 3);
@@ -3644,3 +4011,14 @@ bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img,
 projector_type clip_get_projector_type(const struct clip_ctx * ctx) {
    return ctx->proj_type;
 }
+
+void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel) {
+    clip_image_f32 * audio = new clip_image_f32;
+    audio->nx = n_frames;
+    audio->ny = n_mel;
+    audio->buf.resize(n_frames * n_mel);
+    std::memcpy(audio->buf.data(), mel, n_frames * n_mel * sizeof(float));
+
+    batch->entries.push_back(clip_image_f32_ptr(audio));
+    batch->is_audio = true;
+}
@@ -47,10 +47,6 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
 // this should be equal to the embedding dimension of the text model
 int clip_n_mmproj_embd(const struct clip_ctx * ctx);

-int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
-void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
-struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
-
 struct clip_image_size      * clip_image_size_init(void);
 struct clip_image_u8        * clip_image_u8_init (void);
 struct clip_image_f32       * clip_image_f32_init(void);
@@ -97,3 +93,9 @@ bool clip_is_llava(const struct clip_ctx * ctx);
 bool clip_is_gemma3(const struct clip_ctx * ctx);

 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
+
+// use by audio input
+void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel);
+
+bool clip_has_vision_encoder(const struct clip_ctx * ctx);
+bool clip_has_audio_encoder(const struct clip_ctx * ctx);
@@ -0,0 +1,855 @@
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include "mtmd-audio.h"
+
+//#define MTMD_AUDIO_DEBUG
+
+#define MINIAUDIO_IMPLEMENTATION
+#ifndef MTMD_AUDIO_DEBUG
+#   define MA_NO_ENCODING
+#endif
+#define MA_NO_DEVICE_IO
+#define MA_NO_RESOURCE_MANAGER
+#define MA_NO_NODE_GRAPH
+#define MA_NO_ENGINE
+#define MA_NO_GENERATION
+#define MA_API static
+#include "miniaudio.h"
+
+#define _USE_MATH_DEFINES // for M_PI
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <thread>
+#include <vector>
+#include <fstream>
+#include <algorithm>
+
+// most of the code here is copied from whisper.cpp
+
+// align x to upper multiple of n
+#define _ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
+
+namespace whisper_preprocessor {
+
+#define SIN_COS_N_COUNT WHISPER_N_FFT
+namespace {
+struct whisper_global_cache {
+    // In FFT, we frequently use sine and cosine operations with the same values.
+    // We can use precalculated values to speed up the process.
+    float sin_vals[SIN_COS_N_COUNT];
+    float cos_vals[SIN_COS_N_COUNT];
+
+    // Hann window (Use cosf to eliminate difference)
+    // ref: https://pytorch.org/docs/stable/generated/torch.hann_window.html
+    // ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L147
+    float hann_window[WHISPER_N_FFT];
+
+    whisper_global_cache() {
+        fill_sin_cos_table();
+        fill_hann_window(sizeof(hann_window)/sizeof(hann_window[0]), true, hann_window);
+    }
+
+    void fill_sin_cos_table() {
+        for (int i = 0; i < SIN_COS_N_COUNT; i++) {
+            double theta = (2 * M_PI * i) / SIN_COS_N_COUNT;
+            sin_vals[i] = sinf(theta);
+            cos_vals[i] = cosf(theta);
+        }
+    }
+
+    void fill_hann_window(int length, bool periodic, float * output) {
+        int offset = -1;
+        if (periodic) {
+            offset = 0;
+        }
+        for (int i = 0; i < length; i++) {
+            output[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
+        }
+    }
+} global_cache;
+}
+
+// naive Discrete Fourier Transform
+// input is real-valued
+// output is complex-valued
+static void dft(const float* in, int N, float* out) {
+    const int sin_cos_step = SIN_COS_N_COUNT / N;
+
+    for (int k = 0; k < N; k++) {
+        float re = 0;
+        float im = 0;
+
+        for (int n = 0; n < N; n++) {
+            int idx = (k * n * sin_cos_step) % (SIN_COS_N_COUNT); // t = 2*M_PI*k*n/N
+            re += in[n]*global_cache.cos_vals[idx]; // cos(t)
+            im -= in[n]*global_cache.sin_vals[idx]; // sin(t)
+        }
+
+        out[k*2 + 0] = re;
+        out[k*2 + 1] = im;
+    }
+}
+
+// Cooley-Tukey FFT
+// poor man's implementation - use something better
+// input is real-valued
+// output is complex-valued
+static void fft(float* in, int N, float* out) {
+    if (N == 1) {
+        out[0] = in[0];
+        out[1] = 0;
+        return;
+    }
+
+    const int half_N = N / 2;
+    if (N - half_N*2 == 1) {
+        dft(in, N, out);
+        return;
+    }
+
+    float* even = in + N;
+    for (int i = 0; i < half_N; ++i) {
+        even[i]= in[2*i];
+    }
+    float* even_fft = out + 2 * N;
+    fft(even, half_N, even_fft);
+
+    float* odd = even;
+    for (int i = 0; i < half_N; ++i) {
+        odd[i] = in[2*i + 1];
+    }
+    float* odd_fft = even_fft + N;
+    fft(odd, half_N, odd_fft);
+
+    const int sin_cos_step = SIN_COS_N_COUNT / N;
+    for (int k = 0; k < half_N; k++) {
+        int idx = k * sin_cos_step; // t = 2*M_PI*k/N
+        float re = global_cache.cos_vals[idx]; // cos(t)
+        float im = -global_cache.sin_vals[idx]; // sin(t)
+
+        float re_odd = odd_fft[2*k + 0];
+        float im_odd = odd_fft[2*k + 1];
+
+        out[2*k + 0] = even_fft[2*k + 0] + re*re_odd - im*im_odd;
+        out[2*k + 1] = even_fft[2*k + 1] + re*im_odd + im*re_odd;
+
+        out[2*(k + half_N) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
+        out[2*(k + half_N) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
+    }
+}
+
+static void log_mel_spectrogram_worker_thread(int ith, const float * hann, const std::vector<float> & samples,
+                                              int n_samples, int frame_size, int frame_step, int n_threads,
+                                              const whisper_filters & filters, whisper_mel & mel) {
+    std::vector<float> fft_in(frame_size * 2, 0.0);
+    std::vector<float> fft_out(frame_size * 2 * 2 * 2);
+
+    int n_fft = filters.n_fft;
+    int i = ith;
+
+    // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
+    WHISPER_ASSERT(n_fft == 1 + (frame_size / 2));
+
+    // calculate FFT only when fft_in are not all zero
+    for (; i < std::min(n_samples / frame_step + 1, mel.n_len); i += n_threads) {
+        const int offset = i * frame_step;
+
+        // apply Hann window (~10% faster)
+        for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
+            fft_in[j] = hann[j] * samples[offset + j];
+        }
+
+        // fill the rest with zeros
+        if (n_samples - offset < frame_size) {
+            std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
+        }
+
+        // FFT
+        fft(fft_in.data(), frame_size, fft_out.data());
+
+        // Calculate modulus^2 of complex numbers
+        // Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
+        for (int j = 0; j < n_fft; j++) {
+            fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
+        }
+
+        // mel spectrogram
+        for (int j = 0; j < mel.n_mel; j++) {
+            double sum = 0.0;
+            // unroll loop (suggested by GH user @lunixbochs)
+            int k = 0;
+            for (k = 0; k < n_fft - 3; k += 4) {
+                sum +=
+                        fft_out[k + 0] * filters.data[j * n_fft + k + 0] +
+                        fft_out[k + 1] * filters.data[j * n_fft + k + 1] +
+                        fft_out[k + 2] * filters.data[j * n_fft + k + 2] +
+                        fft_out[k + 3] * filters.data[j * n_fft + k + 3];
+            }
+            // handle n_fft remainder
+            for (; k < n_fft; k++) {
+                sum += fft_out[k] * filters.data[j * n_fft + k];
+            }
+            sum = log10(std::max(sum, 1e-10));
+            mel.data[j * mel.n_len + i] = sum;
+        }
+    }
+
+    // Otherwise fft_out are all zero
+    double sum = log10(1e-10);
+    for (; i < mel.n_len; i += n_threads) {
+        for (int j = 0; j < mel.n_mel; j++) {
+            mel.data[j * mel.n_len + i] = sum;
+        }
+    }
+}
+
+// ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L110-L157
+static bool log_mel_spectrogram(
+        const float * samples,
+        const int   n_samples,
+        const int   /*sample_rate*/,
+        const int   frame_size,
+        const int   frame_step,
+        const int   n_mel,
+        const int   n_threads,
+        const whisper_filters & filters,
+        const bool   debug,
+        whisper_mel & mel) {
+    //const int64_t t_start_us = ggml_time_us();
+
+    // Hann window
+    WHISPER_ASSERT(frame_size == WHISPER_N_FFT && "Unsupported frame_size");
+    const float * hann = global_cache.hann_window;
+
+    // Calculate the length of padding
+    int64_t stage_1_pad = WHISPER_SAMPLE_RATE * 30;
+    int64_t stage_2_pad = frame_size / 2;
+
+    // Initialize a vector and copy data from C array to it.
+    std::vector<float> samples_padded;
+    samples_padded.resize(n_samples + stage_1_pad + stage_2_pad * 2);
+    std::copy(samples, samples + n_samples, samples_padded.begin() + stage_2_pad);
+
+    // pad 30 seconds of zeros at the end of audio (480,000 samples) + reflective pad 200 samples at the end of audio
+    std::fill(samples_padded.begin() + n_samples + stage_2_pad, samples_padded.begin() + n_samples + stage_1_pad + 2 * stage_2_pad, 0);
+
+    // reflective pad 200 samples at the beginning of audio
+    std::reverse_copy(samples + 1, samples + 1 + stage_2_pad, samples_padded.begin());
+
+    mel.n_mel     = n_mel;
+    // https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/SpectralOps.cpp#L936
+    // Calculate number of frames + remove the last frame
+    mel.n_len     = (samples_padded.size() - frame_size) / frame_step;
+    // Calculate semi-padded sample length to ensure compatibility
+    mel.n_len_org = 1 + (n_samples + stage_2_pad - frame_size) / frame_step;
+    mel.data.resize(mel.n_mel * mel.n_len);
+
+    {
+        std::vector<std::thread> workers(n_threads - 1);
+        for (int iw = 0; iw < n_threads - 1; ++iw) {
+            workers[iw] = std::thread(
+                    log_mel_spectrogram_worker_thread, iw + 1, hann, std::cref(samples_padded),
+                    n_samples + stage_2_pad, frame_size, frame_step, n_threads,
+                    std::cref(filters), std::ref(mel));
+        }
+
+        // main thread
+        log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples + stage_2_pad, frame_size, frame_step, n_threads, filters, mel);
+
+        for (int iw = 0; iw < n_threads - 1; ++iw) {
+            workers[iw].join();
+        }
+    }
+
+    // clamping and normalization
+    double mmax = -1e20;
+    for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
+        if (mel.data[i] > mmax) {
+            mmax = mel.data[i];
+        }
+    }
+
+    mmax -= 8.0;
+
+    for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
+        if (mel.data[i] < mmax) {
+            mel.data[i] = mmax;
+        }
+
+        mel.data[i] = (mel.data[i] + 4.0)/4.0;
+    }
+
+    // Dump log_mel_spectrogram
+    if (debug) {
+        std::ofstream outFile("log_mel_spectrogram.json");
+        outFile << "[";
+        for (uint64_t i = 0; i < mel.data.size() - 1; i++) {
+            outFile << mel.data[i] << ", ";
+        }
+        outFile << mel.data[mel.data.size() - 1] << "]";
+        outFile.close();
+    }
+
+    return true;
+}
+
+bool preprocess_audio(
+        const float * samples,
+        size_t n_samples,
+        const whisper_filters & filters,
+        std::vector<whisper_mel> & output) {
+
+    if (n_samples == 0) {
+        // empty audio
+        return false;
+    }
+
+    whisper_mel out_full;
+    bool ok = log_mel_spectrogram(
+                samples,
+                n_samples,
+                COMMON_SAMPLE_RATE,
+                WHISPER_N_FFT,
+                WHISPER_HOP_LENGTH,
+                filters.n_mel,
+                4, // n_threads
+                filters,
+                false, // debug
+                out_full);
+    if (!ok) {
+        return false;
+    }
+
+    // because the cgraph in clip.cpp only accepts 3000 frames each, we need to split the mel
+    // we always expect the mel to have 3000 silent frames at the end
+    // printf("n_len %d\n", out_full.n_len);
+    const size_t frames_per_chunk = 3000;
+    GGML_ASSERT((size_t)out_full.n_len > frames_per_chunk);
+    for (size_t off = 0; off < (size_t)out_full.n_len; off += frames_per_chunk) {
+        int n_len = std::min(frames_per_chunk, (size_t)out_full.n_len - off);
+        if ((size_t)n_len < frames_per_chunk) {
+            break; // last uncomplete chunk will always be a padded chunk, safe to ignore
+        }
+
+        whisper_mel out_chunk;
+        out_chunk.n_len     = n_len;
+        out_chunk.n_mel     = out_full.n_mel;
+        out_chunk.n_len_org = out_full.n_mel; // unused
+        out_chunk.data.reserve(out_chunk.n_mel * out_chunk.n_len);
+
+        for (int i = 0; i < out_full.n_mel; i++) {
+            auto src = out_full.data.begin() + i*out_full.n_len + off;
+            out_chunk.data.insert(out_chunk.data.end(), src, src + frames_per_chunk);
+        }
+
+        output.push_back(std::move(out_chunk));
+    }
+
+    return true;
+}
+
+} // namespace whisper_preprocessor
+
+
+namespace audio_helpers {
+
+bool is_audio_file(const char * buf, size_t len) {
+    if (len < 12) {
+        return false;
+    }
+
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    bool is_wav = memcmp(buf, "RIFF", 4) == 0 && memcmp(buf + 8, "WAVE", 4) == 0;
+    bool is_mp3 = len >= 3 && (
+        memcmp(buf, "ID3", 3) == 0 ||
+        // Check for MPEG sync word (simplified check)
+        ((unsigned char)buf[0] == 0xFF && ((unsigned char)buf[1] & 0xE0) == 0xE0)
+    );
+    bool is_flac = memcmp(buf, "fLaC", 4) == 0;
+
+    return is_wav || is_mp3 || is_flac;
+}
+
+// returns true if the buffer is a valid audio file
+bool decode_audio_from_buf(const unsigned char * buf_in, size_t len, int target_sampler_rate, std::vector<float> & pcmf32_mono) {
+    ma_result result;
+    const int channels = 1;
+    ma_decoder_config decoder_config = ma_decoder_config_init(ma_format_f32, channels, target_sampler_rate);
+    ma_decoder decoder;
+
+    result = ma_decoder_init_memory(buf_in, len, &decoder_config, &decoder);
+    if (result != MA_SUCCESS) {
+        return false;
+    }
+
+    ma_uint64 frame_count;
+    ma_uint64 frames_read;
+    result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+    pcmf32_mono.resize(frame_count);
+    result = ma_decoder_read_pcm_frames(&decoder, pcmf32_mono.data(), frame_count, &frames_read);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+#ifdef MTMD_AUDIO_DEBUG
+    // save audio to wav file
+    ma_encoder_config config = ma_encoder_config_init(ma_encoding_format_wav, ma_format_f32, 1, target_sampler_rate);
+    ma_encoder encoder;
+    ma_encoder_init_file("output.wav", &config, &encoder);
+    ma_encoder_write_pcm_frames(&encoder, pcmf32_mono.data(), pcmf32_mono.size(), &frames_read);
+    ma_encoder_uninit(&encoder);
+#endif
+
+    ma_decoder_uninit(&decoder);
+    return true;
+}
+
+} // namespace wav_utils
+
+
+// precalculated mel filter banks
+// values are multiplied by 1000.0 to save space, and will be divided by 1000.0 in the end of the function
+//
+// generated from python code:
+//
+// from numpy import load
+// data = load('mel_filters.npz')
+// lst = data.files
+// for item in lst:
+//   print(item)
+//   print(data[item].shape)
+//   n_mel = data[item].shape[0]
+//   n_fft = data[item].shape[1]
+//   for i, row in enumerate(data[item]):
+//     for j, val in enumerate(row):
+//       val = val * 1000.0
+//       if val != 0:
+//         print(f"data[{i*n_fft + j}] = {val:.6f};")
+
+namespace whisper_precalc_filters {
+
+whisper_preprocessor::whisper_filters get_128_bins() {
+    whisper_preprocessor::whisper_filters filters;
+    filters.n_mel = 128;
+    filters.n_fft = 201;
+    std::vector data(filters.n_mel * filters.n_fft, 0.0f);
+
+    data[1] = 12.37398665;
+    data[202] = 30.39256483;
+    data[404] = 24.74797331;
+    data[605] = 18.01857911;
+    data[807] = 37.12195903;
+    data[1008] = 5.64459199;
+    data[1009] = 6.72939420;
+    data[1210] = 36.03715822;
+    data[1412] = 19.10337992;
+    data[1613] = 23.66316877;
+    data[1815] = 31.47736564;
+    data[2016] = 11.28918398;
+    data[2017] = 1.08480197;
+    data[2218] = 41.68175161;
+    data[2420] = 13.45878839;
+    data[2621] = 29.30776216;
+    data[2823] = 25.83277412;
+    data[3024] = 16.93377644;
+    data[3226] = 38.20675984;
+    data[3427] = 4.55979025;
+    data[3428] = 7.81419594;
+    data[3629] = 34.95235741;
+    data[3831] = 20.18818259;
+    data[4032] = 22.57836796;
+    data[4234] = 32.56217018;
+    data[4435] = 10.20438317;
+    data[4436] = 2.16960395;
+    data[4637] = 40.59694707;
+    data[4839] = 14.54358920;
+    data[5040] = 28.22295949;
+    data[5242] = 26.91757679;
+    data[5443] = 15.84897563;
+    data[5645] = 39.29156065;
+    data[5846] = 3.47498828;
+    data[5847] = 8.89899861;
+    data[6048] = 33.86755288;
+    data[6250] = 21.27298526;
+    data[6451] = 21.49356715;
+    data[6653] = 33.64697099;
+    data[6854] = 9.11958050;
+    data[6855] = 3.25440569;
+    data[7056] = 39.51214626;
+    data[7258] = 15.62839188;
+    data[7459] = 27.13815868;
+    data[7661] = 28.00237760;
+    data[7862] = 14.76417296;
+    data[8064] = 40.37636518;
+    data[8265] = 2.38068704;
+    data[8266] = 10.20263787;
+    data[8467] = 31.61146119;
+    data[8669] = 24.54700135;
+    data[8870] = 15.32919332;
+    data[8871] = 1.66583748;
+    data[9072] = 36.72905266;
+    data[9274] = 20.09709924;
+    data[9475] = 16.93102531;
+    data[9476] = 2.90265540;
+    data[9677] = 32.84499049;
+    data[9879] = 23.52004871;
+    data[10080] = 11.03894413;
+    data[10081] = 10.72582975;
+    data[10282] = 22.71829173;
+    data[10484] = 32.27872774;
+    data[10685] = 0.11626833;
+    data[10686] = 22.85348251;
+    data[10887] = 8.56344029;
+    data[10888] = 14.97978810;
+    data[11089] = 15.51398356;
+    data[11090] = 8.51490628;
+    data[11291] = 21.10680379;
+    data[11292] = 3.32652032;
+    data[11493] = 25.47064796;
+    data[11695] = 27.35907957;
+    data[11896] = 0.65853616;
+    data[11897] = 23.83812517;
+    data[12098] = 3.44359246;
+    data[12099] = 21.22455277;
+    data[12300] = 5.35842171;
+    data[12301] = 19.42555793;
+    data[12502] = 6.49324711;
+    data[12503] = 18.35542172;
+    data[12704] = 6.93138083;
+    data[12705] = 17.93504693;
+    data[12906] = 6.74968259;
+    data[12907] = 18.09151843;
+    data[13108] = 6.01899112;
+    data[13109] = 18.75767298;
+    data[13310] = 4.80452832;
+    data[13311] = 19.87172849;
+    data[13512] = 3.16627859;
+    data[13513] = 21.37690969;
+    data[13514] = 1.25317345;
+    data[13714] = 1.15934468;
+    data[13715] = 20.80361731;
+    data[13716] = 4.04486805;
+    data[13917] = 17.55363122;
+    data[13918] = 7.08320038;
+    data[14119] = 14.07538634;
+    data[14120] = 10.32655034;
+    data[14321] = 10.40921453;
+    data[14322] = 13.73696327;
+    data[14523] = 6.59187697;
+    data[14524] = 17.27988198;
+    data[14525] = 1.46804214;
+    data[14725] = 2.65681883;
+    data[14726] = 18.09193194;
+    data[14727] = 5.85655728;
+    data[14928] = 13.34277913;
+    data[14929] = 10.28267574;
+    data[15130] = 8.56800377;
+    data[15131] = 14.72230814;
+    data[15132] = 1.04039861;
+    data[15332] = 3.79085587;
+    data[15333] = 17.14678481;
+    data[15334] = 6.11609267;
+    data[15535] = 11.75929047;
+    data[15536] = 11.13393717;
+    data[15737] = 6.43857848;
+    data[15738] = 16.07806236;
+    data[15739] = 4.23917221;
+    data[15939] = 1.19989377;
+    data[15940] = 12.75671553;
+    data[15941] = 9.65298992;
+    data[16142] = 7.06935255;
+    data[16143] = 14.94054683;
+    data[16144] = 4.19024844;
+    data[16344] = 1.51483389;
+    data[16345] = 12.00899947;
+    data[16346] = 9.84823331;
+    data[16547] = 6.10224018;
+    data[16548] = 15.33857174;
+    data[16549] = 5.57676842;
+    data[16749] = 0.36827257;
+    data[16750] = 9.89749376;
+    data[16751] = 11.35340426;
+    data[16752] = 2.05122307;
+    data[16952] = 3.89297144;
+    data[16953] = 12.97352277;
+    data[16954] = 8.06631614;
+    data[17155] = 6.74493238;
+    data[17156] = 13.85874674;
+    data[17157] = 5.41190524;
+    data[17357] = 0.74220158;
+    data[17358] = 8.98779090;
+    data[17359] = 11.37871388;
+    data[17360] = 3.32958088;
+    data[17560] = 2.82313535;
+    data[17561] = 10.68049297;
+    data[17562] = 9.43340641;
+    data[17563] = 1.76325557;
+    data[17763] = 4.39018616;
+    data[17764] = 11.87758986;
+    data[17765] = 7.97005836;
+    data[17766] = 0.66104700;
+    data[17966] = 5.49466675;
+    data[17967] = 12.62953598;
+    data[17968] = 6.93987962;
+    data[18169] = 6.18401915;
+    data[18170] = 12.93473132;
+    data[18171] = 6.29778765;
+    data[18371] = 0.02325210;
+    data[18372] = 6.50206627;
+    data[18373] = 12.32661773;
+    data[18374] = 6.00216538;
+    data[18574] = 0.31548753;
+    data[18575] = 6.48925547;
+    data[18576] = 12.04130240;
+    data[18577] = 6.01462880;
+    data[18777] = 0.29979556;
+    data[18778] = 6.18288014;
+    data[18779] = 12.04272825;
+    data[18780] = 6.29981188;
+    data[18781] = 0.55689598;
+    data[18980] = 0.01120471;
+    data[18981] = 5.61729167;
+    data[18982] = 11.22337859;
+    data[18983] = 6.82516303;
+    data[18984] = 1.35264499;
+    data[19184] = 4.82410006;
+    data[19185] = 10.16623247;
+    data[19186] = 7.56075513;
+    data[19187] = 2.34590308;
+    data[19387] = 3.83235747;
+    data[19388] = 8.92296247;
+    data[19389] = 8.47910438;
+    data[19390] = 3.50978645;
+    data[19590] = 2.66873185;
+    data[19591] = 7.51965167;
+    data[19592] = 9.55500547;
+    data[19593] = 4.81966138;
+    data[19594] = 0.08431751;
+    data[19793] = 1.35767367;
+    data[19794] = 5.98019501;
+    data[19795] = 10.60271543;
+    data[19796] = 6.25298498;
+    data[19797] = 1.74059917;
+    data[19997] = 4.32644226;
+    data[19998] = 8.73131864;
+    data[19999] = 7.78916525;
+    data[20000] = 3.48923868;
+    data[20200] = 2.57835095;
+    data[20201] = 6.77582854;
+    data[20202] = 9.40941647;
+    data[20203] = 5.31194592;
+    data[20204] = 1.21447595;
+    data[20403] = 0.75411191;
+    data[20404] = 4.75395704;
+    data[20405] = 8.75380263;
+    data[20406] = 7.19209015;
+    data[20407] = 3.28754401;
+    data[20607] = 2.68179690;
+    data[20608] = 6.49331464;
+    data[20609] = 9.11457930;
+    data[20610] = 5.39387390;
+    data[20611] = 1.67316827;
+    data[20810] = 0.57394296;
+    data[20811] = 4.20600036;
+    data[20812] = 7.83805829;
+    data[20813] = 7.52023002;
+    data[20814] = 3.97470826;
+    data[20815] = 0.42918732;
+    data[21014] = 1.90464477;
+    data[21015] = 5.36569161;
+    data[21016] = 8.82673822;
+    data[21017] = 6.27609482;
+    data[21018] = 2.89750961;
+    data[21218] = 2.89885257;
+    data[21219] = 6.19694078;
+    data[21220] = 8.56699049;
+    data[21221] = 5.34748193;
+    data[21222] = 2.12797290;
+    data[21421] = 0.44750227;
+    data[21422] = 3.59030394;
+    data[21423] = 6.73310598;
+    data[21424] = 7.77023612;
+    data[21425] = 4.70231380;
+    data[21426] = 1.63439126;
+    data[21625] = 1.01536023;
+    data[21626] = 4.01018746;
+    data[21627] = 7.00501446;
+    data[21628] = 7.23442994;
+    data[21629] = 4.31095669;
+    data[21630] = 1.38748321;
+    data[21829] = 1.33348850;
+    data[21830] = 4.18730825;
+    data[21831] = 7.04112789;
+    data[21832] = 6.93188375;
+    data[21833] = 4.14605811;
+    data[21834] = 1.36023236;
+    data[22033] = 1.42879714;
+    data[22034] = 4.14824858;
+    data[22035] = 6.86769979;
+    data[22036] = 6.83705276;
+    data[22037] = 4.18239459;
+    data[22038] = 1.52773573;
+    data[22237] = 1.32610439;
+    data[22238] = 3.91751388;
+    data[22239] = 6.50892360;
+    data[22240] = 6.92639686;
+    data[22241] = 4.39672917;
+    data[22242] = 1.86706171;
+    data[22441] = 1.04827771;
+    data[22442] = 3.51767405;
+    data[22443] = 5.98707050;
+    data[22444] = 7.17824046;
+    data[22445] = 4.76767914;
+    data[22446] = 2.35711760;
+    data[22645] = 0.61636406;
+    data[22646] = 2.96949223;
+    data[22647] = 5.32262027;
+    data[22648] = 7.57265091;
+    data[22649] = 5.27558755;
+    data[22650] = 2.97852419;
+    data[22651] = 0.68146095;
+    data[22849] = 0.04971400;
+    data[22850] = 2.29204819;
+    data[22851] = 4.53438237;
+    data[22852] = 6.77671656;
+    data[22853] = 5.90240723;
+    data[22854] = 3.71349836;
+    data[22855] = 1.52458926;
+    data[23054] = 1.50285335;
+    data[23055] = 3.63961048;
+    data[23056] = 5.77636715;
+    data[23057] = 6.63159089;
+    data[23058] = 4.54574358;
+    data[23059] = 2.45989650;
+    data[23060] = 0.37404924;
+    data[23258] = 0.61795861;
+    data[23259] = 2.65410915;
+    data[23260] = 4.69025923;
+    data[23261] = 6.72641024;
+    data[23262] = 5.46034705;
+    data[23263] = 3.47270933;
+    data[23264] = 1.48507138;
+    data[23463] = 1.59233576;
+    data[23464] = 3.53261665;
+    data[23465] = 5.47289755;
+    data[23466] = 6.44368259;
+    data[23467] = 4.54962999;
+    data[23468] = 2.65557761;
+    data[23469] = 0.76152512;
+    data[23667] = 0.46749352;
+    data[23668] = 2.31641904;
+    data[23669] = 4.16534441;
+    data[23670] = 6.01426978;
+    data[23671] = 5.67844696;
+    data[23672] = 3.87357362;
+    data[23673] = 2.06870004;
+    data[23674] = 0.26382666;
+    data[23872] = 1.05349103;
+    data[23873] = 2.81536230;
+    data[23874] = 4.57723346;
+    data[23875] = 6.33910485;
+    data[23876] = 5.12815686;
+    data[23877] = 3.40826320;
+    data[23878] = 1.68837002;
+    data[24077] = 1.43350090;
+    data[24078] = 3.11241671;
+    data[24079] = 4.79133241;
+    data[24080] = 6.40943693;
+    data[24081] = 4.77052201;
+    data[24082] = 3.13160778;
+    data[24083] = 1.49269309;
+    data[24281] = 0.02932359;
+    data[24282] = 1.62918994;
+    data[24283] = 3.22905602;
+    data[24284] = 4.82892245;
+    data[24285] = 6.14671456;
+    data[24286] = 4.58496623;
+    data[24287] = 3.02321767;
+    data[24288] = 1.46146910;
+    data[24486] = 0.13601698;
+    data[24487] = 1.66055572;
+    data[24488] = 3.18509457;
+    data[24489] = 4.70963307;
+    data[24490] = 6.04072399;
+    data[24491] = 4.55250870;
+    data[24492] = 3.06429295;
+    data[24493] = 1.57607743;
+    data[24494] = 0.08786193;
+    data[24691] = 0.09328097;
+    data[24692] = 1.54603878;
+    data[24693] = 2.99879676;
+    data[24694] = 4.45155473;
+    data[24695] = 5.90431225;
+    data[24696] = 4.65566106;
+    data[24697] = 3.23751615;
+    data[24698] = 1.81937125;
+    data[24699] = 0.40122634;
+    data[24897] = 1.30262633;
+    data[24898] = 2.68698297;
+    data[24899] = 4.07133950;
+    data[24900] = 5.45569602;
+    data[24901] = 4.87832492;
+    data[24902] = 3.52695142;
+    data[24903] = 2.17557792;
+    data[24904] = 0.82420459;
+    data[25102] = 0.94595028;
+    data[25103] = 2.26512621;
+    data[25104] = 3.58430226;
+    data[25105] = 4.90347855;
+    data[25106] = 5.20569785;
+    data[25107] = 3.91795207;
+    data[25108] = 2.63020652;
+    data[25109] = 1.34246063;
+    data[25110] = 0.05471494;
+    data[25307] = 0.49037894;
+    data[25308] = 1.74744334;
+    data[25309] = 3.00450763;
+    data[25310] = 4.26157191;
+    data[25311] = 5.51863620;
+    data[25312] = 4.39707236;
+    data[25313] = 3.16995848;
+    data[25314] = 1.94284460;
+    data[25315] = 0.71573065;
+    data[25513] = 1.14698056;
+    data[25514] = 2.34485767;
+    data[25515] = 3.54273478;
+    data[25516] = 4.74061165;
+    data[25517] = 4.95198462;
+    data[25518] = 3.78264743;
+    data[25519] = 2.61331047;
+    data[25520] = 1.44397374;
+    data[25521] = 0.27463681;
+    data[25718] = 0.47569509;
+    data[25719] = 1.61717169;
+    data[25720] = 2.75864848;
+    data[25721] = 3.90012516;
+    data[25722] = 5.04160160;
+    data[25723] = 4.45712078;
+    data[25724] = 3.34284059;
+    data[25725] = 2.22856039;
+    data[25726] = 1.11428020;
+
+    for (auto & val : data) {
+        val /= 1000.0f;
+    }
+
+    filters.data = std::move(data);
+    return filters;
+}
+
+} // namespace whisper_precalc_filters
@@ -0,0 +1,62 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <cstdint>
+#include <vector>
+#include <string>
+
+#define WHISPER_ASSERT GGML_ASSERT
+
+#define WHISPER_SAMPLE_RATE 16000
+#define WHISPER_N_FFT       400
+#define WHISPER_HOP_LENGTH  160
+#define WHISPER_CHUNK_SIZE  30
+
+#define COMMON_SAMPLE_RATE 16000
+
+namespace whisper_preprocessor {
+
+struct whisper_mel {
+    int n_len;
+    int n_len_org;
+    int n_mel;
+
+    std::vector<float> data;
+};
+
+struct whisper_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+extern bool preprocess_audio(
+        const float * samples,
+        size_t n_samples,
+        const whisper_filters & filters,
+        std::vector<whisper_mel> & output);
+
+} // namespace whisper_preprocessor
+
+
+// TODO @ngxson : move this helper to mtmd-helpers.cpp
+namespace audio_helpers {
+
+extern bool is_audio_file(const char * buf, size_t len);
+
+extern bool decode_audio_from_buf(
+        const unsigned char * buf_in,
+        size_t len,
+        int target_sampler_rate,
+        std::vector<float> & pcmf32_mono);
+
+} // namespace audio_helpers
+
+
+namespace whisper_precalc_filters {
+
+extern whisper_preprocessor::whisper_filters get_128_bins();
+
+} // namespace whisper_precalc_filters
@@ -37,10 +37,10 @@ static volatile bool g_is_interrupted = false;
 static void show_additional_info(int /*argc*/, char ** argv) {
    LOG(
        "Experimental CLI for multimodal\n\n"
-        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
+        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> --audio <audio> -p <prompt>\n\n"
        "  -m and --mmproj are required\n"
        "  -hf user/repo can replace both -m and --mmproj in most cases\n"
-        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n"
+        "  --image, --audio and -p are optional, if NOT provided, the CLI will run in chat mode\n"
        "  to disable using GPU for mmproj model, add --no-mmproj-offload\n",
        argv[0]
    );
@@ -142,7 +142,7 @@ struct mtmd_cli_context {
        );
    }

-    bool load_image(const std::string & fname) {
+    bool load_media(const std::string & fname) {
        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(fname.c_str()));
        if (!bmp.ptr) {
            return false;
@@ -243,7 +243,7 @@ int main(int argc, char ** argv) {
    common_params params;
    params.sampling.temp = 0.2; // lower temp by default for better quality

-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MTMD, show_additional_info)) {
        return 1;
    }

@@ -283,14 +283,14 @@ int main(int argc, char ** argv) {

    if (is_single_turn) {
        g_is_generating = true;
-        if (params.prompt.find("<__image__>") == std::string::npos) {
-            params.prompt += " <__image__>";
+        if (params.prompt.find(mtmd_default_marker()) == std::string::npos) {
+            params.prompt += mtmd_default_marker();
        }
        common_chat_msg msg;
        msg.role = "user";
        msg.content = params.prompt;
        for (const auto & image : params.image) {
-            if (!ctx.load_image(image)) {
+            if (!ctx.load_media(image)) {
                return 1; // error is already printed by libmtmd
            }
        }
@@ -303,7 +303,12 @@ int main(int argc, char ** argv) {

    } else {
        LOG("\n Running in chat mode, available commands:");
-        LOG("\n   /image <path>    load an image");
+        if (mtmd_support_vision(ctx.ctx_vision.get())) {
+            LOG("\n   /image <path>    load an image");
+        }
+        if (mtmd_support_audio(ctx.ctx_vision.get())) {
+            LOG("\n   /audio <path>    load an audio");
+        }
        LOG("\n   /clear           clear the chat history");
        LOG("\n   /quit or /exit   exit the program");
        LOG("\n");
@@ -333,15 +338,17 @@ int main(int argc, char ** argv) {
                continue;
            }
            g_is_generating = true;
-            if (line == "/image" || line.find("/image ") == 0) {
+            bool is_image = line == "/image" || line.find("/image ") == 0;
+            bool is_audio = line == "/audio" || line.find("/audio ") == 0;
+            if (is_image || is_audio) {
                if (line.size() < 8) {
-                    LOG_ERR("ERR: Missing image filename\n");
+                    LOG_ERR("ERR: Missing media filename\n");
                    continue;
                }
-                std::string image = line.substr(7);
-                if (ctx.load_image(image)) {
-                    LOG("Image %s loaded\n", image.c_str());
-                    content += "<__image__>";
+                std::string media_path = line.substr(7);
+                if (ctx.load_media(media_path)) {
+                    LOG("%s %s loaded\n", media_path.c_str(), is_image ? "image" : "audio");
+                    content += mtmd_default_marker();
                }
                // else, error is already printed by libmtmd
                continue;
@@ -149,13 +149,10 @@ int32_t mtmd_helper_decode_image_chunk(
        llama_seq_id seq_id,
        int32_t n_batch,
        llama_pos * new_n_past) {
-    if (mtmd_input_chunk_get_type(chunk) != MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-        LOG_ERR("failed to decode image chunk: input chunk not of image type\n");
-        return -1;
-    }
-    const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
-    if (!image_tokens) {
-        LOG_ERR("failed to decode image chunk: image tokens are null\n");
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+    const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_ERR("failed to decode chunk: input chunk not of image/audio type\n");
        return -1;
    }

@@ -163,15 +160,23 @@ int32_t mtmd_helper_decode_image_chunk(
    int n_mmproj_embd = llama_model_n_embd(model);
    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;

-    int32_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens);
+    int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
    int32_t i_batch = 0;
    int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
    decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);

-    const int nx = mtmd_image_tokens_get_nx(image_tokens);
-    const int ny = mtmd_image_tokens_get_ny(image_tokens);
-
    if (mtmd_decode_use_mrope(ctx)) {
+        const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+        if (chunk_type != MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            LOG_ERR("failed to decode chunk: M-RoPE only accepts image chunk\n");
+            return -1;
+        }
+        if (!image_tokens) {
+            LOG_ERR("failed to decode chunk: image tokens are null\n");
+            return -1;
+        }
+        const int nx = mtmd_image_tokens_get_nx(image_tokens);
+        const int ny = mtmd_image_tokens_get_ny(image_tokens);
        batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
    } else {
        batch_embd.set_position_normal(n_past, seq_id);
@@ -187,22 +192,22 @@ int32_t mtmd_helper_decode_image_chunk(
        int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
        llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);

-        LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+        LOG_INF("decoding %s batch %d/%d, n_tokens_batch = %d\n", name, i_batch+1, n_img_batches, n_tokens_batch);

        int64_t t1 = ggml_time_ms();
        int32_t ret = llama_decode(lctx, batch_embd_view);
        if (ret != 0) {
-            LOG_ERR("failed to decode image\n");
+            LOG_ERR("failed to decode %s\n", name);
            llama_set_causal_attn(lctx, true); // restore causal attn
            return ret;
        }

-        LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
+        LOG_INF("%s decoded (batch %d/%d) in %" PRId64 " ms\n", name, i_batch+1, n_img_batches, ggml_time_ms() - t1);

        i_batch++;
    }

-    n_past += mtmd_image_tokens_get_n_pos(image_tokens);
+    n_past += mtmd_input_chunk_get_n_pos(chunk);
    *new_n_past = n_past;

    if (mtmd_decode_use_non_causal(ctx)) {
@@ -231,12 +236,14 @@ int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
        while (i < n_tokens) { // split into batches
            text_batch.n_tokens = 0; // clear the batch
            for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
+                int32_t j = text_batch.n_tokens;
+                text_batch.token   [j]    = tokens[i];
+                text_batch.pos     [j]    = n_past++;
+                text_batch.n_seq_id[j]    = 1;
+                text_batch.seq_id  [j][0] = seq_id;
+                text_batch.logits  [j]    = false;
+
                text_batch.n_tokens++;
-                text_batch.token   [i]    = tokens[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;
            }
            bool is_last_token = (i == n_tokens);
            if (logits_last && is_last_token) {
@@ -251,25 +258,25 @@ int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
            *new_n_past += text_batch.n_tokens;
        }

-    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-        const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE || chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
        int64_t t0 = ggml_time_ms();

-        LOG_INF("encoding image or slice...\n");
+        LOG_INF("encoding %s slice...\n", name);

-        ret = mtmd_encode(ctx, image_tokens);
+        ret = mtmd_encode_chunk(ctx, chunk);
        if (ret != 0) {
-            LOG_ERR("failed to encode image\n");
+            LOG_ERR("failed to encode %s slice\n", name);
            llama_batch_free(text_batch);
            return ret;
        }

-        LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+        LOG_INF("%s slice encoded in %" PRId64 " ms\n", name, ggml_time_ms() - t0);

        float * embd = mtmd_get_output_embd(ctx);
        ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past);
        if (ret != 0) {
-            LOG_ERR("failed to decode image\n");
+            LOG_ERR("failed to decode %s\n", name);
            llama_batch_free(text_batch);
            return ret;
        }
@@ -1,6 +1,7 @@
 #include "clip.h"
 #include "clip-impl.h"
 #include "mtmd.h"
+#include "mtmd-audio.h"

 #include "llama.h"

@@ -19,17 +20,49 @@ struct mtmd_bitmap {
    uint32_t ny;
    std::vector<unsigned char> data;
    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
+    bool is_audio = false; // true if the bitmap is audio
 };

-struct mtmd_image_tokens_deleter {
-    void operator()(mtmd_image_tokens * val); // forward declaration
+struct mtmd_image_tokens {
+    uint32_t nx; // number of tokens in x direction
+    uint32_t ny; // number of tokens in y direction
+    bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
+    uint32_t n_tokens() const { return nx * ny; }
+    clip_image_f32_batch batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
+
+    mtmd_image_tokens clone() {
+        return mtmd_image_tokens{
+            nx,
+            ny,
+            use_mrope_pos,
+            batch_f32.clone(),
+            id
+        };
+    }
 };
-using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
+using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens>;
+
+struct mtmd_audio_tokens {
+    uint32_t n_tokens; // number of tokens
+    clip_image_f32_batch batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
+
+    mtmd_audio_tokens clone() {
+        return mtmd_audio_tokens{
+            n_tokens,
+            batch_f32.clone(),
+            id
+        };
+    }
+};
+using mtmd_audio_tokens_ptr = std::unique_ptr<mtmd_audio_tokens>;

 struct mtmd_input_chunk {
    mtmd_input_chunk_type type;
    std::vector<llama_token> tokens_text;
    mtmd_image_tokens_ptr tokens_image;
+    mtmd_audio_tokens_ptr tokens_audio;
 };

 struct mtmd_input_chunks {
@@ -42,9 +75,14 @@ enum mtmd_slice_tmpl {
    MTMD_SLICE_TMPL_NONE,
    MTMD_SLICE_TMPL_MINICPMV_2_5,
    MTMD_SLICE_TMPL_MINICPMV_2_6,
+    MTMD_SLICE_TMPL_LLAMA4,
    // TODO @ngxson : add support for idefics (SmolVLM)
 };

+const char * mtmd_default_marker() {
+    return "<__media__>";
+}
+
 mtmd_context_params mtmd_context_params_default() {
    mtmd_context_params params;
    params.use_gpu = true;
@@ -52,6 +90,7 @@ mtmd_context_params mtmd_context_params_default() {
    params.n_threads = 4;
    params.verbosity = GGML_LOG_LEVEL_INFO;
    params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
+    params.media_marker = mtmd_default_marker();
    return params;
 }

@@ -62,20 +101,29 @@ struct mtmd_context {

    bool print_timings;
    int n_threads;
-    std::string image_marker;
+    std::string media_marker;
+    bool has_vision;
+    bool has_audio;

-    // for minicpmv, we need special tokens in-between slices
+    // for llava-uhd style models, we need special tokens in-between slices
+    // minicpmv calls them "slices", llama 4 calls them "tiles"
    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
-    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
-    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice start
+    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice end
+    llama_token tok_sli_img_mid   = LLAMA_TOKEN_NULL; // between 2 slices
    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
+    bool        tok_row_end_trail = false;
+    bool        ov_img_first      = false;

    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE

+    // for whisper, we pre-calculate the mel filter bank
+    whisper_preprocessor::whisper_filters w_filters;
+
    // TODO @ngxson : add timings

    mtmd_context(const char * mmproj_fname,
@@ -84,8 +132,12 @@ struct mtmd_context {
        text_model   (text_model),
        print_timings(ctx_params.print_timings),
        n_threads    (ctx_params.n_threads),
-        image_marker (ctx_params.image_marker)
+        media_marker (ctx_params.media_marker)
    {
+        if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
+            throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
+        }
+
        clip_context_params ctx_clip_params;
        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
        ctx_clip_params.verbosity = ctx_params.verbosity;
@@ -94,8 +146,11 @@ struct mtmd_context {
            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
        }

-        use_mrope = clip_is_qwen2vl(ctx_clip);
+        has_vision = clip_has_vision_encoder(ctx_clip);
+        has_audio  = clip_has_audio_encoder(ctx_clip);
+        use_mrope  = clip_is_qwen2vl(ctx_clip);

+        projector_type proj = clip_get_projector_type(ctx_clip);
        int minicpmv_version = clip_is_minicpmv(ctx_clip);
        if (minicpmv_version == 2) {
            // minicpmv 2.5 format:
@@ -108,6 +163,8 @@ struct mtmd_context {
            tok_sli_img_start = tok_ov_img_start;
            tok_sli_img_end   = tok_ov_img_end;
            tok_row_end       = lookup_token("\n");
+            tok_row_end_trail = false; // no trailing end-of-row token
+            ov_img_first      = true;

        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
            // minicpmv 2.6 format:
@@ -118,9 +175,40 @@ struct mtmd_context {
            tok_sli_img_start = lookup_token("<slice>");
            tok_sli_img_end   = lookup_token("</slice>");
            tok_row_end       = lookup_token("\n");
+            tok_row_end_trail = false; // no trailing end-of-row token
+            ov_img_first      = true;

        } else if (minicpmv_version != 0) {
            GGML_ASSERT(false && "unsupported minicpmv version");
+        } else if (proj == PROJECTOR_TYPE_LLAMA4) {
+            // llama 4 format:
+            // <|image_start|>
+            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
+            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
+            //     ... <|tile_y_separator|>   <-- trailing end-of-row token
+            // <|image|> (overview)           <-- overview image is last
+            // <|image_end|>
+            slice_tmpl        = MTMD_SLICE_TMPL_LLAMA4;
+            tok_ov_img_start  = lookup_token("<|image|>");
+            tok_sli_img_mid   = lookup_token("<|tile_x_separator|>");
+            tok_row_end       = lookup_token("<|tile_y_separator|>");
+            tok_row_end_trail = true; // add trailing end-of-row token
+            ov_img_first      = false; // overview image is last
+        }
+
+        if (proj == PROJECTOR_TYPE_ULTRAVOX) {
+            // TODO @ngxson : check if model n_mel is 128 or 80
+            w_filters = whisper_precalc_filters::get_128_bins();
+        }
+
+        // warning messages
+        if (proj == PROJECTOR_TYPE_LLAMA4) {
+            LOG_WRN("%s: llama 4 vision is known to have degraded quality:\n"
+                    "    https://github.com/ggml-org/llama.cpp/pull/13282\n", __func__);
+        }
+        if (has_audio) {
+            LOG_WRN("%s: audio input is in experimental stage and may have reduced quality:\n"
+                    "    https://github.com/ggml-org/llama.cpp/pull/13623\n", __func__);
        }
    }

@@ -155,29 +243,6 @@ private:
    }
 };

-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
-    bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
-    uint32_t n_tokens() const { return nx * ny; }
-    clip_image_f32_batch batch_f32; // preprocessed image patches
-    std::string id; // optional user-defined ID, useful for KV cache tracking
-
-    mtmd_image_tokens clone() {
-        return mtmd_image_tokens{
-            nx,
-            ny,
-            use_mrope_pos,
-            batch_f32.clone(),
-            id
-        };
-    }
-};
-
 mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
        const struct llama_model * text_model,
        const struct mtmd_context_params ctx_params) {
@@ -223,57 +288,63 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
    auto vocab = llama_model_get_vocab(ctx->text_model);

    std::string prompt_modified(text->text);
-    std::string marker_modified(ctx->image_marker);
+    std::string marker_modified(ctx->media_marker);
    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);

+    // for compatibility, we convert image marker to media marker
+    string_replace_all(prompt_modified, MTMD_DEFAULT_IMAGE_MARKER, ctx->media_marker);
+
    // a bit hacky here, but works for now
    // for some models, we need to add prefix and suffix to the image embeddings
    if (clip_is_gemma3(ctx->ctx_clip)) {
        // gemma 3
        // <start_of_image> ... (image embeddings) ... <end_of_image>
-        marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+        marker_modified = "<start_of_image>" + ctx->media_marker + "<end_of_image>";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
        // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
-        marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+        marker_modified = "<fake_token_around_image><global-img>" + ctx->media_marker + "<fake_token_around_image>";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

    } else if (proj_type == PROJECTOR_TYPE_PIXTRAL) {
        // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
-        marker_modified = ctx->image_marker + "[IMG_END]";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-    }
+        marker_modified = ctx->media_marker + "[IMG_END]";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

-    else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
+    } else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
        // <|vision_start|> ... (image embeddings) ... <|vision_end|>
-        marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+        marker_modified = "<|vision_start|>" + ctx->media_marker + "<|vision_end|>";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

-    }
+    } else if (proj_type == PROJECTOR_TYPE_LLAMA4) {
+        // (more details in mtmd_context constructor)
+        marker_modified = "<|image_start|>" + ctx->media_marker + "<|image_end|>";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

-    else if (proj_type == PROJECTOR_TYPE_INTERNVL) {
+    } else if (proj_type == PROJECTOR_TYPE_INTERNVL) {
        // <img> ... (image embeddings) ... </img>
-        marker_modified = "<img>" + ctx->image_marker + "</img>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+        marker_modified = "<img>" + ctx->media_marker + "</img>";
+        string_replace_all(prompt_modified, ctx->media_marker, marker_modified);

    }

    // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
    // for glm-edge, BOI and EOI token's embeddings are not present in the text model

-    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
+    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->media_marker);
    output->entries.clear();
    output->entries.reserve(parts.size());

-    size_t i_img = 0;
+    size_t i_bm = 0;

    // utility for adding raw tokens
    auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
        mtmd_input_chunk chunk{
            MTMD_INPUT_CHUNK_TYPE_TEXT,
            std::move(tokens),
-            {},
+            nullptr, // image tokens
+            nullptr, // audio tokens
        };
        output->entries.emplace_back(std::move(chunk));
    };
@@ -291,8 +362,9 @@ int32_t mtmd_tokenize(mtmd_context * ctx,

            mtmd_input_chunk chunk{
                MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                {},
+                {}, // text tokens
                std::move(image_tokens),
+                nullptr, // audio tokens
            };
            chunks.emplace_back(std::move(chunk));
        }
@@ -310,25 +382,36 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
        mtmd_input_chunk chunk{
            MTMD_INPUT_CHUNK_TYPE_TEXT,
            std::move(tokens),
-            {},
+            nullptr, // image tokens
+            nullptr, // audio tokens
        };
        output->entries.emplace_back(std::move(chunk));

-        if (&parts.back() != &part) {
-            // add image token to middle of 2 parts
+        // only add image/audio tokens to middle of 2 parts
+        // therefore, we skip handling image/audio if this is the last part
+        if (&parts.back() == &part) {
+            continue;
+        }

-            if (i_img >= n_bitmaps) {
+        if (!bitmaps[i_bm]->is_audio) {
+            // handle image
+
+            if (i_bm >= n_bitmaps) {
                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
                return 1;
            }

+            if (!ctx->has_vision) {
+                LOG_ERR("%s: error: model does not support vision input\n", __func__);
+                return 2;
+            }
+
            // convert mtmd_bitmap to clip_image_u8
            clip_image_u8_ptr img_u8(clip_image_u8_init());
-            img_u8->nx = bitmaps[i_img]->nx;
-            img_u8->ny = bitmaps[i_img]->ny;
-            img_u8->buf.resize(bitmaps[i_img]->data.size());
-            std::memcpy(img_u8->buf.data(), bitmaps[i_img]->data.data(), img_u8->nx * img_u8->ny * 3);
-            clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
+            img_u8->nx = bitmaps[i_bm]->nx;
+            img_u8->ny = bitmaps[i_bm]->ny;
+            img_u8->buf.resize(bitmaps[i_bm]->data.size());
+            std::memcpy(img_u8->buf.data(), bitmaps[i_bm]->data.data(), img_u8->nx * img_u8->ny * 3);

            // preprocess image
            clip_image_f32_batch batch_f32;
@@ -338,28 +421,40 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                return 2;
            }

-            if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
+            // handle llava-uhd style preprocessing
+            if (
+                ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5
+                || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6
+                || ctx->slice_tmpl == MTMD_SLICE_TMPL_LLAMA4
+            ) {
                // split batch into chunks of single images
-                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img]->id);
+                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_bm]->id);
                GGML_ASSERT(chunks.size() > 0);

-                // add overview image
-                add_text_chunk({ctx->tok_ov_img_start});
-                output->entries.emplace_back(std::move(chunks.front()));
+                auto ov_chunk = std::move(chunks.front());
                chunks.erase(chunks.begin());
-                add_text_chunk({ctx->tok_ov_img_end});

-                // add slices
+                // add overview image (first)
+                if (ctx->ov_img_first) {
+                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_ov_img_start});
+                    }
+                    output->entries.emplace_back(std::move(ov_chunk));
+                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_ov_img_end});
+                    }
+                }
+
+                // add slices (or tiles)
                if (!chunks.empty()) {
-                    clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
-                    int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
-                    int n_row = (int)chunks.size() / n_col;
-                    GGML_ASSERT(n_row * n_col == (int)chunks.size());
+                    const int n_col = batch_f32.grid_x;
+                    const int n_row = batch_f32.grid_y;
                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
                        add_text_chunk({ctx->tok_slices_start});
                    }
                    for (int y = 0; y < n_row; y++) {
                        for (int x = 0; x < n_col; x++) {
+                            const bool is_last_in_row = (x == n_col - 1);
                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
                                add_text_chunk({ctx->tok_sli_img_start});
                            }
@@ -367,8 +462,11 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
                                add_text_chunk({ctx->tok_sli_img_end});
                            }
+                            if (!is_last_in_row && ctx->tok_sli_img_mid != LLAMA_TOKEN_NULL) {
+                                add_text_chunk({ctx->tok_sli_img_mid});
+                            }
                        }
-                        if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
+                        if ((y != n_row - 1 || ctx->tok_row_end_trail) && ctx->tok_row_end != LLAMA_TOKEN_NULL) {
                            add_text_chunk({ctx->tok_row_end});
                        }
                    }
@@ -377,6 +475,17 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                    }
                }

+                // add overview image (last)
+                if (!ctx->ov_img_first) {
+                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_ov_img_start});
+                    }
+                    output->entries.emplace_back(std::move(ov_chunk));
+                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_ov_img_end});
+                    }
+                }
+
            } else {
                size_t n_tokens = 0;
                for (const auto & entry : batch_f32.entries) {
@@ -395,7 +504,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                    image_tokens->ny = 1;
                }
                image_tokens->batch_f32 = std::move(batch_f32);
-                image_tokens->id = bitmaps[i_img]->id; // optional
+                image_tokens->id = bitmaps[i_bm]->id; // optional

                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
@@ -403,23 +512,101 @@ int32_t mtmd_tokenize(mtmd_context * ctx,

                mtmd_input_chunk chunk{
                    MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                    {},
+                    {}, // text tokens
                    std::move(image_tokens),
+                    nullptr, // audio tokens
                };
                output->entries.emplace_back(std::move(chunk));
            }

-            i_img++; // move to next image
+            i_bm++; // move to next image
+            continue;
+
+        } else {
+            // handle audio
+
+            if (i_bm >= n_bitmaps) {
+                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
+                return 1;
+            }
+
+            if (!ctx->has_audio) {
+                LOG_ERR("%s: error: model does not support audio input\n", __func__);
+                return 2;
+            }
+
+            if (bitmaps[i_bm]->data.size() == 0) {
+                LOG_ERR("%s: error: empty audio data\n", __func__);
+                return 2;
+            }
+
+            // preprocess audio
+            GGML_ASSERT(ctx->w_filters.n_mel); // make sure we have filter preloaded
+            std::vector<whisper_preprocessor::whisper_mel> mel_spec_chunks;
+            const float * samples = (const float *)bitmaps[i_bm]->data.data();
+            size_t n_samples = bitmaps[i_bm]->data.size() / sizeof(float);
+            bool ok = whisper_preprocessor::preprocess_audio(samples, n_samples, ctx->w_filters, mel_spec_chunks);
+            if (!ok) {
+                LOG_ERR("Unable to preprocess audio\n");
+                return 2;
+            }
+
+            // consider each mel_spec as a separate audio chunk
+            // TODO: maybe support batching, but this may come with memory cost
+            for (auto & mel_spec : mel_spec_chunks) {
+                clip_image_f32_ptr mel_f32(clip_image_f32_init());
+                mel_f32->nx  = mel_spec.n_len;
+                mel_f32->ny  = mel_spec.n_mel;
+                mel_f32->buf = std::move(mel_spec.data);
+                size_t n_tokens = clip_n_output_tokens(ctx->ctx_clip, mel_f32.get());
+
+                clip_image_f32_batch batch_f32;
+                batch_f32.is_audio = true;
+                batch_f32.entries.push_back(std::move(mel_f32));
+
+                mtmd_audio_tokens_ptr audio_tokens(new mtmd_audio_tokens);
+                audio_tokens->n_tokens = n_tokens;
+                audio_tokens->batch_f32 = std::move(batch_f32);
+                audio_tokens->id = bitmaps[i_bm]->id; // optional
+
+                LOG_DBG("audio_tokens->n_tokens = %d\n", audio_tokens->n_tokens);
+
+                mtmd_input_chunk chunk{
+                    MTMD_INPUT_CHUNK_TYPE_AUDIO,
+                    {}, // text tokens
+                    nullptr, // image tokens
+                    std::move(audio_tokens),
+                };
+                output->entries.emplace_back(std::move(chunk));
+            }
+
+            i_bm++;
+            continue;
        }
    }

    return 0;
 }

-static void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
-    if (image_tokens) {
-        delete image_tokens;
+int32_t mtmd_encode_chunk(mtmd_context * ctx, const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_WRN("mtmd_encode_chunk has no effect for text chunks\n");
+        return 0;
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return mtmd_encode(ctx, chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+        ctx->image_embd_v.resize(chunk->tokens_audio->n_tokens * n_mmproj_embd);
+        bool ok = clip_image_batch_encode(
+            ctx->ctx_clip,
+            ctx->n_threads,
+            &chunk->tokens_audio->batch_f32,
+            ctx->image_embd_v.data());
+        return ok ? 0 : 1;
    }
+
+    LOG_ERR("mtmd_encode_chunk: unknown chunk type %d\n", (int)chunk->type);
+    return 1;
 }

 int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
@@ -427,14 +614,6 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
    bool ok = false;

-    // only effective for minicpmv and qwen2vl, other models will ignore load_image_size
-    {
-        clip_image_size slice_size{
-            image_tokens->batch_f32.entries[0]->nx,
-            image_tokens->batch_f32.entries[0]->ny};
-        clip_add_load_image_size(ctx->ctx_clip, &slice_size);
-    }
-
    if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
        const auto & entries = image_tokens->batch_f32.entries;
@@ -473,8 +652,12 @@ bool mtmd_decode_use_mrope(mtmd_context * ctx) {
    return ctx->use_mrope;
 }

-void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
-    mtmd_image_tokens_free(val);
+bool mtmd_support_vision(mtmd_context * ctx) {
+    return ctx->has_vision;
+}
+
+bool mtmd_support_audio(mtmd_context * ctx) {
+    return ctx->has_audio;
 }

 // these 2 helpers below use internal clip_image_u8_ptr,
@@ -483,6 +666,15 @@ void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
 // whichever library they want, and then use mtmd_bitmap_init() to create bitmap

 mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len) {
+    if (audio_helpers::is_audio_file((const char *)buf, len)) {
+        std::vector<float> pcmf32;
+        if (!audio_helpers::decode_audio_from_buf(buf, len, COMMON_SAMPLE_RATE, pcmf32)) {
+            LOG_ERR("Unable to read WAV audio file from buffer\n");
+            return nullptr;
+        }
+        return mtmd_bitmap_init_from_audio(pcmf32.size(), pcmf32.data());
+    }
+
    clip_image_u8_ptr img_u8(clip_image_u8_init());
    bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
    if (!ok) {
@@ -495,15 +687,26 @@ mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t
 }

 mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname) {
-    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);
+    std::vector<unsigned char> buf;
+    FILE * f = fopen(fname, "rb");
+    if (!f) {
+        LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
        return nullptr;
    }
-    uint32_t nx, ny;
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &nx, &ny);
-    return mtmd_bitmap_init(nx, ny, data);
+
+    fseek(f, 0, SEEK_END);
+    long file_size = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    buf.resize(file_size);
+
+    size_t n_read = fread(buf.data(), 1, file_size, f);
+    fclose(f);
+    if (n_read != (size_t)file_size) {
+        LOG_ERR("Failed to read entire file %s", fname);
+        return nullptr;
+    }
+
+    return mtmd_helper_bitmap_init_from_buf(buf.data(), buf.size());
 }

 //
@@ -524,6 +727,18 @@ mtmd_bitmap * mtmd_bitmap_init(uint32_t nx,
    return bitmap;
 }

+mtmd_bitmap * mtmd_bitmap_init_from_audio(size_t n_samples,
+                                          const float * data) {
+    mtmd_bitmap * bitmap = new mtmd_bitmap;
+    bitmap->nx = n_samples;
+    bitmap->ny = 1;
+    bitmap->is_audio = true;
+    size_t data_size = n_samples * sizeof(float);
+    bitmap->data.resize(data_size);
+    std::memcpy(bitmap->data.data(), data, data_size);
+    return bitmap;
+}
+
 uint32_t mtmd_bitmap_get_nx(const mtmd_bitmap * bitmap) {
    return bitmap->nx;
 }
@@ -536,6 +751,10 @@ const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap) {
    return bitmap->data.data();
 }

+bool mtmd_bitmap_is_audio(const mtmd_bitmap * bitmap) {
+    return bitmap->is_audio;
+}
+
 const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap) {
    return bitmap->id.c_str();
 }
@@ -599,17 +818,56 @@ const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chu
    return nullptr;
 }

+size_t mtmd_input_chunk_get_n_tokens(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        return chunk->tokens_text.size();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return mtmd_image_tokens_get_n_tokens(chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->n_tokens;
+    } else {
+        GGML_ABORT("invalid chunk type");
+    }
+}
+
+llama_pos mtmd_input_chunk_get_n_pos(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        return chunk->tokens_text.size();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return mtmd_image_tokens_get_n_pos(chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->n_tokens;
+    } else {
+        GGML_ABORT("invalid chunk type");
+    }
+}
+
+const char * mtmd_input_chunk_get_id(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return chunk->tokens_image->id.c_str();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->id.c_str();
+    }
+    return nullptr;
+}
+
 mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk) {
    mtmd_input_chunk * copy = new mtmd_input_chunk{
        chunk->type,
        chunk->tokens_text,
-        mtmd_image_tokens_ptr(),
+        nullptr,
+        nullptr,
    };
    if (chunk->tokens_image) {
        // copy the image tokens
        copy->tokens_image = mtmd_image_tokens_ptr(new mtmd_image_tokens());
        *copy->tokens_image = chunk->tokens_image->clone();
    }
+    if (chunk->tokens_audio) {
+        // copy the audio tokens
+        copy->tokens_audio = mtmd_audio_tokens_ptr(new mtmd_audio_tokens());
+        *copy->tokens_audio = chunk->tokens_audio->clone();
+    }
    return copy;
 }

@@ -657,7 +915,8 @@ mtmd_input_chunks * mtmd_test_create_input_chunks() {
    mtmd_input_chunk chunk_text{
        MTMD_INPUT_CHUNK_TYPE_TEXT,
        std::move(tokens_text),
-        {},
+        nullptr, // image tokens
+        nullptr, // audio tokens
    };
    chunks->entries.emplace_back(std::move(chunk_text));

@@ -669,8 +928,9 @@ mtmd_input_chunks * mtmd_test_create_input_chunks() {
    image_tokens->id = "image_1";
    mtmd_input_chunk chunk_image{
        MTMD_INPUT_CHUNK_TYPE_IMAGE,
-        {},
+        {}, // text tokens
        std::move(image_tokens),
+        nullptr, // audio tokens
    };
    chunks->entries.emplace_back(std::move(chunk_image));

@@ -39,6 +39,7 @@
 #    define MTMD_API
 #endif

+// deprecated marker, use mtmd_default_marker() instead
 #define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"

 #ifdef __cplusplus
@@ -48,6 +49,7 @@ extern "C" {
 enum mtmd_input_chunk_type {
    MTMD_INPUT_CHUNK_TYPE_TEXT,
    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+    MTMD_INPUT_CHUNK_TYPE_AUDIO,
 };

 // opaque types
@@ -79,9 +81,12 @@ struct mtmd_context_params {
    bool print_timings;
    int n_threads;
    enum ggml_log_level verbosity;
-    const char * image_marker;
+    const char * image_marker; // deprecated, use media_marker instead
+    const char * media_marker;
 };

+MTMD_API const char * mtmd_default_marker(void);
+
 MTMD_API struct mtmd_context_params mtmd_context_params_default(void);

 // initialize the mtmd context
@@ -98,17 +103,26 @@ MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
 // whether the current model use M-RoPE for llama_decode
 MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);

+// whether the current model supports vision input
+MTMD_API bool mtmd_support_vision(mtmd_context * ctx);
+
+// whether the current model supports audio input
+MTMD_API bool mtmd_support_audio(mtmd_context * ctx);

 // mtmd_bitmap
 //
-// length of data must be nx * ny * 3
-// the data is in RGBRGBRGB... format
-MTMD_API mtmd_bitmap *         mtmd_bitmap_init    (uint32_t nx,
-                                                    uint32_t ny,
-                                                    const unsigned char * data);
+// if bitmap is image:
+//     length of data must be nx * ny * 3
+//     the data is in RGBRGBRGB... format
+// if bitmap is audio:
+//     length of data must be n_samples * sizeof(float)
+//     the data is in float format (PCM F32)
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init           (uint32_t nx, uint32_t ny, const unsigned char * data);
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init_from_audio(size_t n_samples,         const float         * data);
 MTMD_API uint32_t              mtmd_bitmap_get_nx  (const mtmd_bitmap * bitmap);
 MTMD_API uint32_t              mtmd_bitmap_get_ny  (const mtmd_bitmap * bitmap);
 MTMD_API const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap);
+MTMD_API bool                  mtmd_bitmap_is_audio(const mtmd_bitmap * bitmap);
 MTMD_API void                  mtmd_bitmap_free    (mtmd_bitmap * bitmap);
 // bitmap ID is optional, but useful for KV cache tracking
 // these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
@@ -132,6 +146,11 @@ MTMD_API void                     mtmd_input_chunks_free(mtmd_input_chunks * chu
 MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type        (const mtmd_input_chunk * chunk);
 MTMD_API const llama_token *        mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output);
 MTMD_API const mtmd_image_tokens *  mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
+MTMD_API size_t                     mtmd_input_chunk_get_n_tokens    (const mtmd_input_chunk * chunk);
+// returns nullptr for ID on text chunk
+MTMD_API const char *               mtmd_input_chunk_get_id          (const mtmd_input_chunk * chunk);
+// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd_input_chunk * chunk);

 // in case you want to use custom logic to handle the chunk (i.e. KV cache management)
 // you can move the chunk ownership to your own code by copying it
@@ -144,27 +163,28 @@ MTMD_API void               mtmd_input_chunk_free(mtmd_input_chunk * chunk);
 //
 // the instance will be constructed via mtmd_tokenize()
 // it will be freed along with mtmd_input_chunk
-MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
 MTMD_API size_t       mtmd_image_tokens_get_nx      (const mtmd_image_tokens * image_tokens);
 MTMD_API size_t       mtmd_image_tokens_get_ny      (const mtmd_image_tokens * image_tokens);
-MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens);
+MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens); // TODO: deprecate
 // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
-MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens);
+MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens); // TODO: deprecate

-// 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
+// tokenize an input text prompt and a list of bitmaps (images/audio)
+// the prompt must have the input image marker (default: "<__media__>") in it
+// the default marker is defined by mtmd_default_marker()
+// the marker will be replaced with the image/audio chunk
 // for example:
-//   "here is an image: <__image__>\ndescribe it in detail."
+//   "here is an image: <__media__>\ndescribe it in detail."
 //   this will gives 3 chunks:
 //   1. "here is an image: <start_of_image>"
-//   2. (image tokens)
+//   2. (image/audio tokens)
 //   3. "<end_of_image>\ndescribe it in detail."
-// number of bitmaps must be equal to the number of image markers in the prompt
+// number of bitmaps must be equal to the number of markers in the prompt
 // this function is thread-safe (shared ctx)
 // return values:
 //   0 on success
-//   1 on number of images not matching the number of markers
+//   1 on number of bitmaps not matching the number of markers
 //   2 on image preprocessing error
 MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                               mtmd_input_chunks * output,
@@ -173,9 +193,14 @@ MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                               size_t n_bitmaps);

 // returns 0 on success
+// TODO: deprecate
 MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
                             const mtmd_image_tokens * image_tokens);

+// returns 0 on success
+MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
+                                   const mtmd_input_chunk * chunk);
+
 // get output embeddings from the last encode pass
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);

@@ -189,12 +214,16 @@ MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 //

 // helper function to construct a mtmd_bitmap from a file
+// it calls mtmd_helper_bitmap_init_from_buf() internally
 // returns nullptr on failure
 // this function is thread-safe
 MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname);

 // helper function to construct a mtmd_bitmap from a buffer containing a file
-// the file content must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
+// supported formats:
+//     image: formats supported by stb_image: jpg, png, bmp, gif, etc.
+//     audio: formats supported by miniaudio: wav, mp3, flac
+// note: audio files will be auto-detected based on magic bytes
 // returns nullptr on failure
 // this function is thread-safe
 MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len);
@@ -21,6 +21,13 @@ if [ "${1:-}" = "big" ]; then
    echo "Include BIG models..."
 fi

+RUN_HUGE_TESTS=false
+if [ "${1:-}" = "huge" ]; then
+    RUN_HUGE_TESTS=true
+    RUN_BIG_TESTS=true
+    echo "Include BIG models..."
+fi
+
 ###############

 arr_bin=()
@@ -42,7 +49,7 @@ add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
 add_test "llama-mtmd-cli"  "second-state/Llava-v1.5-7B-GGUF:Q2_K"            "vicuna"
-add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"           "vicuna"
+add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K_M"         "vicuna"
 add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
@@ -60,10 +67,17 @@ if [ "$RUN_BIG_TESTS" = true ]; then
    add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M"
    add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
    add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M"
-    add_test "llama-mtmd-cli"  "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M"
-    add_test "llama-mtmd-cli"  "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M"
    # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra
-    # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M" # too big
+fi
+
+# to test the huge models, run: ./tests.sh huge
+# this will run both the big and huge models
+# huge models are > 32B parameters
+if [ "$RUN_HUGE_TESTS" = true ]; then
+    add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF:IQ1_S"
 fi

 # these models always give the wrong answer, not sure why
@@ -936,7 +936,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                           std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_seq_pos_max(llama_data.context.get(), 0) == 0;

    const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
    prompt_tokens.resize(n_prompt_tokens);
@@ -952,7 +952,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
    const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_seq_pos_max(ctx.get(), 0);
    if (n_ctx_used + batch.n_tokens > n_ctx) {
        printf(LOG_COL_DEFAULT "\n");
        printe("context size exceeded\n");
@@ -13,6 +13,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 * Multimodal ([documentation](../../docs/multimodal.md)) / with OpenAI-compatible API support
 * Monitoring endpoints
 * Schema-constrained JSON response format
+ * Prefilling of assistant messages similar to the Claude API
 * [Function calling](../../docs/function-calling.md) / tool use for ~any model
 * Speculative decoding
 * Easy-to-use web UI
@@ -175,6 +176,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--reasoning-format FORMAT` | reasoning format (default: deepseek; allowed values: deepseek, none)<br/>controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).<br/>only supported for non-streamed responses<br/>(env: LLAMA_ARG_THINK) |
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/>(env: LLAMA_ARG_NO_PREFILL_ASSISTANT) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 | `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
@@ -951,7 +951,7 @@ struct server_task_result_cmpl_partial : server_task_result {
    }

    json to_json_oaicompat_chat() {
-        bool first = n_decoded == 0;
+        bool first = n_decoded == 1;
        std::time_t t = std::time(0);
        json choices;

@@ -962,15 +962,18 @@ struct server_task_result_cmpl_partial : server_task_result {
                                            {"delta", json{{"role", "assistant"}}}}});
            } else {
                // We have to send this as two updates to conform to openai behavior
+                // initial_ret is the role message for stream=True
                json initial_ret = json{{"choices", json::array({json{
                                        {"finish_reason", nullptr},
                                        {"index", 0},
                                        {"delta", json{
-                                            {"role", "assistant"}
+                                            {"role", "assistant"},
+                                            {"content", ""}
                                        }}}})},
                            {"created", t},
                            {"id", oaicompat_cmpl_id},
                            {"model", oaicompat_model},
+                            {"system_fingerprint", build_info},
                            {"object", "chat.completion.chunk"}};

                json second_ret = json{
@@ -982,8 +985,19 @@ struct server_task_result_cmpl_partial : server_task_result {
                            {"created", t},
                            {"id", oaicompat_cmpl_id},
                            {"model", oaicompat_model},
+                            {"system_fingerprint", build_info},
                            {"object", "chat.completion.chunk"}};

+                if (prob_output.probs.size() > 0) {
+                    second_ret["choices"][0]["logprobs"] = json{
+                        {"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
+                    };
+                }
+
+                if (timings.prompt_n >= 0) {
+                    second_ret.push_back({"timings", timings.to_json()});
+                }
+
                return std::vector<json>({initial_ret, second_ret});
            }
        } else {
@@ -1137,9 +1151,6 @@ struct server_task_result_metrics : server_task_result {
    int n_tasks_deferred;
    int64_t t_start;

-    int32_t kv_cache_tokens_count;
-    int32_t kv_cache_used_cells;
-
    // TODO: somehow reuse server_metrics in the future, instead of duplicating the fields
    uint64_t n_prompt_tokens_processed_total = 0;
    uint64_t t_prompt_processing_total       = 0;
@@ -1179,9 +1190,6 @@ struct server_task_result_metrics : server_task_result {
            { "n_decode_total",                  n_decode_total },
            { "n_busy_slots_total",              n_busy_slots_total },

-            { "kv_cache_tokens_count",           kv_cache_tokens_count },
-            { "kv_cache_used_cells",             kv_cache_used_cells },
-
            { "slots",                           slots_data },
        };
    }
@@ -2004,6 +2012,23 @@ struct server_context {
            }
        }

+        if (!llama_kv_self_can_shift(ctx)) {
+            if (params_base.ctx_shift) {
+                params_base.ctx_shift = false;
+                SRV_WRN("%s\n", "ctx_shift is not supported by this context, it will be disabled");
+            }
+
+            if (params_base.n_cache_reuse) {
+                params_base.n_cache_reuse = 0;
+                SRV_WRN("%s\n", "cache_reuse is not supported by this context, it will be disabled");
+            }
+
+            if (!params_base.speculative.model.path.empty()) {
+                SRV_ERR("%s\n", "err: speculative decode is not supported by this context");
+                return false;
+            }
+        }
+
        return true;
    }

@@ -2754,9 +2779,6 @@ struct server_context {
                    res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                    res->t_start             = metrics.t_start;

-                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
-                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);
-
                    res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                    res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
                    res->n_tokens_predicted_total        = metrics.n_tokens_predicted_total;
@@ -3181,7 +3203,15 @@ struct server_context {
                                // if we don't cache the prompt, we have to remove the entire KV cache
                                llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                slot.n_past = 0;
-                                slot.cache_tokens.clear();
+                                slot.cache_tokens.clear(); // TODO: not needed, will be cleared later via "keep_first()"
+                            }
+
+                            if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
+                                if (llama_kv_self_seq_pos_min(ctx, slot.id) > 0) {
+                                    SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
+                                            "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+                                    slot.n_past = 0;
+                                }
                            }
                        }

@@ -3311,6 +3341,37 @@ struct server_context {
            common_set_adapter_lora(ctx, slot_batched->lora);
        }

+        const bool do_encode = (params_base.embedding || params_base.reranking);
+
+        // pad the batch so that batch.n_tokens >= n_slots
+        // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689
+        if (do_encode) {
+            const int n_slots = slots.size();
+
+            if (batch.n_tokens < n_slots) {
+                std::set<llama_seq_id> seq_ids;
+                for (int j = 0; j < batch.n_tokens; ++j) {
+                    seq_ids.insert(batch.seq_id[j][0]);
+                }
+
+                // find unused sequence id
+                llama_seq_id seq_id = -1;
+                for (int i = 0; i < n_slots; ++i) {
+                    if (seq_ids.find(i) == seq_ids.end()) {
+                        seq_id = i;
+                    }
+                }
+
+                const int n_add = n_slots - batch.n_tokens;
+
+                SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id);
+
+                for (int j = 0; j < n_add; ++j) {
+                    common_batch_add(batch, 0, j, { seq_id }, false);
+                }
+            }
+        }
+
        // process the created batch of tokens
        for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
@@ -3327,7 +3388,7 @@ struct server_context {

            int ret = 0;

-            if (params_base.embedding || params_base.reranking) {
+            if (do_encode) {
                ret = llama_encode(ctx, batch_view);
            } else {
                ret = llama_decode(ctx, batch_view);
@@ -3336,14 +3397,29 @@ struct server_context {
            metrics.on_decoded(slots);

            if (ret != 0) {
-                if (n_batch == 1 || ret < 0) {
-                    // if you get here, it means the KV cache is full - try increasing it via the context size
-                    SRV_ERR("failed to decode the batch: KV cache is full - try increasing it via the context size, i = %d, n_batch = %d, ret = %d\n", i, n_batch, ret);
-                    for (auto & slot : slots) {
-                        slot.release();
-                        send_error(slot, "Input prompt is too big compared to KV size. Please try increasing KV size.");
+                {
+                    std::string err;
+
+                    if (n_batch == 1 && ret == 1) {
+                        err = "Context size has been exceeded.";
+                    }
+
+                    if (ret == -1) {
+                        err = "Invalid input batch.";
+                    }
+
+                    if (ret < -1) {
+                        err = "Compute error.";
+                    }
+
+                    if (!err.empty()) {
+                        SRV_ERR("%s, i = %d, n_batch = %d, ret = %d\n", err.c_str(), i, n_batch, ret);
+                        for (auto & slot : slots) {
+                            slot.release();
+                            send_error(slot, err);
+                        }
+                        break;
                    }
-                    break; // break loop of n_batch
                }

                // retry with half the batch size to try to find a free slot in the KV cache
@@ -3677,6 +3753,7 @@ int main(int argc, char ** argv) {
            "/health",
            "/models",
            "/v1/models",
+            "/api/tags"
        };

        // If API key is not set, skip validation
@@ -3715,7 +3792,7 @@ int main(int argc, char ** argv) {
            if (req.path == "/" || tmp.back() == "html") {
                res.set_content(reinterpret_cast<const char*>(loading_html), loading_html_len, "text/html; charset=utf-8");
                res.status = 503;
-            } else if (req.path == "/models" || req.path == "/v1/models") {
+            } else if (req.path == "/models" || req.path == "/v1/models" || req.path == "/api/tags") {
                // allow the models endpoint to be accessed during loading
                return true;
            } else {
@@ -3858,14 +3935,6 @@ int main(int argc, char ** argv) {
                    {"name",  "predicted_tokens_seconds"},
                    {"help",  "Average generation throughput in tokens/s."},
                    {"value",  res_metrics->n_tokens_predicted ? 1.e3 / res_metrics->t_tokens_generation * res_metrics->n_tokens_predicted : 0.}
-            },{
-                    {"name",  "kv_cache_usage_ratio"},
-                    {"help",  "KV-cache usage. 1 means 100 percent usage."},
-                    {"value",  1. * res_metrics->kv_cache_used_cells / params.n_ctx}
-            },{
-                    {"name",  "kv_cache_tokens"},
-                    {"help",  "KV-cache tokens."},
-                    {"value",  (uint64_t) res_metrics->kv_cache_tokens_count}
            },{
                    {"name",  "requests_processing"},
                    {"help",  "Number of requests processing."},
@@ -4061,6 +4130,19 @@ int main(int argc, char ** argv) {
                    { "llama.context_length", ctx_server.slots.back().n_ctx, },
                }
            },
+            {"modelfile", ""},
+            {"parameters", ""},
+            {"template", common_chat_templates_source(ctx_server.chat_templates.get())},
+            {"details", {
+                {"parent_model", ""},
+                {"format", "gguf"},
+                {"family", ""},
+                {"families", {""}},
+                {"parameter_size", ""},
+                {"quantization_level", ""}
+            }},
+            {"model_info", ""},
+            {"capabilities", {"completion"}}
        };

        res_ok(res, data);
@@ -4348,6 +4430,7 @@ int main(int argc, char ** argv) {
        json data = oaicompat_completion_params_parse(
            body,
            params.use_jinja,
+            params.prefill_assistant,
            params.reasoning_format,
            ctx_server.chat_templates.get(),
            ctx_server.mctx,
@@ -4369,6 +4452,7 @@ int main(int argc, char ** argv) {
        json data = oaicompat_completion_params_parse(
            body,
            params.use_jinja,
+            params.prefill_assistant,
            params.reasoning_format,
            ctx_server.chat_templates.get(),
            ctx_server.mctx,
@@ -4384,6 +4468,28 @@ int main(int argc, char ** argv) {
        }

        json models = {
+            {"models", {
+                {
+                    {"name", params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"model", params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"modified_at", ""},
+                    {"size", ""},
+                    {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
+                    {"type", "model"},
+                    {"description", ""},
+                    {"tags", {""}},
+                    {"capabilities", {"completion"}},
+                    {"parameters", ""},
+                    {"details", {
+                        {"parent_model", ""},
+                        {"format", "gguf"},
+                        {"family", ""},
+                        {"families", {""}},
+                        {"parameter_size", ""},
+                        {"quantization_level", ""}
+                    }}
+                }
+            }},
            {"object", "list"},
            {"data", {
                {
@@ -4393,7 +4499,7 @@ int main(int argc, char ** argv) {
                    {"owned_by", "llamacpp"},
                    {"meta",     model_meta},
                },
-             }}
+            }}
        };

        res_ok(res, models);
@@ -4721,11 +4827,13 @@ int main(int argc, char ** argv) {
    svr->Post("/api/show",            handle_api_show);
    svr->Get ("/models",              handle_models); // public endpoint (no API key check)
    svr->Get ("/v1/models",           handle_models); // public endpoint (no API key check)
+    svr->Get ("/api/tags",            handle_models); // ollama specific endpoint. public endpoint (no API key check)
    svr->Post("/completion",          handle_completions); // legacy
    svr->Post("/completions",         handle_completions);
    svr->Post("/v1/completions",      handle_completions_oai);
    svr->Post("/chat/completions",    handle_chat_completions);
    svr->Post("/v1/chat/completions", handle_chat_completions);
+    svr->Post("/api/chat",            handle_chat_completions); // ollama specific endpoint
    svr->Post("/infill",              handle_infill);
    svr->Post("/embedding",           handle_embeddings); // legacy
    svr->Post("/embeddings",          handle_embeddings);
@@ -71,8 +71,14 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
    })
    content = ""
    last_cmpl_id = None
-    for data in res:
+    for i, data in enumerate(res):
        choice = data["choices"][0]
+        if i == 0:
+            # Check first role message for stream=True
+            assert choice["delta"]["content"] == ""
+            assert choice["delta"]["role"] == "assistant"
+        else:
+            assert "role" not in choice["delta"]
        assert data["system_fingerprint"].startswith("b")
        assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
        if last_cmpl_id is None:
@@ -242,12 +248,18 @@ def test_chat_completion_with_timings_per_token():
        "stream": True,
        "timings_per_token": True,
    })
-    for data in res:
-        assert "timings" in data
-        assert "prompt_per_second" in data["timings"]
-        assert "predicted_per_second" in data["timings"]
-        assert "predicted_n" in data["timings"]
-        assert data["timings"]["predicted_n"] <= 10
+    for i, data in enumerate(res):
+        if i == 0:
+            # Check first role message for stream=True
+            assert data["choices"][0]["delta"]["content"] == ""
+            assert data["choices"][0]["delta"]["role"] == "assistant"
+        else:
+            assert "role" not in data["choices"][0]["delta"]
+            assert "timings" in data
+            assert "prompt_per_second" in data["timings"]
+            assert "predicted_per_second" in data["timings"]
+            assert "predicted_n" in data["timings"]
+            assert data["timings"]["predicted_n"] <= 10


 def test_logprobs():
@@ -295,17 +307,23 @@ def test_logprobs_stream():
    )
    output_text = ''
    aggregated_text = ''
-    for data in res:
+    for i, data in enumerate(res):
        choice = data.choices[0]
-        if choice.finish_reason is None:
-            if choice.delta.content:
-                output_text += choice.delta.content
-            assert choice.logprobs is not None
-            assert choice.logprobs.content is not None
-            for token in choice.logprobs.content:
-                aggregated_text += token.token
-                assert token.logprob <= 0.0
-                assert token.bytes is not None
-                assert token.top_logprobs is not None
-                assert len(token.top_logprobs) > 0
+        if i == 0:
+            # Check first role message for stream=True
+            assert choice.delta.content == ""
+            assert choice.delta.role == "assistant"
+        else:
+            assert choice.delta.role is None
+            if choice.finish_reason is None:
+                if choice.delta.content:
+                    output_text += choice.delta.content
+                assert choice.logprobs is not None
+                assert choice.logprobs.content is not None
+                for token in choice.logprobs.content:
+                    aggregated_text += token.token
+                    assert token.logprob <= 0.0
+                    assert token.bytes is not None
+                    assert token.top_logprobs is not None
+                    assert len(token.top_logprobs) > 0
    assert aggregated_text == output_text
@@ -583,6 +583,7 @@ static json oaicompat_completion_params_parse(const json & body) {
 static json oaicompat_completion_params_parse(
    const json & body, /* openai api json semantics */
    bool use_jinja,
+    bool prefill_assistant,
    common_reasoning_format reasoning_format,
    const struct common_chat_templates * tmpls,
    bool allow_non_text,
@@ -709,7 +710,7 @@ static json oaicompat_completion_params_parse(

                // replace this chunk with a marker
                p["type"] = "text";
-                p["text"] = MTMD_DEFAULT_IMAGE_MARKER;
+                p["text"] = mtmd_default_marker();
                p.erase("image_url");
            }
        }
@@ -732,7 +733,7 @@ static json oaicompat_completion_params_parse(

    // if the assistant message appears at the end of list, we do not add end-of-turn token
    // for ex. this can be useful to modify the reasoning process in reasoning models
-    bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant";
+    bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant" && prefill_assistant;
    common_chat_msg last_message;
    if (prefill_assistant_message) {
        last_message = inputs.messages.back();
@@ -579,6 +579,8 @@ int main(int argc, char ** argv) {

    params.model = params.vocoder.model;
    params.embedding = true;
+    params.ctx_shift = false; // silence warning
+    params.n_ubatch = params.n_batch;

    common_init_result llama_init_cts = common_init_from_params(params);

@@ -1020,8 +1022,8 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
    }
    GGML_ASSERT(batch.n_tokens == n_codes);

-    if (llama_decode(ctx_cts, batch) != 0) {
-        LOG_ERR("%s: llama_decode() failed\n", __func__);
+    if (llama_encode(ctx_cts, batch) != 0) {
+        LOG_ERR("%s: llama_encode() failed\n", __func__);
        return 1;
    }
Author	SHA1	Message	Date
Jeff Bolz	1dcd01960c	vulkan: support CPY from any type to itself (#13695 ) Reuse the f16/f32 copy shaders, and just scale the number of elements according to the type size.	2025-05-23 06:45:02 +02:00
Jeff Bolz	c10ed6cbcc	vulkan: Disable coopmat/coopmat2/bfloat extensions if glslc doesn't support it (#13696 )	2025-05-23 06:33:45 +02:00
Judd	a127ff1780	use LOG_WARN to replace `std::cerr` (#13657 )	2025-05-23 06:33:08 +02:00
Diego Devesa	3079e9ac8e	release : fix windows hip release (#13707 ) * release : fix windows hip release * make single hip release with multiple targets	2025-05-23 00:21:37 +02:00
Georgi Gerganov	8a1d206f1d	tts : fix n_ubatch + make WavTokenizer cache-less (#13713 ) ggml-ci	2025-05-22 22:21:07 +03:00
Xuan-Son Nguyen	797990c4bc	mtmd : add ultravox audio input (#13623 ) * convert ok, load ok * warmup ok * test * still does not work? * fix padding * temporary give up * fix merge conflict * build_ultravox() * rm test * fix merge conflict * add necessary mtmd APIs * first working version (only 4s of audio) * will this monster compile? * fix compile * please compile * fPIC * fix windows * various fixes * clean up audio_helpers * fix conversion * add some debug stuff * long audio input ok * adapt the api * add --audio arg * final touch UX * add miniaudio to readme * fix typo * refactor kv metadata * mtmd_default_marker()	2025-05-22 20:42:48 +02:00
Aaron Teo	ab86335760	common: Include torch package for s390x (#13699 ) * common: update requirements.txt to include pytorch nightly for s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * common: fix torch installation via pip for s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> --------- Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2025-05-22 21:31:29 +03:00
Georgi Gerganov	cc74d5be99	server : pad small embedding batches (#13692 ) ggml-ci	2025-05-22 16:33:39 +03:00
Sigbjørn Skjæret	5be24af73d	gguf-py : correct charsmap parameter typing (#13701 )	2025-05-22 14:25:05 +02:00
Nicolò Scipione	d394a9aedc	sycl : Remove waits from function calls (#13702 ) * removes the waits in async memcpy functions	2025-05-22 12:54:43 +01:00
Ewan Crawford	6b56a64690	SYCL: Avoid using with SYCL-Graph for unsupported nodes (#13587 ) Currently on a CUDA backend to SYCL when running `GGML_SYCL_DISABLE_GRAPH=0 ./bin/test-backend-ops -b SYCL0` there are two operations that throw an exception from the blocking waits during queue recording. * `-o CONCAT` : Use of blocking waits on a queue that's being recorded https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-sycl/concat.cpp#L185-L187 * `-o MUL_MAT_ID`: Blocking wait on a recording queue for a copy to host memory https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-sycl/ggml-sycl.cpp#L3072-L3074 We've noticed that `ggml-cuda.cu` has the [check_node_graph_compatibility_and_refresh_copy_ops](https://github.com/ggml-org/llama.cpp/blob/39e73ae0d69f882d7e29cecc6dd8f5052fca6731/ggml/src/ggml-cuda/ggml-cuda.cu#L2458-L2458) method for checking if a graph can be used, even if enabled. I've taken a similar approach in this PR by adding a method to `ggml-sycl.cpp` for checking if a graph can be used for the operations even if a user has asked for it to be enabled.	2025-05-22 16:24:09 +08:00
Henry Linjamäki	a4e8912dfd	opencl: Add support for multiple devices (#12622 ) * opencl: Add support for multiple devices ... but limited to one platform. A platform with a GPU will be preferred. Additionally: * Filter out devices that lack capabilities needed by the backend implementation (half support, OpenCL 2.0+, etc). * Make ggml_backend_opencl_reg() thread-safe. * fixup: fix an error in sync_with_other_backends ... when there is only one OpenCL device available.	2025-05-21 16:21:45 -07:00
Henry Linjamäki	edbf42edfd	opencl: fix couple crashes (#12795 ) * opencl: fix couple crashes * fix kernel launches failed on devices which do not support non-uniform work-groups. When non-uniform work-groups are not supported, set `local_work_size` to NULL (= let driver choose the work-group sizes). This patch does not cover everything - just the cases tested by test-backend-ops. * fix sub-buffer creation failed due to `cl_buffer_region::origin` not being aligned to `CL_DEVICE_MEM_BASE_ADDR_ALIGN`. * OpenCL: query non-uniform WG sizes only on OpenCL 3.0+	2025-05-21 13:21:17 -07:00
Diego Devesa	d643bb2c79	releases : build CPU backend separately (windows) (#13642 )	2025-05-21 22:09:57 +02:00
Georgi Gerganov	8e186ef0e7	hparams : support models for which all layers use SWA (#13682 ) ggml-ci	2025-05-21 20:00:49 +03:00
Georgi Gerganov	5fbfe384d4	server : improve error reporting (#13680 )	2025-05-21 19:46:56 +03:00
antichristHater	c76532e7ba	convert : add qwen2vl support for unsloth merges (#13686 )	2025-05-21 18:40:35 +02:00
Sigbjørn Skjæret	2aa777d86d	examples : switch retrieval to llama_encode (#13685 ) * switch retrieval to llama_encode * enable --no-warmup for retrieval	2025-05-21 16:57:38 +02:00
Emmanuel Ferdman	eb0f5c28d3	gguf-py : display the invalid gguf type (#13687 ) Signed-off-by: Emmanuel Ferdman <emmanuelferdman@gmail.com>	2025-05-21 16:33:54 +02:00
Xuan-Son Nguyen	cf4cb59e64	ggml : add ggml_gelu_erf() (#13667 ) * ggml : add ggml_gelu_na (not approximated) * fix naming order * rename na --> erf * apply review suggesions * revert naming order	2025-05-21 16:26:33 +02:00
Robin Davidsson	0d5c742161	server : Add the endpoints /api/tags and /api/chat (#13659 ) * Add the endpoints /api/tags and /api/chat Add the endpoints /api/tags and /api/chat, and improved the model metadata response * Remove trailing whitespaces * Removed code that is not needed for copilot to work.	2025-05-21 15:15:27 +02:00
Dorin-Andrei Geman	42158ae2e8	server : fix first message identification (#13634 ) * server : fix first message identification When using the OpenAI SDK (https://github.com/openai/openai-node/blob/master/src/lib/ChatCompletionStream.ts#L623-L626) we noticed that the expected assistant role is missing in the first streaming message. Fix this by correctly checking for the first message. Co-authored-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com> Signed-off-by: Dorin Geman <dorin.geman@docker.com> * server : Fix checks for first role message for stream=True Co-authored-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com> Signed-off-by: Dorin Geman <dorin.geman@docker.com> --------- Signed-off-by: Dorin Geman <dorin.geman@docker.com> Co-authored-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com>	2025-05-21 15:07:57 +02:00
Georgi Gerganov	797f2ac062	kv-cache : simplify the interface (#13660 ) * kv-cache : simplify the interface ggml-ci * context : revert llama_batch_allocr position change ggml-ci	2025-05-21 15:11:13 +03:00
Georgi Gerganov	b44890df2e	model : disable SWA for Phi models (#13676 ) * model : disable SWA for Phi models ggml-ci * model : update warning message * model : print warning only if n_swa > 0 * model : fix typo	2025-05-21 13:09:21 +03:00
R0CKSTAR	33983057d0	musa: Upgrade MUSA SDK version to rc4.0.1 and use mudnn::Unary::IDENTITY op to accelerate D2D memory copy (#13647 ) * musa: fix build warning (unused parameter) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * musa: upgrade MUSA SDK version to rc4.0.1 Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * musa: use mudnn::Unary::IDENTITY op to accelerate D2D memory copy Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * Update ggml/src/ggml-cuda/cpy.cu Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * musa: remove MUDNN_CHECK_GEN and use CUDA_CHECK_GEN instead in MUDNN_CHECK Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> --------- Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2025-05-21 09:58:49 +08:00
Eve	fb1cab201c	vulkan: fix warnings (#13626 ) * small fixes * remove ifdef	2025-05-20 21:35:16 +00:00
l3utterfly	b7a17463ec	mtmd-helper : bug fix to token batching in mtmd (#13650 ) * Update mtmd-helper.cpp * Update tools/mtmd/mtmd-helper.cpp Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> --------- Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>	2025-05-20 18:55:30 +02:00
Georgi Gerganov	be0239693c	model : fix llama4 graph (#13663 ) ggml-ci	2025-05-20 19:21:04 +03:00
Georgi Gerganov	a4090d1174	llama : remove llama_kv_cache_view API + remove deprecated (#13653 ) ggml-ci	2025-05-20 16:13:16 +03:00
Johannes Gäßler	b69f1647f9	CUDA: skip fully masked-out KV in FA vec kernel (#13584 ) * CUDA: skip fully masked-out KV in FA vec kernel	2025-05-20 14:45:07 +02:00
Sigbjørn Skjæret	759e37b0d8	tests : avoid github urls due to throttling (#13654 )	2025-05-20 12:03:17 +02:00
Svetlozar Georgiev	4245e622e0	sycl: disable reorder for sycl mulmat (#13536 )	2025-05-20 11:34:15 +02:00
0cc4m	c9c64dee57	Set GLM4 blk..attn_output.weight, kqv_out- matmul to GGML_PREC_F32 to fix infinity values in output (#13639 )	2025-05-20 10:11:56 +02:00
Georgi Gerganov	c00a2634be	metal : fix typo in FA kernel comments (#13651 )	2025-05-20 10:41:40 +03:00
Georgi Gerganov	e298d2fbd0	kv-cache : add SWA support (#13194 ) * kv-cache : prepare for SWA ggml-ci * kv-cache : initial iSWA implementation ggml-ci * kv-cache : rework error recovery logic ggml-ci * models : fix Phi-3 SWA parameters ggml-ci * model : adjust Granite to rope factor changes ggml-ci * server : check if context can do shifts ggml-ci * iswa : for now, always enable shifts (experiment) ggml-ci * kv-cache : simplify SWA logic ggml-ci * kv-cache : apply defrag when we fail to find slots for the batch ggml-ci * llama : update docs about llama_decode ggml-ci * kv-cache : update warning logs when no space for the batch is available ggml-ci * llama : add llama_kv_self_seq_pos_min() * kv-cache : keep track of partial SWA computes and print warnings * server : disallow use cases involving partial SWA context ggml-ci * llama : add param to control SWA cache size ggml-ci * minor : clean-up ggml-ci	2025-05-20 08:05:46 +03:00
Xinpeng Dou	f0adb80bf7	CANN: Update CANN model support (#13162 ) * Update CANN model support status * Update of model support * update * update * update * fix format of CANN.md * fix format of CANN.md * fix format of CANN.md	2025-05-20 11:43:43 +08:00
Nicolò Scipione	f7c9429c85	sycl : Overcoming workaround for mmap() allocation on Windows (#13482 ) * Remove mmap workaround on windows After some testing I found that mmap is supported on windows and for many GPUs on Linux. Therefore I remove the workaround for windows since it is not necessary. * Update llama-bench README SYCL backend introduced a workaround that allows execution of llama-bench also without specifying `--mmp 0` flag	2025-05-20 08:54:43 +08:00
psocolovsky	1dfbf2cf3a	common : add load_progress_callback (#13617 )	2025-05-19 21:17:36 +02:00
0cc4m	8960efd0a6	Vulkan: Add f32 accumulator support to quantized mul mat to fix GLM4 32B incoherence (#13607 )	2025-05-19 17:54:08 +02:00
Alberto Cabrera Pérez	725f23f1f3	sycl : backend documentation review (#13544 ) * sycl: reviewing and updating docs * Updates Runtime error codes * Improves OOM troubleshooting entry * Added a llama 3 sample * Updated supported models * Updated releases table	2025-05-19 14:38:20 +01:00
Xuan-Son Nguyen	92ecdcc06a	mtmd : add vision support for llama 4 (#13282 ) * wip llama 4 conversion * rm redundant __init__ * fix conversion * fix conversion * test impl * try this * reshape patch_embeddings_0 * fix view * rm ffn_post_norm * cgraph ok * f32 for pos embd * add image marker tokens * Llama4UnfoldConvolution * correct pixel shuffle * fix merge conflicts * correct * add debug_graph * logits matched, but it still preceives the image incorrectly * fix style * add image_grid_pinpoints * handle llama 4 preprocessing * rm load_image_size * rm unused line * fix * small fix 2 * add test & docs * fix llava-1.6 test * test: add notion of huge models * add comment * add warn about degraded quality	2025-05-19 13:04:14 +02:00
Alberto Cabrera Pérez	f71f40a284	ci : upgraded oneAPI version in SYCL workflows and dockerfile (#13532 )	2025-05-19 11:46:09 +01:00
Georgi Gerganov	d30cb5a7fa	sync : ggml ggml-ci	2025-05-19 13:29:56 +03:00
Johannes Gäßler	6c35981a64	mnist: fix segmentation fault (ggml/1227)	2025-05-19 13:29:56 +03:00
Diego Devesa	8b5e19aea6	ggml : fix apple OS check in ggml_print_backtrace (ggml/1229)	2025-05-19 13:29:56 +03:00
Daniel Tang	60aea028b5	ggml : Fix missing backtrace on Linux (ggml/1228) * Modern Linux defaults /proc/sys/kernel/yama/ptrace_scope to 1 * Fixed lldb attach * Simplify by having the child do ggml_print_backtrace_symbols	2025-05-19 13:29:56 +03:00
Nick	9c55e5c5c2	fix: check model pointer validity before use (#13631 )	2025-05-19 13:25:41 +03:00
Chenguang Li	33d7aed4a8	CANN: Support MOE Model MUL_MAT_ID (#13042 ) Signed-off-by: noemotiovon <757486878@qq.com>	2025-05-19 14:21:17 +08:00
Isaac McFadyen	6a2bc8bfb7	server : added --no-prefill-assistant flag (#13608 ) * added no-prefill-assistant flag * reworded documentation comment * updated server README.md	2025-05-17 23:59:48 +02:00
Gilad S.	e3a7cf6c5b	cmake: use the current build config for vulkan-shaders-gen (#13595 ) * fix: use the current build config for `vulkan-shaders-gen` * fix: only pass a valid build type to `--config`	2025-05-17 15:26:43 -03:00
Georgi Gerganov	518329b2d4	parallel : add option for non-shared and larger prompts (#13598 ) * parallel : add option for non-shared and larger prompts * parallel : update readme [no ci] * cont : add note about base models [no ci] * parallel : better var name ggml-ci	2025-05-17 12:58:55 +03:00
Jeff Bolz	2f5a4e1e09	vulkan: move common FA code to flash_attn_base.comp (#13556 ) * vulkan: move common FA code to flash_attn_base.comp * vulkan: move common FA index/stride setup code to flash_attn_base.comp * build fix	2025-05-17 09:14:55 +02:00
Jeff Bolz	4f41ee11d6	vulkan: use scalar FA rather than coopmat2 when N==1 (#13554 )	2025-05-17 08:35:47 +02:00