fit : wrap llama_device_memory_data (#24522)

* fix sycl links in release notes

* remove extra line

.devops/cuda.Dockerfile

@@ -1,6 +1,7 @@
 ARG UBUNTU_VERSION=24.04
 # This needs to generally match the container host's environment.
 ARG CUDA_VERSION=12.8.1
+ARG GCC_VERSION=14
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
@@ -12,13 +13,14 @@ ARG APP_REVISION=N/A
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
+ARG GCC_VERSION
 # CUDA architecture to build for (defaults to all supported archs)
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y gcc-14 g++-14 build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y gcc-${GCC_VERSION} g++-${GCC_VERSION} build-essential cmake python3 python3-pip git libssl-dev libgomp1
 
-ENV CC=gcc-14 CXX=g++-14 CUDAHOSTCXX=g++-14
+ENV CC=gcc-${GCC_VERSION} CXX=g++-${GCC_VERSION} CUDAHOSTCXX=g++-${GCC_VERSION}
 
 WORKDIR /app
 

.github/workflows/build-sycl.yml

@@ -34,129 +34,108 @@ env:
   LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
 
-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  ubuntu-24-sycl:
-#    strategy:
-#      matrix:
-#        build: [fp32]
-#        include:
-#          - build: fp32
-#            fp16: OFF
-#
-#    runs-on: ubuntu-24.04
-#
-#    env:
-#      ONEAPI_ROOT: /opt/intel/oneapi/
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#      LEVEL_ZERO_VERSION: "1.28.2"
-#      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
-#
-#    continue-on-error: true
-#
-#    steps:
-#      - uses: actions/checkout@v6
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          cd /tmp
-#          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-#          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
-#
-#      - name: Install Level Zero SDK
-#        shell: bash
-#        run: |
-#          cd /tmp
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
-#          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
-#
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: sycl-ubuntu-24-${{ matrix.build }}
-#          evict-old-files: 1d
-#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          source /opt/intel/oneapi/setvars.sh
-#          cmake -B build \
-#            -G "Ninja" \
-#            -DCMAKE_BUILD_TYPE=Release \
-#            -DGGML_SYCL=ON \
-#            -DCMAKE_C_COMPILER=icx \
-#            -DCMAKE_CXX_COMPILER=icpx \
-#            -DLLAMA_OPENSSL=OFF \
-#            -DGGML_NATIVE=OFF \
-#            -DGGML_SYCL_F16=${{ matrix.fp16 }}
-#          time cmake --build build --config Release -j $(nproc)
+    runs-on: ubuntu-24.04
 
-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  windows-latest-sycl:
-#    runs-on: windows-2022
-#
-#    defaults:
-#      run:
-#        shell: bash
-#
-#    env:
-#      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_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
-#      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
-#      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-#
-#      - name: Install Level Zero SDK
-#        shell: pwsh
-#        run: |
-#          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
-#          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
-#          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: sycl-windows-latest
-#          variant: ccache
-#          evict-old-files: 1d
-#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-#
-#      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-#
-#      - name: Build
-#        id: cmake_build
-#        run:  examples/sycl/win-build-sycl.bat
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      LEVEL_ZERO_VERSION: "1.28.2"
+      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Download & Install oneAPI
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: Install Level Zero SDK
+        shell: bash
+        run: |
+          cd /tmp
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: sycl-ubuntu-24-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+  windows-latest-sycl:
+    runs-on: windows-2022
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_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
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Download & Install oneAPI
+        shell: bash
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Install Level Zero SDK
+        shell: pwsh
+        run: |
+          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
+          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
+          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: sycl-windows-latest
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      # TODO: add ssl 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

.github/workflows/release.yml

@@ -59,8 +59,31 @@ jobs:
             echo "should_release=false" >> $GITHUB_OUTPUT
           fi
 
+  get-version:
+    runs-on: ubuntu-slim
+    outputs:
+      ui_version: ${{ steps.version.outputs.ui_version }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - id: version
+        run: |
+          # Resolve UI version: BUILD_NUMBER from cmake/build-info.cmake > git hash + epoch > fallback
+          version=""
+          if grep -q "BUILD_NUMBER" cmake/build-info.cmake; then
+            build_number=$(grep "set(BUILD_NUMBER" cmake/build-info.cmake | grep -oP '\d+')
+            if [ -n "$build_number" ] && [ "$build_number" -gt 0 ]; then
+              version="b${build_number}"
+            fi
+          fi
+          if [ -z "$version" ]; then
+            version=$(git rev-parse --short HEAD)-$(date +%s)
+          fi
+          echo "ui_version=${version}" >> $GITHUB_OUTPUT
+
   macos-cpu:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -116,6 +139,7 @@ jobs:
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -141,7 +165,7 @@ jobs:
           name: llama-bin-macos-${{ matrix.build }}.tar.gz
 
   ubuntu-cpu:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -201,6 +225,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -227,7 +252,7 @@ jobs:
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
 
   ubuntu-vulkan:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     strategy:
@@ -287,6 +312,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -312,7 +338,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
   android-arm64:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-latest
@@ -379,6 +405,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_OPENMP=OFF \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -404,7 +431,7 @@ jobs:
           name: llama-bin-android-arm64.tar.gz
 
   ubuntu-24-openvino:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-24.04
@@ -476,7 +503,8 @@ jobs:
           source ./openvino_toolkit/setupvars.sh
           cmake -B build/ReleaseOV -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_OPENVINO=ON
+            -DGGML_OPENVINO=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
           cmake --build build/ReleaseOV --config Release -j $(nproc)
 
       - name: ccache-clear
@@ -754,213 +782,205 @@ jobs:
           path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  windows-sycl:
-#
-#    runs-on: windows-2022
-#
-#    defaults:
-#      run:
-#        shell: bash
-#
-#    env:
-#      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_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
-#      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
-#      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-#
-#      - name: Install Level Zero SDK
-#        shell: pwsh
-#        run: |
-#          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
-#          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
-#          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
-#
-#      - name: Setup Node.js
-#        uses: actions/setup-node@v6
-#        with:
-#          node-version: "24"
-#          cache: "npm"
-#          cache-dependency-path: "tools/ui/package-lock.json"
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: release-windows-2022-x64-sycl
-#
-#      - name: Build
-#        id: cmake_build
-#        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_BUILD_BORINGSSL=ON
-#          cmake --build build --target ggml-sycl -j
-#
-#      - name: Build the release package
-#        id: pack_artifacts
-#        run: |
-#          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-#
-#          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
-#
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
-#          ZE_LOADER_DLL=$(find "${{ env.ONEAPI_ROOT }}" "$LEVEL_ZERO_V1_SDK_PATH" -iname ze_loader.dll -print -quit 2>/dev/null || true)
-#          if [ -n "$ZE_LOADER_DLL" ]; then
-#            echo "Using Level Zero loader: $ZE_LOADER_DLL"
-#            cp "$ZE_LOADER_DLL" ./build/bin
-#          else
-#            echo "Level Zero loader DLL not found in oneAPI or SDK; relying on system driver/runtime"
-#          fi
-#
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin
-#
-#          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
-#
-#          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/tcm.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/libhwloc-15.dll" ./build/bin
-#          cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
-#
-#          echo "cp oneAPI running time dll files to ./build/bin done"
-#          7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
-#
-#      - name: Upload the release package
-#        uses: actions/upload-artifact@v6
-#        with:
-#          path: llama-bin-win-sycl-x64.zip
-#          name: llama-bin-win-sycl-x64.zip
+  windows-sycl:
 
-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  ubuntu-24-sycl:
-#
-#    strategy:
-#      matrix:
-#        build: [fp32]
-#        include:
-#          - build: fp32
-#            fp16: OFF
-#
-#    runs-on: ubuntu-24.04
-#
-#    env:
-#      ONEAPI_ROOT: /opt/intel/oneapi/
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#      LEVEL_ZERO_VERSION: "1.28.2"
-#      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#        with:
-#          fetch-depth: 0
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          cd /tmp
-#          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-#          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
-#
-#      - name: Install Level Zero SDK
-#        shell: bash
-#        run: |
-#          cd /tmp
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
-#          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
-#
-#      - name: Setup Node.js
-#        uses: actions/setup-node@v6
-#        with:
-#          node-version: "24"
-#          cache: "npm"
-#          cache-dependency-path: "tools/ui/package-lock.json"
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: release-ubuntu-24.04-sycl
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          source /opt/intel/oneapi/setvars.sh
-#          cmake -B build \
-#            -G "Ninja" \
-#            -DCMAKE_BUILD_TYPE=Release \
-#            -DGGML_SYCL=ON \
-#            -DCMAKE_C_COMPILER=icx \
-#            -DCMAKE_CXX_COMPILER=icpx \
-#            -DLLAMA_OPENSSL=OFF \
-#            -DGGML_NATIVE=OFF \
-#            -DGGML_SYCL_F16=${{ matrix.fp16 }}
-#          time cmake --build build --config Release -j $(nproc)
-#
-#      - name: Determine tag name
-#        id: tag
-#        uses: ./.github/actions/get-tag-name
-#
-#      - name: Pack artifacts
-#        id: pack_artifacts
-#        run: |
-#          cp LICENSE ./build/bin/
-#          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .
-#
-#      - name: Upload artifacts
-#        uses: actions/upload-artifact@v6
-#        with:
-#          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
-#          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+    runs-on: windows-2022
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_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
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Download & Install oneAPI
+        shell: bash
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Install Level Zero SDK
+        shell: pwsh
+        run: |
+          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
+          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
+          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-windows-2022-x64-sycl
+
+      - name: Build
+        id: cmake_build
+        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_BUILD_BORINGSSL=ON
+          cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
+      - name: Build the release package
+        id: pack_artifacts
+        run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
+          ZE_LOADER_DLL=$(find "${{ env.ONEAPI_ROOT }}" "$LEVEL_ZERO_V1_SDK_PATH" -iname ze_loader.dll -print -quit 2>/dev/null || true)
+          if [ -n "$ZE_LOADER_DLL" ]; then
+            echo "Using Level Zero loader: $ZE_LOADER_DLL"
+            cp "$ZE_LOADER_DLL" ./build/bin
+          else
+            echo "Level Zero loader DLL not found in oneAPI or SDK; relying on system driver/runtime"
+          fi
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/tcm.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/libhwloc-15.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
+
+          echo "cp oneAPI running time dll files to ./build/bin done"
+          7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
+
+      - name: Upload the release package
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-bin-win-sycl-x64.zip
+          name: llama-bin-win-sycl-x64.zip
+
+  ubuntu-24-sycl:
+
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
+
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      LEVEL_ZERO_VERSION: "1.28.2"
+      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Download & Install oneAPI
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: Install Level Zero SDK
+        shell: bash
+        run: |
+          cd /tmp
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
 
   ubuntu-22-rocm:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-22.04
@@ -1052,6 +1072,7 @@ jobs:
             -DGGML_HIP=ON \
             -DHIP_PLATFORM=amd \
             -DGGML_HIP_ROCWMMA_FATTN=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -1080,7 +1101,7 @@ jobs:
           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 
   windows-hip:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2022
@@ -1176,6 +1197,7 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} `
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
@@ -1203,7 +1225,7 @@ jobs:
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     runs-on: macos-26
 
@@ -1232,7 +1254,8 @@ jobs:
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=iOS \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=16.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
       - name: xcodebuild for swift package
@@ -1352,10 +1375,12 @@ jobs:
 #          path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
 #          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
 
-  ui:
-    needs: [check-release]
+  ui-build:
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     uses: ./.github/workflows/ui-build.yml
+    with:
+      hf_ui_version: ${{ needs.get-version.outputs.ui_version }}
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1368,6 +1393,7 @@ jobs:
     runs-on: ubuntu-slim
 
     needs:
+      - get-version
       - windows
       - windows-cpu
       - windows-cuda
@@ -1382,7 +1408,7 @@ jobs:
       - macos-cpu
       - ios-xcode
       #- openEuler-cann
-      - ui
+      - ui-build
 
     outputs:
       tag_name: ${{ steps.tag.outputs.name }}
@@ -1482,7 +1508,8 @@ jobs:
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
             - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
-            - Ubuntu x64 (SYCL FP32) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)
+            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
+            - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
 
             **Android:**
             - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)
@@ -1493,7 +1520,7 @@ jobs:
             - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
             - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.3-x64.zip) - [CUDA 13.3 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
-            - Windows x64 (SYCL) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)
+            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
             - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
 
             **openEuler:**

.github/workflows/ui-build.yml

@@ -2,6 +2,11 @@ name: UI Build
 
 on:
   workflow_call:
+    inputs:
+      hf_ui_version:
+        description: 'Version string for version.json (e.g. 12345)'
+        required: false
+        type: string
 
 jobs:
   build:
@@ -25,9 +30,16 @@ jobs:
         working-directory: tools/ui
 
       - name: Build application
+        env:
+          HF_UI_VERSION: ${{ inputs.hf_ui_version || '' }}
+          LLAMA_UI_VERSION: ${{ inputs.hf_ui_version || 'b0000' }}
         run: npm run build
         working-directory: tools/ui
 
+      - name: Run PWA unit tests (versioned build output)
+        run: npx vitest --project=unit --run tests/unit/pwa.spec.ts
+        working-directory: tools/ui
+
       - name: Generate checksums
         run: |
           cd tools/ui/dist

.github/workflows/ui-self-hosted.yml

@@ -1,8 +1,8 @@
 name: UI (self-hosted)
 
 # these are the same as ui.yml, but with self-hosted runners
-# the runners come with pre-installed Playwright browsers version: 1.56.1
-# the jobs are much lighter because they don't need to install node and playwright browsers
+# the jobs are lighter because they don't need to install Node.js or Playwright browsers
+# the runner has pre-installed Playwright browsers for @playwright/test (1.56.1) at /ms-playwright/
 
 on:
   workflow_dispatch:
@@ -61,6 +61,12 @@ jobs:
         run: npm ci
         working-directory: tools/ui
 
+      - name: Download built UI artifacts
+        uses: actions/download-artifact@v6
+        with:
+          name: ui-build
+          path: tools/ui/dist/
+
       - name: Run type checking
         if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run check
@@ -72,12 +78,12 @@ jobs:
         working-directory: tools/ui
 
       - name: Run Client tests
-        if: ${{ always() }}
+        if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run test:client
         working-directory: tools/ui
 
       - name: Run Unit tests
-        if: ${{ always() }}
+        if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run test:unit
         working-directory: tools/ui
 
@@ -97,22 +103,23 @@ jobs:
         run: npm ci
         working-directory: tools/ui
 
-      - name: Build application
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run build
-        working-directory: tools/ui
+      - name: Download built UI artifacts
+        uses: actions/download-artifact@v6
+        with:
+          name: ui-build
+          path: tools/ui/dist/
 
       - name: Build Storybook
-        if: ${{ always() }}
+        if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run build-storybook
         working-directory: tools/ui
 
       - name: Run UI tests
-        if: ${{ always() }}
+        if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run test:ui -- --testTimeout=60000
         working-directory: tools/ui
 
       - name: Run E2E tests
-        if: ${{ always() }}
+        if: ${{ always() && steps.setup.conclusion == 'success' }}
         run: npm run test:e2e
         working-directory: tools/ui

.github/workflows/ui.yml

@@ -43,7 +43,7 @@ jobs:
   ui-checks:
     name: Checks
     needs: ui-build
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
     continue-on-error: true
     steps:
       - name: Checkout code
@@ -60,6 +60,12 @@ jobs:
           cache: "npm"
           cache-dependency-path: "tools/ui/package-lock.json"
 
+      - name: Download built UI artifacts
+        uses: actions/download-artifact@v6
+        with:
+          name: ui-build
+          path: tools/ui/dist/
+
       - name: Install dependencies
         id: setup
         if: ${{ steps.node.conclusion == 'success' }}
@@ -87,7 +93,7 @@ jobs:
         run: npm run test:client
         working-directory: tools/ui
 
-      - name: Run Unit tests
+      - name: Run Unit tests (uses pre-built dist/ from ui-build)
         if: ${{ always() && steps.playwright.conclusion == 'success' }}
         run: npm run test:unit
         working-directory: tools/ui
@@ -95,7 +101,7 @@ jobs:
   e2e-tests:
     name: E2E Tests
     needs: ui-build
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
     steps:
       - name: Checkout code
         uses: actions/checkout@v6
@@ -117,10 +123,11 @@ jobs:
         run: npm ci
         working-directory: tools/ui
 
-      - name: Build application
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run build
-        working-directory: tools/ui
+      - name: Download built UI artifacts (reuses ui-build)
+        uses: actions/download-artifact@v6
+        with:
+          name: ui-build
+          path: tools/ui/dist/
 
       - name: Install Playwright browsers
         id: playwright
@@ -138,7 +145,7 @@ jobs:
         run: npm run test:ui -- --testTimeout=60000
         working-directory: tools/ui
 
-      - name: Run E2E tests
+      - name: Run E2E tests (uses pre-built dist/ from ui-build)
         if: ${{ always() && steps.playwright.conclusion == 'success' }}
         run: npm run test:e2e
         working-directory: tools/ui

.gitignore

@@ -92,13 +92,6 @@
 !/examples/sycl/*.bat
 !/examples/sycl/*.sh
 
-# Server Web UI temporary files (+ legacy directory)
-
-/tools/server/webui/node_modules
-/tools/server/webui/dist
-/tools/ui/node_modules
-/tools/ui/dist
-
 # Python
 
 /.venv

README.md

@@ -1,6 +1,6 @@
 # llama.cpp
 
-![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
+![llama](https://raw.githubusercontent.com/ggml-org/llama.brand/refs/heads/master/cover/llama-cpp/cover-llama-cpp-dark.svg)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)

common/arg.cpp

@@ -2243,6 +2243,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.image_max_tokens = value;
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_IMAGE_MAX_TOKENS"));
+    add_opt(common_arg(
+        {"--mtmd-batch-max-tokens"}, "N",
+        string_format("maximum number of image tokens per batch when encoding images (default: %d)", params.mtmd_batch_max_tokens),
+        [](common_params & params, int value) {
+            params.mtmd_batch_max_tokens = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",

common/common.h

@@ -575,6 +575,7 @@ struct common_params {
     std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
+    int mtmd_batch_max_tokens = 1024;
 
     // finetune
     struct lr_opt lr;

common/fit.cpp

@@ -26,7 +26,7 @@ class common_params_fit_exception : public std::runtime_error {
     using std::runtime_error::runtime_error;
 };
 
-std::vector<llama_device_memory_data> common_get_device_memory_data(
+static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
         const char * path_model,
         const llama_model_params * mparams,
         const llama_context_params * cparams,
@@ -150,6 +150,29 @@ std::vector<llama_device_memory_data> common_get_device_memory_data(
     return ret;
 }
 
+common_device_memory_data_vec common_get_device_memory_data(
+        const char * path_model,
+        const llama_model_params * mparams,
+        const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs,
+        uint32_t & hp_ngl,
+        uint32_t & hp_n_ctx_train,
+        uint32_t & hp_n_expert,
+        ggml_log_level log_level) {
+    std::vector<llama_device_memory_data> impl = common_get_device_memory_data_impl(
+            path_model, mparams, cparams, devs, hp_ngl, hp_n_ctx_train, hp_n_expert, log_level);
+
+    common_device_memory_data_vec ret(impl.size());
+    for (size_t i = 0; i < impl.size(); i++) {
+        ret[i].total   = impl[i].total;
+        ret[i].free    = impl[i].free;
+        ret[i].model   = impl[i].mb.model;
+        ret[i].context = impl[i].mb.context;
+        ret[i].compute = impl[i].mb.compute;
+    }
+    return ret;
+}
+
 static void common_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
@@ -169,7 +192,7 @@ static void common_params_fit_impl(
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    const dmds_t dmds_full = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
     const size_t nd = devs.size(); // number of devices
 
     std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
@@ -304,7 +327,7 @@ static void common_params_fit_impl(
 
                     int64_t sum_projected_used_min_ctx = 0;
                     cparams->n_ctx = n_ctx_min;
-                    const dmds_t dmds_min_ctx = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
                     if (nd == 0) {
                         sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
                     } else {
@@ -482,7 +505,7 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        const dmds_t dmd_nl = common_get_device_memory_data(
+        const dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
@@ -510,7 +533,7 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        const dmds_t dmds_cpu_moe = common_get_device_memory_data(
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
 
         for (size_t id = 0; id < nd; id++) {
@@ -940,7 +963,7 @@ void common_fit_print(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
-    auto dmd = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
+    auto dmd = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
     GGML_ASSERT(dmd.size() == devs.size() + 1);
 
     for (size_t id = 0; id < devs.size(); id++) {

common/fit.h

@@ -1,9 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-backend.h"
 #include "llama.h"
-#include "../src/llama-ext.h"
 
 #include <vector>
 
@@ -18,31 +16,41 @@ enum common_params_fit_status {
 //   - this function is NOT thread safe because it modifies the global llama logger state
 //   - only parameters that have the same value as in llama_default_model_params are modified
 //     with the exception of the context size which is modified if and only if equal to 0
-enum common_params_fit_status common_fit_params(
-                               const char   * path_model,
-                struct llama_model_params   * mparams,
-                struct llama_context_params * cparams,
-                                      float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
-    struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
-                                     size_t * margins,               // margins of memory to leave per device in bytes
-                                   uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
-                        enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
+common_params_fit_status common_fit_params(
+                         const char * path_model,
+                 llama_model_params * mparams,
+               llama_context_params * cparams,
+                              float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
+   llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
+                             size_t * margins,               // margins of memory to leave per device in bytes
+                           uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+                     ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout
 void common_fit_print(
-                               const char   * path_model,
-                struct llama_model_params   * mparams,
-                struct llama_context_params * cparams);
+                         const char * path_model,
+                 llama_model_params * mparams,
+               llama_context_params * cparams);
 
-void common_memory_breakdown_print(const struct llama_context * ctx);
+void common_memory_breakdown_print(const llama_context * ctx);
+
+struct common_device_memory_data {
+    int64_t total;
+    int64_t free;
+    size_t  model;
+    size_t  context;
+    size_t  compute;
+};
+
+using common_device_memory_data_vec = std::vector<common_device_memory_data>;
 
 // Load a model + context with no_alloc and return the per-device memory breakdown.
-std::vector<llama_device_memory_data> common_get_device_memory_data(
-                                  const char   * path_model,
-        const struct llama_model_params         * mparams,
-        const struct llama_context_params       * cparams,
-        std::vector<ggml_backend_dev_t>         & devs,
-                                      uint32_t  & hp_ngl,
-                                      uint32_t  & hp_n_ctx_train,
-                                      uint32_t  & hp_n_expert,
-                           enum ggml_log_level    log_level);
+common_device_memory_data_vec common_get_device_memory_data(
+                         const char * path_model,
+           const llama_model_params * mparams,
+         const llama_context_params * cparams,
+    std::vector<ggml_backend_dev_t> & devs,
+                           uint32_t & hp_ngl,
+                           uint32_t & hp_n_ctx_train,
+                           uint32_t & hp_n_expert,
+                     ggml_log_level   log_level);

common/speculative.cpp

@@ -375,31 +375,437 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
     }
 };
 
+
+// EAGLE3 speculative decoding state
+//
+// Input of draft decoder: (This is different compared to MTP)
+//   At "pos P", the decoder takes input pair (t_{P+1}, g_P), with RoPE at P.
+//     - t_{P+1} = token at sequence pos P+1 (the *next* token after P)
+//     - g_P     = encoder output = projection of target's extracted hidden states at P
+//
+// Deferred boundary (MTP doesn't have this issue):
+//   Within a single process() call with n_tokens, we can only write decoder KV for
+//   training pos 0..n_tokens-2. The last training pos (n_tokens-1) needs t_{n_tokens}
+//   which lies *outside* this batch — it is the token target will sample next or the first token from next ubatch.
+//   So the last training pos of each process() call is *deferred* to whichever next call has
+//   the missing token in hand:
+//     - multi-ubatch prefill: the next process()'s first token completes the pair
+//                              (handled by the per-seq "cross-ubatch bridge")
+//     - single-ubatch prefill / after verify: draft()'s seed step uses "dp.id_last"
+//                              (target's freshest sample) to complete the pair
+//
+// Per-seq carry-over state:
+//   pending_g_last    [n_embd_dec]  ┐  the deferred boundary's (g, pos). Set by
+//   pending_pos_last  llama_pos     ┘  process() at end of ubatch (= last row);
+//                                       rebased by accept() to first-non-accepted pos.
+//   verify_g          [N × n_embd_dec] snapshot of process()'s encoder output;
+//   verify_pos_first  llama_pos         consumed by accept() to recover the right
+//   verify_g_rows     int32_t           pending_g_last row for any n_accepted value.
+//
+// Performance is overall good but there is waste in verify cycle:
+//   process() runs encoder + decoder on the *full* verify batch including rows for
+//   rejected drafts. The KV at those positions is then dropped.
+//
+// TODO: Not sure if we need optimization for this waste?
+// If so we may need hybrid stash:
+//      in verify mode, have process() only stash features and let draft() seed run
+//      encoder+decoder on n_accepted+1 rows).
 struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
-    //common_params_speculative_eagle3 params;
+    common_params_speculative_draft params;
+    llama_batch batch;
+
+    std::vector<common_sampler_ptr> smpls;
+
+    int32_t n_embd_dec = 0;       // draft hidden size
+    int32_t n_embd_enc = 0;       // target_layer_ids_n * target_hidden_size
+    int32_t n_embd_tgt = 0;       // target model hidden size
+
+    const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
+    uint32_t        target_layer_ids_n = 0;
+
+    // [per-seq] deferred boundary state
+    std::vector<std::vector<float>> pending_g_last;
+    std::vector<llama_pos>          pending_pos_last;
+
+    // [per-seq] snapshot of the most recent process()'s encoder output
+    std::vector<std::vector<float>> verify_g;         // [n_seq][n_rows * n_embd_dec]
+    std::vector<llama_pos>          verify_pos_first; // [n_seq] — pos of verify_g[seq][0]
+    std::vector<int32_t>            verify_g_rows;    // [n_seq] — number of rows
+
+    // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
+    std::vector<float> features_buf;
+    std::vector<float> g_embd_buf;
 
     common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq)
+        , params(params.draft)
     {
         LOG_INF("%s: adding speculative implementation 'draft-eagle3'\n", __func__);
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, params.draft.n_max, params.draft.n_min, params.draft.p_min);
+
+        auto * ctx_tgt = this->params.ctx_tgt;
+        auto * ctx_dft = this->params.ctx_dft;
+        GGML_ASSERT(ctx_tgt && ctx_dft && "EAGLE3 requires ctx_tgt and ctx_dft to be set");
+
+        const llama_model * model_dft = llama_get_model(ctx_dft);
+        const llama_model * model_tgt = llama_get_model(ctx_tgt);
+
+        target_layer_ids   = llama_model_target_layer_ids  (model_dft);
+        target_layer_ids_n = llama_model_target_layer_ids_n(model_dft);
+        if (target_layer_ids_n != 3) {
+            throw std::runtime_error("draft model is not eagle3 (expected 3 extract layers, got " +
+                                     std::to_string(target_layer_ids_n) + ")");
+        }
+
+        n_embd_tgt = llama_model_n_embd(model_tgt);
+        n_embd_dec = llama_model_n_embd(model_dft);
+        n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt;
+
+        const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
+        batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd_dec, /*n_seq_max=*/ 1);
+        // llama_batch_init allocates only one of token/embd; eagle3 decoder needs both.
+        // TODO: fix, how to call without malloc
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_b);
+
+        smpls.resize(n_seq);
+        for (auto & s : smpls) {
+            common_params_sampling sparams;
+            sparams.no_perf  = false;
+            sparams.top_k    = 10;
+            sparams.samplers = { COMMON_SAMPLER_TYPE_TOP_K };
+            s.reset(common_sampler_init(llama_get_model(ctx_dft), sparams));
+        }
+
+        // turn on extraction of the target layers' input embeddings
+        for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
+            llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
+        }
+
+        // turn on extraction of the draft model's pre-norm hidden state
+        // (used both for the encoder output g_embd and the decoder pre-norm output).
+        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
+
+        pending_g_last.assign(n_seq, std::vector<float>(n_embd_dec, 0.0f));
+        pending_pos_last.assign(n_seq, -1);
+
+        verify_g.assign(n_seq, std::vector<float>());
+        verify_pos_first.assign(n_seq, -1);
+        verify_g_rows.assign(n_seq, 0);
     }
 
-    void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
-        // noop
+    ~common_speculative_impl_draft_eagle3() override {
+        if (batch.token != nullptr) {
+            free(batch.token);
+            batch.token = nullptr;
+        }
+        llama_batch_free(batch);
     }
 
-    bool process(const llama_batch & /*batch*/) override {
-        // TODO: implement
+    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
+        const int32_t N = (int32_t) prompt.size();
+        if (N <= 0) {
+            return;
+        }
+        // expected state after prefill: ctx_dft has pos 0..N-2 (last position is deferred to
+        // draft()'s seed step). Warn only if more than one position is missing.
+        auto * ctx_dft = this->params.ctx_dft;
+        const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
+        if (pos_max < N - 2) {
+            LOG_WRN("%s: ctx_dft pos_max=%d < N-2=%d — process() did not run on every prefill ubatch. "
+                    "Drafts may degrade.\n",
+                    __func__, (int) pos_max, N - 2);
+        }
+    }
+
+    bool process(const llama_batch & batch_in) override {
+        if (batch_in.n_tokens <= 0) {
+            return true;
+        }
+
+        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
+            return true;
+        }
+
+        const int32_t n_tokens = batch_in.n_tokens;
+
+        // i_batch_beg[seq] / i_batch_end[seq]: inclusive batch indices of this seq's
+        // first/last token in batch_in. Assumes per-seq tokens are contiguous within
+        // the ubatch (server's default ordering).
+        std::vector<int32_t> i_batch_beg(n_seq, -1);
+        std::vector<int32_t> i_batch_end(n_seq, -1);
+        for (int k = 0; k < n_tokens; ++k) {
+            GGML_ASSERT(batch_in.n_seq_id[k] == 1);
+            const llama_seq_id seq_id = batch_in.seq_id[k][0];
+            if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
+                continue;
+            }
+            i_batch_end[seq_id] = k;
+            if (i_batch_beg[seq_id] < 0) {
+                i_batch_beg[seq_id] = k;
+            }
+        }
+
+        auto * ctx_tgt = this->params.ctx_tgt;
+        auto * ctx_dft = this->params.ctx_dft;
+
+        // Interleave each extract_layer's hidden state into a contiguous buffer of
+        // shape [n_tokens, target_layer_ids_n * n_embd_tgt]. Then run EAGLE3 encoder
+        // to get one g_embd row per token.
+        features_buf.resize((size_t) n_tokens * n_embd_enc, 0.0f);
+
+        for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
+            const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
+            if (!layer) {
+                GGML_ABORT("EAGLE3: target layer %d input not extracted.", target_layer_ids[k]);
+            }
+            for (int32_t i = 0; i < n_tokens; ++i) {
+                float * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
+                const float * src = layer + (size_t) i * n_embd_tgt;
+                std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float));
+            }
+        }
+
+        g_embd_buf.resize((size_t) n_tokens * n_embd_dec);
+
+        // llama_encode() requires the full encoder batch to fit in n_ubatch.
+        // Allow batch > ubatch: eagle3's per-token encoder can be chunked safely.
+        const int32_t n_ubatch_dft = (int32_t) llama_n_ubatch(ctx_dft);
+        for (int32_t i = 0; i < n_tokens; i += n_ubatch_dft) {
+            const int32_t n_chunk = std::min(n_ubatch_dft, n_tokens - i);
+
+            llama_batch enc_batch = {
+                /*.n_tokens =*/ n_chunk,
+                /*.token    =*/ nullptr,
+                /*.embd     =*/ features_buf.data() + (size_t) i * n_embd_enc,
+                /*.pos      =*/ nullptr,
+                /*.n_seq_id =*/ nullptr,
+                /*.seq_id   =*/ nullptr,
+                /*.logits   =*/ nullptr,
+            };
+            const int32_t rc = llama_encode(ctx_dft, enc_batch);
+            if (rc != 0) {
+                LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                        __func__, rc, (int) n_chunk, (int) i);
+                return false;
+            }
+
+            // g_embd has shape [n_chunk, n_embd_dec] in ctx_dft's pre-norm embeddings buffer.
+            const float * g_embd_chunk = llama_get_embeddings_nextn(ctx_dft);
+            GGML_ASSERT(g_embd_chunk && "EAGLE3 encoder produced no output.");
+            std::memcpy(g_embd_buf.data() + (size_t) i * n_embd_dec,
+                        g_embd_chunk,
+                        (size_t) n_chunk * n_embd_dec * sizeof(float));
+        }
+
+        const float * g_embd = g_embd_buf.data();
+
+        const size_t row_bytes = (size_t) n_embd_dec * sizeof(float);
+
+        // EAGLE3 decoder input convention: at memory pos P the input pair is
+        // (token[P+1], g_embd[P]). This shifts the token index "left by one" relative to g_embd.
+        //
+        // Per seq, in order:
+        //   (a) cross-ubatch bridge — when applicable, write the previously-deferred
+        //       pos using this ubatch's first token + pending_g_last.
+        //   (b) main write loop — for k in [beg, end-1], write (token[k+1], g_embd[k])
+        //       at pos[k]. The last training pos (k=end) is left unwritten = new
+        //       deferred boundary, completed by the next process() or draft() call.
+        //   (c) refresh deferred state — stash this ubatch's full g_embd into verify_g,
+        //       update pending_g_last / pending_pos_last to the last row.
+        common_batch_clear(batch);
+
+        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+            const int32_t beg = i_batch_beg[seq_id];
+            const int32_t end = i_batch_end[seq_id];
+            if (beg < 0 || end < 0) {
+                continue;
+            }
+
+            // cross-ubatch bridge — complete the prior ubatch's deferred boundary.
+            // Fires iff all three preconditions hold:
+            //   1) pending_pos_last >= 0
+            //   2) pending_pos_last + 1 == pos[beg]
+            //   3) pending_pos_last > dft_pos_max // TODO: is this check needed?
+            const llama_pos pending_pos = pending_pos_last[seq_id];
+            if (pending_pos >= 0 && pending_pos + 1 == batch_in.pos[beg]) {
+                const llama_pos dft_pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
+                if (pending_pos > dft_pos_max) {
+                    common_batch_add(batch, batch_in.token[beg], pending_pos, { seq_id }, /*logits=*/ false);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                                pending_g_last[seq_id].data(), row_bytes);
+                }
+            }
+
+            for (int32_t k = beg; k < end; ++k) {
+                common_batch_add(batch, batch_in.token[k + 1], batch_in.pos[k], { seq_id }, /*logits=*/ false);
+                std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                            g_embd + (size_t) k * n_embd_dec, row_bytes);
+            }
+
+            // refresh deferred state
+            const int32_t n_rows = end - beg + 1;
+            verify_pos_first[seq_id] = batch_in.pos[beg];
+            pending_pos_last[seq_id] = batch_in.pos[end];
+            verify_g_rows[seq_id]    = n_rows;
+            verify_g[seq_id].resize((size_t) n_rows * n_embd_dec, 0.0f);
+            std::memcpy(verify_g[seq_id].data(),       g_embd + (size_t) beg * n_embd_dec, row_bytes * n_rows);
+            std::memcpy(pending_g_last[seq_id].data(), g_embd + (size_t) end * n_embd_dec, row_bytes);
+        }
+
+        if (batch.n_tokens > 0) {
+            const int32_t rc = llama_decode(ctx_dft, batch);
+            if (rc != 0) {
+                LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, ubatch_pos[0]=%d)\n",
+                        __func__, rc, (int) batch.n_tokens, (int) batch_in.pos[0]);
+                return false;
+            }
+        }
+
         return true;
     }
 
-    void draft(common_speculative_draft_params_vec & /*dparams*/) override {
-        // TODO: implement
+    void draft(common_speculative_draft_params_vec & dparams) override {
+        auto & ctx_dft = params.ctx_dft;
+
+        common_batch_clear(batch);
+
+        // keep track of which sequences are still drafting
+        int n_drafting = 0;
+        std::vector<bool> drafting(n_seq);
+
+        const size_t row_bytes = (size_t) n_embd_dec * sizeof(float);
+
+        // Complete the deferred boundary pair (dp.id_last, pending_g_last) at memory
+        // pos pending_pos_last. dp.id_last is target's freshest sample (= corrected
+        // token after verify, or first generated token after prefill), matching the
+        // EAGLE3 input convention (token[P+1], g_embd[P]) at pos P.
+        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+            auto & dp = dparams[seq_id];
+
+            if (!dp.drafting) {
+                continue;
+            }
+            if (pending_pos_last[seq_id] < 0) {
+                continue;
+            }
+
+            n_drafting++;
+            drafting[seq_id] = true;
+            common_sampler_reset(smpls[seq_id].get());
+
+            llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, pending_pos_last[seq_id], -1);
+
+            common_batch_add(batch, dp.id_last, pending_pos_last[seq_id], { seq_id }, true);
+            std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                        pending_g_last[seq_id].data(),
+                        row_bytes);
+        }
+
+        if (batch.n_tokens == 0) {
+            return;
+        }
+
+        int ret = llama_decode(ctx_dft, batch);
+        if (ret != 0) {
+            LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
+            return;
+        }
+
+        int i = 0;
+
+        while (n_drafting > 0) {
+            int i_batch = 0;
+
+            common_batch_clear(batch);
+
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                if (!drafting[seq_id]) {
+                    continue;
+                }
+
+                auto * smpl = smpls[seq_id].get();
+
+                common_sampler_sample(smpl, ctx_dft, i_batch, true);
+                // pre-norm hidden state of this position becomes g_embd for the next step
+                const float * prenorm = llama_get_embeddings_nextn_ith(ctx_dft, i_batch);
+                ++i_batch;
+
+                const auto * cur_p = common_sampler_get_candidates(smpl, true);
+
+                for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                    LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                            seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p,
+                            common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                }
+
+                const llama_token id = cur_p->data[0].id;
+
+                // only collect very high-confidence draft tokens
+                // (configurable via --spec-draft-p-min, set to 0.0 to disable early-stop)
+                if (cur_p->data[0].p < params.p_min) {
+                    drafting[seq_id] = false;
+                    n_drafting--;
+
+                    continue;
+                }
+
+                common_sampler_accept(smpl, id, true);
+
+                auto & dp = dparams.at(seq_id);
+                auto & result = *dp.result;
+
+                result.push_back(id);
+
+                if (params.n_max <= (int) result.size()) {
+                    drafting[seq_id] = false;
+                    n_drafting--;
+                    continue;
+                }
+
+                common_batch_add(batch, id, pending_pos_last[seq_id] + (i + 1), { seq_id }, true);
+                std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec, prenorm, row_bytes);
+            }
+
+            if (batch.n_tokens == 0) {
+                break;
+            }
+
+            ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
+                break;
+            }
+
+            ++i;
+        }
+
+        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+            auto & dp = dparams[seq_id];
+            if (!dp.drafting) {
+                continue;
+            }
+
+            if (dp.result->size() < (size_t) params.n_min) {
+                dp.result->clear();
+            }
+        }
     }
 
-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
-        // noop
+    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool /*is_other*/) override {
+        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
+            return;
+        }
+
+        const int32_t n_rows = verify_g_rows[seq_id];
+        if (n_rows <= 0) {
+            return;
+        }
+
+        const int32_t i_g = std::min<int32_t>(n_accepted, n_rows - 1);
+        pending_pos_last[seq_id] = verify_pos_first[seq_id] + i_g;
+        std::memcpy(pending_g_last[seq_id].data(),
+                    verify_g[seq_id].data() + (size_t) i_g * n_embd_dec,
+                    (size_t) n_embd_dec * sizeof(float));
     }
 
     bool need_embd() const override {
@@ -1370,9 +1776,11 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);
 
         bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
-        bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
+        bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
         bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;
 
+
+
         bool has_ngram_cache   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_CACHE));
         bool has_ngram_simple  = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE));
         bool has_ngram_map_k   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K));

conversion/__init__.py

@@ -130,6 +130,9 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "LlamaBidirectionalModel": "llama",
     "LlamaForCausalLM": "llama",
     "LlamaModel": "llama",
+    "Eagle3DraftModel": "llama",
+    "Eagle3Speculator": "llama",
+    "LlamaForCausalLMEagle3": "llama",
     "LlavaForConditionalGeneration": "llama",
     "LlavaStableLMEpochForCausalLM": "stablelm",
     "MPTForCausalLM": "mpt",

conversion/base.py

@@ -94,6 +94,7 @@ class ModelBase:
     metadata: gguf.Metadata
     dir_model_card: Path
     remote_hf_model_id: str | None
+    target_model_dir: Path | None
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
@@ -119,6 +120,7 @@ class ModelBase:
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
                  disable_mistral_community_chat_template: bool = False,
                  sentence_transformers_dense_modules: bool = False,
+                 target_model_dir: Path | None = None,
                  fuse_gate_up_exps: bool = False,
                  fp8_as_q8: bool = False):
         if type(self) is ModelBase or \
@@ -139,6 +141,7 @@ class ModelBase:
         self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
         self.sentence_transformers_dense_modules = sentence_transformers_dense_modules
+        self.target_model_dir = target_model_dir
         self.fuse_gate_up_exps = fuse_gate_up_exps
         self._gate_exp_buffer: dict[int, Tensor] = {}
         self._up_exp_buffer: dict[int, Tensor] = {}
@@ -2481,6 +2484,7 @@ class LazyTorchTensor(gguf.LazyBase):
         torch.float16: np.float16,
         torch.float32: np.float32,
         torch.uint8: np.uint8,
+        torch.int64: np.int64,
     }
 
     # only used when byteswapping data. Only correct size is needed

conversion/llama.py

@@ -5,12 +5,13 @@ import math
 
 from typing import Callable, Iterable, TYPE_CHECKING
 
+import numpy as np
 import torch
 
 if TYPE_CHECKING:
     from torch import Tensor
 
-from .base import ModelBase, TextModel, gguf
+from .base import ModelBase, TextModel, gguf, logger
 
 
 @ModelBase.register(
@@ -21,6 +22,9 @@ from .base import ModelBase, TextModel, gguf
     "VLlama3ForCausalLM",
     "LlavaForConditionalGeneration",
     "VoxtralForConditionalGeneration",
+    "LlamaForCausalLMEagle3",
+    "Eagle3Speculator",
+    "Eagle3DraftModel",
     "IQuestCoderForCausalLM",
     "LlamaModel")
 class LlamaModel(TextModel):
@@ -39,7 +43,61 @@ class LlamaModel(TextModel):
             hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
             self.origin_hf_arch = hparams.get('architectures', [None])[0]
 
+        # Detect eagle3 draft checkpoint by hparams (some models don't use a distinct HF arch name)
+        if "draft_vocab_size" in self.hparams and self.hparams["num_hidden_layers"] == 1:
+            self.is_eagle3 = True
+            self.model_arch = gguf.MODEL_ARCH.EAGLE3
+            logger.info("Detected EAGLE-3 draft model, switching to EAGLE3 architecture")
+            # Re-initialize tensor_map with eagle3 architecture
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+            # Update gguf_writer architecture
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
+            self.gguf_writer.add_architecture()
+            if self.target_model_dir is None:
+                raise ValueError(
+                    "EAGLE-3 model requires --target-model-dir to be specified. "
+                    "Please provide the path to the target model directory to read config.json"
+                )
+            # Read both eagle3 raw config and target model config
+            with open(self.dir_model / "config.json", 'r', encoding='utf-8') as f:
+                eagle3_raw_config = json.load(f)
+            with open(self.target_model_dir / "config.json", 'r', encoding='utf-8') as f:
+                target_config = json.load(f)
+
+            if "text_config" in target_config:
+                target_config = {**target_config, **target_config["text_config"]}
+            self.target_vocab_size = target_config["vocab_size"]
+
+            # target_layers: derived from target model layer count (low/mid/high)
+            target_num_layers = target_config["num_hidden_layers"]
+            target_layers = [2, target_num_layers // 2, target_num_layers - 3]
+            logger.info(f"EAGLE-3: target_layers = {target_layers} (target model has {target_num_layers} layers)")
+            self.gguf_writer.add_array(f"{self.gguf_writer.arch}.target_layers", target_layers)
+
+            # target_hidden_size: prefer eagle3 config, fallback to target config
+            if eagle3_raw_config.get("target_hidden_size") is not None:
+                target_hidden_size = eagle3_raw_config["target_hidden_size"]
+                src = "EAGLE-3 config"
+            else:
+                target_hidden_size = target_config["hidden_size"]
+                src = "target model config"
+            logger.info(f"EAGLE-3: target_hidden_size = {target_hidden_size} (from {src})")
+            self.gguf_writer.add_uint32(f"{self.gguf_writer.arch}.target_hidden_size", target_hidden_size)
+
+            # norm_before_residual (RedHat-style eagle3 specific)
+            norm_before_residual = eagle3_raw_config.get("norm_before_residual", False)
+            logger.info(f"EAGLE-3: norm_before_residual = {norm_before_residual}")
+            self.gguf_writer.add_bool(f"{self.gguf_writer.arch}.norm_before_residual", norm_before_residual)
+
     def set_vocab(self):
+        # eagle3: use tokenizer from target model if provided
+        original_dir_model = None
+        if getattr(self, 'is_eagle3', False):
+            assert self.target_model_dir is not None
+            logger.info(f"EAGLE-3: Using tokenizer from target model: {self.target_model_dir}")
+            original_dir_model = self.dir_model
+            self.dir_model = self.target_model_dir
+
         if self.origin_hf_arch == "GlmasrModel":
             return self._set_vocab_glmedge()
 
@@ -85,6 +143,10 @@ class LlamaModel(TextModel):
         if self.hparams.get("vocab_size", 32000) == 49152:
             self.gguf_writer.add_add_bos_token(False)
 
+        # eagle3: Restore original dir_model
+        if original_dir_model is not None:
+            self.dir_model = original_dir_model
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -129,7 +191,49 @@ class LlamaModel(TextModel):
 
         return super().filter_tensors((name, gen))
 
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        tensors = super().index_tensors(remote_hf_model_id)
+
+        # Handle Eagle3Speculator nested config
+        if "transformer_layer_config" in self.hparams:
+            self.hparams = {**self.hparams, **self.hparams["transformer_layer_config"]}
+
+        # eagle3 detection
+        if "draft_vocab_size" in self.hparams and self.hparams["num_hidden_layers"] == 1:
+            logger.info("EAGLE-3: renaming midlayer.* / layers.0.* to model.layers.0.*")
+            new_tensors = {}
+            for name, gen in tensors.items():
+                if name.startswith("midlayer."):
+                    new_name = "model.layers.0." + name[len("midlayer."):]
+                    new_tensors[new_name] = gen
+                elif name.startswith("layers.0."):  # Eagle3Speculator format
+                    new_name = "model." + name
+                    new_tensors[new_name] = gen
+                else:
+                    new_tensors[name] = gen
+            return new_tensors
+
+        return tensors
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # eagle3: special tensors that bypass standard llama mapping
+        if getattr(self, 'is_eagle3', False):
+            if name == "fc.weight":
+                yield (name, data_torch)
+                return
+            if name == "d2t":
+                # store for manual int64 handling in prepare_tensors (avoid F32 conversion)
+                if not hasattr(self, '_eagle3_int_tensors'):
+                    self._eagle3_int_tensors = {}
+                self._eagle3_int_tensors[name] = data_torch
+                return
+            if name == "t2d":
+                # not used at runtime, skip
+                return
+            if name.endswith(".hidden_norm.weight"):
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_NORM_2, bid), data_torch)
+                return
+
         n_head = self.find_hparam(["n_heads", "num_attention_heads"])
         n_kv_head = self.find_hparam(["n_kv_heads", "num_key_value_heads"])
 
@@ -205,8 +309,33 @@ class LlamaModel(TextModel):
                 yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
     def prepare_tensors(self):
+        # eagle3: collect d2t original dtype before parent converts tensors to F32
+        eagle3_original_dtypes = {}
+        if getattr(self, 'is_eagle3', False):
+            for name, data_torch in self.get_tensors():
+                if name == "d2t":
+                    eagle3_original_dtypes[name] = data_torch.dtype
+
         super().prepare_tensors()
 
+        # eagle3: write d2t as absolute target token ids
+        if getattr(self, 'is_eagle3', False) and hasattr(self, '_eagle3_int_tensors'):
+            for name, data_torch in self._eagle3_int_tensors.items():
+                old_dtype = eagle3_original_dtypes.get(name, data_torch.dtype)
+                data = data_torch.to(torch.int64).cpu().numpy()
+                if name == "d2t":
+                    data = data.reshape(-1)
+                    data = data + np.arange(data.size, dtype=np.int64)
+                    if np.any((data < 0) | (data >= self.target_vocab_size)):
+                        raise ValueError(f"EAGLE-3 d2t target ids out of range for target vocab size {self.target_vocab_size}")
+                    if np.unique(data).size != data.size:
+                        raise ValueError("EAGLE-3 d2t contains duplicate target ids")
+                data_qtype = gguf.GGMLQuantizationType.I64
+
+                shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
+                logger.info(f"{name + ',':<30} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
+                self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
+
         if self._experts is not None:
             # flatten `list[dict[str, Tensor]]` into `list[str]`
             experts = [k for d in self._experts for k in d.keys()]

convert_hf_to_gguf.py

@@ -153,6 +153,15 @@ def parse_args() -> argparse.Namespace:
         help="Store tensors dequantized from FP8 as Q8_0 instead of BF16/F16.",
     )
 
+    parser.add_argument(
+        "--target-model-dir", type=str, default=None,
+        help=(
+            "path to the target model directory; required when converting a standalone draft model "
+            "(e.g. EAGLE3 / DFlash) that needs target-model metadata such as tokenizer, hidden size, and "
+            "layer count to populate its GGUF."
+        ),
+    )
+
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
         parser.error("the following arguments are required: model")
@@ -269,6 +278,7 @@ def main() -> None:
                                      small_first_shard=args.no_tensor_first_split,
                                      remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
                                      sentence_transformers_dense_modules=args.sentence_transformers_dense_modules,
+                                     target_model_dir=Path(args.target_model_dir) if args.target_model_dir else None,
                                      fuse_gate_up_exps=args.fuse_gate_up_exps,
                                      fp8_as_q8=args.fp8_as_q8,
                                      )

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 14)
-set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_MINOR 15)
+set(GGML_VERSION_PATCH 1)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/src/ggml-cuda/concat.cu

@@ -1,16 +1,18 @@
 #include "concat.cuh"
 
+#include <stdint.h>
+
 // contiguous kernels
-template <int dim>
-static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_f32_cont(const float * x,
-                                                                                 const float * y,
-                                                                                 float *       dst,
-                                                                                 int64_t       ne00,
-                                                                                 int64_t       ne01,
-                                                                                 int64_t       ne02,
-                                                                                 int64_t       ne0,
-                                                                                 int64_t       ne1,
-                                                                                 int64_t       ne2) {
+template <typename T, int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_cont(const T * x,
+                                                                             const T * y,
+                                                                             T *       dst,
+                                                                             int64_t   ne00,
+                                                                             int64_t   ne01,
+                                                                             int64_t   ne02,
+                                                                             int64_t   ne0,
+                                                                             int64_t   ne1,
+                                                                             int64_t   ne2) {
     static_assert(dim >= 0 && dim <= 2, "dim must be in [0, 2]");
 
     const int64_t n = ne0 * ne1 * ne2;
@@ -50,37 +52,37 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_f32_cont
     }
 }
 
-static void concat_f32_cuda(const float * x,
-                            const float * y,
-                            float *       dst,
-                            int64_t       ne00,
-                            int64_t       ne01,
-                            int64_t       ne02,
-                            int64_t       ne0,
-                            int64_t       ne1,
-                            int64_t       ne2,
-                            int           dim,
-                            cudaStream_t  stream) {
+template <typename T>
+static void concat_cont_cuda(const T * x,
+                             const T * y,
+                             T *       dst,
+                             int64_t   ne00,
+                             int64_t   ne01,
+                             int64_t   ne02,
+                             int64_t   ne0,
+                             int64_t   ne1,
+                             int64_t   ne2,
+                             int       dim,
+                             cudaStream_t stream) {
     const int64_t n          = ne0 * ne1 * ne2;
     const int     num_blocks = (n + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
 
     if (dim == 0) {
         const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream);
-        ggml_cuda_kernel_launch(concat_f32_cont<0>, launch_params,x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
+        ggml_cuda_kernel_launch(concat_cont<T, 0>, launch_params, x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
     if (dim == 1) {
-        concat_f32_cont<1>
-            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
+        concat_cont<T, 1><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
-    concat_f32_cont<2><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
+    concat_cont<T, 2><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
 }
 
 // non-contiguous kernel (slow)
-template <int dim>
+template <typename T, int dim>
 static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
-    concat_f32_non_cont(
+    concat_non_cont(
         const char * src0,
         const char * src1,
               char * dst,
@@ -107,61 +109,49 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
           uint64_t   nb0,
           uint64_t   nb1,
           uint64_t   nb2,
-          uint64_t   nb3){
+          uint64_t   nb3) {
     static_assert(dim >= 0 && dim <= 3, "dim must be in [0, 3]");
 
     const int64_t i3 = blockIdx.z;
     const int64_t i2 = blockIdx.y;
     const int64_t i1 = blockIdx.x;
 
-    const float * x;
+    const T * x;
 
     for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+            x = (const T *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
         } else {
             if constexpr (dim == 0) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + i1 * nb11 + (i0 - ne00) * nb10);
+                x = (const T *)(src1 + i3*nb13 + i2*nb12 + i1*nb11 + (i0 - ne00)*nb10);
             } else if constexpr (dim == 1) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + (i1 - ne01) * nb11 + i0 * nb10);
+                x = (const T *)(src1 + i3*nb13 + i2*nb12 + (i1 - ne01)*nb11 + i0*nb10);
             } else if constexpr (dim == 2) {
-                x = (const float *) (src1 + i3 * nb13 + (i2 - ne02) * nb12 + i1 * nb11 + i0 * nb10);
+                x = (const T *)(src1 + i3*nb13 + (i2 - ne02)*nb12 + i1*nb11 + i0*nb10);
             } else if constexpr (dim == 3) {
-                x = (const float *) (src1 + (i3 - ne03) * nb13 + i2 * nb12 + i1 * nb11 + i0 * nb10);
+                x = (const T *)(src1 + (i3 - ne03)*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
             }
         }
 
-        float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+        T * y = (T *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
         *y = *x;
     }
 }
 
-
-void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    cudaStream_t stream = ctx.stream();
-
-    const int32_t dim = ((int32_t *) dst->op_params)[0];
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
-
+template <typename T>
+static void concat_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, int dim, cudaStream_t stream) {
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        const float * src0_d = (const float *)src0->data;
-        const float * src1_d = (const float *)src1->data;
-
-        float * dst_d = (float *)dst->data;
+        const T * src0_d = (const T *) src0->data;
+        const T * src1_d = (const T *) src1->data;
+        T *       dst_d  = (T *) dst->data;
 
         if (dim != 3) {
-            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-                concat_f32_cuda(
-                        src0_d + i3 * (src0->nb[3] / 4),
-                        src1_d + i3 * (src1->nb[3] / 4),
-                        dst_d + i3 * ( dst->nb[3] / 4),
+            for (int64_t i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_cont_cuda(
+                        src0_d + i3*(src0->nb[3] / sizeof(T)),
+                        src1_d + i3*(src1->nb[3] / sizeof(T)),
+                        dst_d  + i3*( dst->nb[3] / sizeof(T)),
                         src0->ne[0], src0->ne[1], src0->ne[2],
                         dst->ne[0],  dst->ne[1],  dst->ne[2], dim, stream);
             }
@@ -169,13 +159,13 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             const size_t size0 = ggml_nbytes(src0);
             const size_t size1 = ggml_nbytes(src1);
 
-            CUDA_CHECK(cudaMemcpyAsync(dst_d,           src0_d, size0, cudaMemcpyDeviceToDevice, stream));
-            CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync((char *) dst->data,         src0->data, size0, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync((char *) dst->data + size0, src1->data, size1, cudaMemcpyDeviceToDevice, stream));
         }
     } else {
         dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
         auto launch_kernel = [&](auto dim) {
-            concat_f32_non_cont<dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+            concat_non_cont<T, dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
                 (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
                 src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
                 src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
@@ -203,3 +193,35 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         }
     }
 }
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    cudaStream_t stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(dst->type  == src0->type);
+    GGML_ASSERT(!ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_blck_size(src0->type) == 1);
+
+    switch (ggml_type_size(src0->type)) {
+        case 1:
+            concat_cuda<uint8_t>(src0, src1, dst, dim, stream);
+            break;
+        case 2:
+            concat_cuda<uint16_t>(src0, src1, dst, dim, stream);
+            break;
+        case 4:
+            concat_cuda<uint32_t>(src0, src1, dst, dim, stream);
+            break;
+        case 8:
+            concat_cuda<uint64_t>(src0, src1, dst, dim, stream);
+            break;
+        default:
+            GGML_ABORT("Unsupported type size: %zu", ggml_type_size(src0->type));
+            break;
+    }
+}

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

@@ -5345,7 +5345,15 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONCAT:
             {
                 ggml_type src0_type = op->src[0]->type;
-                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+                ggml_type src1_type = op->src[1]->type;
+                return src0_type == src1_type &&
+                       src0_type == op->type &&
+                       !ggml_is_quantized(src0_type) &&
+                       ggml_blck_size(src0_type) == 1 &&
+                       (ggml_type_size(src0_type) == 1 ||
+                        ggml_type_size(src0_type) == 2 ||
+                        ggml_type_size(src0_type) == 4 ||
+                        ggml_type_size(src0_type) == 8);
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {

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

@@ -1120,8 +1120,17 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_VIEW:
         case GGML_OP_TRANSPOSE:
         case GGML_OP_PERMUTE:
-        case GGML_OP_CONCAT:
             return true;
+        case GGML_OP_CONCAT:
+            {
+                // kernel_concat copies one float-sized value per element.
+                // Other scalar types need a type-generic copy kernel first.
+                const enum ggml_type src0_type = op->src[0]->type;
+                const enum ggml_type src1_type = op->src[1]->type;
+                return src0_type == src1_type &&
+                       src0_type == op->type &&
+                       (src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_I32);
+            }
         case GGML_OP_ADD:
         case GGML_OP_SUB:
         case GGML_OP_MUL:

ggml/src/ggml-opencl/CMakeLists.txt

@@ -142,6 +142,10 @@ set(GGML_OPENCL_KERNELS
     gemm_noshuffle_q4_0_f32
     gemv_noshuffle_q4_1_f32
     gemm_noshuffle_q4_1_f32
+    gemv_noshuffle_q5_0_f32
+    gemm_noshuffle_q5_0_f32
+    gemv_noshuffle_q5_1_f32
+    gemm_noshuffle_q5_1_f32
     gemv_noshuffle_iq4_nl_f32
     gemm_noshuffle_iq4_nl_f32
     gemv_noshuffle_q8_0_f32

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

@@ -593,6 +593,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_restore_block_q4_0_noshuffle;
     cl_kernel kernel_convert_block_q4_1_noshuffle;
     cl_kernel kernel_restore_block_q4_1_noshuffle;
+    cl_kernel kernel_convert_block_q5_0_noshuffle;
+    cl_kernel kernel_restore_block_q5_0_noshuffle;
+    cl_kernel kernel_convert_block_q5_1_noshuffle;
+    cl_kernel kernel_restore_block_q5_1_noshuffle;
     cl_kernel kernel_convert_block_q4_K_noshuffle;
     cl_kernel kernel_restore_block_q4_K_noshuffle;
     cl_kernel kernel_convert_block_q4_K, kernel_restore_block_q4_K;
@@ -829,6 +833,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_noshuffle_q6_K_f32;
     cl_kernel kernel_gemv_noshuffle_q5_k_f32;
     cl_kernel kernel_gemm_noshuffle_q5_k_f32;
+    cl_kernel kernel_gemv_noshuffle_q5_0_f32;
+    cl_kernel kernel_gemm_noshuffle_q5_0_f32;
+    cl_kernel kernel_gemv_noshuffle_q5_1_f32;
+    cl_kernel kernel_gemm_noshuffle_q5_1_f32;
     cl_kernel kernel_gemv_noshuffle_iq4_nl_f32;
     cl_kernel kernel_gemm_noshuffle_iq4_nl_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
@@ -1152,6 +1160,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         CL_CHECK((backend_ctx->kernel_restore_block_q4_1_trans4_ns = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_1_trans4_ns", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q5_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q5_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q5_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_0_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q5_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_0_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q5_1_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_1_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q5_1_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_1_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q5_0_trans4_ns = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_0_trans4_ns", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q5_0_trans4_ns = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_0_trans4_ns", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q5_1  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_1", &err), err));
@@ -3065,6 +3077,80 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         GGML_LOG_CONT(".");
     }
 
+    // gemm_noshuffle_q5_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q5_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q5_0_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q5_0_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q5_0_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_q5_0_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAST ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q5_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q5_0_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q5_0_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q5_0_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemm_noshuffle_q5_1_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q5_1_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q5_1_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q5_1_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q5_1_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_q5_1_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAST ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q5_1_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q5_1_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q5_1_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q5_1_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // gemm_noshuffle_iq4_nl_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -6107,15 +6193,16 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             return;
         }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
-            cl_kernel kernel = backend_ctx->kernel_convert_block_q5_0;
-            cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_convert_block_q5_0_noshuffle;
             CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
             CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qs));
             CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
             CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
-            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &n_blk));
 
-            size_t global_work_size[] = {(size_t)CEIL_DIV(n_blk, 64) * 64, 1, 1};
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
             size_t local_work_size[] = {64, 1, 1};
 
             cl_event evt;
@@ -6124,7 +6211,39 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
 
             tensor->extra = extra;
+
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose qs as ushort
+            transpose_2d_as_16b(backend_ctx, extra->qs, extra->qs, size_qs, K/4, M);
+            // Transpose qh as uchar
+            transpose_2d_as_8b(backend_ctx, extra->qh, extra->qh, size_qh, K/8, M);
+            // Transpose d as ushort
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/32, M);
+
             return;
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q5_0;
+        cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qs));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &n_blk));
+
+        size_t global_work_size[] = {(size_t)CEIL_DIV(n_blk, 64) * 64, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+        return;
     }
     if (tensor->type == GGML_TYPE_Q5_1) {
         ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
@@ -6225,6 +6344,42 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             return;
         }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_convert_block_q5_1_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qs));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->m));
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {64, 1, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+            CL_CHECK(clReleaseMemObject(data_device));
+
+            tensor->extra = extra;
+
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose qs as ushort
+            transpose_2d_as_16b(backend_ctx, extra->qs, extra->qs, size_qs, K/4, M);
+            // Transpose qh as uchar
+            transpose_2d_as_8b(backend_ctx, extra->qh, extra->qh, size_qh, K/8, M);
+            // Transpose d as ushort
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/32, M);
+            // Transpose m as ushort
+            transpose_2d_as_16b(backend_ctx, extra->m, extra->m, size_m, K/32, M);
+
+            return;
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
         cl_kernel kernel = backend_ctx->kernel_convert_block_q5_1;
         cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
@@ -7299,6 +7454,48 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            ggml_cl_buffer buf_trans_qs;
+            ggml_cl_buffer buf_trans_qh;
+            ggml_cl_buffer buf_trans_d;
+            ggml_cl_buffer buf_unpacked;
+
+            cl_int M = tensor->ne[1];
+            cl_int K = tensor->ne[0];
+
+            GGML_ASSERT(K % 32 == 0);
+
+            size_t size_qs = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*ggml_blck_size(tensor->type)/2;
+            size_t size_qh = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(int32_t);
+            size_t size_d = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+
+            buf_trans_qs.allocate(backend_ctx->context, size_qs);
+            buf_trans_qh.allocate(backend_ctx->context, size_qh);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));
+
+            transpose_2d_as_16b(backend_ctx, extra->qs, buf_trans_qs.buffer, size_qs, M, K/4);
+            transpose_2d_as_8b(backend_ctx, extra->qh, buf_trans_qh.buffer, size_qh, M, K/8);
+            transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/32);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q5_0_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_qs.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_qh.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &buf_unpacked.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_F0));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
+            return;
+        }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
         cl_int err;
@@ -7362,6 +7559,54 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
+
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            ggml_cl_buffer buf_trans_qs;
+            ggml_cl_buffer buf_trans_qh;
+            ggml_cl_buffer buf_trans_d;
+            ggml_cl_buffer buf_trans_m;
+            ggml_cl_buffer buf_unpacked;
+
+            cl_int M = tensor->ne[1];
+            cl_int K = tensor->ne[0];
+            GGML_ASSERT(K % 32 == 0);
+
+            size_t size_qs = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*ggml_blck_size(tensor->type)/2;
+            size_t size_qh = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(int32_t);
+            size_t size_d  = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            size_t size_m  = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+
+            buf_trans_qs.allocate(backend_ctx->context, size_qs);
+            buf_trans_qh.allocate(backend_ctx->context, size_qh);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_trans_m.allocate(backend_ctx->context, size_m);
+            buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));
+
+            // Transpose back: from col-major to row-major
+            transpose_2d_as_16b(backend_ctx, extra->qs, buf_trans_qs.buffer, size_qs, M, K/4);
+            transpose_2d_as_8b(backend_ctx, extra->qh, buf_trans_qh.buffer, size_qh, M, K/8);
+            transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/32);
+            transpose_2d_as_16b(backend_ctx, extra->m,  buf_trans_m.buffer,  size_m,  M, K/32);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q5_1_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_qs.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_qh.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &buf_trans_m.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &buf_unpacked.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_F0));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
+            return;
+        }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
         cl_int err;
         cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
@@ -12205,6 +12450,368 @@ static void ggml_cl_mul_mat_q4_1_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }
 
+static void ggml_cl_mul_mat_q5_0_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl_q5_0 * extra0_q5_0 = (ggml_tensor_extra_cl_q5_0 *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+
+    const int ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
+    cl_buffer_region    region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (ne1 == 1) {
+        cl_mem qs_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for qs
+        img_fmt = { CL_R, CL_UNSIGNED_INT32 };
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer = extra0_q5_0->qs;
+        CL_CHECK((qs_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q5_0_f32;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &qs_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q5_0->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q5_0->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne01));
+
+        size_t local_work_size[3] = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(qs_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
+        cl_mem d_sub_buf = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for output
+        region.origin = extrad->offset;
+        region.size = M * N * sizeof(float);
+        CL_CHECK((d_sub_buf = clCreateSubBuffer(extrad->data_device, CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // transpose activations
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q5_0_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_q5_0->qs));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q5_0->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q5_0->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &d_sub_buf));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne1));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+        CL_CHECK(clReleaseMemObject(d_sub_buf));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
+static void ggml_cl_mul_mat_q5_1_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl_q5_1 * extra0_q5_1 = (ggml_tensor_extra_cl_q5_1 *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+
+    const int ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
+    cl_buffer_region    region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (ne1 == 1) {
+        cl_mem qs_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for qs
+        img_fmt = { CL_R, CL_UNSIGNED_INT32 };
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer = extra0_q5_1->qs;
+        CL_CHECK((qs_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q5_1_f32;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &qs_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q5_1->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q5_1->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0_q5_1->m));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne01));
+
+        size_t local_work_size[3] = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(qs_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
+        cl_mem d_sub_buf = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for output
+        region.origin = extrad->offset;
+        region.size = M * N * sizeof(float);
+        CL_CHECK((d_sub_buf = clCreateSubBuffer(extrad->data_device, CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // transpose activations
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q5_1_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_q5_1->qs));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q5_1->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q5_1->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0_q5_1->m));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &d_sub_buf));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_int),   &ne1));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+        CL_CHECK(clReleaseMemObject(d_sub_buf));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat_iq4_nl_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     GGML_ASSERT(src0);
@@ -13243,6 +13850,18 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             return;
         }
 
+        // q5_0 x fp32
+        if (src0t == GGML_TYPE_Q5_0 && src1t == GGML_TYPE_F32) {
+            ggml_cl_mul_mat_q5_0_f32_adreno(backend, src0, src1, dst);
+            return;
+        }
+
+        // q5_1 x fp32
+        if (src0t == GGML_TYPE_Q5_1 && src1t == GGML_TYPE_F32) {
+            ggml_cl_mul_mat_q5_1_f32_adreno(backend, src0, src1, dst);
+            return;
+        }
+
         // iq4_nl x fp32
         if (src0t == GGML_TYPE_IQ4_NL && src1t == GGML_TYPE_F32) {
             ggml_cl_mul_mat_iq4_nl_f32_adreno(backend, src0, src1, dst);

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

@@ -584,6 +584,60 @@ kernel void kernel_restore_block_q5_0(
     }
 }
 
+kernel void kernel_convert_block_q5_0_noshuffle(
+    global struct block_q5_0 * src0,
+    global uchar * dst_q,
+    global uint  * dst_qh,
+    global half  * dst_d
+) {
+    global struct block_q5_0 * b = (global struct block_q5_0 *) src0 + get_global_id(0);
+    global uchar * q  = (global uchar *) dst_q + QK5_0/2*get_global_id(0);
+    global uint  * qh = (global uint  *) dst_qh + get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    *d = b->d;
+    *qh = *((global uint *)(b->qh));
+
+    for (int i = 0; i < QK5_0/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0      ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
+        q[i + QK5_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
+
+#ifdef ADRENO_GPU
+        if (get_global_id(0) == 65536*4096) {
+            printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
+        }
+#endif
+    }
+}
+
+kernel void kernel_restore_block_q5_0_noshuffle(
+    global uchar * src_q,
+    global uint  * src_qh,
+    global half  * src_d,
+    global struct block_q5_0 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q5_0 * b = (global struct block_q5_0 *) dst + get_global_id(0);
+    global uchar * q  = (global uchar *) src_q + QK5_0/2*get_global_id(0);
+    global uint  * qh = (global uint  *) src_qh + get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+
+    b->d = *d;
+    *((global uint *)(b->qh)) = *qh;
+
+    for (int i = 0; i < QK5_0/4; ++i) {
+        uchar x0 = q[i + 0      ];
+        uchar x1 = q[i + QK5_0/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
 kernel void kernel_convert_block_q5_0_trans4_ns(
     __global struct block_q5_0 * src0,
     __global uint * dst_qs,
@@ -736,6 +790,66 @@ kernel void kernel_restore_block_q5_1(
     }
 }
 
+kernel void kernel_convert_block_q5_1_noshuffle(
+    global struct block_q5_1 * src0,
+    global uchar * dst_q,
+    global uint  * dst_qh,
+    global half  * dst_d,
+    global half  * dst_m
+) {
+    global struct block_q5_1 * b = (global struct block_q5_1 *) src0 + get_global_id(0);
+    global uchar * q  = (global uchar *) dst_q + QK5_1/2*get_global_id(0);
+    global uint  * qh = (global uint  *) dst_qh + get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+    global half  * m  = (global half  *) dst_m  + get_global_id(0);
+
+    *d = b->d;
+    *m = b->m;
+    *qh = *((global uint *)(b->qh));
+
+    for (int i = 0; i < QK5_1/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0      ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
+        q[i + QK5_1/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
+
+#ifdef ADRENO_GPU
+        if (get_global_id(0) == 65536*4096) {
+            printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
+        }
+#endif
+    }
+}
+
+kernel void kernel_restore_block_q5_1_noshuffle(
+    global uchar * src_q,
+    global uint  * src_qh,
+    global half  * src_d,
+    global half  * src_m,
+    global struct block_q5_1 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q5_1 * b = (global struct block_q5_1 *) dst + get_global_id(0);
+    global uchar * q  = (global uchar *) src_q + QK5_1/2*get_global_id(0);
+    global uint  * qh = (global uint  *) src_qh + get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+    global half  * m  = (global half  *) src_m  + get_global_id(0);
+
+    b->d = *d;
+    b->m = *m;
+    *((global uint *)(b->qh)) = *qh;
+
+    for (int i = 0; i < QK5_1/4; ++i) {
+        uchar x0 = q[i + 0      ];
+        uchar x1 = q[i + QK5_1/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
 kernel void kernel_convert_block_q5_1_trans4_ns(
     __global struct block_q5_1 * src0,
     __global uint * dst_qs,

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl

@@ -0,0 +1,131 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_gemm_noshuffle_q5_0_f32(
+        global const ushort * src0_qs,      // quantized A
+        global const uchar  * src0_qh,      // 5th bits
+        global const half   * src0_d,       // A scales
+        __read_only image1d_buffer_t src1,  // B (1d image)
+        global float * dst,                 // C
+        int m,                              // M
+        int n,                              // N with padding
+        int k,                              // K
+        int n_no_padding                    // N without padding
+) {
+
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    global const ushort * weight_ptr = src0_qs + gx_2;
+    global const uchar  * qh_ptr    = src0_qh + gx_2;
+    global const half   * scale_ptr = src0_d  + gx_2;
+
+    for (int i = 0; i < k; i += 4) {
+
+        B.s0123 = read_imageh(src1, gy*2 + i*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + i*n_4 + 1);
+
+        ushort4 bits4 = vload4(0, weight_ptr + (i >> 2)*m);
+        uchar4  bits1 = vload4(0, qh_ptr + (i >> 3)*m);
+        uchar4  qh = bits1 >> (uchar4)(i & 4);
+
+        half4 scale = vload4(0, scale_ptr + (i >> 5)*m);
+
+        // j=0
+        dequantized_weights.s0 = (convert_half((bits4.s0 & 0x000F) | ((qh.s0 & 0x01) << 4)) - 16.0h) * scale.s0;
+        dequantized_weights.s1 = (convert_half((bits4.s1 & 0x000F) | ((qh.s1 & 0x01) << 4)) - 16.0h) * scale.s1;
+        dequantized_weights.s2 = (convert_half((bits4.s2 & 0x000F) | ((qh.s2 & 0x01) << 4)) - 16.0h) * scale.s2;
+        dequantized_weights.s3 = (convert_half((bits4.s3 & 0x000F) | ((qh.s3 & 0x01) << 4)) - 16.0h) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half(((bits4.s0 & 0x00F0) >> 4) | ((qh.s0 & 0x02) << 3)) - 16.0h) * scale.s0;
+        dequantized_weights.s1 = (convert_half(((bits4.s1 & 0x00F0) >> 4) | ((qh.s1 & 0x02) << 3)) - 16.0h) * scale.s1;
+        dequantized_weights.s2 = (convert_half(((bits4.s2 & 0x00F0) >> 4) | ((qh.s2 & 0x02) << 3)) - 16.0h) * scale.s2;
+        dequantized_weights.s3 = (convert_half(((bits4.s3 & 0x00F0) >> 4) | ((qh.s3 & 0x02) << 3)) - 16.0h) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half(((bits4.s0 & 0x0F00) >> 8) | ((qh.s0 & 0x04) << 2)) - 16.0h) * scale.s0;
+        dequantized_weights.s1 = (convert_half(((bits4.s1 & 0x0F00) >> 8) | ((qh.s1 & 0x04) << 2)) - 16.0h) * scale.s1;
+        dequantized_weights.s2 = (convert_half(((bits4.s2 & 0x0F00) >> 8) | ((qh.s2 & 0x04) << 2)) - 16.0h) * scale.s2;
+        dequantized_weights.s3 = (convert_half(((bits4.s3 & 0x0F00) >> 8) | ((qh.s3 & 0x04) << 2)) - 16.0h) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*n_4 + 1);
+        dequantized_weights.s0 = (convert_half(((bits4.s0 & 0xF000) >> 12) | ((qh.s0 & 0x08) << 1)) - 16.0h) * scale.s0;
+        dequantized_weights.s1 = (convert_half(((bits4.s1 & 0xF000) >> 12) | ((qh.s1 & 0x08) << 1)) - 16.0h) * scale.s1;
+        dequantized_weights.s2 = (convert_half(((bits4.s2 & 0xF000) >> 12) | ((qh.s2 & 0x08) << 1)) - 16.0h) * scale.s2;
+        dequantized_weights.s3 = (convert_half(((bits4.s3 & 0xF000) >> 12) | ((qh.s3 & 0x08) << 1)) - 16.0h) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl

@@ -0,0 +1,134 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_gemm_noshuffle_q5_1_f32(
+        global const ushort * src0_qs,      // quantized A
+        global const uchar  * src0_qh,      // 5th bits
+        global const half   * src0_d,       // A scales
+        global const half   * src0_m,       // A mins
+        __read_only image1d_buffer_t src1,  // B (1d image)
+        global float * dst,                 // C
+        int m,                              // M
+        int n,                              // N with padding
+        int k,                              // K
+        int n_no_padding                    // N without padding
+) {
+
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    global const ushort * weight_ptr = src0_qs + gx_2;
+    global const uchar  * qh_ptr    = src0_qh + gx_2;
+    global const half   * scale_ptr = src0_d  + gx_2;
+    global const half   * min_ptr   = src0_m  + gx_2;
+
+    for (int i = 0; i < k; i += 4) {
+
+        B.s0123 = read_imageh(src1, gy*2 + i*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + i*n_4 + 1);
+
+        ushort4 bits4 = vload4(0, weight_ptr + (i >> 2)*m);
+        uchar4  bits1 = vload4(0, qh_ptr + (i >> 3)*m);
+        uchar4  qh = bits1 >> (uchar4)(i & 4);
+
+        half4 scale = vload4(0, scale_ptr + (i >> 5)*m);
+        half4 minv  = vload4(0, min_ptr   + (i >> 5)*m);
+
+        // j=0
+        dequantized_weights.s0 = convert_half((bits4.s0 & 0x000F) | ((qh.s0 & 0x01) << 4)) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = convert_half((bits4.s1 & 0x000F) | ((qh.s1 & 0x01) << 4)) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = convert_half((bits4.s2 & 0x000F) | ((qh.s2 & 0x01) << 4)) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = convert_half((bits4.s3 & 0x000F) | ((qh.s3 & 0x01) << 4)) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*n_4 + 1);
+        dequantized_weights.s0 = convert_half(((bits4.s0 & 0x00F0) >> 4) | ((qh.s0 & 0x02) << 3)) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = convert_half(((bits4.s1 & 0x00F0) >> 4) | ((qh.s1 & 0x02) << 3)) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = convert_half(((bits4.s2 & 0x00F0) >> 4) | ((qh.s2 & 0x02) << 3)) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = convert_half(((bits4.s3 & 0x00F0) >> 4) | ((qh.s3 & 0x02) << 3)) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*n_4 + 1);
+        dequantized_weights.s0 = convert_half(((bits4.s0 & 0x0F00) >> 8) | ((qh.s0 & 0x04) << 2)) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = convert_half(((bits4.s1 & 0x0F00) >> 8) | ((qh.s1 & 0x04) << 2)) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = convert_half(((bits4.s2 & 0x0F00) >> 8) | ((qh.s2 & 0x04) << 2)) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = convert_half(((bits4.s3 & 0x0F00) >> 8) | ((qh.s3 & 0x04) << 2)) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*n_4);
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*n_4 + 1);
+        dequantized_weights.s0 = convert_half(((bits4.s0 & 0xF000) >> 12) | ((qh.s0 & 0x08) << 1)) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = convert_half(((bits4.s1 & 0xF000) >> 12) | ((qh.s1 & 0x08) << 1)) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = convert_half(((bits4.s2 & 0xF000) >> 12) | ((qh.s2 & 0x08) << 1)) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = convert_half(((bits4.s3 & 0xF000) >> 12) | ((qh.s3 & 0x08) << 1)) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl

@@ -0,0 +1,291 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK5_0 32
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_hi(total_sums, bits4, bits1, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0      ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4      ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1      ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5      ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_lo(total_sums, bits4, bits1, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2      ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6      ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3      ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7      ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_hi(total_sums, bits4, bits1, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | (((bits1.s0     ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s0 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s0 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s0 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s0 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s0 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | (((bits1.s4     ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s4 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s4 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s4 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s4 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s4 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | (((bits1.s1     ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s1 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s1 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s1 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s1 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s1 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | (((bits1.s5     ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s5 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s5 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s5 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s5 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s5 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_lo(total_sums, bits4, bits1, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | (((bits1.s2     ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s2 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s2 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s2 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s2 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s2 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | (((bits1.s6     ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s6 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s6 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s6 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s6 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s6 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | (((bits1.s3     ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s3 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s3 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s3 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s3 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s3 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | (((bits1.s7     ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s7 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s7 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s7 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s7 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s7 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle_q5_0_f32(
+        __read_only  image1d_buffer_t src0_qs,  // quantized A
+        global ushort * src0_qh,                 // 5th bits
+        global half2  * src0_d,                  // A scales
+        __read_only  image1d_buffer_t src1,      // B activations
+        global float * dst,
+        ulong offsetd,
+        int ne00,               // K
+        int ne01)               // M
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A    = M / 2;
+    uint BLOCK_STRIDE_A  = NSUBGROUPS * M;
+
+    private uint4     regA;
+    private half2     regS;
+    private float8    regB;
+
+    private float2 totalSum = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / QK5_0); k += NSUBGROUPS) {
+        regS = src0_d[gid + k * LINE_STRIDE_A];
+
+        ushort4 qh_raw;
+        qh_raw.s0 = src0_qh[gid + (4*k + 0) * LINE_STRIDE_A];
+        qh_raw.s1 = src0_qh[gid + (4*k + 1) * LINE_STRIDE_A];
+        qh_raw.s2 = src0_qh[gid + (4*k + 2) * LINE_STRIDE_A];
+        qh_raw.s3 = src0_qh[gid + (4*k + 3) * LINE_STRIDE_A];
+
+        uchar8 raw = as_uchar8(qh_raw);
+        uchar8 qh_bytes = (uchar8)(raw.s0, raw.s2, raw.s4, raw.s6,
+                                    raw.s1, raw.s3, raw.s5, raw.s7);
+
+        // Load activations
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
+#else
+        dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
+#else
+        dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl

@@ -0,0 +1,294 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK5_1 32
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_hi(total_sums, bits4, bits1, scale, minv, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0      ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4      ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1      ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5      ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_lo(total_sums, bits4, bits1, scale, minv, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2      ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6      ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3      ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7      ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_hi(total_sums, bits4, bits1, scale, minv, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | (((bits1.s0     ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s0 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s0 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s0 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s0 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s0 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | (((bits1.s4     ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s4 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s4 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s4 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s4 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s4 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | (((bits1.s1     ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s1 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s1 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s1 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s1 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s1 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | (((bits1.s5     ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s5 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s5 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s5 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s5 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s5 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_lo(total_sums, bits4, bits1, scale, minv, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | (((bits1.s2     ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s2 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s2 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s2 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s2 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s2 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | (((bits1.s6     ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s6 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s6 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s6 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s6 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s6 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | (((bits1.s3     ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s3 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s3 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s3 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s3 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s3 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | (((bits1.s7     ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s7 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s7 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s7 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s7 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s7 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle_q5_1_f32(
+        __read_only  image1d_buffer_t src0_qs,  // quantized A
+        global ushort * src0_qh,                 // 5th bits
+        global half2  * src0_d,                  // A scales
+        global half2  * src0_m,                  // A mins
+        __read_only  image1d_buffer_t src1,      // B activations
+        global float * dst,
+        ulong offsetd,
+        int ne00,               // K
+        int ne01)               // M
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A    = M / 2;
+   uint BLOCK_STRIDE_A  = NSUBGROUPS * M;
+
+    __private uint4     regA;
+    __private half2     regS;
+    __private half2     regM;
+    __private float8    regB;
+
+    __private float2 totalSum = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / QK5_1); k += NSUBGROUPS) {
+        regS = src0_d[gid + k * LINE_STRIDE_A];
+        regM = src0_m[gid + k * LINE_STRIDE_A];
+
+        ushort4 qh_raw;
+        qh_raw.s0 = src0_qh[gid + (4*k + 0) * LINE_STRIDE_A];
+        qh_raw.s1 = src0_qh[gid + (4*k + 1) * LINE_STRIDE_A];
+        qh_raw.s2 = src0_qh[gid + (4*k + 2) * LINE_STRIDE_A];
+        qh_raw.s3 = src0_qh[gid + (4*k + 3) * LINE_STRIDE_A];
+
+        uchar8 raw = as_uchar8(qh_raw);
+        uchar8 qh_bytes = (uchar8)(raw.s0, raw.s2, raw.s4, raw.s6,
+                                    raw.s1, raw.s3, raw.s5, raw.s7);
+
+        // Load activations
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -6202,6 +6202,19 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 break;
             }
 
+#if VK_HEADER_VERSION >= 287
+            // Honeykrisp driver for Asahi Linux doesn't report VK_VENDOR_ID_APPLE.
+            // Check for Honeykrisp driver and force same configuration as the VK_VENDOR_ID_APPLE case.
+            if (device->driver_id == vk::DriverId::eMesaHoneykrisp) {
+                device->mul_mat_l[i] = false;
+                device->mul_mat_m[i] = true;
+                device->mul_mat_s[i] = false;
+                device->mul_mat_id_l[i] = false;
+                device->mul_mat_id_m[i] = true;
+                device->mul_mat_id_s[i] = false;
+            }
+#endif
+
             device->mul_mat_l_int[i]    = device->mul_mat_l[i];
             device->mul_mat_m_int[i]    = device->mul_mat_m[i];
             device->mul_mat_s_int[i]    = device->mul_mat_s[i];
@@ -7604,8 +7617,12 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
     if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         GGML_ASSERT(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
-        for (size_t i = 0; i < height; i++) {
-            memcpy((uint8_t *)dst->ptr + offset + i * dpitch, (const uint8_t *) src + i * spitch, width);
+        if (width == spitch && width == dpitch) {
+            memcpy((uint8_t *)dst->ptr + offset, src, width * height);
+        } else {
+            for (size_t i = 0; i < height; i++) {
+                memcpy((uint8_t *)dst->ptr + offset + i * dpitch, (const uint8_t *) src + i * spitch, width);
+            }
         }
     } else {
         std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
@@ -7724,8 +7741,29 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
     if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
         GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
-        for (size_t i = 0; i < height; i++) {
-            memcpy((uint8_t *) dst + i * dpitch, (const uint8_t *) src->ptr + offset + i * spitch, width);
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->compute_queue.cmd_pool);
+        ggml_vk_ctx_begin(src->device, subctx);
+        subctx->s->buffer->buf.pipelineBarrier(
+            vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer,
+            vk::PipelineStageFlagBits::eHost,
+            {},
+            { { vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferWrite,
+                vk::AccessFlagBits::eHostRead } },
+            {}, {});
+        ggml_vk_ctx_end(subctx);
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX),
+                 "vk_buffer_read_2d uma waitForFences");
+        src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
+
+        if (width == spitch && width == dpitch) {
+            memcpy(dst, (const uint8_t *) src->ptr + offset, width * height);
+        } else {
+            for (size_t i = 0; i < height; i++) {
+                memcpy((uint8_t *) dst + i * dpitch, (const uint8_t *) src->ptr + offset + i * spitch, width);
+            }
         }
     } else {
         std::lock_guard<std::recursive_mutex> guard(src->device->mutex);

gguf-py/gguf/constants.py

@@ -154,6 +154,9 @@ class Keys:
         HIDDEN_ACT                        = "{arch}.hidden_activation"
         DENSE_FEAT_IN_SIZE                = "{arch}.{dense}_feat_in"
         DENSE_FEAT_OUT_SIZE               = "{arch}.{dense}_feat_out"
+        TARGET_LAYERS                     = "{arch}.target_layers"
+        TARGET_HIDDEN_SIZE                = "{arch}.target_hidden_size"
+        NORM_BEFORE_RESIDUAL              = "{arch}.norm_before_residual"
 
     class Attention:
         HEAD_COUNT                   = "{arch}.attention.head_count"
@@ -272,7 +275,8 @@ class Keys:
         CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # Normalizer constants
-        NORMALIZER_LOWERCASE = "tokenizer.ggml.normalizer.lowercase"
+        NORMALIZER_LOWERCASE     = "tokenizer.ggml.normalizer.lowercase"
+        NORMALIZER_STRIP_ACCENTS = "tokenizer.ggml.normalizer.strip_accents"
         # FIM/Infill special tokens constants
         FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
         FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
@@ -510,6 +514,7 @@ class MODEL_ARCH(IntEnum):
     RND1             = auto()
     PANGU_EMBED      = auto()
     MISTRAL3         = auto()
+    EAGLE3           = auto()
     MISTRAL4         = auto()
     PADDLEOCR        = auto()
     MIMO2            = auto()
@@ -900,14 +905,17 @@ class MODEL_TENSOR(IntEnum):
     A_PER_DIM_K_SCALE     = auto() # gemma4
     A_PER_DIM_SCALE       = auto() # gemma4
     # nextn/mtp
-    NEXTN_PROJ_PRE       = auto()
-    NEXTN_PROJ_POST      = auto()
-    NEXTN_EH_PROJ        = auto()
-    NEXTN_EMBED_TOKENS   = auto()
-    NEXTN_ENORM          = auto()
-    NEXTN_HNORM          = auto()
+    NEXTN_PROJ_PRE         = auto()
+    NEXTN_PROJ_POST        = auto()
+    NEXTN_EH_PROJ          = auto()
+    NEXTN_EMBED_TOKENS     = auto()
+    NEXTN_ENORM            = auto()
+    NEXTN_HNORM            = auto()
     NEXTN_SHARED_HEAD_HEAD = auto()
     NEXTN_SHARED_HEAD_NORM = auto()
+    # eagle3
+    FC                     = auto()  # feature fusion layer
+    D2T                    = auto()  # draft to target vocabulary mapping
     # lfm2 audio
     A_ENC_NORM_CONV        = auto()
     A_ENC_LINEAR_POS       = auto()
@@ -1062,6 +1070,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
     MODEL_ARCH.MISTRAL3:         "mistral3",
+    MODEL_ARCH.EAGLE3:           "eagle3",
     MODEL_ARCH.MISTRAL4:         "mistral4",
     MODEL_ARCH.PADDLEOCR:        "paddleocr",
     MODEL_ARCH.MIMO2:            "mimo2",
@@ -1094,8 +1103,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.POS_EMBD:                  "position_embd",
     MODEL_TENSOR.OUTPUT_NORM:               "output_norm",
     MODEL_TENSOR.OUTPUT:                    "output",
-    MODEL_TENSOR.DENSE_2_OUT:                "dense_2", # embeddinggemma 2_Dense
-    MODEL_TENSOR.DENSE_3_OUT:                "dense_3", # embeddinggemma 2_Dense
+    MODEL_TENSOR.DENSE_2_OUT:               "dense_2", # embeddinggemma 2_Dense
+    MODEL_TENSOR.DENSE_3_OUT:               "dense_3", # embeddinggemma 2_Dense
     MODEL_TENSOR.ROPE_FREQS:                "rope_freqs",
     MODEL_TENSOR.ROPE_FACTORS_LONG:         "rope_factors_long",
     MODEL_TENSOR.ROPE_FACTORS_SHORT:        "rope_factors_short",
@@ -1487,6 +1496,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.NEXTN_HNORM:               "blk.{bid}.nextn.hnorm",
     MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD:    "blk.{bid}.nextn.shared_head_head",
     MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM:    "blk.{bid}.nextn.shared_head_norm",
+    MODEL_TENSOR.FC:                        "fc",
+    MODEL_TENSOR.D2T:                       "d2t",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -4027,6 +4038,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.EAGLE3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FC,
+        MODEL_TENSOR.D2T,
+    ],
     MODEL_ARCH.MISTRAL4: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -1124,6 +1124,9 @@ class GGUFWriter:
     def add_normalizer_lowercase(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.NORMALIZER_LOWERCASE, value)
 
+    def add_normalizer_strip_accents(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.NORMALIZER_STRIP_ACCENTS, value)
+
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 

gguf-py/gguf/vocab.py

@@ -53,6 +53,7 @@ class SpecialVocab:
     special_token_ids: dict[str, int]
     chat_template: str | Sequence[Mapping[str, str]] | None
     normalizer_lowercase: bool | None
+    normalizer_strip_accents: bool | None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
@@ -66,6 +67,7 @@ class SpecialVocab:
         self.merges = []
         self.chat_template = None
         self.normalizer_lowercase = None
+        self.normalizer_strip_accents = None
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
@@ -108,6 +110,10 @@ class SpecialVocab:
             if not quiet:
                 logger.info(f'Setting normalizer_lowercase to {self.normalizer_lowercase}')
             gw.add_normalizer_lowercase(self.normalizer_lowercase)
+        if self.normalizer_strip_accents is not None:
+            if not quiet:
+                logger.info(f'Setting normalizer_strip_accents to {self.normalizer_strip_accents}')
+            gw.add_normalizer_strip_accents(self.normalizer_strip_accents)
 
     def _load(self, path: Path) -> None:
         self._try_load_from_tokenizer_json(path)
@@ -155,17 +161,21 @@ class SpecialVocab:
     def _parse_normalizer(self, normalizer: dict) -> None:
         # ref: https://huggingface.co/docs/tokenizers/api/normalizers
         #
-        # Detects lowercase normalization in three possible formats:
-        # 1. Standalone: {"type": "Lowercase"}
-        # 2. BertNormalizer attribute: {"type": "BertNormalizer", "lowercase": true, ...}
-        # 3. Nested in Sequence: {"type": "Sequence", "normalizers": [...]}
+        # Extracts normalizer flags from three possible formats:
+        # 1. Standalone:           {"type": "Lowercase"}
+        # 2. BertNormalizer attrs: {"type": "BertNormalizer", ...}
+        # 3. Nested in Sequence:   {"type": "Sequence", "normalizers": [...]}
 
         normalizer_type = normalizer.get('type')
         if normalizer_type == 'Lowercase':
             self.normalizer_lowercase = True
+        elif normalizer_type == 'StripAccents':
+            self.normalizer_strip_accents = True
         elif normalizer_type == 'BertNormalizer':
             if 'lowercase' in normalizer:
                 self.normalizer_lowercase = normalizer['lowercase']
+            if 'strip_accents' in normalizer:
+                self.normalizer_strip_accents = normalizer['strip_accents']
         elif normalizer_type == 'Sequence':
             for norm in normalizer.get('normalizers', []):
                 self._parse_normalizer(norm)
@@ -246,6 +256,11 @@ class SpecialVocab:
                             if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
                                 if not tokenizer_config:
                                     special_bos = special_first
+                                elif special_first not in (special_bos, special_cls):
+                                    if not special_bos:
+                                        tokenizer_config['bos_token'] = special_bos = special_first
+                                    if not special_cls:
+                                        tokenizer_config['cls_token'] = special_cls = special_first
                                 self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
                                 if special_first not in (special_bos, special_cls):
                                     logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')

scripts/sync-ggml.last

@@ -1 +1 @@
-7142aa6bf9fcaeec0fef8d80fcd90afe4268adf1
+3af5f5760e19a96427f5f7a93b79cbdf3d4b265b

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.46.1"
+HTTPLIB_VERSION = "refs/tags/v0.47.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

scripts/ui-assets.cmake

@@ -16,11 +16,80 @@ set(HF_ENABLED        "" CACHE STRING "Whether to allow HF Bucket download (ON/O
 set(BUILD_UI          "" CACHE STRING "Build UI via npm (ON/OFF)")
 set(LLAMA_UI_EMBED    "" CACHE STRING "Path to llama-ui-embed helper")
 
+# IMPORTANT: When adding PWA assets, sync across all 3 places:
+#   1. tools/ui/src/lib/constants/pwa.ts   (APPLE_DEVICES, PUBLIC_ENDPOINTS)
+#   2. tools/server/server-http.cpp        (public_endpoints)
+#   3. scripts/ui-assets.cmake             (ASSETS list)
+# - C++ (server-http.cpp) - public endpoints (splash screens generated via helper)
+# - TypeScript (constants/pwa.ts) - APPLE_DEVICES, PWA_MANIFEST, PUBLIC_ENDPOINTS
+#
+# When adding/changing PWA assets, update tools/ui/src/lib/constants/pwa.ts first,
+# then sync any new file names here and in server-http.cpp.
 set(ASSETS
-    bundle.css
-    bundle.js
     index.html
     loading.html
+    # PWA assets
+    favicon.ico
+    favicon-dark.ico
+    favicon.svg
+    favicon-dark.svg
+    pwa-64x64.png
+    pwa-192x192.png
+    pwa-512x512.png
+    maskable-icon-512x512.png
+    apple-touch-icon-180x180.png
+    # iOS splash screens
+    apple-splash-portrait-640x1136.png
+    apple-splash-landscape-1136x640.png
+    apple-splash-portrait-750x1334.png
+    apple-splash-landscape-1334x750.png
+    apple-splash-portrait-1170x2532.png
+    apple-splash-landscape-2532x1170.png
+    apple-splash-portrait-1179x2556.png
+    apple-splash-landscape-2556x1179.png
+    apple-splash-portrait-1206x2622.png
+    apple-splash-landscape-2622x1206.png
+    apple-splash-portrait-1284x2778.png
+    apple-splash-landscape-2778x1284.png
+    apple-splash-portrait-1290x2796.png
+    apple-splash-landscape-2796x1290.png
+    apple-splash-portrait-1320x2868.png
+    apple-splash-landscape-2868x1320.png
+    apple-splash-portrait-1488x2266.png
+    apple-splash-landscape-2266x1488.png
+    apple-splash-portrait-1640x2360.png
+    apple-splash-landscape-2360x1640.png
+    apple-splash-portrait-1668x2388.png
+    apple-splash-landscape-2388x1668.png
+    apple-splash-portrait-2048x2732.png
+    apple-splash-landscape-2732x2048.png
+    # iOS dark splash screens
+    apple-splash-portrait-dark-640x1136.png
+    apple-splash-landscape-dark-1136x640.png
+    apple-splash-portrait-dark-750x1334.png
+    apple-splash-landscape-dark-1334x750.png
+    apple-splash-portrait-dark-1170x2532.png
+    apple-splash-landscape-dark-2532x1170.png
+    apple-splash-portrait-dark-1179x2556.png
+    apple-splash-landscape-dark-2556x1179.png
+    apple-splash-portrait-dark-1206x2622.png
+    apple-splash-landscape-dark-2622x1206.png
+    apple-splash-portrait-dark-1284x2778.png
+    apple-splash-landscape-dark-2778x1284.png
+    apple-splash-portrait-dark-1290x2796.png
+    apple-splash-landscape-dark-2796x1290.png
+    apple-splash-portrait-dark-1320x2868.png
+    apple-splash-landscape-dark-2868x1320.png
+    apple-splash-portrait-dark-1640x2360.png
+    apple-splash-landscape-dark-2360x1640.png
+    apple-splash-portrait-dark-1668x2388.png
+    apple-splash-landscape-dark-2388x1668.png
+    apple-splash-portrait-dark-2048x2732.png
+    apple-splash-landscape-dark-2732x2048.png
+    manifest.webmanifest
+    sw.js
+    _app/version.json
+    build.json
 )
 
 set(DIST_DIR     "${UI_BINARY_DIR}/dist")
@@ -159,7 +228,7 @@ function(npm_build out_var)
 
     message(STATUS "UI: running npm run build, output -> ${DIST_DIR}")
     execute_process(
-        COMMAND ${CMAKE_COMMAND} -E env "LLAMA_UI_OUT_DIR=${DIST_DIR}"
+        COMMAND ${CMAKE_COMMAND} -E env "LLAMA_UI_OUT_DIR=${DIST_DIR}" "LLAMA_UI_VERSION=${HF_VERSION}" "LLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}"
                 ${NPM_EXECUTABLE} run build
         WORKING_DIRECTORY "${UI_SOURCE_DIR}"
         RESULT_VARIABLE rc
@@ -274,8 +343,35 @@ function(emit_files)
         foreach(asset ${ASSETS})
             list(APPEND args "${asset}" "${DIST_DIR}/${asset}")
         endforeach()
+
+        # Bundle files live in _app/immutable/ — vanilla SvelteKit output, no plugin
+        # rewriting. Embedded names must match the exact _app/ paths that index.html
+        # and sw.js reference.
+        file(GLOB_RECURSE detected_bundle_js "${DIST_DIR}/_app/immutable/bundle.*.js")
+        file(GLOB_RECURSE detected_bundle_css "${DIST_DIR}/_app/immutable/assets/bundle.*.css")
+        file(GLOB_RECURSE detected_workbox "${DIST_DIR}/workbox-*.js")
+        # Compute relative path from DIST_DIR to each found file.
+        # e.g. /path/to/build/tools/ui/dist/_app/immutable/bundle.XXX.js
+        #      -> _app/immutable/bundle.XXX.js
+        foreach(f ${detected_bundle_js})
+            string(REPLACE "${DIST_DIR}/" "" rel "${f}")
+            list(APPEND args "${rel}" "${f}")
+        endforeach()
+        foreach(f ${detected_bundle_css})
+            string(REPLACE "${DIST_DIR}/" "" rel "${f}")
+            list(APPEND args "${rel}" "${f}")
+        endforeach()
+        foreach(f ${detected_workbox})
+            string(REPLACE "${DIST_DIR}/" "" rel "${f}")
+            list(APPEND args "${rel}" "${f}")
+        endforeach()
     endif()
 
+    # Create build.json with the llama.cpp build number for UI version display.
+    # This is separate from SvelteKit's _app/version.json (used for SW cache invalidation).
+    # build.json is generated by the vite plugin (buildInfoPlugin) during npm build.
+    # CMake just embeds it from the dist that npm produced.
+
     execute_process(
         COMMAND "${LLAMA_UI_EMBED}" ${args}
         RESULT_VARIABLE rc
@@ -300,6 +396,8 @@ endif()
 set(provisioned FALSE)
 
 if(BUILD_UI)
+    # Resolve version from git build-info if not explicitly set
+    resolve_version(HF_VERSION)
     npm_build(NPM_OK)
     if(NPM_OK)
         set(provisioned TRUE)

src/llama-arch.cpp

@@ -3,7 +3,6 @@
 #include "llama-impl.h"
 
 #include <map>
-#include <set>
 #include <vector>
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
@@ -128,6 +127,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_EAGLE3,           "eagle3"           },
     { LLM_ARCH_MISTRAL4,         "mistral4"         },
     { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
     { LLM_ARCH_MIMO2,            "mimo2"            },
@@ -292,46 +292,51 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
 
+    { LLM_KV_TARGET_LAYERS,         "%s.target_layers"        },
+    { LLM_KV_TARGET_HIDDEN_SIZE,    "%s.target_hidden_size"   },
+    { LLM_KV_NORM_BEFORE_RESIDUAL,  "%s.norm_before_residual" },
+
     { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
     // sentence-transformers dense modules feature dims
     { LLM_KV_DENSE_2_FEAT_IN,        "%s.dense_2_feat_in"  },
-    { LLM_KV_DENSE_2_FEAT_OUT,       "%s.dense_2_feat_out"  },
-    { LLM_KV_DENSE_3_FEAT_IN,        "%s.dense_3_feat_in"   },
-    { LLM_KV_DENSE_3_FEAT_OUT,       "%s.dense_3_feat_out"  },
+    { LLM_KV_DENSE_2_FEAT_OUT,       "%s.dense_2_feat_out" },
+    { LLM_KV_DENSE_3_FEAT_IN,        "%s.dense_3_feat_in"  },
+    { LLM_KV_DENSE_3_FEAT_OUT,       "%s.dense_3_feat_out" },
 
-    { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
-    { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
-    { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE,           "tokenizer.ggml.token_type"               },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,     "tokenizer.ggml.token_type_count"         },
-    { LLM_KV_TOKENIZER_SCORES,               "tokenizer.ggml.scores"                   },
-    { LLM_KV_TOKENIZER_MERGES,               "tokenizer.ggml.merges"                   },
-    { LLM_KV_TOKENIZER_BOS_ID,               "tokenizer.ggml.bos_token_id"             },
-    { LLM_KV_TOKENIZER_EOS_ID,               "tokenizer.ggml.eos_token_id"             },
-    { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
-    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
-    { LLM_KV_TOKENIZER_UNK_ID,               "tokenizer.ggml.unknown_token_id"         },
-    { LLM_KV_TOKENIZER_SEP_ID,               "tokenizer.ggml.seperator_token_id"       },
-    { LLM_KV_TOKENIZER_PAD_ID,               "tokenizer.ggml.padding_token_id"         },
-    { LLM_KV_TOKENIZER_CLS_ID,               "tokenizer.ggml.cls_token_id"             },
-    { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
-    { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
-    { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
-    { LLM_KV_TOKENIZER_ADD_SEP,              "tokenizer.ggml.add_sep_token"            },
-    { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
-    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
-    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
-    { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
-    { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
-    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
-    { LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE, "tokenizer.ggml.normalizer.lowercase"     },
-    { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_PAD_ID,           "tokenizer.ggml.fim_pad_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_REP_ID,           "tokenizer.ggml.fim_rep_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_SEP_ID,           "tokenizer.ggml.fim_sep_token_id"         },
-    { LLM_KV_TOKENIZER_SUPPRESS_TOKENS,      "tokenizer.ggml.suppress_tokens"          },
+    { LLM_KV_TOKENIZER_MODEL,                    "tokenizer.ggml.model"                    },
+    { LLM_KV_TOKENIZER_PRE,                      "tokenizer.ggml.pre"                      },
+    { LLM_KV_TOKENIZER_LIST,                     "tokenizer.ggml.tokens"                   },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE,               "tokenizer.ggml.token_type"               },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,         "tokenizer.ggml.token_type_count"         },
+    { LLM_KV_TOKENIZER_SCORES,                   "tokenizer.ggml.scores"                   },
+    { LLM_KV_TOKENIZER_MERGES,                   "tokenizer.ggml.merges"                   },
+    { LLM_KV_TOKENIZER_BOS_ID,                   "tokenizer.ggml.bos_token_id"             },
+    { LLM_KV_TOKENIZER_EOS_ID,                   "tokenizer.ggml.eos_token_id"             },
+    { LLM_KV_TOKENIZER_EOT_ID,                   "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,                   "tokenizer.ggml.eom_token_id"             },
+    { LLM_KV_TOKENIZER_UNK_ID,                   "tokenizer.ggml.unknown_token_id"         },
+    { LLM_KV_TOKENIZER_SEP_ID,                   "tokenizer.ggml.seperator_token_id"       },
+    { LLM_KV_TOKENIZER_PAD_ID,                   "tokenizer.ggml.padding_token_id"         },
+    { LLM_KV_TOKENIZER_CLS_ID,                   "tokenizer.ggml.cls_token_id"             },
+    { LLM_KV_TOKENIZER_MASK_ID,                  "tokenizer.ggml.mask_token_id"            },
+    { LLM_KV_TOKENIZER_ADD_BOS,                  "tokenizer.ggml.add_bos_token"            },
+    { LLM_KV_TOKENIZER_ADD_EOS,                  "tokenizer.ggml.add_eos_token"            },
+    { LLM_KV_TOKENIZER_ADD_SEP,                  "tokenizer.ggml.add_sep_token"            },
+    { LLM_KV_TOKENIZER_ADD_PREFIX,               "tokenizer.ggml.add_space_prefix"         },
+    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,          "tokenizer.ggml.remove_extra_whitespaces" },
+    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,     "tokenizer.ggml.precompiled_charsmap"     },
+    { LLM_KV_TOKENIZER_HF_JSON,                  "tokenizer.huggingface.json"              },
+    { LLM_KV_TOKENIZER_RWKV,                     "tokenizer.rwkv.world"                    },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,            "tokenizer.chat_template"                 },
+    { LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     "tokenizer.ggml.normalizer.lowercase"     },
+    { LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, "tokenizer.ggml.normalizer.strip_accents" },
+    { LLM_KV_TOKENIZER_FIM_PRE_ID,               "tokenizer.ggml.fim_pre_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SUF_ID,               "tokenizer.ggml.fim_suf_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_MID_ID,               "tokenizer.ggml.fim_mid_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_PAD_ID,               "tokenizer.ggml.fim_pad_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_REP_ID,               "tokenizer.ggml.fim_rep_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SEP_ID,               "tokenizer.ggml.fim_sep_token_id"         },
+    { LLM_KV_TOKENIZER_SUPPRESS_TOKENS,          "tokenizer.ggml.suppress_tokens"          },
 
     { LLM_KV_ADAPTER_TYPE,                    "adapter.type"               },
     { LLM_KV_ADAPTER_LORA_ALPHA,              "adapter.lora.alpha"         },
@@ -561,6 +566,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_INDEXER_ATTN_Q_B,                       "blk.%d.indexer.attn_q_b" },
     { LLM_TENSOR_MASKED_EMBD_CENTROIDS,                  "masked_embd_centroids" },
     { LLM_TENSOR_MASKED_EMBD_ORDERING,                   "masked_embd_ordering" },
+    { LLM_TENSOR_FC,                                     "fc" },
+    { LLM_TENSOR_D2T,                                    "d2t" },
 };
 
 // declare information about the model weight tensors:
@@ -787,6 +794,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_LATENT_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_MASKED_EMBD_CENTROIDS,      {LLM_TENSOR_LAYER_INPUT,     GGML_OP_NONE}},
     {LLM_TENSOR_MASKED_EMBD_ORDERING,       {LLM_TENSOR_LAYER_INPUT,     GGML_OP_NONE}},
+    // eagle3
+    {LLM_TENSOR_FC,                         {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_D2T,                        {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

src/llama-arch.h

@@ -141,6 +141,7 @@ enum llm_arch {
     LLM_ARCH_KIMI_LINEAR,
     LLM_ARCH_TALKIE,
     LLM_ARCH_MELLUM,
+    LLM_ARCH_EAGLE3,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -314,6 +315,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
     LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,
+    LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -336,6 +338,10 @@ enum llm_kv {
 
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
 
+    LLM_KV_TARGET_LAYERS,
+    LLM_KV_TARGET_HIDDEN_SIZE,
+    LLM_KV_NORM_BEFORE_RESIDUAL,
+
     LLM_KV_SHORTCONV_L_CACHE,
 
     LLM_KV_XIELU_ALPHA_N,
@@ -568,6 +574,8 @@ enum llm_tensor {
     LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
     LLM_TENSOR_MASKED_EMBD_CENTROIDS,
     LLM_TENSOR_MASKED_EMBD_ORDERING,
+    LLM_TENSOR_FC,
+    LLM_TENSOR_D2T,
 };
 
 

src/llama-context.cpp

@@ -71,6 +71,9 @@ llama_context::llama_context(
     cparams.no_perf                 = params.no_perf;
     cparams.warmup                  = false;
 
+    cparams.embeddings_layer_inp.resize(hparams.n_layer(), false);
+    embd_layer_inp.resize(hparams.n_layer());
+
     cparams.ctx_type     = params.ctx_type;
     cparams.pooling_type = params.pooling_type;
 
@@ -91,12 +94,21 @@ llama_context::llama_context(
     if (model.arch == LLM_ARCH_GEMMA4_ASSISTANT) {
         if (params.ctx_other == nullptr) {
             // TODO: change from runtime_error to llama_exception to avoid printing error message
-            throw std::runtime_error("Gemma4Assistant requires ctx_other to be set (this is normal during memory fitting)");
+            throw std::runtime_error("Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)");
         }
 
         cparams.ctx_other = params.ctx_other;
     }
 
+    if (model.arch == LLM_ARCH_EAGLE3) {
+        if (model.tok_embd == nullptr || model.output == nullptr) {
+            if (params.ctx_other == nullptr) {
+                throw std::runtime_error("EAGLE3 requires ctx_other to be set (this warning is normal during memory fitting)");
+            }
+            cparams.ctx_other = params.ctx_other;
+        }
+    }
+
     // Initialize backend samplers here so they are part of the sampling graph
     // before the reserve passes run later in this function. This avoids a later
     // re-reserve when graph nodes change.
@@ -194,7 +206,7 @@ llama_context::llama_context(
 
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
-    cparams.n_outputs_max = params.n_outputs_max == 0 ? cparams.n_batch : params.n_outputs_max;
+    cparams.n_outputs_max = params.n_outputs_max == 0 || llama_model_has_encoder(&model) ? cparams.n_batch : params.n_outputs_max;
 
     cparams.op_offload = params.op_offload;
     cparams.kv_unified = params.kv_unified;
@@ -938,6 +950,14 @@ float * llama_context::get_embeddings_nextn_ith(int32_t i) {
     }
 }
 
+float * llama_context::get_embeddings_layer_inp(uint32_t lid) {
+    output_reorder();
+
+    GGML_ASSERT(lid < embd_layer_inp.size() && embd_layer_inp[lid].has_data());
+
+    return embd_layer_inp[lid].data;
+}
+
 llama_token llama_context::get_sampled_token_ith(int32_t idx) {
     output_reorder();
 
@@ -1125,6 +1145,17 @@ void llama_context::set_embeddings_nextn(bool value, bool masked) {
     cparams.embeddings_nextn_masked = masked;
 }
 
+void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
+    LLAMA_LOG_DEBUG("%s: lid = %d, enable = %d\n", __func__, lid, enable);
+
+    GGML_ASSERT(lid < model.hparams.n_layer());
+
+    cparams.embeddings_layer_inp[lid] = enable;
+
+    // note: without this reserve, the draft acceptance drops to zero. not sure why - this is unexpected
+    sched_need_reserve = true;
+}
+
 void llama_context::set_causal_attn(bool value) {
     LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);
 
@@ -1350,7 +1381,8 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd_inp();
+    // eagle3/DFlash: features as encoder input, and non-draft paths fall back to model's input dim
+    const int64_t n_embd = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -1925,6 +1957,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
             }
         }
 
+        extract_layer_inputs(res, n_tokens_prev, ubatch.n_tokens);
+
         // extract nextn embeddings before
         // only meaningful in LLAMA_POOLING_TYPE_NONE (per-token); other pooling modes are ignored.
         {
@@ -2029,6 +2063,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     const auto n_batch    = cparams.n_batch;
     const auto n_vocab    = vocab.n_tokens();
+    const auto n_embd     = hparams.n_embd;
     const auto n_embd_out = hparams.n_embd_out();
 
     bool has_logits     = true;
@@ -2041,9 +2076,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         has_embd   = true;
     }
 
-
     size_t backend_float_count = 0;
     size_t backend_token_count = 0;
+    size_t embd_layer_inp_float_count = 0;
 
     logits.size     = has_logits     ? n_vocab*n_outputs_max     : 0;
     embd.size       = has_embd       ? n_embd_out*n_outputs_max  : 0;
@@ -2055,6 +2090,12 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         embd_nextn.size = (size_t) n_embd_out * n_batch;
     }
 
+    for (bool enabled : cparams.embeddings_layer_inp) {
+        if (enabled) {
+            embd_layer_inp_float_count += (size_t) n_embd * n_batch;
+        }
+    }
+
     // Allocate backend sampling output buffers if there are backend samplers configured.
     const bool has_sampling = !sampling.samplers.empty();
     if (has_sampling) {
@@ -2069,8 +2110,8 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0;
     const size_t new_size  =
-        (logits.size + embd.size + embd_nextn.size + backend_float_count) * sizeof(float) +
-        (                                               backend_token_count) * sizeof(llama_token);
+        (logits.size + embd.size + embd_nextn.size + embd_layer_inp_float_count + backend_float_count) * sizeof(float) +
+        (                                                                         backend_token_count) * sizeof(llama_token);
 
     // alloc only when more than the current capacity is required
     // TODO: also consider shrinking the buffer
@@ -2087,6 +2128,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
             logits.data = nullptr;
             embd.data = nullptr;
             embd_nextn.data = nullptr;
+            for (auto & layer_inp : embd_layer_inp) {
+                layer_inp = {nullptr, 0};
+            }
         }
 
         auto * buft = ggml_backend_cpu_buffer_type();
@@ -2118,6 +2162,15 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     embd_nextn = has_embd_nextn ? buffer_view<float>{(float *) (base + offset), embd_nextn.size} : buffer_view<float>{nullptr, 0};
     offset += embd_nextn.size * sizeof(float);
 
+    for (uint32_t il = 0; il < embd_layer_inp.size(); ++il) {
+        if (cparams.embeddings_layer_inp[il]) {
+            embd_layer_inp[il] = buffer_view<float>{(float *) (base + offset), (size_t) n_embd * n_batch};
+            offset += embd_layer_inp[il].size * sizeof(float);
+        } else {
+            embd_layer_inp[il] = buffer_view<float>{nullptr, 0};
+        }
+    }
+
     if (has_sampling) {
         sampling.logits = {(float *) (base + offset), (size_t)(n_vocab*n_outputs_max)};
         offset += sampling.logits.size * sizeof(float);
@@ -2164,6 +2217,34 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     return n_outputs_max;
 }
 
+void llama_context::extract_layer_inputs(const llm_graph_result * res, size_t token_offset, size_t n_tokens) {
+    for (uint32_t il = 0; il < cparams.embeddings_layer_inp.size(); ++il) {
+        if (!cparams.embeddings_layer_inp[il]) {
+            continue;
+        }
+        if (!embd_layer_inp[il].has_data()) {
+            GGML_ABORT("output layer input buffer not allocated");
+        }
+        ggml_tensor * t = res->get_layer_inp((int) il);
+        if (!t) {
+            GGML_ABORT("layer input tensor not found");
+        }
+
+        const size_t nbytes = ggml_nbytes(t);
+        const size_t nfloats = nbytes / sizeof(float);
+        GGML_ASSERT(n_tokens > 0);
+        GGML_ASSERT(nfloats % n_tokens == 0);
+
+        const size_t row_floats = nfloats / n_tokens;
+        const size_t dst_offset = token_offset * row_floats;
+        GGML_ASSERT(dst_offset + nfloats <= embd_layer_inp[il].size);
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched.get(), t);
+        GGML_ASSERT(backend != nullptr);
+        ggml_backend_tensor_get_async(backend, t, embd_layer_inp[il].data + dst_offset, 0, nbytes);
+    }
+}
+
 void llama_context::output_reorder() {
     const uint64_t n_vocab = model.vocab.n_tokens();
     const uint64_t n_embd  = model.hparams.n_embd;
@@ -2190,6 +2271,16 @@ void llama_context::output_reorder() {
             }
         }
 
+        if (embd_layer_inp.size() > 0) {
+            for (int lid = 0; lid < (int) embd_layer_inp.size(); ++lid) {
+                if (embd_layer_inp[lid].size > 0) {
+                    for (uint64_t k = 0; k < n_embd; ++k) {
+                        std::swap(embd_layer_inp[lid].data[i0*n_embd + k], embd_layer_inp[lid].data[i1*n_embd + k]);
+                    }
+                }
+            }
+        }
+
         if (!sampling.samplers.empty()) {
             assert(sampling.logits.size > 0);
             assert(sampling.probs.size > 0);
@@ -3604,6 +3695,10 @@ void llama_set_embeddings_nextn(llama_context * ctx, bool value, bool masked) {
     ctx->set_embeddings_nextn(value, masked);
 }
 
+void llama_set_embeddings_layer_inp(llama_context * ctx, uint32_t lid, bool value) {
+    ctx->set_embeddings_layer_inp(lid, value);
+}
+
 llama_memory_t llama_get_memory(const struct llama_context * ctx) {
     if (!ctx) {
         return nullptr;
@@ -3624,6 +3719,12 @@ float * llama_get_embeddings_nextn_ith(llama_context * ctx, int32_t i) {
     return ctx->get_embeddings_nextn_ith(i);
 }
 
+float * llama_get_embeddings_layer_inp(llama_context * ctx, uint32_t lid) {
+    ctx->synchronize();
+
+    return ctx->get_embeddings_layer_inp(lid);
+}
+
 bool llama_set_sampler(llama_context * ctx, llama_seq_id seq_id, llama_sampler * smpl) {
     return ctx->set_sampler(seq_id, smpl);
 }

src/llama-context.h

@@ -88,6 +88,8 @@ struct llama_context {
     float * get_embeddings_nextn();
     float * get_embeddings_nextn_ith(int32_t i);
 
+    float * get_embeddings_layer_inp(uint32_t lid);
+
     llama_token * get_sampled_tokens() const;
     llama_token   get_sampled_token_ith(int32_t idx);
 
@@ -112,6 +114,7 @@ struct llama_context {
 
     void set_embeddings (bool value);
     void set_embeddings_nextn(bool value, bool masked);
+    void set_embeddings_layer_inp(uint32_t lid, bool enable);
     void set_causal_attn(bool value);
     void set_warmup(bool value);
 
@@ -226,6 +229,10 @@ private:
     // map the output row index `i` to batch index
     int64_t output_resolve_row(int32_t i) const;
 
+    // async-copy enabled layer-input tensors (per cparams.output_layer_inp)
+    // from backend into host-side embd_layer_inp buffers
+    void extract_layer_inputs(const llm_graph_result * res, size_t token_offset, size_t n_tokens);
+
     //
     // graph
     //
@@ -288,6 +295,10 @@ private:
     // sets llm_graph_result::t_h_nextn
     buffer_view<float> embd_nextn = {nullptr, 0};
 
+    // host buffers for output layer input embeddings, per layer
+    // populated when cparams.output_layer_inp[il] is true
+    std::vector<buffer_view<float>> embd_layer_inp;
+
     struct sampling_info {
         // !samplers.empty() to check if any samplers are active
         std::map<llama_seq_id, llama_sampler *> samplers;

src/llama-cparams.h

@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <cstdint>
+#include <vector>
 
 #define LLAMA_MAX_SEQ 256
 
@@ -44,6 +45,8 @@ struct llama_cparams {
     bool kv_unified;
     bool pipeline_parallel;
 
+    std::vector<bool> embeddings_layer_inp; // [n_layer()] extract input embeddings for layer
+
     enum llama_context_type ctx_type;
     enum llama_pooling_type pooling_type;
 

src/llama-ext.h

@@ -2,6 +2,7 @@
 
 // this is a staging header for new llama.cpp API
 // breaking changes and C++ are allowed. everything here should be considered WIP
+// try as much as possible to not include this header in the rest of the codebase
 
 #include "llama.h"
 
@@ -101,4 +102,20 @@ LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);
 // LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
 LLAMA_API float * llama_get_embeddings_nextn_ith(struct llama_context * ctx, int32_t i);
 
+// Set whether the context outputs the input embeddings of a specific layer
+LLAMA_API void llama_set_embeddings_layer_inp(struct llama_context * ctx, uint32_t lid, bool value);
+
+// mirrors:
+// LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+LLAMA_API float * llama_get_embeddings_layer_inp(struct llama_context * ctx, uint32_t lid);
+
 LLAMA_API llama_context * llama_get_ctx_other(struct llama_context * ctx);
+
+//
+// model/context data extraction
+//
+
+// returns pointer to the target-model layer indices
+LLAMA_API const int32_t * llama_model_target_layer_ids  (const struct llama_model * model);
+// returns the number of extracted layers from target model
+LLAMA_API uint32_t        llama_model_target_layer_ids_n(const struct llama_model * model);

src/llama-graph.cpp

@@ -904,6 +904,10 @@ void llm_graph_result::reset() {
     t_logits      = nullptr;
     t_embd        = nullptr;
     t_embd_pooled = nullptr;
+
+    t_layer_inp.resize(LLAMA_MAX_LAYERS);
+    std::fill(t_layer_inp.begin(), t_layer_inp.end(), nullptr);
+
     t_sampled.clear();
     t_sampled_probs.clear();
     t_sampled_logits.clear();
@@ -932,7 +936,7 @@ void llm_graph_result::set_inputs(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_result::set_outputs() {
+void llm_graph_result::set_outputs(const llm_graph_params & params) {
     if (t_logits != nullptr) {
         ggml_set_output(t_logits);
     }
@@ -945,6 +949,15 @@ void llm_graph_result::set_outputs() {
     if (t_h_nextn != nullptr) {
         ggml_set_output(t_h_nextn);
     }
+    {
+        const auto & embeddings_layer_inp = params.cparams.embeddings_layer_inp;
+        for (size_t il = 0; il < embeddings_layer_inp.size(); ++il) {
+            if (embeddings_layer_inp[il]) {
+                GGML_ASSERT(t_layer_inp[il] != nullptr && "layer input tensor is null");
+                ggml_set_output(t_layer_inp[il]);
+            }
+        }
+    }
     for (auto & [seq_id, t] : t_sampled) {
         if (t != nullptr) {
             ggml_set_output(t);

src/llama-graph.h

@@ -705,6 +705,8 @@ public:
     ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
     ggml_tensor * get_h_nextn()     const { return t_h_nextn; }
 
+    ggml_tensor * get_layer_inp(int il) const { return t_layer_inp[il]; }
+
     ggml_cgraph  * get_gf()  const { return gf; }
     ggml_context * get_ctx() const { return ctx_compute.get(); }
 
@@ -713,7 +715,7 @@ public:
     void reset();
 
     void set_inputs(const llama_ubatch * ubatch);
-    void set_outputs();
+    void set_outputs(const llm_graph_params & params);
 
     // try to update the existing graph result using the new graph parameters in order to reuse it
     // this can only be done if we determine that the resulting graph using the new graph parameters
@@ -734,10 +736,12 @@ public:
     ggml_tensor * t_embd_pooled = nullptr;
     ggml_tensor * t_h_nextn     = nullptr; // [n_embd, n_outputs] hidden state before final output norm
 
-    std::map<llama_seq_id, ggml_tensor*> t_sampled_logits;
-    std::map<llama_seq_id, ggml_tensor*> t_candidates;
-    std::map<llama_seq_id, ggml_tensor*> t_sampled;
-    std::map<llama_seq_id, ggml_tensor*> t_sampled_probs;
+    std::vector<ggml_tensor *> t_layer_inp;
+
+    std::map<llama_seq_id, ggml_tensor *> t_sampled_logits;
+    std::map<llama_seq_id, ggml_tensor *> t_candidates;
+    std::map<llama_seq_id, ggml_tensor *> t_sampled;
+    std::map<llama_seq_id, ggml_tensor *> t_sampled_probs;
 
     std::vector<llm_graph_input_ptr> inputs;
 

src/llama-hparams.h

@@ -45,6 +45,7 @@ struct llama_hparams {
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;
+    bool norm_before_residual = false;
 
     uint32_t n_ctx_train; // context size the model was trained on
     uint32_t n_embd;

src/llama-model-loader.cpp

@@ -394,6 +394,7 @@ namespace GGUFMeta {
 
     template bool llama_model_loader::get_arr<std::vector<std::string>>(enum llm_kv kid, std::vector<std::string> & result, bool required);
     template bool llama_model_loader::get_arr<std::array<int32_t, 512>>(enum llm_kv kid, std::array<int32_t, 512> & result, bool required);
+    template bool llama_model_loader::get_arr<std::vector<int32_t>>(enum llm_kv kid, std::vector<int32_t> & result, bool required);
 
     template<typename T>
     bool llama_model_loader::get_key(const std::string & key, T & result, bool required) {

src/llama-model.cpp

@@ -287,6 +287,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_qwen35moe(params);
         case LLM_ARCH_MISTRAL3:
             return new llama_model_mistral3(params);
+        case LLM_ARCH_EAGLE3:
+            return new llama_model_eagle3(params);
         case LLM_ARCH_MIMO2:
             return new llama_model_mimo2(params);
         case LLM_ARCH_KIMI_LINEAR:
@@ -2238,7 +2240,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     // TODO: move reranking logic here and generalize
     llm->build_dense_out(dense_2_out_layers, dense_2_out_layers_b, dense_3_out_layers);
 
-    llm->res->set_outputs();
+    llm->res->set_outputs(params);
 
     return llm->res->get_gf();
 }
@@ -2406,6 +2408,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
         case LLM_ARCH_MISTRAL3:
+        case LLM_ARCH_EAGLE3:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_LLAMA_EMBED:
         case LLM_ARCH_MAINCODER:
@@ -2600,8 +2603,9 @@ uint64_t llama_model_n_params(const llama_model * model) {
 
 bool llama_model_has_encoder(const llama_model * model) {
     switch (model->arch) {
-        case LLM_ARCH_T5:        return true;
-        case LLM_ARCH_T5ENCODER: return true;
+        case LLM_ARCH_T5:
+        case LLM_ARCH_T5ENCODER:
+        case LLM_ARCH_EAGLE3:    return true;
         default:                 return false;
     }
 }
@@ -2687,3 +2691,12 @@ void llama_model_base::create_tensor_qkv(llama_layer & layer, int bid,
         layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", bid), {n_embd_v_}, TENSOR_NOT_REQUIRED);
     }
 }
+
+const int32_t * llama_model_target_layer_ids(const struct llama_model * model) {
+    const auto & v = model->target_layer_ids;
+    return v.empty() ? nullptr : v.data();
+}
+
+uint32_t llama_model_target_layer_ids_n(const struct llama_model * model) {
+    return (uint32_t) model->target_layer_ids.size();
+}

src/llama-model.h

@@ -569,6 +569,13 @@ struct llama_model {
     struct ggml_tensor * per_layer_model_proj = nullptr;
     struct ggml_tensor * per_layer_proj_norm  = nullptr;
 
+    // eagle3
+    struct ggml_tensor * fc  = nullptr;  // feature fusion layer
+    struct ggml_tensor * d2t = nullptr;  // draft to target vocabulary mapping
+
+    // unified vector to store target-model extracted layer ids in eagle3, dflash, etc.
+    std::vector<int32_t> target_layer_ids;
+
     std::vector<llama_layer> layers;
 
     //Dense linear projections for SentenceTransformers models like embeddinggemma

src/llama-vocab.cpp

@@ -764,7 +764,7 @@ struct llm_tokenizer_wpm_session {
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) {
         // normalize and split by whitespace
-        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_lowercase());
+        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_opts());
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -809,11 +809,14 @@ struct llm_tokenizer_wpm_session {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    static std::vector<std::string> preprocess(const std::string & text, bool lowercase)  {
-        const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
+    static std::vector<std::string> preprocess(const std::string & text, const llama_vocab::normalizer_options & normalizer_opts)  {
+        std::vector<uint32_t> cpts = unicode_cpts_from_utf8(text);
+        if (normalizer_opts.strip_accents) {
+            cpts = unicode_cpts_normalize_nfd(cpts);
+        }
         std::vector<std::string> words(1, "");
 
-        for (const uint32_t cpt : cpts_nfd) {
+        for (const uint32_t cpt : cpts) {
             const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
@@ -828,7 +831,11 @@ struct llm_tokenizer_wpm_session {
                 continue;
             }
 
-            const std::string s = unicode_cpt_to_utf8(lowercase ? unicode_tolower(cpt) : cpt);
+            if (normalizer_opts.strip_accents && flags.is_accent_mark) {
+                continue;
+            }
+
+            const std::string s = unicode_cpt_to_utf8(normalizer_opts.lowercase ? unicode_tolower(cpt) : cpt);
             if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) {
                 if (words.back().size()) {  // finish previous word if any
                     words.emplace_back();
@@ -1692,7 +1699,7 @@ struct llm_tokenizer_whitespace_session : llm_tokenizer_bpe_session {
     llm_tokenizer_whitespace_session(const llama_vocab & vocab, const llm_tokenizer_bpe & tokenizer) : llm_tokenizer_bpe_session{vocab, tokenizer}, vocab{vocab} {}
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) override {
-        const bool lowercase = vocab.get_normalizer_lowercase();
+        const bool lowercase = vocab.get_normalizer_opts().lowercase;
 
         std::string segment;
         auto flush = [&]() {
@@ -1797,7 +1804,9 @@ struct llama_vocab::impl {
     bool remove_extra_whitespaces   = false;
     bool escape_whitespaces         = true;
     bool treat_whitespace_as_suffix = false;
-    bool normalizer_lowercase       = true; // Lowercase normalizer (tokenizer.json)
+
+    // BertNormalizer options
+    llama_vocab::normalizer_options normalizer_opts;
 
     std::unordered_map<std::string, llama_token> token_to_id;
     std::vector<token_data>                      id_to_token;
@@ -2172,7 +2181,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                     tokenizer_pre == "whitespace") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_WHITESPACE;
-                normalizer_lowercase = false;
+                normalizer_opts.lowercase = false;
             } else if (
                     tokenizer_pre == "refact") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT;
@@ -2532,8 +2541,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
-        // Lowercase normalizer flag (consulted by WPM / whitespace BPE)
-        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE, normalizer_lowercase, false);
+        // BertNormalizer options
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     normalizer_opts.lowercase,     false);
+        normalizer_opts.strip_accents = normalizer_opts.lowercase;
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, normalizer_opts.strip_accents, false);
 
         // suppress tokens
         {
@@ -3969,8 +3980,8 @@ bool llama_vocab::get_treat_whitespace_as_suffix() const {
     return pimpl->treat_whitespace_as_suffix;
 }
 
-bool llama_vocab::get_normalizer_lowercase() const {
-    return pimpl->normalizer_lowercase;
+const llama_vocab::normalizer_options & llama_vocab::get_normalizer_opts() const {
+    return pimpl->normalizer_opts;
 }
 
 const std::vector<llama_token> & llama_vocab::get_suppress_tokens() const {

src/llama-vocab.h

@@ -76,6 +76,12 @@ struct llama_vocab {
         llama_token_attr attr;
     };
 
+    struct normalizer_options {
+        bool lowercase     = true;
+        bool strip_accents = true;
+        // TODO: clean_text, handle_chinese_chars
+    };
+
     llama_vocab();
     ~llama_vocab();
 
@@ -141,7 +147,7 @@ struct llama_vocab {
     bool get_remove_extra_whitespaces  () const;
     bool get_escape_whitespaces        () const;
     bool get_treat_whitespace_as_suffix() const;
-    bool get_normalizer_lowercase      () const;
+    const normalizer_options & get_normalizer_opts() const;
 
     const std::vector<llama_token> & get_suppress_tokens() const;
 

src/models/eagle3.cpp

@@ -0,0 +1,323 @@
+#include "models.h"
+
+void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    if (!ml.get_arr(LLM_KV_TARGET_LAYERS, target_layer_ids, false)) {
+        throw std::runtime_error("EAGLE3 model requires 'extract_layers' in GGUF metadata");
+    }
+    if (target_layer_ids.size() != 3) {
+        throw std::runtime_error("EAGLE3 requires exactly 3 entries in 'extract_layers'");
+    }
+    LLAMA_LOG_INFO("%s: EAGLE3 extract_layers = [%d, %d, %d]\n", __func__,
+            target_layer_ids[0],
+            target_layer_ids[1],
+            target_layer_ids[2]);
+
+    uint32_t n_embd_tgt = 0;
+
+    ml.get_key(LLM_KV_TARGET_HIDDEN_SIZE, n_embd_tgt);
+    LLAMA_LOG_INFO("%s: EAGLE3 n_embd_tgt = %u (draft n_embd = %u)\n", __func__, n_embd_tgt, hparams.n_embd);
+
+    hparams.n_embd_inp_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
+
+    // eagle3 norm_before_residual (optional, default false)
+    // compatible with Readhat eagle3 speculator model
+    ml.get_key(LLM_KV_NORM_BEFORE_RESIDUAL, hparams.norm_before_residual, false);
+    if (hparams.norm_before_residual) {
+        LLAMA_LOG_INFO("%s: EAGLE3gnorm_before_residual = true\n", __func__);
+    }
+
+    type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_embd_inp = hparams.n_embd_inp();
+    const int64_t n_embd_attn_input = 2 * n_embd;
+
+    // Get vocab size from the d2t tensor in the GGUF file (optional - only needed if eagle3 has different vocab_size than target)
+    // d2t: draft to target vocabulary mapping
+    int64_t n_draft_vocab = n_vocab;  // Default: same as target vocab
+    const struct ggml_tensor * d2t_meta = ml->get_tensor_meta("d2t");
+    if (d2t_meta) {
+        n_draft_vocab = d2t_meta->ne[0]; // update draft vocab size
+        d2t = create_tensor(tn(LLM_TENSOR_D2T), {n_draft_vocab}, 0);
+        LLAMA_LOG_INFO("%s: EAGLE3 using d2t mapping (draft_vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);
+    } else {
+        d2t = nullptr; // no d2t, use default vocab size
+        LLAMA_LOG_INFO("%s: EAGLE3 without d2t - sharing same vocab_size with target (vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);
+    }
+
+    // Feature fusion layer: projects 3 target layers to draft hidden size
+    fc = create_tensor(tn(LLM_TENSOR_FC, "weight"), {n_embd_inp, n_embd}, 0);
+
+    // Output layer (uses draft vocab size)
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_draft_vocab}, TENSOR_NOT_REQUIRED);
+
+    // Token embeddings (optional - Llama 3.3 70B EAGLE3 has its own)
+    const struct ggml_tensor * tok_embd_meta = ml->get_tensor_meta(tn(LLM_TENSOR_TOKEN_EMBD, "weight").str().c_str());
+    if (tok_embd_meta) {
+        const int64_t n_target_vocab = tok_embd_meta->ne[1];
+        tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_target_vocab}, 0);
+        LLAMA_LOG_INFO("%s: EAGLE3 using its own token_embd (vocab = %lld)\n", __func__, (long long)n_target_vocab);
+    }
+
+    // Single decoder layer
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // input_layernorm: applied to token embeddings
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        // eagle3 specific: hidden_norm applied to fused target features
+        layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, 0);
+
+        // Attention takes input_embeds_normed + fused_target_normed as input
+        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd_attn_input, n_embd_head_k * n_head}, 0);
+        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd_attn_input, n_embd_k_gqa}, 0);
+        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd_attn_input, n_embd_v_gqa}, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+        // rope_freqs for llama3 rope scaling (optional - only if eagle3 config has rope_scaling)
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_eagle3::build_arch_graph(const llm_graph_params & params) const {
+    switch (params.gtype) {
+        case LLM_GRAPH_TYPE_ENCODER:
+            return std::make_unique<graph<true>>(*this, params);
+        case LLM_GRAPH_TYPE_DEFAULT:
+        case LLM_GRAPH_TYPE_DECODER:
+            return std::make_unique<graph<false>>(*this, params);
+        default:
+            GGML_ABORT("invalid graph type");
+    };
+}
+
+template <>
+ggml_tensor * llama_model_eagle3::graph<true>::build_inp_embd_enc() const {
+    ggml_tensor * cur = nullptr;
+
+    // Input: Target model features (3 layers concatenated: low, mid, high)
+    // Data will be provided via ubatch->embd in encode_eagle3_features()
+    auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp());
+    inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32,hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp_target->embd);
+
+    cur = inp_target->embd;
+    cb(cur, "inp_embd", -1);
+
+    res->add_input(std::move(inp_target));
+
+    return cur;
+}
+
+// eagle3 Encoder: processes target model features through feature fusion layer
+// Input: target_features e.g. [12288, n_tokens] from target model layers low, middle, high
+// Output: g_embeddings e.g. [4096, n_tokens] stored in context
+template <>
+llama_model_eagle3::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur = nullptr;
+
+    cur = build_inp_embd_enc();
+
+    // Feature fusion layer
+    cur = build_lora_mm(model.fc, cur);
+    cb(cur, "fc_out", -1);
+
+    // Output: g_embeddings e.g. [4096, n_tokens]
+    // store in t_h_nextn (same as MTP) so can be read via llama_get_embeddings_nextn(ctx_dft)
+    ggml_set_output(cur);
+    res->t_h_nextn = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// eagle3 Decoder: processes draft tokens using g_embeddings from encoder
+// Input: draft tokens + g_embeddings from encoder
+// Output: draft logits
+template <>
+llama_model_eagle3::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : 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_layer == 1);  // eagle3 has only one decoder layer
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // eagle3 Decoder receives:
+    // 1. Token embeddings (e.g.from eagle3's own tok_embd for Llama 3.3 70B, or target model for Llama 3.1 8B)
+    // 2. g_embeddings from encoder
+    auto * tok_embd = model.tok_embd;
+    if (model.tok_embd == nullptr) {
+        GGML_ASSERT(cparams.ctx_other != nullptr);
+        const auto * model_other = llama_get_model(cparams.ctx_other);
+
+        GGML_ASSERT(model_other->tok_embd != nullptr && "EAGLE3 decoder requires token embeddings (own or from target model)");
+        tok_embd = model_other->tok_embd;
+    }
+
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * inp_embd = ggml_get_rows(ctx0, tok_embd, inp->tokens);
+    cb(inp_embd, "inp_embd", -1);
+
+    ggml_tensor * inp_g = inp->embd;
+    cb(inp_g, "inp_g_embeddings", -1);
+
+    res->add_input(std::move(inp));
+
+    inpL = inp_g;
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
+
+    // Single decoder layer (il = 0)
+    const int il = 0;
+    {
+        // Apply input_layernorm to the token embeddings
+        ggml_tensor * embd_norm = build_norm(inp_embd,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(embd_norm, "embd_norm", il);
+
+        // Apply hidden_norm to inp_g
+        ggml_tensor * g_norm = build_norm(inp_g,
+                model.layers[il].attn_norm_2, NULL,
+                LLM_NORM_RMS, -1);
+        cb(g_norm, "g_norm", il);
+
+        // norm_before_residual: determines what goes into the residual connection (compatible with Readhat eagle3 speculator model)
+        // - false (default): use raw inp_g for residual
+        // - true: use normalized g_norm for residual
+        // inpL is the concatenated input (normalized inp_embd + normalized inp_g)
+        ggml_tensor * inpSA = hparams.norm_before_residual ? g_norm : inpL;
+
+        // Concatenate normalized inp_embd and normalized inp_g
+        cur = ggml_concat(ctx0, embd_norm, g_norm, il);
+        cb(cur, "concat_embd", il);
+
+        // Self-attention with concatenated input
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        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);
+
+        // rope freq factors, returns nullptr if not available
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+        // RoPE
+        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_rope", il);
+        cb(Kcur, "Kcur_rope", il);
+
+        cur = build_attn(inp_attn,
+                model.layers[il].wo, NULL, nullptr,
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+
+        // Add residual and update it
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // Apply FFN norm to the sum
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        // Output norm with residual
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "eagle3_prenorm", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // Output prenorm state (for next token's g_embeddings in autoregressive generation)
+    ggml_set_output(cur);
+    res->t_h_nextn = cur;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    // lm_head - projects to draft vocabulary
+    // if the draft has no own output projection, inherit the target model's lm_head
+    auto * output = model.output;
+    if (output == nullptr) {
+        GGML_ASSERT(cparams.ctx_other != nullptr);
+        const auto * model_other = llama_get_model(cparams.ctx_other);
+
+        GGML_ASSERT(model_other->output != nullptr && "EAGLE3 decoder requires an output projection (own or from target model)");
+        output = model_other->output;
+    }
+    cur = build_lora_mm(output, cur);
+
+    if (model.d2t) {
+        const int64_t n_draft_vocab = cur->ne[0];
+        const int64_t n_outputs     = cur->ne[1];
+        const int64_t n_vocab       = (int64_t) model.vocab.n_tokens();
+
+        GGML_ASSERT(model.d2t->type == GGML_TYPE_I64);
+        GGML_ASSERT(model.d2t->ne[0] == n_draft_vocab);
+
+        ggml_tensor * logits = ggml_fill(ctx0, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, n_vocab, n_outputs), -INFINITY);
+        cur = ggml_set_rows(ctx0, logits,
+                ggml_reshape_3d(ctx0, cur,       1,             n_draft_vocab, n_outputs),
+                ggml_reshape_3d(ctx0, model.d2t, n_draft_vocab, 1,             1));
+        cur = ggml_reshape_2d(ctx0, cur, n_vocab, n_outputs);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma4.cpp

@@ -210,6 +210,8 @@ llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_para
         const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
         const int   n_rot_l      = hparams.n_rot(il);
 
+        res->t_layer_inp[il] = inpL;
+
         // norm
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);

src/models/llama.cpp

@@ -124,6 +124,8 @@ llama_model_llama::graph<embed>::graph(const llama_model & model, const llm_grap
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         ggml_tensor * inpSA = inpL;
 
         // norm

src/models/models.h

@@ -1089,6 +1089,21 @@ struct llama_model_glm_dsa : public llama_model_base {
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
+struct llama_model_eagle3 : public llama_model_base {
+    llama_model_eagle3(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    template <bool is_enc>
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+
+        ggml_tensor * build_inp_embd_enc() const;
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
 
 struct llama_model_mistral4 : public llama_model_deepseek2 {
     llama_model_mistral4(const struct llama_model_params & params) : llama_model_deepseek2(params) {}

src/models/openai-moe.cpp

@@ -75,6 +75,8 @@ llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         const float freq_base_l  = model.get_rope_freq_base (cparams, il);
         const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 

src/models/qwen3.cpp

@@ -69,6 +69,8 @@ llama_model_qwen3::graph::graph(const llama_model & model, const llm_graph_param
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         ggml_tensor * inpSA = inpL;
 
         // norm

src/models/qwen35.cpp

@@ -173,7 +173,7 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
         }
 
         if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
-            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            cur   = ggml_get_rows(ctx0, cur,   inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
 

src/models/qwen3moe.cpp

@@ -78,6 +78,8 @@ llama_model_qwen3moe::graph::graph(const llama_model & model, const llm_graph_pa
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         ggml_tensor * inpSA = inpL;
 
         // norm

tests/test-backend-ops.cpp

@@ -8849,7 +8849,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     for (int v : { 0, 1, 2, 3 }) {
         for (int dim : { 0, 1, 2, 3, }) {
             test_cases.emplace_back(new test_concat(GGML_TYPE_F32, {11, 12, 13, 14}, 7, dim, v));
+            test_cases.emplace_back(new test_concat(GGML_TYPE_F16, {11, 12, 13, 14}, 7, dim, v));
+            test_cases.emplace_back(new test_concat(GGML_TYPE_BF16, {11, 12, 13, 14}, 7, dim, v));
+            test_cases.emplace_back(new test_concat(GGML_TYPE_I8, {11, 12, 13, 14}, 7, dim, v));
+            test_cases.emplace_back(new test_concat(GGML_TYPE_I16, {11, 12, 13, 14}, 7, dim, v));
             test_cases.emplace_back(new test_concat(GGML_TYPE_I32, {11, 12, 13, 14}, 7, dim, v));
+            test_cases.emplace_back(new test_concat(GGML_TYPE_I64, {11, 12, 13, 14}, 7, dim, v));
         }
     }
 

tests/test-llama-archs.cpp

@@ -450,6 +450,9 @@ static int save_models(const llm_arch target_arch, const size_t seed, const ggml
         if (arch == LLM_ARCH_GEMMA4 || arch == LLM_ARCH_GEMMA4_ASSISTANT) {
             continue; // FIXME: ISWA KV cache initialization needs more fixture params
         }
+        if (arch == LLM_ARCH_EAGLE3) {
+            continue;
+        }
         for (bool moe : {false, true}) {
             if (moe && !moe_implemented(arch)) {
                 continue;
@@ -553,6 +556,9 @@ static int test_backends(const llm_arch target_arch, const size_t seed, const gg
         if (arch == LLM_ARCH_GEMMA4 || arch == LLM_ARCH_GEMMA4_ASSISTANT) {
             continue; // FIXME: ISWA KV cache initialization needs more fixture params
         }
+        if (arch == LLM_ARCH_EAGLE3) {
+            continue;
+        }
 
         const bool encode = arch == LLM_ARCH_T5 || arch == LLM_ARCH_DREAM || arch == LLM_ARCH_LLADA || arch == LLM_ARCH_LLADA_MOE || arch == LLM_ARCH_RND1;
         for (bool moe : {false, true}) {

tools/mtmd/clip-graph.h

@@ -54,6 +54,10 @@ struct clip_graph {
     virtual ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const;
     // TODO: build_mm(w, b, x) to support bias
 
+    virtual bool support_batch() const {
+        return false;
+    }
+
     //
     // utility functions
     //

tools/mtmd/clip.cpp

@@ -171,6 +171,8 @@ struct clip_ctx {
     std::map<ggml_backend_dev_t, size_t> mem_usage;
     std::map<ggml_backend_dev_t, size_t> mem_compute;
 
+    bool support_batch = false;
+
     clip_ctx(clip_context_params & ctx_params) {
         flash_attn_type = ctx_params.flash_attn_type;
         no_alloc = ctx_params.no_alloc;
@@ -314,7 +316,7 @@ ggml_tensor * clip_graph::build_vit(
             std::function<ggml_tensor *(ggml_tensor *, const clip_layer &)> add_pos,
             const build_vit_opts & opts
         ) {
-    // batch dim: inp is [n_embd, n_pos] (B==1) or [n_embd, n_pos, B] (multi-tile encode)
+    // batch dim: inp is [n_embd, n_pos, B]
     const int64_t B = inp->ne[2];
 
     if (learned_pos_embd) {
@@ -862,7 +864,7 @@ ggml_tensor * clip_graph::build_patch_merge_permute(ggml_tensor * cur, int scale
     return cur;
 }
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
+static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
     const clip_image_f32 & img = *imgs.entries[0];
     std::unique_ptr<clip_graph> builder;
 
@@ -1025,7 +1027,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     // TODO [QWEN_VIDEO]: improve this in the future
     builder->n_batch = imgs.entries.size();
 
-    return builder->build();
+    return builder;
 }
 
 //
@@ -2819,7 +2821,7 @@ struct clip_model_loader {
         std::vector<support_info_op> ops;
     };
 
-    static void warmup(clip_ctx & ctx_clip) {
+    static clip_image_f32_batch get_dummy_batch(clip_ctx & ctx_clip) {
         // create a fake batch
         const auto & hparams = ctx_clip.model.hparams;
         clip_image_f32_batch batch;
@@ -2833,6 +2835,20 @@ struct clip_model_loader {
             LOG_INF("%s: warmup with audio size = %d\n", __func__, hparams.warmup_audio_size);
         }
         batch.entries.push_back(std::move(img));
+        return batch;
+    }
+
+    static void init_ctx(clip_ctx & ctx_clip) {
+        ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
+
+        // check batching support
+        auto batch = get_dummy_batch(ctx_clip);
+        auto builder = clip_get_graph_builder(&ctx_clip, batch);
+        ctx_clip.support_batch = builder->support_batch();
+    }
+
+    static void warmup(clip_ctx & ctx_clip) {
+        auto batch = get_dummy_batch(ctx_clip);
         warmup(ctx_clip, batch);
     }
 
@@ -2905,9 +2921,7 @@ struct clip_model_loader {
 
     // only initialize backend buffers, but do not allocate them yet
     static support_info_graph reserve_compute_meta(clip_ctx & ctx_clip, const clip_image_f32_batch & batch) {
-        ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
-
-        ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch);
+        ggml_cgraph * gf = clip_get_graph_builder(&ctx_clip, batch)->build();
         ggml_backend_sched_reserve(ctx_clip.sched.get(), gf);
 
         ctx_clip.mem_compute.clear();
@@ -3070,6 +3084,7 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
             ctx_vision = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_vision->model, CLIP_MODALITY_VISION);
             loader.load_tensors(*ctx_vision);
+            loader.init_ctx(*ctx_vision);
             if (ctx_params.warmup) {
                 loader.warmup(*ctx_vision);
             }
@@ -3083,6 +3098,7 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
             ctx_audio = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
             loader.load_tensors(*ctx_audio);
+            loader.init_ctx(*ctx_audio);
             if (ctx_params.warmup) {
                 loader.warmup(*ctx_audio);
             }
@@ -3484,25 +3500,22 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     return n_patches;
 }
 
-bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
+bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, std::vector<float> & out_vec) {
     clip_image_f32_batch imgs;
     clip_image_f32_ptr img_copy(clip_image_f32_init());
     *img_copy = *img;
     imgs.entries.push_back(std::move(img_copy));
 
-    return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
+    return clip_image_batch_encode(ctx, n_threads, &imgs, out_vec);
 }
 
-bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, float * vec) {
+bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, std::vector<float> & out_batch_embd) {
     const clip_image_f32_batch & imgs = *imgs_c_ptr;
     int n_batch_cur = imgs.entries.size();
 
-    // maximum supported batch size, usually == 2 for qwen-vl-based models
-    int n_batch_max = clip_model_n_batch_max(ctx);
-
-    // TODO @ngxson : implement batch size > 1 as a loop
-    //                we don't need true batching support because the cgraph will gonna be big anyway
-    if (n_batch_cur > n_batch_max) {
+    // [QWEN_VIDEO] for video models, the batch dimension is used as temporal dimension for merged frames
+    if (!ctx->support_batch && n_batch_cur > clip_model_n_temporal_merge(ctx)) {
+        LOG_ERR("%s: batch size %d exceeds maximum supported batch/temporal-merge size %d\n", __func__, n_batch_cur, clip_model_n_temporal_merge(ctx));
         return false;
     }
 
@@ -3513,7 +3526,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 
     // build the inference graph
     ggml_backend_sched_reset(ctx->sched.get());
-    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
+    ggml_cgraph * gf = clip_get_graph_builder(ctx, imgs)->build();
     ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);
 
     // set inputs
@@ -3582,6 +3595,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             const int n  = nx * ny;
 
             for (int b = 0; b < n_batch_cur; b++) {
+                LOG_DBG("%s: copying image %d/%d to input buffer (nx=%d, ny=%d)\n", __func__, b+1, n_batch_cur, nx, ny);
                 const auto & buf = imgs.entries[b]->get_ro_buf();
                 float * batch_entry = inp_raw.data() + b * (3*n);
                 for (int y = 0; y < ny; y++) {
@@ -4416,7 +4430,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     // the last node is the embedding tensor
     ggml_tensor * embeddings = ggml_graph_node(gf, -1);
 
-    // sanity check (only support batch size of 1 for now)
+    // sanity check (assuming that all images in batch have the same number of tokens, so we only check the first one)
     const int n_tokens_out = embeddings->ne[1];
     const int expected_n_tokens_out = clip_n_output_tokens(ctx, imgs.entries[0].get());
     if (n_tokens_out != expected_n_tokens_out) {
@@ -4424,16 +4438,26 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         GGML_ABORT("Invalid number of output tokens");
     }
 
-    // copy the embeddings to the location passed by the user
-    if (vec != nullptr) {
-        ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+    LOG_DBG("%s: output embedding shape [%d, %d, %d]\n", __func__,
+        (int)embeddings->ne[0], (int)embeddings->ne[1], (int)embeddings->ne[2]);
+
+    // copy output to user buffer if provided
+    // if output is empty, skip the copy
+    if (!out_batch_embd.empty()) {
+        if (out_batch_embd.size() != (size_t)ggml_nelements(embeddings)) {
+            LOG_ERR("%s: output buffer has %zu elements but expected %zu\n", __func__, out_batch_embd.size(), (size_t)ggml_nelements(embeddings));
+            GGML_ABORT("Output buffer size mismatch");
+        }
+        ggml_backend_tensor_get(embeddings, out_batch_embd.data(), 0, ggml_nbytes(embeddings));
+    } else {
+        LOG_WRN("%s: output buffer is empty, skipping copy\n", __func__);
     }
 
     // Debug: dump final embeddings if MTMD_DEBUG_EMBEDDINGS is set
     if (ctx->debug_output_embeddings) {
         const int64_t n_embd = embeddings->ne[0];
         const int64_t n_tokens = embeddings->ne[1];
-        std::vector<float> emb_data(n_embd * n_tokens);
+        std::vector<float> emb_data(ggml_nelements(embeddings));
         ggml_backend_tensor_get(embeddings, emb_data.data(), 0, ggml_nbytes(embeddings));
 
         LOG_INF("\n=== MTMD_DEBUG_EMBEDDINGS ===\n");
@@ -4570,7 +4594,14 @@ bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
     return ctx->model.modality == CLIP_MODALITY_AUDIO;
 }
 
-int clip_model_n_batch_max(const struct clip_ctx * ctx) {
+bool clip_support_batch(const struct clip_ctx * ctx) {
+    return ctx->support_batch;
+}
+
+// TODO @ngxson : this is no longer correct with mtmd_batch API
+// this was only meant to be used by qwen-vl-based models, to fuse 2 input images into one (qwen-vl video support)
+// this logic should be refactored in near future to distinctly handle "merge frames" and "batching"
+int clip_model_n_temporal_merge(const struct clip_ctx * ctx) {
     switch (ctx->proj_type()) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:

tools/mtmd/clip.h

@@ -97,8 +97,8 @@ size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int id
 size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
 struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
 
-bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
-bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, std::vector<float> & out_vec);
+bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, std::vector<float> & out_batch_embd);
 
 bool clip_is_llava(const struct clip_ctx * ctx);
 // note for contributor: this clip_is_(model) pattern is deprecated
@@ -107,7 +107,9 @@ bool clip_is_llava(const struct clip_ctx * ctx);
 bool clip_has_vision_encoder(const struct clip_ctx * ctx);
 bool clip_has_audio_encoder(const struct clip_ctx * ctx);
 
-int clip_model_n_batch_max(const struct clip_ctx * ctx);
+bool clip_support_batch(const struct clip_ctx * ctx);
+
+int clip_model_n_temporal_merge(const struct clip_ctx * ctx); // TODO @ngxson : remove, refactor this
 
 std::map<ggml_backend_dev_t, size_t> clip_get_mem_usage(const struct clip_ctx * ctx);
 

tools/mtmd/models/gemma4v.cpp

@@ -10,7 +10,7 @@ ggml_cgraph * clip_graph_gemma4v::build() {
     ggml_set_name(inp_raw, "inp_raw_scaled");
 
     ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
-    inp = ggml_reshape_2d(ctx0, inp, n_patches, n_embd);
+    inp = ggml_reshape_3d(ctx0, inp, n_patches, n_embd, n_batch);
     inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
     ggml_set_name(inp, "inp");
     // note: no patch bias
@@ -51,10 +51,11 @@ ggml_cgraph * clip_graph_gemma4v::build() {
         // first half
         ggml_tensor * first;
         {
-            first = ggml_view_3d(ctx0, cur,
-                n_dim/2, n_head, n_pos,
+            first = ggml_view_4d(ctx0, cur,
+                n_dim/2, n_head, n_pos, n_batch,
                 cur->nb[1],
                 cur->nb[2],
+                cur->nb[3],
                 0);
             first = ggml_rope_ext(
                 ctx0,
@@ -70,10 +71,11 @@ ggml_cgraph * clip_graph_gemma4v::build() {
         // second half
         ggml_tensor * second;
         {
-            second = ggml_view_3d(ctx0, cur,
-                n_dim/2, n_head, n_pos,
+            second = ggml_view_4d(ctx0, cur,
+                n_dim/2, n_head, n_pos, n_batch,
                 cur->nb[1],
                 cur->nb[2],
+                cur->nb[3],
                 n_dim/2 * ggml_element_size(cur));
             second = ggml_rope_ext(
                 ctx0,
@@ -103,14 +105,14 @@ ggml_cgraph * clip_graph_gemma4v::build() {
         const int kernel_size = hparams.n_merge;
         GGML_ASSERT(kernel_size > 0);
 
-        // [n_embd, n_patches] -> [n_patches_x, n_patches_y, n_embd, 1]
-        cur = ggml_cont_4d(ctx0, ggml_transpose(ctx0, cur), n_patches_x, n_patches_y, n_embd, 1);
+        // [n_embd, n_patches] -> [n_patches_x, n_patches_y, n_embd, n_batch]
+        cur = ggml_cont_4d(ctx0, ggml_transpose(ctx0, cur), n_patches_x, n_patches_y, n_embd, n_batch);
         cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG,
                            kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
         const int out_x = n_patches_x / kernel_size;
         const int out_y = n_patches_y / kernel_size;
-        // [out_x, out_y, n_embd, 1] -> [n_embd, out_x * out_y]
-        cur = ggml_reshape_3d(ctx0, cur, out_x * out_y, n_embd, 1);
+        // [out_x, out_y, n_embd, n_batch] -> [n_embd, out_x * out_y, n_batch]
+        cur = ggml_reshape_3d(ctx0, cur, out_x * out_y, n_embd, n_batch);
         cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
         cur = ggml_scale(ctx0, cur, sqrtf((float)n_embd));
         cb(cur, "pooled", -1);

tools/mtmd/models/models.h

@@ -16,6 +16,7 @@ struct clip_graph_gemma4v : clip_graph {
     clip_graph_gemma4v(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
     ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const override;
+    bool support_batch() const override { return true; }
 };
 
 struct clip_graph_gemma4uv : clip_graph {

tools/mtmd/mtmd-helper.h

@@ -67,8 +67,8 @@ MTMD_API void mtmd_helper_image_get_decoder_pos(const mtmd_image_tokens * image,
 
 // helper function that automatically:
 // 1. run llama_decode() on text chunks
-// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
-// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
+// 2. run mtmd_encode_chunk() on image chunks, then mtmd_get_output_embd() and then llama_decode()
+// if any of the mtmd_encode_chunk() or llama_decode() calls return non-zero, stop and forward the error
 // otherwise, returns 0 on success
 // this function is NOT thread-safe
 MTMD_API int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
@@ -157,13 +157,16 @@ MTMD_API int32_t mtmd_helper_video_read_next(mtmd_helper_video * ctx,
 } // extern "C"
 #endif
 
+#ifdef __cplusplus
+#include <set>
+#include <memory>
+
+namespace mtmd_helper {
+
 //
 // C++ wrappers
 //
 
-#ifdef __cplusplus
-namespace mtmd_helper {
-
 // video-related C++ wrappers
 struct mtmd_helper_video_deleter {
     void operator()(mtmd_helper_video * val) { mtmd_helper_video_free(val); }

tools/mtmd/mtmd.cpp

@@ -69,8 +69,8 @@ struct mtmd_bitmap {
         return data.size();
     }
 
-    bool can_batch_with(const mtmd_bitmap & other) const {
-        // [QWEN_VIDEO] can batch if both are images with same size
+    bool can_merge_with(const mtmd_bitmap & other) const {
+        // [QWEN_VIDEO] can (temporal) merge if both are images with same size
         return !is_audio && !other.is_audio && nx == other.nx && ny == other.ny;
     }
 
@@ -90,12 +90,24 @@ struct mtmd_image_tokens {
     uint32_t ny = 0; // number of tokens in y direction
     mtmd_pos_type pos = MTMD_POS_TYPE_NORMAL;
     uint32_t image_idx = 0; // 0-based position of this image among image chunks in the prompt(used by pos == MTMD_POS_TYPE_HUNYUANVL)
+    uint32_t n_temporal_merge = 1; // for qwen-vl style temporal merge
     uint32_t n_tokens() const {
         if (pos == MTMD_POS_TYPE_HUNYUANVL) {
             // [BOI] [row0 tokens + newline] ... [row(ny-1) tokens + newline] [EOI]
             return (nx + 1) * ny + 2;
         }
-        return nx * ny;
+        // [QWEN_VIDEO] this logic is quite ugly, it's mostly to make qwen-vl temporal merge work, can be improved in the future
+        if (batch_f32.entries.size() == 1 || n_temporal_merge == 1) {
+            return nx * ny;
+        }
+        uint32_t nz = batch_f32.entries.size();
+        // TODO: simplify this by repeating the last frame until it fits the temporal merge
+        if (nz % n_temporal_merge != 0) {
+            nz = nz / n_temporal_merge + 1;
+        } else {
+            nz = nz / n_temporal_merge;
+        }
+        return nx * ny * nz;
     }
     clip_image_f32_batch batch_f32; // preprocessed image patches
     std::string id; // optional user-defined ID, useful for KV cache tracking
@@ -110,12 +122,17 @@ struct mtmd_image_tokens {
         return false;
     }
 
+    bool can_batch_with(const mtmd_image_tokens & other) {
+        return nx == other.nx && ny == other.ny && pos == other.pos;
+    }
+
     mtmd_image_tokens clone() {
         return mtmd_image_tokens{
             nx,
             ny,
             pos,
             image_idx,
+            n_temporal_merge,
             batch_f32.clone(),
             id
         };
@@ -153,12 +170,49 @@ struct mtmd_input_chunk {
     std::vector<llama_token> tokens_text;
     mtmd_image_tokens_ptr tokens_image;
     mtmd_audio_tokens_ptr tokens_audio;
+
+    bool can_batch_with(const mtmd_input_chunk & other) const {
+        if (type != other.type) {
+            return false;
+        }
+
+        if (tokens_image && other.tokens_image) {
+            return tokens_image->can_batch_with(*other.tokens_image);
+        }
+
+        // TODO: allow batching audio chunks of the same size
+
+        return false;
+    }
+
+    bool is_placeholder() const {
+        if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            return tokens_image && tokens_image->is_placeholder();
+        } else if (type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+            return tokens_audio && tokens_audio->is_placeholder();
+        }
+        return false;
+    }
 };
 
 struct mtmd_input_chunks {
     std::vector<mtmd_input_chunk> entries;
 };
 
+struct mtmd_batch {
+    mtmd_context * ctx;
+    std::vector<const mtmd_input_chunk *> entries;
+    std::vector<float> output_embd; // aggregated output embedding for the whole batch
+    mtmd_batch(mtmd_context * ctx): ctx(ctx) {}
+    int32_t n_tokens() const {
+        int32_t n = 0;
+        for (const auto * chunk : entries) {
+            n += mtmd_input_chunk_get_n_tokens(chunk);
+        }
+        return n;
+    }
+};
+
 // slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
 // models not having it (llava-1.6) will process embeddings without any special tokens in-between
 enum mtmd_slice_tmpl {
@@ -197,6 +251,7 @@ mtmd_context_params mtmd_context_params_default() {
         /* image_max_tokens  */ -1,
         /* cb_eval           */ nullptr,
         /* cb_eval_user_data */ nullptr,
+        /* batch_max_tokens  */ 1024,
     };
     return params;
 }
@@ -204,7 +259,7 @@ mtmd_context_params mtmd_context_params_default() {
 struct mtmd_context {
     struct clip_ctx * ctx_v; // vision
     struct clip_ctx * ctx_a; // audio
-    std::vector<float> image_embd_v; // image embedding vector
+    std::vector<float> out_embd; // image embedding vector
 
     bool print_timings;
     int n_threads;
@@ -239,17 +294,21 @@ struct mtmd_context {
     std::unique_ptr<mtmd_audio_preprocessor> audio_preproc;
     std::unique_ptr<mtmd_image_preprocessor> image_preproc;
 
+    // batching
+    int32_t batch_max_tokens;
+
     // TODO @ngxson : add timings
 
     mtmd_context(const char * mmproj_fname,
                    const llama_model * text_model,
                    const mtmd_context_params & ctx_params,
                    bool no_alloc = false) :
-        print_timings(ctx_params.print_timings),
-        n_threads    (ctx_params.n_threads),
-        media_marker (ctx_params.media_marker),
-        n_embd_text  (text_model ? llama_model_n_embd_inp(text_model) : -1),
-        vocab        (text_model ? llama_model_get_vocab(text_model) : nullptr)
+        print_timings   (ctx_params.print_timings),
+        n_threads       (ctx_params.n_threads),
+        media_marker    (ctx_params.media_marker),
+        n_embd_text     (text_model ? llama_model_n_embd_inp(text_model) : -1),
+        vocab           (text_model ? llama_model_get_vocab(text_model) : nullptr),
+        batch_max_tokens(ctx_params.batch_max_tokens)
     {
         if (ctx_params.image_marker != nullptr) {
             throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
@@ -680,6 +739,16 @@ struct mtmd_context {
         return ctx_a ? clip_get_projector_type(ctx_a) : PROJECTOR_TYPE_UNKNOWN;
     }
 
+    int64_t n_embd_out() const {
+        if (ctx_v) {
+            return clip_n_mmproj_embd(ctx_v);
+        } else if (ctx_a) {
+            return clip_n_mmproj_embd(ctx_a);
+        } else {
+            throw std::runtime_error("no CLIP model loaded");
+        }
+    }
+
     ~mtmd_context() {
         clip_free(ctx_a);
         clip_free(ctx_v);
@@ -845,7 +914,7 @@ struct mtmd_tokenizer {
         // [QWEN_VIDEO] handle frame merging for models that support it (i.e. qwen-vl)
         int n_merge_frames = 1;
         if (ctx->ctx_v) {
-            n_merge_frames = clip_model_n_batch_max(ctx->ctx_v);
+            n_merge_frames = clip_model_n_temporal_merge(ctx->ctx_v);
             GGML_ASSERT(n_merge_frames <= 2 && "we only support merging maximum 2 images for now; open an issue if this model supports merging more");
         }
 
@@ -860,7 +929,7 @@ struct mtmd_tokenizer {
                 if (i + 1 < parts.size() && parts[i + 1].bitmap != nullptr) {
                     const mtmd_bitmap * bm_a = parts[i].bitmap;
                     const mtmd_bitmap * bm_b = parts[i + 1].bitmap;
-                    if (bm_a->can_batch_with(*bm_b)) {
+                    if (bm_a->can_merge_with(*bm_b)) {
                         LOG_DBG("%s: merging 2 frames at part index %zu and %zu\n", __func__, i, i + 1);
                         merged_bitmaps.push_back({bm_a, bm_b});
                         parts.erase(parts.begin() + i + 1); // collapse the second bitmap part
@@ -1103,13 +1172,17 @@ struct mtmd_tokenizer {
                 size_t n_tokens = 0;
                 for (const auto & e : batch_f32.entries) {
                     n_tokens += clip_n_output_tokens(ctx->ctx_v, e.get());
-                    if (clip_model_n_batch_max(ctx->ctx_v) == 2) {
+                    if (clip_model_n_temporal_merge(ctx->ctx_v) == 2) {
                         // [QWEN_VIDEO] pair input is merged to the same embd, so only count as one image
                         break;
                     }
                 }
 
                 mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+
+                // [QWEN_VIDEO] improve this in the future
+                image_tokens->n_temporal_merge = clip_model_n_temporal_merge(ctx->ctx_v);
+
                 if (mtmd_decode_use_mrope(ctx)) {
                     // for Qwen2VL, we need this information for M-RoPE decoding positions
                     image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
@@ -1327,60 +1400,18 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
     }
 }
 
-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) {
-        if (!ctx->ctx_v) {
-            LOG_ERR("%s: model does not support vision input\n", __func__);
-            return 1;
-        }
-        if (chunk->tokens_image == nullptr) {
-            LOG_ERR("%s: image tokens are null\n", __func__);
-            return 1;
-        }
-        if (chunk->tokens_image->is_placeholder()) {
-            LOG_ERR("%s: image tokens batch is placeholder\n", __func__);
-            return 1;
-        }
-        return mtmd_encode(ctx, chunk->tokens_image.get());
-    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
-        if (!ctx->ctx_a) {
-            LOG_ERR("%s: model does not support audio input\n", __func__);
-            return 1;
-        }
-        if (chunk->tokens_audio == nullptr) {
-            LOG_ERR("%s: audio tokens are null\n", __func__);
-            return 1;
-        }
-        if (chunk->tokens_audio->is_placeholder()) {
-            LOG_ERR("%s: audio tokens batch is placeholder\n", __func__);
-            return 1;
-        }
-        int n_mmproj_embd = ctx->n_embd_text;
-        ctx->image_embd_v.resize(chunk->tokens_audio->n_tokens * n_mmproj_embd);
-        bool ok = clip_image_batch_encode(
-            ctx->ctx_a,
-            ctx->n_threads,
-            &chunk->tokens_audio->batch_f32,
-            ctx->image_embd_v.data());
-        return ok ? 0 : 1;
-    }
-
-    LOG_ERR("%s: unknown chunk type %d\n", __func__, (int)chunk->type);
-    return 1;
-}
-
-int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
+static int32_t mtmd_encode_impl(mtmd_context * ctx, const mtmd_image_tokens * image_tokens, std::vector<float> & out_embd) {
     clip_ctx * ctx_clip = ctx->ctx_v;
     if (!ctx_clip) {
         LOG_ERR("%s: this API does not support non-vision input, please use mtmd_encode_chunk instead\n", __func__);
         return 1;
     }
     auto proj_type = clip_get_projector_type(ctx_clip);
-    int n_mmproj_embd = clip_n_mmproj_embd(ctx_clip);
-    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
+
+    int n_embd_out = ctx->n_embd_out();
+    auto n_tokens_out = image_tokens->n_tokens();
+    out_embd.resize((size_t)n_embd_out * n_tokens_out);
+
     bool ok = false;
 
     if (clip_is_llava(ctx_clip)
@@ -1400,12 +1431,19 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
                 return 1;
             }
             int n_tokens_per_image = clip_n_output_tokens(ctx_clip, entries[i].get());
-            ok = clip_image_encode(
+            std::vector<float> tmp_embd((size_t)n_tokens_per_image * n_embd_out);
+            bool ok_i = clip_image_encode(
                 ctx_clip,
                 ctx->n_threads,
                 entries[i].get(),
-                ctx->image_embd_v.data() + offset);
-            offset += static_cast<size_t>(n_mmproj_embd) * n_tokens_per_image;
+                tmp_embd);
+            if (!ok_i) {
+                LOG_ERR("%s: failed to encode image %zu\n", __func__, i);
+                return 1;
+            }
+            ok = true;
+            std::copy(tmp_embd.begin(), tmp_embd.end(), out_embd.begin() + offset);
+            offset += static_cast<size_t>(n_embd_out) * n_tokens_per_image;
         }
     } else {
         if (image_tokens->is_placeholder()) {
@@ -1416,14 +1454,206 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
             ctx_clip,
             ctx->n_threads,
             &image_tokens->batch_f32,
-            ctx->image_embd_v.data());
+            out_embd);
     }
 
     return ok ? 0 : 1;
 }
 
+static int32_t mtmd_encode_chunk_impl(mtmd_context * ctx, const mtmd_input_chunk * chunk, std::vector<float> & out_embd) {
+    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) {
+        if (!ctx->ctx_v) {
+            LOG_ERR("%s: model does not support vision input\n", __func__);
+            return 1;
+        }
+        if (chunk->tokens_image == nullptr) {
+            LOG_ERR("%s: image tokens are null\n", __func__);
+            return 1;
+        }
+        if (chunk->tokens_image->is_placeholder()) {
+            LOG_ERR("%s: image tokens batch is placeholder\n", __func__);
+            return 1;
+        }
+        return mtmd_encode_impl(ctx, chunk->tokens_image.get(), out_embd);
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        if (!ctx->ctx_a) {
+            LOG_ERR("%s: model does not support audio input\n", __func__);
+            return 1;
+        }
+        if (chunk->tokens_audio == nullptr) {
+            LOG_ERR("%s: audio tokens are null\n", __func__);
+            return 1;
+        }
+        if (chunk->tokens_audio->is_placeholder()) {
+            LOG_ERR("%s: audio tokens batch is placeholder\n", __func__);
+            return 1;
+        }
+        int n_mmproj_embd = ctx->n_embd_out();
+        out_embd.resize((size_t)chunk->tokens_audio->n_tokens * n_mmproj_embd);
+        bool ok = clip_image_batch_encode(
+            ctx->ctx_a,
+            ctx->n_threads,
+            &chunk->tokens_audio->batch_f32,
+            out_embd);
+        return ok ? 0 : 1;
+    }
+
+    LOG_ERR("%s: unknown chunk type %d\n", __func__, (int)chunk->type);
+    return 1;
+}
+
+int32_t mtmd_encode_chunk(mtmd_context * ctx, const mtmd_input_chunk * chunk) {
+    // this is the non-batching version
+    try {
+        return mtmd_encode_chunk_impl(ctx, chunk, ctx->out_embd);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return 1;
+    }
+}
+
+int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
+    try {
+        return mtmd_encode_impl(ctx, image_tokens, ctx->out_embd);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return 1;
+    }
+}
+
 float * mtmd_get_output_embd(mtmd_context * ctx) {
-    return ctx->image_embd_v.data();
+    return ctx->out_embd.data();
+}
+
+mtmd_batch * mtmd_batch_init(mtmd_context * ctx) {
+    return new mtmd_batch(ctx);
+}
+
+void mtmd_batch_free(mtmd_batch * batch) {
+    if (batch) {
+        delete batch;
+    }
+}
+
+int32_t mtmd_batch_add_chunk(mtmd_batch * batch, const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_ERR("%s: text chunk is not supported in batch\n", __func__);
+        return 1;
+    }
+
+    auto * ctx = batch->ctx->get_clip_ctx(chunk);
+    if (!ctx) {
+        LOG_ERR("%s: model does not support input chunk type %d\n", __func__, (int)chunk->type);
+        return 1;
+    }
+
+    if (batch->entries.empty()) {
+        // batch must have at least one chunk
+        batch->entries.push_back(chunk);
+        return 0;
+    }
+
+    if (!clip_support_batch(ctx)) {
+        // if no batching support, batch can only have one single chunk
+        return 2; // "batch too large" error code
+    }
+
+    int32_t new_n_tokens = batch->n_tokens() + (int32_t)mtmd_input_chunk_get_n_tokens(chunk);
+    if (new_n_tokens > batch->ctx->batch_max_tokens) {
+        return 2; // "batch too large" error code
+    }
+
+    auto & first_chunk = batch->entries[0];
+    if (first_chunk->can_batch_with(*chunk)) {
+        batch->entries.push_back(chunk);
+        return 0;
+    }
+
+    return 3; // "cannot batch" error code
+}
+
+static int32_t mtmd_batch_encode_impl(mtmd_batch * batch) {
+    if (batch->entries.empty()) {
+        LOG_ERR("%s: batch is empty\n", __func__);
+        return 1;
+    }
+    for (const auto * chunk : batch->entries) {
+        if (chunk->is_placeholder()) {
+            LOG_ERR("%s: chunk is placeholder\n", __func__);
+            return 1;
+        }
+    }
+
+    // represent the whole batch as one single chunk
+    mtmd::input_chunk_ptr batch_chunk(mtmd_input_chunk_copy(batch->entries[0]));
+    if (batch_chunk->tokens_image) {
+        auto & b0_f32 = batch_chunk->tokens_image->batch_f32;
+        // copy all entries from other chunks into the first chunk's batch_f32
+        // note: skip first entry because it's already in batch_chunk
+        for (size_t ic = 1; ic < batch->entries.size(); ic++) {
+            auto & chunk = batch->entries[ic];
+            GGML_ASSERT(chunk->tokens_image);
+            auto b1_f32 = chunk->tokens_image->batch_f32.clone();
+            for (size_t i = 0; i < b1_f32.entries.size(); i++) {
+                b0_f32.entries.push_back(std::move(b1_f32.entries[i]));
+            }
+        }
+    } else if (batch_chunk->tokens_audio) {
+        auto & b0_f32 = batch_chunk->tokens_audio->batch_f32;
+        // copy all entries from other chunks into the first chunk's batch_f32
+        // note: skip first entry because it's already in batch_chunk
+        for (size_t ic = 1; ic < batch->entries.size(); ic++) {
+            auto & chunk = batch->entries[ic];
+            GGML_ASSERT(chunk->tokens_audio);
+            auto b1_f32 = chunk->tokens_audio->batch_f32.clone();
+            for (size_t i = 0; i < b1_f32.entries.size(); i++) {
+                b0_f32.entries.push_back(std::move(b1_f32.entries[i]));
+            }
+        }
+    } else {
+        LOG_ERR("%s: unsupported chunk type\n", __func__);
+        return 1;
+    }
+
+    LOG_DBG("%s: encoding batch with %zu entries and total %zu tokens\n",
+            __func__, batch->entries.size(), mtmd_input_chunk_get_n_tokens(batch_chunk.get()));
+    int32_t res = mtmd_encode_chunk_impl(
+        batch->ctx,
+        batch_chunk.get(),
+        batch->output_embd);
+    return res;
+}
+
+int32_t mtmd_batch_encode(mtmd_batch * batch) {
+    try {
+        return mtmd_batch_encode_impl(batch);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return 1;
+    }
+}
+
+float * mtmd_batch_get_output_embd(mtmd_batch * batch, const mtmd_input_chunk * chunk) {
+    if (batch->output_embd.empty()) {
+        LOG_ERR("%s: batch has not been encoded yet\n", __func__);
+        return nullptr;
+    }
+    size_t offset = 0;
+    const size_t n_embd = batch->ctx->n_embd_out();
+    for (const auto * c : batch->entries) {
+        size_t offset_prev = offset;
+        size_t n_tokens = mtmd_input_chunk_get_n_tokens(c);
+        offset += n_tokens * n_embd;
+        GGML_ASSERT(offset_prev <  batch->output_embd.size());
+        GGML_ASSERT(offset      <= batch->output_embd.size());
+        if (c == chunk) {
+            return &batch->output_embd.data()[offset_prev];
+        }
+    }
+    return nullptr; // not found
 }
 
 bool mtmd_decode_use_non_causal(const mtmd_context * ctx, const mtmd_input_chunk * chunk) {
@@ -1801,7 +2031,7 @@ static void mtmd_debug_encode_impl(mtmd_context * ctx, clip_ctx * ctx_clip, clip
         ctx_clip,
         ctx->n_threads,
         &image,
-        embd_output.data());
+        embd_output);
     if (!ok) {
         LOG_ERR("%s: failed to encode image\n", __func__);
     }

tools/mtmd/mtmd.h

@@ -63,6 +63,7 @@ struct mtmd_bitmap;
 struct mtmd_image_tokens;
 struct mtmd_input_chunk;
 struct mtmd_input_chunks;
+struct mtmd_batch;
 
 struct mtmd_input_text {
     const char * text;
@@ -80,6 +81,7 @@ typedef struct mtmd_image_tokens mtmd_image_tokens;
 typedef struct mtmd_input_chunk  mtmd_input_chunk;
 typedef struct mtmd_input_chunks mtmd_input_chunks;
 typedef struct mtmd_input_text   mtmd_input_text;
+typedef struct mtmd_batch        mtmd_batch;
 
 struct mtmd_context_params {
     bool use_gpu;
@@ -97,6 +99,11 @@ struct mtmd_context_params {
     // callback function passed over to mtmd proper
     ggml_backend_sched_eval_callback cb_eval;
     void * cb_eval_user_data;
+
+    // batching params
+    int32_t batch_max_tokens; // maximum number of output tokens in a batch
+                              // (note: this is not a hard-limit, the first image will always be added even if it exceeds this limit)
+                              // (default: 1024)
 };
 
 MTMD_API const char * mtmd_default_marker(void);
@@ -265,12 +272,12 @@ MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                                const mtmd_bitmap ** bitmaps,
                                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);
+DEPRECATED(MTMD_API int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens),
+           "use mtmd_encode_chunk() instead");
 
+// text chunk will be ignored silently, only media chunk will be encoded
 // returns 0 on success
+// returns 1 on generic error
 MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
                                    const mtmd_input_chunk * chunk);
 
@@ -279,6 +286,26 @@ MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
 // llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
+
+// batch encoding API
+// chunks are not owned by the batch, they will not be freed by mtmd_batch_free()
+// batch is valid for a given context, cannot be shared across contexts
+MTMD_API mtmd_batch * mtmd_batch_init(mtmd_context * ctx);
+MTMD_API void         mtmd_batch_free(mtmd_batch * batch);
+
+// only media chunks are allowed, text chunks will be rejected
+// returns 0 on success
+// returns 1 on generic error
+// returns 2 if the batch is too large (chunk won't be added)
+// returns 3 if it cannot be batched with the existing chunks in the batch
+MTMD_API int32_t mtmd_batch_add_chunk(mtmd_batch * batch, const mtmd_input_chunk * chunk);
+
+// returns 0 on success
+// returns 1 on generic error
+MTMD_API int32_t mtmd_batch_encode(mtmd_batch * batch);
+MTMD_API float * mtmd_batch_get_output_embd(mtmd_batch * batch, const mtmd_input_chunk * chunk);
+
+
 // Set callback for all future logging events.
 // If this is not called, or NULL is supplied, everything is output on stderr.
 MTMD_API void mtmd_log_set(ggml_log_callback log_callback, void * user_data);
@@ -336,6 +363,11 @@ struct mtmd_input_chunk_deleter {
 };
 using input_chunk_ptr = std::unique_ptr<mtmd_input_chunk, mtmd_input_chunk_deleter>;
 
+struct mtmd_batch_deleter {
+    void operator()(mtmd_batch * val) { mtmd_batch_free(val); }
+};
+using batch_ptr = std::unique_ptr<mtmd_batch, mtmd_batch_deleter>;
+
 struct bitmap {
     bitmap_ptr ptr;
     bitmap() : ptr(nullptr) {}

tools/server/server-common.cpp

@@ -344,6 +344,14 @@ const mtmd::input_chunk_ptr & server_tokens::find_chunk(size_t idx) const {
     throw std::runtime_error("Chunk not found");
 }
 
+std::pair<const mtmd::input_chunk_ptr *, size_t> server_tokens::find_next_media_chunk(size_t idx) const {
+    auto it = map_idx_to_media.upper_bound(idx);
+    if (it != map_idx_to_media.end()) {
+        return { &it->second, it->first };
+    }
+    return { nullptr, 0 };
+}
+
 void server_tokens::push_back(llama_token tok) {
     if (tok == LLAMA_TOKEN_NULL) {
         throw std::runtime_error("Invalid token");
@@ -1126,9 +1134,9 @@ json oaicompat_chat_params_parse(
 
     // Reasoning budget: pass parameters through to sampling layer
     {
-        int reasoning_budget = opt.reasoning_budget;
-        if (reasoning_budget == -1 && body.contains("thinking_budget_tokens")) {
-            reasoning_budget = json_value(body, "thinking_budget_tokens", -1);
+        int reasoning_budget = json_value(body, "thinking_budget_tokens", -1);
+        if (reasoning_budget == -1) {
+            reasoning_budget = opt.reasoning_budget;
         }
 
         if (!chat_params.thinking_end_tag.empty()) {

tools/server/server-common.h

@@ -180,6 +180,10 @@ public:
 
     const mtmd::input_chunk_ptr & find_chunk(size_t idx) const;
 
+    // find next media chunk after idx
+    // returns a pair of pointer to the chunk (nullptr if not found) and its start index in tokens
+    std::pair<const mtmd::input_chunk_ptr *, size_t> find_next_media_chunk(size_t idx) const;
+
     void push_back(llama_token tok);
 
     // will create a copy of the chunk if it contains non-text data

tools/server/server-context.cpp

@@ -80,6 +80,8 @@ struct server_slot {
 
     // multimodal
     mtmd_context * mctx = nullptr;
+    mtmd::batch_ptr mbatch = nullptr;
+    std::array<llama_context *, 2> mtgt = {nullptr, nullptr}; // [0] for main context, [1] for optional draft context
 
     // speculative decoding
     common_speculative * spec;
@@ -239,6 +241,18 @@ struct server_slot {
 
         // clear alora start
         alora_invocation_start = -1;
+
+        // clear multimodal state
+        mbatch.reset();
+        mtgt[0] = ctx_tgt;
+        mtgt[1] = nullptr;
+        if (ctx_dft && llama_get_ctx_other(ctx_dft) != ctx_tgt) {
+            // TODO: in the future, figure out how to infuse target embeddings to the images
+            //       for now, we re-decode the same chunk in both ctx_tgt and ctx_dft
+            //       maybe we simply need to call `common_speculative_process()` ?
+            //       [TAG_MTMD_DRAFT_PROCESSING]
+            mtgt[1] = ctx_dft;
+        }
     }
 
     void init_sampler() const {
@@ -578,6 +592,87 @@ struct server_slot {
         other.prompt = prompt.clone();
         other.init_sampler();
     }
+
+    // returns 0 on success
+    // caller need to update prompt.tokens after a successful call to keep track of the processing progress
+    int process_mtmd_chunk(size_t idx, size_t & n_tokens_out) {
+        GGML_ASSERT(mctx);
+        const auto & input_tokens = task->tokens;
+        auto & chunk = input_tokens.find_chunk(idx);
+        int32_t res = 0;
+
+        auto try_decode = [&]() -> int32_t {
+            if (mbatch) {
+                float * embd = mtmd_batch_get_output_embd(mbatch.get(), chunk.get());
+                if (embd) {
+                    for (auto * lctx : mtgt) {
+                        if (lctx == nullptr) {
+                            continue;
+                        }
+                        llama_pos new_n_past; // unused for now
+                        res = mtmd_helper_decode_image_chunk(
+                            mctx,
+                            lctx,
+                            chunk.get(),
+                            embd,
+                            prompt.tokens.pos_next(),
+                            id,
+                            llama_n_batch(lctx),
+                            &new_n_past
+                        );
+                        if (res != 0) {
+                            SLT_ERR(*this, "failed to decode mtmd chunk, idx = %zu, res = %d\n", idx, res);
+                            return -1;
+                        }
+                    }
+                    n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());
+                    return 0; // success
+                }
+            }
+            return 1; // (non-error) need to create & encode batch
+        };
+
+        // if the batch is already exist, try searching & encode
+        res = try_decode();
+        if (res == 0) {
+            return 0;
+        } else if (res < 0) {
+            // fatal error
+            return res;
+        }
+
+        // otherwise, the batch is either uninitialized or is used up
+        // we need to create & encode a new batch
+        mbatch.reset(mtmd_batch_init(mctx));
+        res = mtmd_batch_add_chunk(mbatch.get(), chunk.get());
+        GGML_ASSERT(res == 0); // we should never have an empty batch
+
+        // try batching as much as possible
+        int n_added = 1;
+        size_t idx_cur = idx;
+        while (res == 0) {
+            auto [next_chunk, next_idx] = input_tokens.find_next_media_chunk(idx_cur);
+            if (next_chunk == nullptr) {
+                break;
+            }
+            res = mtmd_batch_add_chunk(mbatch.get(), next_chunk->get());
+            n_added += (res == 0 ? 1 : 0);
+            idx_cur = next_idx;
+            SLT_DBG(*this, "try adding media chunk idx = %zu to batch, res = %d\n", next_idx, res);
+            // if res != 0, batch is full or chunk is not compatible -> this loop breaks
+        }
+
+        // TODO @ngxson : move this log line to debug when it become more stable
+        SLT_INF(*this, "encoding mtmd batch from idx = %zu, n_chunks = %d\n", idx, n_added);
+
+        res = mtmd_batch_encode(mbatch.get());
+        if (res != 0) {
+            SLT_ERR(*this, "failed to encode mtmd batch for chunk idx = %zu, res = %d\n", idx, res);
+            return -1;
+        }
+
+        return try_decode();
+    }
 };
 
 
@@ -781,6 +876,7 @@ private:
             mparams.warmup           = params_base.warmup;
             mparams.image_min_tokens = params_base.image_min_tokens;
             mparams.image_max_tokens = params_base.image_max_tokens;
+            mparams.batch_max_tokens = params_base.mtmd_batch_max_tokens;
             mparams.media_marker     = get_media_marker();
         }
 
@@ -866,10 +962,7 @@ private:
                     }
 
                     for (size_t j = 0; j < devs.size(); ++j) {
-                        const size_t bytes =
-                            (measure_model_bytes ? dmd[j].mb.model : 0) +
-                            dmd[j].mb.context +
-                            dmd[j].mb.compute;
+                        const size_t bytes = (measure_model_bytes ? dmd[j].model : 0) + dmd[j].context + dmd[j].compute;
                         total += bytes;
                         for (size_t i = 0; i < tgt_devices.size(); i++) {
                             if (tgt_devices[i] == devs[j]) {
@@ -2046,6 +2139,9 @@ private:
 
         auto & cur = slot.prompt.checkpoints.emplace_back();
 
+        // [TAG_CHECKPOINTS_FIX_POS_MIN]
+        // TODO: here we incorrectly deterimne that the saved checkpoint data covers the [pos_min, pos_max] range
+        //       this is not true for SWA models: https://github.com/ggml-org/llama.cpp/pull/24411#issuecomment-4677983225
         cur.update_pos(slot.prompt.n_tokens() - n_tokens_cur, pos_min, pos_max);
 
         cur.update_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
@@ -2860,6 +2956,10 @@ private:
                                             // guarantee that a checkpoint will result in at least one token being processed [TAG_PROMPT_LOGITS]
                                             LOG_INF("slot %12.*s: id %2d | task %d | Checking checkpoint with [%d, %d] against %d...\n", 12,
                                                 func_name, (slot).id, ((slot).task ? (slot).task->id : -1), cur.pos_min, cur.pos_max, pos_min_thold);
+                                            // workaround for [TAG_CHECKPOINTS_FIX_POS_MIN]
+                                            if (cur.pos_max > pos_next) {
+                                                return false;
+                                            }
                                             return cur.pos_min < pos_min_thold || cur.pos_min == 0;
                                         }
                                     );
@@ -2921,7 +3021,7 @@ private:
                                 send_partial_response(slot, {}, false, true);
                             }
                         }
-                    }
+                    } // end of SLOT_STATE_STARTED
 
                     if (!slot.can_split()) {
                         // cannot fit the prompt in the current batch - will try next iter
@@ -2976,10 +3076,18 @@ private:
                     bool has_mtmd = false;
 
                     // check if we should process the image
-                    while (slot.prompt.n_tokens() < slot.task->n_tokens() && input_tokens[slot.prompt.n_tokens()] == LLAMA_TOKEN_NULL) {
+                    while (true) {
+                        auto cur_token_idx = slot.prompt.n_tokens();
+                        if (
+                            cur_token_idx >= slot.task->n_tokens() ||
+                            input_tokens[cur_token_idx] != LLAMA_TOKEN_NULL // encountered a text token
+                        ) {
+                            break;
+                        }
+
                         // process the image
                         size_t n_tokens_out = 0;
-                        int32_t res = input_tokens.process_chunk(ctx_tgt, mctx, slot.prompt.n_tokens(), slot.prompt.tokens.pos_next(), slot.id, n_tokens_out);
+                        int32_t res = slot.process_mtmd_chunk(cur_token_idx, n_tokens_out);
                         if (res != 0) {
                             SLT_ERR(slot, "failed to process image, res = %d\n", res);
                             send_error(slot, "failed to process image", ERROR_TYPE_SERVER);
@@ -2987,22 +3095,11 @@ private:
                             continue;
                         }
 
-                        if (ctx_dft && llama_get_ctx_other(ctx_dft.get()) != ctx_tgt) {
-                            // TODO: in the future, figure out how to infuse target embeddings to the images
-                            //       for now, we skip this for simplicity
-                            //       maybe we simply need to call `common_speculative_process()` on the mtmd batches in the `process_chunk` above?
-                            //       [TAG_MTMD_DRAFT_PROCESSING]
-                            res = input_tokens.process_chunk(ctx_dft.get(), mctx, slot.prompt.n_tokens(), slot.prompt.tokens.pos_next(), slot.id, n_tokens_out);
-                            if (res != 0) {
-                                GGML_ABORT("failed to process multi-modal data on draft context\n");
-                            }
-                        }
-
                         slot.n_prompt_tokens_processed += n_tokens_out;
 
                         // add the image chunk to cache
                         {
-                            const auto & chunk = input_tokens.find_chunk(slot.prompt.n_tokens());
+                            const auto & chunk = input_tokens.find_chunk(cur_token_idx);
                             slot.prompt.tokens.push_back(chunk.get()); // copy
                         }
 

tools/server/server-http.cpp

@@ -113,7 +113,7 @@ bool server_http_context::init(const common_params & params) {
 #endif
 
     srv->set_default_headers({{"Server", "llama.cpp"}});
-    srv->set_logger(log_server_request);
+    // srv->set_logger(log_server_request); // TODO @ngxson : this is too spamy, no very useful; improve it in the future
     srv->set_exception_handler([](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) {
         // this is fail-safe; exceptions should already handled by `ex_wrapper`
 
@@ -169,29 +169,108 @@ bool server_http_context::init(const common_params & params) {
         SRV_INF("api_keys: %zu keys loaded\n", params.api_keys.size());
     }
 
+    //
+    // Helper: Generate iOS splash screen paths from device dimensions
+    // This centralizes PWA asset paths to avoid duplication across CMake, C++, and TypeScript.
+    // Source of truth: tools/ui/src/lib/constants/pwa.ts (APPLE_DEVICES)
+    //
+    auto generate_splash_endpoints = []() -> std::vector<std::string> {
+        // Apple device dimensions (width x height) with orientation and color scheme
+        // Format: "orientation-dimension1xdimension2.png" or "orientation-dark-dimension1xdimension2.png"
+        // Based on https://developer.apple.com/design/human-interface-guidelines/app-icons
+        static const std::vector<std::pair<std::string, std::string>> splash_specs = {
+            // Portrait screens (light)
+            {"portrait", "640x1136"},     {"portrait", "750x1334"},
+            {"portrait", "1170x2532"},    {"portrait", "1179x2556"},
+            {"portrait", "1206x2622"},    {"portrait", "1284x2778"},
+            {"portrait", "1290x2796"},    {"portrait", "1320x2868"},
+            {"portrait", "1488x2266"},    {"portrait", "1640x2360"},
+            {"portrait", "1668x2388"},    {"portrait", "2048x2732"},
+            // Landscape screens (light) - dimensions swapped
+            {"landscape", "1136x640"},    {"landscape", "1334x750"},
+            {"landscape", "2532x1170"},   {"landscape", "2556x1179"},
+            {"landscape", "2622x1206"},   {"landscape", "2778x1284"},
+            {"landscape", "2796x1290"},   {"landscape", "2868x1320"},
+            {"landscape", "2266x1488"},   {"landscape", "2360x1640"},
+            {"landscape", "2388x1668"},   {"landscape", "2732x2048"},
+            // Portrait screens (dark)
+            {"portrait-dark", "640x1136"}, {"portrait-dark", "750x1334"},
+            {"portrait-dark", "1170x2532"}, {"portrait-dark", "1179x2556"},
+            {"portrait-dark", "1206x2622"}, {"portrait-dark", "1284x2778"},
+            {"portrait-dark", "1290x2796"}, {"portrait-dark", "1320x2868"},
+            {"portrait-dark", "1488x2266"}, {"portrait-dark", "1640x2360"},
+            {"portrait-dark", "1668x2388"}, {"portrait-dark", "2048x2732"},
+            // Landscape screens (dark)
+            {"landscape-dark", "1136x640"}, {"landscape-dark", "1334x750"},
+            {"landscape-dark", "2532x1170"}, {"landscape-dark", "2556x1179"},
+            {"landscape-dark", "2622x1206"}, {"landscape-dark", "2778x1284"},
+            {"landscape-dark", "2796x1290"}, {"landscape-dark", "2868x1320"},
+            {"landscape-dark", "2266x1488"}, {"landscape-dark", "2360x1640"},
+            {"landscape-dark", "2388x1668"}, {"landscape-dark", "2732x2048"}
+        };
+
+        std::vector<std::string> endpoints;
+        endpoints.reserve(splash_specs.size());
+        for (const auto & [orientation, dimensions] : splash_specs) {
+            endpoints.push_back("/apple-splash-" + orientation + "-" + dimensions + ".png");
+        }
+        return endpoints;
+    };
+
     //
     // Middlewares
     //
 
-    auto middleware_validate_api_key = [api_keys = params.api_keys](const httplib::Request & req, httplib::Response & res) {
-        static const std::unordered_set<std::string> public_endpoints = {
+    // Public endpoints list - includes health, UI, and PWA assets
+    // Source of truth for splash screen paths: tools/ui/src/lib/constants/pwa.ts (APPLE_DEVICES)
+    static const std::unordered_set<std::string> get_public_endpoints = [generate_splash_endpoints]() {
+        std::unordered_set<std::string> endpoints {
             "/health",
             "/v1/health",
             "/models",
             "/v1/models",
             "/",
             "/index.html",
-            "/bundle.js",
-            "/bundle.css",
+            // PWA assets
+            "/favicon.ico",
+            "/favicon-dark.ico",
+            "/favicon.svg",
+            "/favicon-dark.svg",
+            "/pwa-64x64.png",
+            "/pwa-192x192.png",
+            "/pwa-512x512.png",
+            "/maskable-icon-512x512.png",
+            "/apple-touch-icon-180x180.png",
+            // iOS splash screens (generated from APPLE_DEVICES in TypeScript)
+            // PWA runtime files
+            "/manifest.webmanifest",
+            "/sw.js",
+            "/version.json",
+            "/workbox-<hash>.js",
+            "/_app/version.json",
+            "/build.json"
         };
+        // Add all splash screen endpoints
+        auto splash = generate_splash_endpoints();
+        for (const auto & path : splash) {
+            endpoints.insert(path);
+        }
+        return endpoints;
+    }();
 
+    auto middleware_validate_api_key = [api_keys = params.api_keys](const httplib::Request & req, httplib::Response & res) {
         // If API key is not set, skip validation
         if (api_keys.empty()) {
             return true;
         }
 
         // If path is public or static file, skip validation
-        if (public_endpoints.find(req.path) != public_endpoints.end()) {
+        if (get_public_endpoints.find(req.path) != get_public_endpoints.end()) {
+            return true;
+        }
+        // Static assets (_app/ files, workbox runtime). These are embedded at build time
+        // so no API key is needed — browsers fetch them directly.
+        if (req.path.find("/_app/") == 0 || req.path.find("/workbox-") == 0) {
             return true;
         }
 
@@ -315,7 +394,11 @@ bool server_http_context::init(const common_params & params) {
             }
         } else {
 #if defined(LLAMA_UI_HAS_ASSETS)
-            auto serve_asset = [](const std::string & name, const char * mime, bool with_isolation_headers) {
+            // Embedded assets are immutable — cache aggressively for PWA/sw offline support.
+            // PWA runtime files (sw.js, manifest, version.json) use no-cache for revalidation.
+            // Bundle files use Vite content hashes (bundle.<hash>.js/css) so each build
+            // produces a different filename — browsers naturally get a fresh copy on upgrade.
+            auto serve_asset_cached = [](const std::string & name, const char * mime, bool with_isolation_headers) {
                 return [name, mime, with_isolation_headers](const httplib::Request & req, httplib::Response & res) {
                     const llama_ui_asset * a = llama_ui_find_asset(name.c_str());
                     if (!a) {
@@ -334,14 +417,129 @@ bool server_http_context::init(const common_params & params) {
                         res.set_header("Cross-Origin-Embedder-Policy", "require-corp");
                         res.set_header("Cross-Origin-Opener-Policy", "same-origin");
                     }
+                    res.set_header("Cache-Control", "public, max-age=31536000, immutable");
                     res.set_content(reinterpret_cast<const char*>(a->data), a->size, mime);
                     return false;
                 };
             };
 
-            srv->Get(params.api_prefix + "/",           serve_asset("index.html", "text/html; charset=utf-8",              true));
-            srv->Get(params.api_prefix + "/bundle.js",  serve_asset("bundle.js",  "application/javascript; charset=utf-8", false));
-            srv->Get(params.api_prefix + "/bundle.css", serve_asset("bundle.css", "text/css; charset=utf-8",               false));
+            auto serve_asset_nocache = [](const std::string & name, const char * mime, bool with_isolation_headers) {
+                return [name, mime, with_isolation_headers](const httplib::Request & /*req*/, httplib::Response & res) {
+                    const llama_ui_asset * a = llama_ui_find_asset(name.c_str());
+                    if (!a) {
+                        res.status = 404;
+                        return false;
+                    }
+                    if (with_isolation_headers) {
+                        res.set_header("Cross-Origin-Embedder-Policy", "require-corp");
+                        res.set_header("Cross-Origin-Opener-Policy", "same-origin");
+                    }
+                    res.set_header("Cache-Control", "no-cache");
+                    res.set_content(reinterpret_cast<const char*>(a->data), a->size, mime);
+                    return false;
+                };
+            };
+
+            // Bundle files in _app/immutable/ — SvelteKit outputs them here and index.html
+            // and sw.js reference them via these paths (vanilla build, no plugin).
+            auto serve_bundle = [serve_asset_cached](const httplib::Request & req, httplib::Response & res) {
+                std::string path = req.path;
+                std::string name;
+                const char * mime;
+                if (path.rfind("/_app/immutable/bundle.", 0) == 0 && path.size() > 22) {
+                    name = path.substr(1);  // strip leading /
+                    mime = "application/javascript; charset=utf-8";
+                } else if (path.rfind("/_app/immutable/assets/bundle.", 0) == 0 && path.size() > 30) {
+                    name = path.substr(1);  // strip leading /
+                    mime = "text/css; charset=utf-8";
+                } else {
+                    res.status = 404;
+                    return false;
+                }
+                return serve_asset_cached(name, mime, false)(req, res);
+            };
+
+            // _app/ paths — vanilla SvelteKit output, index.html and sw.js reference
+            // bundles and version.json here directly.
+            srv->Get(params.api_prefix + R"(/_app/immutable/bundle\.[^/]+\.js)",  serve_bundle);
+            srv->Get(params.api_prefix + R"(/_app/immutable/assets/bundle\.[^/]+\.css)", serve_bundle);
+            srv->Get(params.api_prefix + "/_app/version.json",                    serve_asset_cached("_app/version.json", "application/json; charset=utf-8", false));
+
+            auto serve_workbox = [serve_asset_cached](const httplib::Request & req, httplib::Response & res) {
+                std::string name = req.path.substr(1);
+                if (name.rfind("workbox-", 0) == 0 && name.size() > 10) {
+                    return serve_asset_cached(name, "application/javascript; charset=utf-8", false)(req, res);
+                }
+                res.status = 404;
+                return false;
+            };
+            srv->Get(params.api_prefix + R"(/workbox-[^/]+\.js)",               serve_workbox);
+            srv->Get(params.api_prefix + R"(/sw\.js)",                          serve_asset_cached("sw.js",               "application/javascript; charset=utf-8", false));
+            srv->Get(params.api_prefix + "/manifest.webmanifest",                serve_asset_cached("manifest.webmanifest", "application/manifest+json; charset=utf-8", false));
+            srv->Get(params.api_prefix + "/version.json",                        serve_asset_cached("_app/version.json",  "application/json; charset=utf-8",       false));
+            srv->Get(params.api_prefix + "/build.json",                          serve_asset_cached("build.json",         "application/json; charset=utf-8",       false));
+
+            // Finally serve index.html for all other routes (SPA fallback)
+            srv->Get(params.api_prefix + "/",                               serve_asset_cached("index.html",                   "text/html; charset=utf-8",                 true));
+            srv->Get(params.api_prefix + "/favicon.ico",                    serve_asset_cached("favicon.ico",                  "image/x-icon",                             false));
+            srv->Get(params.api_prefix + "/favicon-dark.ico",                serve_asset_cached("favicon-dark.ico",              "image/x-icon",                             false));
+            srv->Get(params.api_prefix + "/favicon.svg",                    serve_asset_cached("favicon.svg",                  "image/svg+xml",                            false));
+            srv->Get(params.api_prefix + "/favicon-dark.svg",              serve_asset_cached("favicon-dark.svg",            "image/svg+xml",                            false));
+            srv->Get(params.api_prefix + "/pwa-64x64.png",                  serve_asset_cached("pwa-64x64.png",                "image/png",                                false));
+            srv->Get(params.api_prefix + "/pwa-192x192.png",                serve_asset_cached("pwa-192x192.png",              "image/png",                                false));
+            srv->Get(params.api_prefix + "/pwa-512x512.png",                serve_asset_cached("pwa-512x512.png",              "image/png",                                false));
+            srv->Get(params.api_prefix + "/maskable-icon-512x512.png",      serve_asset_cached("maskable-icon-512x512.png",    "image/png",                                false));
+            srv->Get(params.api_prefix + "/apple-touch-icon-180x180.png",   serve_asset_cached("apple-touch-icon-180x180.png", "image/png",                                false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-640x1136.png",          serve_asset_cached("apple-splash-portrait-640x1136.png",          "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-1136x640.png",         serve_asset_cached("apple-splash-landscape-1136x640.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-750x1334.png",          serve_asset_cached("apple-splash-portrait-750x1334.png",          "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-1334x750.png",         serve_asset_cached("apple-splash-landscape-1334x750.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1170x2532.png",         serve_asset_cached("apple-splash-portrait-1170x2532.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2532x1170.png",        serve_asset_cached("apple-splash-landscape-2532x1170.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1179x2556.png",         serve_asset_cached("apple-splash-portrait-1179x2556.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2556x1179.png",        serve_asset_cached("apple-splash-landscape-2556x1179.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1206x2622.png",         serve_asset_cached("apple-splash-portrait-1206x2622.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2622x1206.png",        serve_asset_cached("apple-splash-landscape-2622x1206.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1284x2778.png",         serve_asset_cached("apple-splash-portrait-1284x2778.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2778x1284.png",        serve_asset_cached("apple-splash-landscape-2778x1284.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1290x2796.png",         serve_asset_cached("apple-splash-portrait-1290x2796.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2796x1290.png",        serve_asset_cached("apple-splash-landscape-2796x1290.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1320x2868.png",         serve_asset_cached("apple-splash-portrait-1320x2868.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2868x1320.png",        serve_asset_cached("apple-splash-landscape-2868x1320.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1488x2266.png",         serve_asset_cached("apple-splash-portrait-1488x2266.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2266x1488.png",        serve_asset_cached("apple-splash-landscape-2266x1488.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1640x2360.png",         serve_asset_cached("apple-splash-portrait-1640x2360.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2360x1640.png",        serve_asset_cached("apple-splash-landscape-2360x1640.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-1668x2388.png",         serve_asset_cached("apple-splash-portrait-1668x2388.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2388x1668.png",        serve_asset_cached("apple-splash-landscape-2388x1668.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-2048x2732.png",         serve_asset_cached("apple-splash-portrait-2048x2732.png",         "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-2732x2048.png",        serve_asset_cached("apple-splash-landscape-2732x2048.png",        "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-640x1136.png",     serve_asset_cached("apple-splash-portrait-dark-640x1136.png",     "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-1136x640.png",    serve_asset_cached("apple-splash-landscape-dark-1136x640.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-750x1334.png",     serve_asset_cached("apple-splash-portrait-dark-750x1334.png",     "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-1334x750.png",    serve_asset_cached("apple-splash-landscape-dark-1334x750.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1170x2532.png",    serve_asset_cached("apple-splash-portrait-dark-1170x2532.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2532x1170.png",   serve_asset_cached("apple-splash-landscape-dark-2532x1170.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1179x2556.png",    serve_asset_cached("apple-splash-portrait-dark-1179x2556.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2556x1179.png",   serve_asset_cached("apple-splash-landscape-dark-2556x1179.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1206x2622.png",    serve_asset_cached("apple-splash-portrait-dark-1206x2622.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2622x1206.png",   serve_asset_cached("apple-splash-landscape-dark-2622x1206.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1284x2778.png",    serve_asset_cached("apple-splash-portrait-dark-1284x2778.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2778x1284.png",   serve_asset_cached("apple-splash-landscape-dark-2778x1284.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1290x2796.png",    serve_asset_cached("apple-splash-portrait-dark-1290x2796.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2796x1290.png",   serve_asset_cached("apple-splash-landscape-dark-2796x1290.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1320x2868.png",    serve_asset_cached("apple-splash-portrait-dark-1320x2868.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2868x1320.png",   serve_asset_cached("apple-splash-landscape-dark-2868x1320.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1640x2360.png",    serve_asset_cached("apple-splash-portrait-dark-1640x2360.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2360x1640.png",   serve_asset_cached("apple-splash-landscape-dark-2360x1640.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-1668x2388.png",    serve_asset_cached("apple-splash-portrait-dark-1668x2388.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2388x1668.png",   serve_asset_cached("apple-splash-landscape-dark-2388x1668.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-portrait-dark-2048x2732.png",    serve_asset_cached("apple-splash-portrait-dark-2048x2732.png",    "image/png", false));
+            srv->Get(params.api_prefix + "/apple-splash-landscape-dark-2732x2048.png",   serve_asset_cached("apple-splash-landscape-dark-2732x2048.png",   "image/png", false));
+            srv->Get(params.api_prefix + "/manifest.webmanifest",           serve_asset_nocache("manifest.webmanifest",        "application/manifest+json",                false));
+            srv->Get(params.api_prefix + "/sw.js",                          serve_asset_nocache("sw.js",                       "application/javascript; charset=utf-8",    false));
+            srv->Get(params.api_prefix + "/version.json",                   serve_asset_nocache("version.json",                 "application/json",                         false));
+
 #endif
         }
     }

tools/server/server.cpp

@@ -94,20 +94,22 @@ int llama_server(int argc, char ** argv) {
     const bool is_router_server = params.model.path.empty();
     common_params_print_info(params, !is_router_server);
 
-    // validate batch size for embeddings
-    // embeddings require all tokens to be processed in a single ubatch
-    // see https://github.com/ggml-org/llama.cpp/issues/12836
-    if (params.embedding && params.n_batch > params.n_ubatch) {
-        SRV_WRN("embeddings enabled with n_batch (%d) > n_ubatch (%d)\n", params.n_batch, params.n_ubatch);
-        SRV_WRN("setting n_batch = n_ubatch = %d to avoid assertion failure\n", params.n_ubatch);
-        params.n_batch = params.n_ubatch;
-    }
+    if (!is_router_server) {
+        // validate batch size for embeddings
+        // embeddings require all tokens to be processed in a single ubatch
+        // see https://github.com/ggml-org/llama.cpp/issues/12836
+        if (params.embedding && params.n_batch > params.n_ubatch) {
+            SRV_WRN("embeddings enabled with n_batch (%d) > n_ubatch (%d)\n", params.n_batch, params.n_ubatch);
+            SRV_WRN("setting n_batch = n_ubatch = %d to avoid assertion failure\n", params.n_ubatch);
+            params.n_batch = params.n_ubatch;
+        }
 
-    if (params.n_parallel < 0) {
-        SRV_INF("%s", "n_parallel is set to auto, using n_parallel = 4 and kv_unified = true\n");
+        if (params.n_parallel < 0) {
+            SRV_INF("%s", "n_parallel is set to auto, using n_parallel = 4 and kv_unified = true\n");
 
-        params.n_parallel = 4;
-        params.kv_unified = true;
+            params.n_parallel = 4;
+            params.kv_unified = true;
+        }
     }
 
     // for consistency between server router mode and single-model mode, we set the same model name as alias

tools/server/tests/unit/test_security.py

@@ -26,7 +26,7 @@ def test_access_static_assets_without_api_key():
     """Static web UI assets should not require API key authentication (issue #21229)"""
     global server
     server.start()
-    for path in ["/", "/bundle.js", "/bundle.css"]:
+    for path in ["/", "/sw.js", "/manifest.webmanifest", "/_app/version.json"]:
         res = server.make_request("GET", path)
         assert res.status_code == 200, f"Expected 200 for {path}, got {res.status_code}"
 

tools/ui/.gitignore

@@ -8,6 +8,8 @@ node_modules
 .wrangler
 /.svelte-kit
 /build
+dev-dist
+dist
 
 # OS
 .DS_Store
@@ -23,6 +25,15 @@ Thumbs.db
 vite.config.js.timestamp-*
 vite.config.ts.timestamp-*
 
+# PWA Artifacts
+apple-splash-*.png
+apple-touch-icon-*.png
+favicon.ico
+favicon-dark.ico
+maskable-icon-*.png
+pwa-*.png
+
+# Storybook
 *storybook.log
 storybook-static
 *.code-workspace

tools/ui/CMakeLists.txt

@@ -77,6 +77,7 @@ add_custom_target(llama-ui-assets ALL
         "-DUI_SOURCE_DIR=${CMAKE_CURRENT_SOURCE_DIR}"
         "-DUI_BINARY_DIR=${CMAKE_CURRENT_BINARY_DIR}"
         "-DLLAMA_SOURCE_DIR=${PROJECT_SOURCE_DIR}"
+        "-DLLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}"
         "-DHF_BUCKET=${LLAMA_UI_HF_BUCKET}"
         "-DHF_VERSION=${HF_UI_VERSION}"
         "-DHF_ENABLED=${LLAMA_USE_PREBUILT_UI}"

tools/ui/package.json

@@ -4,8 +4,9 @@
 	"version": "1.0.0",
 	"type": "module",
 	"scripts": {
+		"build": "npm run build-pwa-assets && vite build",
+		"build-pwa-assets": "npx @vite-pwa/assets-generator --root . --config pwa-assets.config.ts && npx @vite-pwa/assets-generator --root . --config pwa-assets-dark.config.ts && node scripts/make-icons-circular.js",
 		"dev": "bash scripts/dev.sh",
-		"build": "vite build",
 		"preview": "vite preview",
 		"prepare": "svelte-kit sync || echo ''",
 		"check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json",
@@ -15,12 +16,15 @@
 		"lint": "prettier --check . && eslint .",
 		"test": "npm run test:ui -- --run && npm run test:client -- --run && npm run test:unit -- --run && npm run test:e2e",
 		"test:e2e": "playwright test",
+		"test:e2e:pwa": "playwright test tests/e2e/pwa.e2e.ts",
 		"test:client": "vitest --project=client",
 		"test:unit": "vitest --project=unit",
+		"test:unit:pwa": "vitest --project=unit --run tests/unit/pwa.spec.ts",
+		"test:pwa": "npm run test:unit:pwa && npm run test:e2e:pwa",
 		"test:ui": "vitest --project=ui",
 		"storybook": "storybook dev -p 6006",
 		"build-storybook": "storybook build",
-		"cleanup": "rm -rf .svelte-kit build node_modules test-results"
+		"cleanup": "rm -rf .svelte-kit build node_modules test-results dist dev-dist debug-storybook.log static/pwa-*.png static/maskable-icon-*.png static/apple-touch-icon-*.png static/apple-splash-*.png static/favicon*.ico"
 	},
 	"devDependencies": {
 		"@chromatic-com/storybook": "5.0.0",
@@ -41,29 +45,31 @@
 		"@tailwindcss/forms": "0.5.10",
 		"@tailwindcss/typography": "0.5.16",
 		"@tailwindcss/vite": "4.1.11",
-		"@types/node": "^24",
+		"@types/node": "24.13.0",
+		"@vite-pwa/assets-generator": "1.0.2",
+		"@vite-pwa/sveltekit": "1.1.0",
 		"@vitest/browser": "4.1.8",
 		"@vitest/browser-playwright": "4.1.8",
 		"@vitest/coverage-v8": "4.1.8",
 		"bits-ui": "2.18.1",
 		"clsx": "2.1.1",
-		"dexie": "4.0.11",
-		"eslint": "9.39.2",
+		"dexie": "4.4.3",
+		"eslint": "9.39.4",
 		"eslint-config-prettier": "10.1.8",
-		"eslint-plugin-storybook": "10.2.4",
-		"eslint-plugin-svelte": "3.15.0",
-		"globals": "16.3.0",
+		"eslint-plugin-storybook": "10.4.2",
+		"eslint-plugin-svelte": "3.19.0",
+		"globals": "16.5.0",
 		"highlight.js": "11.11.1",
 		"http-server": "14.1.1",
 		"mdast": "3.0.0",
-		"mdsvex": "0.12.6",
+		"mdsvex": "0.12.7",
 		"mermaid": "11.15.0",
 		"mode-watcher": "1.1.0",
 		"pdfjs-dist": "5.4.54",
 		"playwright": "1.56.1",
-		"prettier": "3.6.2",
-		"prettier-plugin-svelte": "3.4.0",
-		"prettier-plugin-tailwindcss": "0.6.14",
+		"prettier": "3.8.3",
+		"prettier-plugin-svelte": "4.1.0",
+		"prettier-plugin-tailwindcss": "0.8.0",
 		"rehype-highlight": "7.0.2",
 		"rehype-katex": "7.0.1",
 		"rehype-stringify": "10.0.1",
@@ -73,25 +79,25 @@
 		"remark-html": "16.0.1",
 		"remark-math": "6.0.0",
 		"remark-rehype": "11.1.2",
-		"sass": "1.93.3",
-		"storybook": "10.3.3",
-		"svelte": "5.55.7",
-		"svelte-check": "4.3.0",
-		"svelte-sonner": "1.0.5",
-		"tailwind-merge": "3.3.1",
+		"sass": "1.100.0",
+		"storybook": "10.4.2",
+		"svelte": "5.56.1",
+		"svelte-check": "4.6.0",
+		"svelte-sonner": "1.1.1",
+		"tailwind-merge": "3.6.0",
 		"tailwind-variants": "3.2.2",
-		"tailwindcss": "4.1.11",
-		"tw-animate-css": "1.3.5",
-		"typescript": "5.8.3",
-		"typescript-eslint": "8.56.0",
+		"tailwindcss": "4.3.0",
+		"tw-animate-css": "1.4.0",
+		"typescript": "5.9.3",
+		"typescript-eslint": "8.60.1",
 		"unified": "11.0.5",
-		"unist-util-visit": "5.0.0",
+		"unist-util-visit": "5.1.0",
 		"uuid": "13.0.2",
-		"vite": "7.3.2",
+		"vite": "7.3.5",
 		"vite-plugin-devtools-json": "0.2.1",
 		"vitest": "4.1.8",
 		"vitest-browser-svelte": "2.1.1",
-		"zod": "4.2.1"
+		"workbox-window": "7.4.1"
 	},
 	"overrides": {
 		"cookie": "1.1.1"

tools/ui/playwright.config.ts

@@ -1,11 +1,31 @@
-import { defineConfig } from '@playwright/test';
+import { defineConfig, devices } from '@playwright/test';
 
 export default defineConfig({
+	testDir: 'tests/e2e',
+	testMatch: ['**/*.e2e.ts'],
+	timeout: 30000,
+	expect: {
+		timeout: 5000
+	},
+	fullyParallel: true,
+	forbidOnly: !!process.env.CI,
+	retries: process.env.CI ? 2 : 0,
+	workers: process.env.CI ? 1 : undefined,
+	reporter: 'line',
+	use: {
+		baseURL: 'http://localhost:8181',
+		trace: 'on-first-retry'
+	},
+	projects: [
+		{
+			name: 'chromium',
+			use: { ...devices['Desktop Chrome'] }
+		}
+	],
 	webServer: {
 		command: 'npm run build && npx http-server ./dist -p 8181',
 		port: 8181,
 		timeout: 120000,
-		reuseExistingServer: false
-	},
-	testDir: 'tests/e2e'
+		reuseExistingServer: !process.env.CI
+	}
 });

tools/ui/pwa-assets-dark.config.ts

@@ -0,0 +1,20 @@
+import { defineConfig } from '@vite-pwa/assets-generator/config';
+
+export default defineConfig({
+	headLinkOptions: {
+		preset: '2023'
+	},
+	preset: {
+		transparent: {
+			sizes: [],
+			favicons: [[48, 'favicon-dark.ico']]
+		},
+		maskable: {
+			sizes: []
+		},
+		apple: {
+			sizes: []
+		}
+	},
+	images: ['static/favicon-dark.svg']
+});

tools/ui/pwa-assets.config.ts

@@ -0,0 +1,51 @@
+import {
+	combinePresetAndAppleSplashScreens,
+	defineConfig,
+	minimal2023Preset
+} from '@vite-pwa/assets-generator/config';
+import { readFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+import { THEME_COLORS, PWA_GENERATOR_DEVICES, PWA_ASSET_GENERATOR } from './src/lib/constants/pwa';
+import { SplashOrientation } from './src/lib/enums/splash.enums';
+
+export default defineConfig({
+	headLinkOptions: {
+		preset: PWA_ASSET_GENERATOR.LINK_PRESET
+	},
+	preset: combinePresetAndAppleSplashScreens(
+		minimal2023Preset,
+		{
+			padding: PWA_ASSET_GENERATOR.SPLASH_PADDING,
+			resizeOptions: {
+				background: THEME_COLORS.BACKGROUND_LIGHT,
+				fit: PWA_ASSET_GENERATOR.FIT_MODE
+			},
+			darkResizeOptions: {
+				background: THEME_COLORS.BACKGROUND_DARK,
+				fit: PWA_ASSET_GENERATOR.FIT_MODE
+			},
+			darkImageResolver: async (imageName: string) => {
+				if (imageName.endsWith('favicon.svg')) {
+					return readFileSync(resolve('static/favicon-dark.svg'));
+				}
+			},
+			linkMediaOptions: {
+				log: true,
+				addMediaScreen: PWA_ASSET_GENERATOR.ADD_MEDIA_SCREEN,
+				basePath: PWA_ASSET_GENERATOR.BASE_PATH,
+				xhtml: PWA_ASSET_GENERATOR.XHTML
+			},
+			png: {
+				compressionLevel: PWA_ASSET_GENERATOR.PNG_COMPRESSION_LEVEL,
+				quality: PWA_ASSET_GENERATOR.PNG_QUALITY
+			},
+			name: (landscape, size, dark) => {
+				const orientation = landscape ? SplashOrientation.LANDSCAPE : SplashOrientation.PORTRAIT;
+				const darkPrefix = dark ? PWA_ASSET_GENERATOR.DARK_PREFIX : '';
+				return `apple-splash-${orientation}-${darkPrefix}${size.width}x${size.height}.png`;
+			}
+		},
+		PWA_GENERATOR_DEVICES
+	),
+	images: ['static/favicon.svg']
+});

tools/ui/scripts/make-icons-circular.js

@@ -0,0 +1,137 @@
+#!/usr/bin/env node
+
+/**
+ * Apply circular mask to pwa-*.png icons.
+ * Uses the maskable icon as source (white bg, full logo) to avoid
+ * the small-colormap pwa icons looking bad when cropped to a circle.
+ *
+ * Usage: node scripts/make-icons-circular.js [--padding-pct <0-50>] [--scale-pct <50-100>]
+ *
+ * - padding-pct: percentage of icon size kept as padding around the circle (default: 25)
+ * - scale-pct: scale down the source image before cropping (default: 85)
+ *
+ * maskable-icon and apple-touch-icon are left untouched.
+ */
+
+import sharp from 'sharp';
+import fs from 'fs';
+import path from 'path';
+import { fileURLToPath } from 'url';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+
+const STATIC_DIR = path.resolve(__dirname, '..', 'static');
+
+const paddingPct = process.argv.reduce((acc, arg, i, args) => {
+	if (arg === '--padding-pct' && args[i + 1]) return parseFloat(args[i + 1]);
+	return acc;
+}, 0);
+
+// Scale down the source image before cropping to circle
+const scalePct = process.argv.reduce((acc, arg, i, args) => {
+	if (arg === '--scale-pct' && args[i + 1]) return parseFloat(args[i + 1]);
+	return acc;
+}, 85); // default 85% - icon fills 85% of the circular area
+
+// Source for circular icons: the maskable icon (white bg, full logo)
+const sourceIcon = 'maskable-icon-512x512.png';
+const targetIcons = ['pwa-64x64.png', 'pwa-192x192.png', 'pwa-512x512.png'];
+
+// maskable-icon and apple-touch-icon stay square
+const untouchedIcons = ['maskable-icon-512x512.png', 'apple-touch-icon-180x180.png'];
+
+async function makeCircle(targetFilename) {
+	const targetPath = path.join(STATIC_DIR, targetFilename);
+	const sourcePath = path.join(STATIC_DIR, sourceIcon);
+
+	if (!fs.existsSync(sourcePath)) {
+		console.log(`⏭️  ${sourceIcon} not found, skipping`);
+		return;
+	}
+	if (!fs.existsSync(targetPath)) {
+		console.log(`⏭️  ${targetFilename} not found, skipping`);
+		return;
+	}
+
+	const metadata = await sharp(targetPath).metadata();
+	const size = Math.max(metadata.width, metadata.height);
+	const radius = Math.floor((size * (1 - paddingPct / 100)) / 2);
+	const center = Math.floor(size / 2);
+
+	// Build circular mask as RGBA buffer: white opaque circle on transparent bg
+	const maskBuf = Buffer.alloc(size * size * 4, 0);
+	for (let y = 0; y < size; y++) {
+		for (let x = 0; x < size; x++) {
+			const dx = x - center;
+			const dy = y - center;
+			const dist = Math.sqrt(dx * dx + dy * dy);
+			if (dist < radius) {
+				const i = (y * size + x) * 4;
+				maskBuf[i] = 255;
+				maskBuf[i + 1] = 255;
+				maskBuf[i + 2] = 255;
+				maskBuf[i + 3] = 255;
+			}
+		}
+	}
+
+	const tmpMask = path.join(STATIC_DIR, '.mask-tmp.png');
+	await sharp(maskBuf, {
+		raw: { width: size, height: size, channels: 4 }
+	})
+		.png()
+		.toFile(tmpMask);
+
+	// Step 1: Scale source relative to circle diameter (not full icon), composite centered onto white canvas of full size
+	const circleDiameter = Math.floor(size * (1 - paddingPct / 100));
+	const scaledSize = Math.floor((circleDiameter * scalePct) / 100);
+	const offset = Math.floor((size - scaledSize) / 2);
+
+	const scaledBuf = await sharp(sourcePath)
+		.resize(scaledSize, scaledSize, {
+			fit: 'cover',
+			background: { r: 255, g: 255, b: 255, alpha: 1 }
+		})
+		.ensureAlpha()
+		.png()
+		.toBuffer();
+
+	// Step 2: Composite scaled image onto white background, then apply circular mask
+	const output = await sharp({
+		create: {
+			width: size,
+			height: size,
+			channels: 4,
+			background: { r: 255, g: 255, b: 255, alpha: 1 }
+		}
+	})
+		.composite([
+			{ input: scaledBuf, top: offset, left: offset },
+			{ input: tmpMask, top: 0, left: 0, blend: 'dest-in' }
+		])
+		.png()
+		.toBuffer();
+
+	fs.writeFileSync(targetPath, output);
+	fs.unlinkSync(tmpMask);
+
+	console.log(
+		`✓ ${targetFilename} → circle from ${sourceIcon}, ${paddingPct}% padding (size=${size}, r=${radius}, scale=${scalePct}%, circleDiameter=${circleDiameter})`
+	);
+}
+
+async function main() {
+	console.log(`Circular mask: ${paddingPct}% padding, ${scalePct}% scale, source=${sourceIcon}\n`);
+	for (const icon of targetIcons) {
+		await makeCircle(icon);
+	}
+
+	console.log('\nUnchanged:');
+	for (const icon of untouchedIcons) {
+		const fp = path.join(STATIC_DIR, icon);
+		console.log(`  ${icon} (${fs.existsSync(fp) ? fs.statSync(fp).size + ' bytes' : 'missing'})`);
+	}
+}
+
+main();

tools/ui/scripts/vite-plugin-build-info.ts

@@ -0,0 +1,42 @@
+import { writeFileSync, existsSync } from 'node:fs';
+import { resolve } from 'path';
+import type { Plugin } from 'vite';
+import { BUILD_CONFIG } from '../src/lib/constants/pwa';
+
+let processed = false;
+
+const OUTPUT_DIR = process.env.LLAMA_UI_OUT_DIR ?? BUILD_CONFIG.OUTPUT_DIR;
+
+/**
+ * Write build.json with the llama.cpp release build number.
+ *
+ * LLAMA_BUILD_NUMBER is passed from CMake -> npm -> vite via env var.
+ * Used for display of the current llama-server release (e.g. "b1234").
+ */
+export function buildInfoPlugin(): Plugin {
+	return {
+		name: 'llamacpp:build-info',
+		apply: 'build',
+		closeBundle() {
+			setTimeout(() => {
+				try {
+					if (processed) return;
+					processed = true;
+
+					const buildNumber = process.env.LLAMA_BUILD_NUMBER;
+					if (!buildNumber) return;
+
+					const outDir = resolve(OUTPUT_DIR);
+					const indexPath = resolve(outDir, 'index.html');
+					if (!existsSync(indexPath)) return;
+
+					const buildJsonPath = resolve(outDir, 'build.json');
+					writeFileSync(buildJsonPath, JSON.stringify({ version: buildNumber }), 'utf-8');
+					console.log(`Created build.json (version: ${buildNumber})`);
+				} catch (error) {
+					console.error('Failed to write build.json:', error);
+				}
+			}, 100);
+		}
+	};
+}

tools/ui/scripts/vite-plugin-llama-cpp-build.ts

@@ -1,105 +0,0 @@
-import {
-	readFileSync,
-	writeFileSync,
-	existsSync,
-	readdirSync,
-	copyFileSync,
-	rmSync,
-	unlinkSync
-} from 'fs';
-import { resolve } from 'path';
-import type { Plugin } from 'vite';
-
-const GUIDE_FOR_FRONTEND = `
-<!--
-  This is a static build of the frontend.
-  It is automatically generated by the build process.
-  Do not edit this file directly.
-  To make changes, refer to the "Web UI" section in the README.
--->
-`.trim();
-
-const OUTPUT_DIR = process.env.LLAMA_UI_OUT_DIR ?? './dist';
-
-export function llamaCppBuildPlugin(): Plugin {
-	return {
-		name: 'llamacpp:build',
-		apply: 'build',
-		closeBundle() {
-			setTimeout(() => {
-				try {
-					const outDir = resolve(OUTPUT_DIR);
-					const indexPath = resolve(outDir, 'index.html');
-					if (!existsSync(indexPath)) return;
-
-					let content = readFileSync(indexPath, 'utf-8');
-
-					// Inline favicon as base64 data URL
-					const faviconPath = resolve('static/favicon.svg');
-					if (existsSync(faviconPath)) {
-						const faviconContent = readFileSync(faviconPath, 'utf-8');
-						const faviconBase64 = Buffer.from(faviconContent).toString('base64');
-						const faviconDataUrl = `data:image/svg+xml;base64,${faviconBase64}`;
-						content = content.replace(/href="[^"]*favicon\.svg"/g, `href="${faviconDataUrl}"`);
-						console.log('✓ Inlined favicon.svg as base64 data URL');
-					}
-
-					content = content.replace(/\r/g, '');
-					content = GUIDE_FOR_FRONTEND + '\n' + content;
-
-					// Keep the Vite hash as a query string so each build busts the browser cache
-					content = content.replace(/\/_app\/immutable\/bundle\.([^".]+)\.js/g, './bundle.js?$1');
-					content = content.replace(
-						/\/_app\/immutable\/assets\/bundle\.([^".]+)\.css/g,
-						'./bundle.css?$1'
-					);
-					content = content.replace(/__sveltekit_[a-z0-9]+/g, '__sveltekit__');
-
-					writeFileSync(indexPath, content, 'utf-8');
-					console.log('✓ Updated index.html');
-
-					// Copy bundle.*.js -> bundle.js at output root
-					const immutableDir = resolve(outDir, '_app/immutable');
-					const bundleDir = resolve(outDir, '_app/immutable/assets');
-
-					if (existsSync(immutableDir)) {
-						const jsFiles = readdirSync(immutableDir).filter((f) => f.match(/^bundle\..+\.js$/));
-						if (jsFiles.length > 0) {
-							copyFileSync(resolve(immutableDir, jsFiles[0]), resolve(outDir, 'bundle.js'));
-							// Normalize __sveltekit_<hash> to __sveltekit__ in bundle.js
-							const bundleJsPath = resolve(outDir, 'bundle.js');
-							let bundleJs = readFileSync(bundleJsPath, 'utf-8');
-							bundleJs = bundleJs.replace(/__sveltekit_[a-z0-9]+/g, '__sveltekit__');
-							writeFileSync(bundleJsPath, bundleJs, 'utf-8');
-							console.log(`✓ Copied ${jsFiles[0]} -> bundle.js`);
-						}
-					}
-
-					// Copy bundle.*.css -> bundle.css at output root
-					if (existsSync(bundleDir)) {
-						const cssFiles = readdirSync(bundleDir).filter((f) => f.match(/^bundle\..+\.css$/));
-						if (cssFiles.length > 0) {
-							copyFileSync(resolve(bundleDir, cssFiles[0]), resolve(outDir, 'bundle.css'));
-							console.log(`✓ Copied ${cssFiles[0]} -> bundle.css`);
-						}
-					}
-
-					// Cleanup: remove _app directory, favicon.svg, and legacy index.html.gz
-					const appDir = resolve(outDir, '_app');
-					if (existsSync(appDir)) {
-						rmSync(appDir, { recursive: true, force: true });
-						console.log('✓ Removed _app directory');
-					}
-
-					const faviconOut = resolve(outDir, 'favicon.svg');
-					if (existsSync(faviconOut)) {
-						unlinkSync(faviconOut);
-						console.log('✓ Removed favicon.svg');
-					}
-				} catch (error) {
-					console.error('Failed to process build output:', error);
-				}
-			}, 100);
-		}
-	};
-}

tools/ui/scripts/vite-plugin-relativize-base.ts

@@ -0,0 +1,61 @@
+import { readFileSync, writeFileSync, existsSync } from 'node:fs';
+import { resolve } from 'path';
+import type { Plugin } from 'vite';
+import { BUILD_CONFIG } from '../src/lib/constants/pwa';
+
+let processed = false;
+
+const OUTPUT_DIR = process.env.LLAMA_UI_OUT_DIR ?? BUILD_CONFIG.OUTPUT_DIR;
+
+function rewrite(path: string, pairs: [string, string][]): void {
+	if (!existsSync(path)) {
+		return;
+	}
+	const text = readFileSync(path, 'utf-8');
+	let out = text;
+	for (const [from, to] of pairs) {
+		out = out.split(from).join(to);
+	}
+	if (out !== text) {
+		writeFileSync(path, out, 'utf-8');
+	}
+}
+
+/**
+ * Relativize SvelteKit absolute base refs so the build is relocatable under any subpath.
+ *
+ * SvelteKit bakes root absolute /_app/ paths into the SPA fallback because paths.relative
+ * does not apply to a depth agnostic fallback page. Rewriting to ./_app/ lets a plain
+ * recursive copy of the output into /any/subdir/ resolve assets against the document URL.
+ * Runs after adapter-static writes index.html and the PWA plugin writes sw.js, deferred the
+ * same way as buildInfoPlugin so the emitted files exist.
+ */
+export function relativizeBasePlugin(): Plugin {
+	return {
+		name: 'llamacpp:relativize-base',
+		apply: 'build',
+		closeBundle() {
+			setTimeout(() => {
+				try {
+					if (processed) return;
+					processed = true;
+
+					const outDir = resolve(OUTPUT_DIR);
+
+					// index.html: modulepreload, stylesheet and bootstrap import reference "/_app/
+					rewrite(resolve(outDir, 'index.html'), [['"/_app/', '"./_app/']]);
+
+					// sw.js: the only absolute entries are the navigate fallback precache key and handler
+					rewrite(resolve(outDir, 'sw.js'), [
+						['{url:"/"', '{url:"./"'],
+						['createHandlerBoundToURL("/"', 'createHandlerBoundToURL("./"']
+					]);
+
+					console.log('Relativized base refs in index.html and sw.js');
+				} catch (error) {
+					console.error('Failed to relativize base refs:', error);
+				}
+			}, 100);
+		}
+	};
+}

tools/ui/scripts/vite-plugin-splash-screen.ts

@@ -0,0 +1,115 @@
+import { readdirSync, readFileSync, writeFileSync, existsSync } from 'node:fs';
+import { resolve } from 'path';
+import type { Plugin } from 'vite';
+import { TAB, NEWLINE } from '../src/lib/constants/code';
+import { APPLE_DEVICES, BUILD_CONFIG, REGEX_PATTERNS, SPLASH_LINK } from '../src/lib/constants/pwa';
+import type { SplashDimensions } from '../src/lib/types';
+import { SplashOrientation } from '../src/lib/enums/splash.enums';
+
+let processed = false;
+
+const OUTPUT_DIR = process.env.LLAMA_UI_OUT_DIR ?? BUILD_CONFIG.OUTPUT_DIR;
+
+/**
+ * Generate iOS splash screen <link> tags from generated apple-splash-*.png files.
+ * Returns an array of HTML link strings to be injected into the page head.
+ */
+export function generateSplashScreenLinks(outDir: string): string[] {
+	const files = readdirSync(outDir).filter((f) => f.match(REGEX_PATTERNS.SPLASH_FILE));
+	if (files.length === 0) return [];
+
+	const dimMap = new Map<string, SplashDimensions>();
+	for (const [dims, spec] of Object.entries(APPLE_DEVICES)) {
+		const [w, h] = dims.split('x').map(Number);
+		// logical-point dimensions
+		dimMap.set(`${w}x${h}`, { deviceW: spec.width, deviceH: spec.height, dpr: spec.dpr });
+		dimMap.set(`${h}x${w}`, { deviceW: spec.width, deviceH: spec.height, dpr: spec.dpr });
+		// pixel dimensions (used by actual generated splash files)
+		dimMap.set(`${w * spec.dpr}x${h * spec.dpr}`, {
+			deviceW: spec.width,
+			deviceH: spec.height,
+			dpr: spec.dpr
+		});
+		dimMap.set(`${h * spec.dpr}x${w * spec.dpr}`, {
+			deviceW: spec.width,
+			deviceH: spec.height,
+			dpr: spec.dpr
+		});
+	}
+
+	const lightLinks: string[] = [];
+	const darkLinks: string[] = [];
+
+	for (const file of files) {
+		const match = file.match(REGEX_PATTERNS.SPLASH_FILE);
+		if (!match) continue;
+		const orientation = match[1] as SplashOrientation;
+		const isDark = !!match[2];
+		const pixelW = parseInt(match[3]);
+		const pixelH = parseInt(match[4]);
+
+		const key = `${pixelW}x${pixelH}`;
+		const spec = dimMap.get(key);
+		if (!spec) {
+			console.warn(`Unknown splash screen dimensions: ${key} (${file})`);
+			continue;
+		}
+
+		const { deviceW, deviceH, dpr } = spec;
+		const media = `screen and (device-width: ${deviceW}px) and (device-height: ${deviceH}px) and (-webkit-device-pixel-ratio: ${dpr}) and (orientation: ${orientation})`;
+		const href = `./${file}`;
+
+		if (isDark) {
+			darkLinks.push(
+				`${SPLASH_LINK.HTML} media="${media}${SPLASH_LINK.DARK_MEDIA_SUFFIX}" href="${href}">`
+			);
+		} else {
+			lightLinks.push(`${SPLASH_LINK.HTML} media="${media}" href="${href}">`);
+		}
+	}
+
+	return [...lightLinks, ...darkLinks];
+}
+
+export function splashScreenPlugin(): Plugin {
+	return {
+		name: 'llamacpp:splash-screen',
+		apply: 'build',
+		closeBundle() {
+			setTimeout(() => {
+				try {
+					if (processed) return;
+					processed = true;
+
+					const outDir = resolve(OUTPUT_DIR);
+					const indexPath = resolve(outDir, 'index.html');
+					if (!existsSync(indexPath)) return;
+
+					let content = readFileSync(indexPath, 'utf-8');
+
+					// Inject iOS splash screen <link> tags into <head>.
+					// The @vite-pwa/assets-generator generates apple-splash-*.png files;
+					// this scans them and creates the <link> tags SvelteKit needs.
+					const splashLinks = generateSplashScreenLinks(outDir);
+					if (splashLinks.length > 0) {
+						console.log(`Generated ${splashLinks.length} apple-splash link tags`);
+						const splashHtml = splashLinks.map((l) => TAB + TAB + l).join(NEWLINE);
+						content = content.replace(
+							REGEX_PATTERNS.HEAD_CLOSE,
+							splashHtml + NEWLINE + TAB + TAB + '</head>'
+						);
+					}
+
+					// Remove trailing \r from Windows line endings
+					content = content.replace(/\r/g, '');
+					content = BUILD_CONFIG.GUIDE_COMMENT + NEWLINE + content;
+
+					writeFileSync(indexPath, content, 'utf-8');
+					console.log('Updated index.html');
+				} catch (error) {
+					console.error('Failed to process build output:', error);
+				}
+			}, 100);
+		}
+	};
+}

tools/ui/src/app.d.ts

@@ -1,6 +1,9 @@
 // See https://svelte.dev/docs/kit/types#app.d.ts
 // for information about these interfaces
 
+import 'vite-plugin-pwa/pwa-assets';
+import 'vite-plugin-pwa/svelte';
+
 // Import chat types from dedicated module
 
 import type {

tools/ui/src/app.html

@@ -2,10 +2,17 @@
 <html lang="en">
 	<head>
 		<meta charset="utf-8" />
-		<link rel="icon" href="%sveltekit.assets%/favicon.svg" />
+		<link rel="icon" href="favicon.ico" sizes="48x48" />
+		<link rel="icon" href="favicon.svg" sizes="any" type="image/svg+xml" />
+
+		<link rel="apple-touch-icon" href="apple-touch-icon-180x180.png" />
+
+		<link rel="manifest" href="./manifest.webmanifest" />
+
 		<meta name="viewport" content="width=device-width, initial-scale=1" />
 		%sveltekit.head%
 	</head>
+
 	<body data-sveltekit-preload-data="hover">
 		<div style="display: contents">%sveltekit.body%</div>
 	</body>

tools/ui/src/lib/components/app/settings/SettingsChat/SettingsChat.svelte

@@ -20,6 +20,8 @@
 	import { ColorMode } from '$lib/enums/ui.enums';
 	import { fade } from 'svelte/transition';
 	import { goto } from '$app/navigation';
+	import { Button } from '$lib/components/ui/button';
+	import { RefreshCw } from '@lucide/svelte';
 	import { page } from '$app/state';
 	import { setChatSettingsConfigContext } from '$lib/contexts';
 	import { settingsReferrer } from '$lib/stores/settings-referrer.svelte';
@@ -164,6 +166,15 @@
 								onConfigChange={handleConfigChange}
 								onThemeChange={handleThemeChange}
 							/>
+
+							{#if currentSection.title === SETTINGS_SECTION_TITLES.GENERAL}
+								<div class="flex justify-end">
+									<Button variant="outline" onclick={() => window.location.reload()}>
+										<RefreshCw class="h-3 w-3" />
+										Reload app
+									</Button>
+								</div>
+							{/if}
 						</div>
 					{/if}
 				</div>

tools/ui/src/lib/components/pwa/PwaMetaTags.svelte

@@ -0,0 +1,23 @@
+<script lang="ts">
+	import { APPLE_META_TAGS, MEDIA_QUERIES, THEME_COLORS } from '$lib/constants/pwa';
+	import { APP_NAME } from '$lib/constants';
+
+	let { appName = APP_NAME } = $props();
+</script>
+
+<svelte:head>
+	<!-- Theme color for light/dark modes -->
+	<meta name="theme-color" content={THEME_COLORS.LIGHT} media={MEDIA_QUERIES.PREFERS_LIGHT} />
+	<meta name="theme-color" content={THEME_COLORS.DARK} media={MEDIA_QUERIES.PREFERS_DARK} />
+
+	<!-- Apple mobile web app meta tags -->
+	<meta
+		name={APPLE_META_TAGS.MOBILE_WEB_APP_CAPABLE.name}
+		content={APPLE_META_TAGS.MOBILE_WEB_APP_CAPABLE.content}
+	/>
+	<meta
+		name={APPLE_META_TAGS.STATUS_BAR_STYLE.name}
+		content={APPLE_META_TAGS.STATUS_BAR_STYLE.content}
+	/>
+	<meta name={APPLE_META_TAGS.MOBILE_WEB_APP_TITLE.name} content={appName} />
+</svelte:head>

tools/ui/src/lib/components/pwa/PwaRefreshAlert.svelte

@@ -0,0 +1,35 @@
+<script lang="ts">
+	import * as Card from '$lib/components/ui/card';
+	import { Button } from '$lib/components/ui/button';
+
+	let { needRefresh: needRefreshProp, updateServiceWorker, forceReload } = $props();
+	let needRefresh = $derived(needRefreshProp ?? false);
+</script>
+
+{#if needRefresh}
+	<Card.Root class="overflow-hidden gap-1 py-5">
+		<Card.Header class="px-5">
+			<Card.Title class="text-sm font-medium">Update available</Card.Title>
+		</Card.Header>
+
+		<Card.Content class="gap-6 grid px-5">
+			<p class="text-xs text-muted-foreground">A new version is available. Reload to update.</p>
+
+			<Button
+				class="justify-self-end-safe"
+				size="sm"
+				onclick={() => {
+					updateServiceWorker();
+
+					if (forceReload) {
+						window.location.reload();
+					}
+
+					needRefresh = false;
+				}}
+			>
+				Reload
+			</Button>
+		</Card.Content>
+	</Card.Root>
+{/if}

tools/ui/src/lib/components/pwa/index.ts

@@ -0,0 +1,2 @@
+export { default as PwaMetaTags } from './PwaMetaTags.svelte';
+export { default as PwaRefreshAlert } from './PwaRefreshAlert.svelte';

tools/ui/src/lib/constants/app.ts

@@ -0,0 +1 @@
+export const APP_NAME = import.meta.env?.VITE_PUBLIC_APP_NAME || 'llama-ui';

tools/ui/src/lib/constants/code.ts

@@ -1,4 +1,5 @@
 export const NEWLINE = '\n';
+export const TAB = '\t';
 export const DEFAULT_LANGUAGE = 'text';
 export const LANG_PATTERN = /^(\w*)\n?/;
 export const AMPERSAND_REGEX = /&/g;

tools/ui/src/lib/constants/index.ts

@@ -3,6 +3,7 @@
 
 export * from './agentic';
 export * from './api-endpoints';
+export * from './app';
 export * from './attachment-labels';
 export * from './database';
 export * from './reasoning-effort';
@@ -36,6 +37,7 @@ export * from './message-export';
 export * from './model-id';
 export * from './precision';
 export * from './processing-info';
+export * from './pwa';
 export * from './routes';
 export * from './sandbox';
 export * from './settings-keys';

tools/ui/src/lib/constants/jpeg-exif.ts

@@ -0,0 +1,30 @@
+/**
+ * JPEG and EXIF binary format constants for orientation parsing.
+ */
+
+/** Bytes of file prefix to scan, the APP1 EXIF segment sits near the start */
+export const EXIF_SCAN_BYTE_LIMIT = 128 * 1024;
+
+/** JPEG start of image marker */
+export const JPEG_SOI_MARKER = 0xffd8;
+
+/** APP1 segment marker byte, carries the EXIF payload */
+export const APP1_MARKER = 0xe1;
+
+/** Start of scan marker byte, compressed data begins and no EXIF follows */
+export const SOS_MARKER = 0xda;
+
+/** "Exif" signature opening the APP1 payload, big endian uint32 */
+export const EXIF_SIGNATURE = 0x45786966;
+
+/** TIFF byte order mark for little endian ("II") */
+export const TIFF_LITTLE_ENDIAN = 0x4949;
+
+/** TIFF magic number following the byte order mark */
+export const TIFF_MAGIC = 42;
+
+/** EXIF tag id holding the orientation value */
+export const EXIF_ORIENTATION_TAG = 0x0112;
+
+/** Size in bytes of one IFD directory entry */
+export const IFD_ENTRY_SIZE = 12;

tools/ui/src/lib/constants/pwa.ts

@@ -0,0 +1,352 @@
+/**
+ * Centralized PWA constants to avoid magic strings, regexes, and duplicated
+ * definitions across the codebase.
+ */
+
+import { APP_NAME } from './app';
+
+export const MEDIA_QUERIES = {
+	PREFERS_DARK: '(prefers-color-scheme: dark)',
+	PREFERS_LIGHT: '(prefers-color-scheme: light)'
+} as const;
+
+export const THEME_COLORS = {
+	LIGHT: '#ffffff',
+	DARK: '#0d0d0d',
+	ACCENT_BLUE: '#2563eb',
+	ACCENT_BLUE_HOVER: '#1d4ed8',
+	BACKGROUND_LIGHT: 'white',
+	BACKGROUND_DARK: '#111111',
+	TITLE_UPDATE_ALERT: {
+		BORDER_LIGHT: 'zinc-200',
+		BORDER_DARK: 'zinc-700',
+		BG_LIGHT: 'white',
+		BG_DARK: 'zinc-800',
+		TEXT_LIGHT: 'zinc-500',
+		TEXT_DARK: 'zinc-400'
+	}
+} as const;
+
+export const FAVICON_PATHS = {
+	ICO_LIGHT: 'favicon.ico',
+	ICO_DARK: 'favicon-dark.ico',
+	SVG_LIGHT: 'favicon.svg',
+	SVG_DARK: 'favicon-dark.svg'
+} as const;
+
+export const FAVICON_SELECTORS = {
+	ICO_48X48: 'link[rel="icon"][sizes="48x48"]',
+	SVG_ANY: 'link[rel="icon"][type="image/svg+xml"]'
+} as const;
+
+export const APPLE_ASSETS = {
+	TOUCH_ICON: 'apple-touch-icon-180x180.png'
+} as const;
+
+export const PWA_MANIFEST = {
+	name: APP_NAME,
+	short_name: APP_NAME,
+	description: 'Local AI chat interface powered by llama.cpp',
+	start_url: './',
+	display: 'standalone' as const,
+	background_color: THEME_COLORS.BACKGROUND_LIGHT,
+	theme_color: THEME_COLORS.BACKGROUND_LIGHT,
+	icons: [
+		{ src: 'pwa-64x64.png', sizes: '64x64', type: 'image/png' },
+		{ src: 'pwa-192x192.png', sizes: '192x192', type: 'image/png' },
+		{ src: 'pwa-512x512.png', sizes: '512x512', type: 'image/png', purpose: 'any' as const },
+		{
+			src: 'maskable-icon-512x512.png',
+			sizes: '512x512',
+			type: 'image/png',
+			purpose: 'maskable' as const
+		}
+	]
+};
+
+export const PWA_ICON_PATHS = {
+	PWA_64: '/pwa-64x64.png',
+	PWA_192: '/pwa-192x192.png',
+	PWA_512: '/pwa-512x512.png',
+	MASKABLE_512: '/maskable-icon-512x512.png'
+} as const;
+
+/** Apple device dimensions (logical points) and DPR, from Apple HIG. */
+export const APPLE_DEVICES = {
+	// iPhones (DPR 3)
+	'1170x2532': { width: 390, height: 844, dpr: 3 }, // iPhone 13, 15
+	'1179x2556': { width: 393, height: 852, dpr: 3 }, // iPhone 14, 15 Pro, 16
+	'1206x2622': { width: 402, height: 874, dpr: 3 }, // iPhone 16 Plus, 16e
+	'1284x2778': { width: 428, height: 926, dpr: 3 }, // iPhone 15 Plus
+	'1290x2796': { width: 430, height: 932, dpr: 3 }, // iPhone 15 Pro Max, 16 Pro
+	'1320x2868': { width: 440, height: 956, dpr: 3 }, // iPhone 16 Pro Max
+	'750x1334': { width: 375, height: 667, dpr: 2 }, // iPhone 6/7/8, 14
+	'640x1136': { width: 320, height: 568, dpr: 2 }, // iPhone 6/7/8 Plus
+	// iPads (DPR 2)
+	'1668x2388': { width: 834, height: 1194, dpr: 2 }, // iPad Air 11", iPad 11"
+	'2048x2732': { width: 1024, height: 1366, dpr: 2 }, // iPad Pro 12.9"
+	'1640x2360': { width: 820, height: 1180, dpr: 2 }, // iPad Air 10.9"
+	'1032x1376': { width: 1032, height: 1376, dpr: 2 }, // iPad Air 13"
+	'744x1133': { width: 376, height: 573, dpr: 2 } // iPad mini 8.3"
+} as const;
+
+export type AppleDeviceKey = keyof typeof APPLE_DEVICES;
+
+export const PWA_FILE_PATHS = {
+	MANIFEST: '/manifest.webmanifest',
+	SERVICE_WORKER: '/sw.js',
+	VERSION: '/version.json',
+	WORKBOX: '/workbox-<hash>.js'
+} as const;
+
+// Used by the server middleware to skip API key validation.
+// Keep in sync with tools/server/server-http.cpp public_endpoints list.
+
+export const PUBLIC_ENDPOINTS = [
+	'/health',
+	'/v1/health',
+	'/models',
+	'/v1/models',
+	'/props',
+	'/metrics',
+	'/',
+	'/index.html',
+
+	'/favicon.ico',
+	'/favicon-dark.ico',
+	'/favicon.svg',
+	'/favicon-dark.svg',
+	'/pwa-64x64.png',
+	'/pwa-192x192.png',
+	'/pwa-512x512.png',
+	'/maskable-icon-512x512.png',
+	'/apple-touch-icon-180x180.png',
+	'/apple-splash-portrait-640x1136.png',
+	'/apple-splash-landscape-640x1136.png',
+	'/apple-splash-portrait-750x1334.png',
+	'/apple-splash-landscape-750x1334.png',
+	'/apple-splash-portrait-1170x2532.png',
+	'/apple-splash-landscape-1170x2532.png',
+	'/apple-splash-portrait-1179x2556.png',
+	'/apple-splash-landscape-1179x2556.png',
+	'/apple-splash-portrait-1206x2622.png',
+	'/apple-splash-landscape-1206x2622.png',
+	'/apple-splash-portrait-1284x2778.png',
+	'/apple-splash-landscape-1284x2778.png',
+	'/apple-splash-portrait-1290x2796.png',
+	'/apple-splash-landscape-1290x2796.png',
+	'/apple-splash-portrait-1320x2868.png',
+	'/apple-splash-landscape-1320x2868.png',
+	'/apple-splash-portrait-1488x2266.png',
+	'/apple-splash-landscape-1488x2266.png',
+	'/apple-splash-portrait-1640x2360.png',
+	'/apple-splash-landscape-1640x2360.png',
+	'/apple-splash-portrait-1668x2388.png',
+	'/apple-splash-landscape-1668x2388.png',
+	'/apple-splash-portrait-2048x2732.png',
+	'/apple-splash-landscape-2048x2732.png',
+	'/apple-splash-portrait-dark-640x1136.png',
+	'/apple-splash-landscape-dark-640x1136.png',
+	'/apple-splash-portrait-dark-750x1334.png',
+	'/apple-splash-landscape-dark-750x1334.png',
+	'/apple-splash-portrait-dark-1170x2532.png',
+	'/apple-splash-landscape-dark-1170x2532.png',
+	'/apple-splash-portrait-dark-1179x2556.png',
+	'/apple-splash-landscape-dark-1179x2556.png',
+	'/apple-splash-portrait-dark-1206x2622.png',
+	'/apple-splash-landscape-dark-1206x2622.png',
+	'/apple-splash-portrait-dark-1284x2778.png',
+	'/apple-splash-landscape-dark-1284x2778.png',
+	'/apple-splash-portrait-dark-1290x2796.png',
+	'/apple-splash-landscape-dark-1290x2796.png',
+	'/apple-splash-portrait-dark-1320x2868.png',
+	'/apple-splash-landscape-dark-1320x2868.png',
+	'/apple-splash-portrait-dark-1488x2266.png',
+	'/apple-splash-landscape-dark-1488x2266.png',
+	'/apple-splash-portrait-dark-1640x2360.png',
+	'/apple-splash-landscape-dark-1640x2360.png',
+	'/apple-splash-portrait-dark-1668x2388.png',
+	'/apple-splash-landscape-dark-1668x2388.png',
+	'/apple-splash-portrait-dark-2048x2732.png',
+	'/apple-splash-landscape-dark-2048x2732.png',
+	'/manifest.webmanifest',
+	'/sw.js',
+	'/version.json',
+	'/workbox-<hash>.js'
+] as const;
+export const BUILD_CONFIG = {
+	OUTPUT_DIR: './dist',
+	GUIDE_COMMENT: `
+<!--
+  This is a static build of the frontend.
+  It is automatically generated by the build process.
+  Do not edit this file directly.
+  To make changes, refer to the "Web UI" section in the README.
+-->
+`.trim()
+} as const;
+
+export const REGEX_PATTERNS = {
+	SPLASH_FILE: /^apple-splash-(portrait|landscape)-(dark-)?(\d+)x(\d+)\.png$/,
+	HEAD_CLOSE: /\t*<\/head>/
+} as const;
+
+// Device names used by @vite-pwa/assets-generator for splash screen generation.
+// Keep in sync with pwa-assets.config.ts.
+export const PWA_GENERATOR_DEVICES = [
+	'iPhone 13',
+	'iPhone 13 Pro',
+	'iPhone 13 Pro Max',
+	'iPhone 14',
+	'iPhone 14 Plus',
+	'iPhone 14 Pro',
+	'iPhone 14 Pro Max',
+	'iPhone 15',
+	'iPhone 15 Plus',
+	'iPhone 15 Pro',
+	'iPhone 15 Pro Max',
+	'iPhone 16',
+	'iPhone 16 Plus',
+	'iPhone 16 Pro',
+	'iPhone 16 Pro Max',
+	'iPhone 16e',
+	'iPhone SE 4"',
+	'iPhone SE 4.7"',
+	'iPad 11"',
+	'iPad Air 10.9"',
+	'iPad Air 11"',
+	'iPad Air 13"',
+	'iPad Pro 11"',
+	'iPad Pro 12.9"',
+	'iPad mini 8.3"'
+] as const;
+
+// PWA assets generator configuration — used by pwa-assets.config.ts
+export const PWA_ASSET_GENERATOR = {
+	LINK_PRESET: '2023',
+	SPLASH_PADDING: 0.75,
+	FIT_MODE: 'contain',
+	ADD_MEDIA_SCREEN: true,
+	BASE_PATH: './',
+	XHTML: false,
+	PNG_COMPRESSION_LEVEL: 9,
+	PNG_QUALITY: 60,
+	DARK_PREFIX: 'dark-'
+} as const;
+
+export const CACHE_SETTINGS = {
+	IMMUTABLE_MAX_AGE_SECONDS: 31536000,
+	API_CACHE_MAX_AGE_SECONDS: 60 * 60 * 24,
+	API_CACHE_MAX_ENTRIES: 50,
+	MAX_FILE_SIZE_BYTES: 10 * 1024 * 1024
+} as const;
+
+export const GLOB_PATTERNS: string[] = [
+	'**/*.{js,css,html,ico,svg,png,webp,woff,woff2,json,webmanifest}'
+];
+
+// loading.html is the model loading page served by llama-server itself.
+// The SvelteKit PWA manifest transform strips the html extension from every
+// precache entry to match clean URLs, but loading.html is a plain static asset
+// with no clean URL, so static servers answer 404 and the SW install fails.
+export const GLOB_IGNORES: string[] = ['**/loading.html'];
+
+export const SW_CONFIG = {
+	CHECK_INTERVAL_MS: 60000,
+	UPDATE_FETCH_OPTIONS: {
+		CACHE: 'no-store',
+		HEADERS: {
+			CACHE: 'no-store',
+			CACHE_CONTROL: 'no-cache'
+		}
+	}
+} as const;
+
+// Runtime caching configuration for Workbox
+export const RUNTIME_CACHING = {
+	HANDLER: 'NetworkFirst',
+	CACHE_NAME: 'api-cache'
+} as const;
+
+// Workbox runtime caching patterns
+export const API_CACHING_PATTERNS = {
+	V1_API: /^\/v1\/.*/,
+	STATIC_API: /^\/(health|props|models|tools|slots|cors-proxy).*/
+} as const;
+
+// SvelteKit PWA plugin options
+export const PWA_KIT_OPTIONS = {
+	NAVIGATE_FALLBACK: './'
+} as const;
+
+export const APPLE_META_TAGS = {
+	MOBILE_WEB_APP_CAPABLE: { name: 'apple-mobile-web-app-capable', content: 'yes' },
+	STATUS_BAR_STYLE: { name: 'apple-mobile-web-app-status-bar-style', content: 'black-translucent' },
+	MOBILE_WEB_APP_TITLE: { name: 'apple-mobile-web-app-title' }
+} as const;
+
+// Splash screen HTML link tag prefix used by generateSplashScreenLinks
+export const SPLASH_LINK = {
+	HTML: '<link rel="apple-touch-startup-image"',
+	DARK_MEDIA_SUFFIX: ' and (prefers-color-scheme: dark)'
+} as const;
+
+// SvelteKit PWA plugin configuration — used by @vite.config.ts
+import type { SvelteKitPWAOptions } from '@vite-pwa/sveltekit';
+
+export const SVELTEKIT_PWA_OPTIONS: SvelteKitPWAOptions = {
+	// Strategy: generateSW - the plugin generates a service worker automatically
+	// using Workbox. For a custom SW, use 'injectManifest' instead.
+	// Manifest configuration
+	manifest: PWA_MANIFEST,
+
+	// Workbox configuration for generateSW strategy
+	workbox: {
+		// Match all static assets in the build output.
+		// Uses '**/' because SvelteKit outputs files under _app/immutable/
+		// subdirectories.
+		globPatterns: GLOB_PATTERNS,
+		globIgnores: GLOB_IGNORES,
+		maximumFileSizeToCacheInBytes: CACHE_SETTINGS.MAX_FILE_SIZE_BYTES,
+
+		// Runtime caching for API calls - use NetworkFirst so APIs are always fresh
+		runtimeCaching: [
+			{
+				urlPattern: API_CACHING_PATTERNS.V1_API,
+				handler: RUNTIME_CACHING.HANDLER,
+				options: {
+					cacheName: RUNTIME_CACHING.CACHE_NAME,
+					expiration: {
+						maxEntries: CACHE_SETTINGS.API_CACHE_MAX_ENTRIES,
+						maxAgeSeconds: CACHE_SETTINGS.API_CACHE_MAX_AGE_SECONDS
+					}
+				}
+			},
+			{
+				urlPattern: API_CACHING_PATTERNS.STATIC_API,
+				handler: RUNTIME_CACHING.HANDLER,
+				options: {
+					cacheName: RUNTIME_CACHING.CACHE_NAME,
+					expiration: {
+						maxEntries: CACHE_SETTINGS.API_CACHE_MAX_ENTRIES,
+						maxAgeSeconds: CACHE_SETTINGS.API_CACHE_MAX_AGE_SECONDS
+					}
+				}
+			}
+		]
+	},
+
+	devOptions: {
+		enabled: true,
+		suppressWarnings: true,
+		// Use PWA_KIT_OPTIONS.NAVIGATE_FALLBACK to match production SW behaviour
+		// (navigateFallback defaults to the configured base path, which is '/' for this SPA).
+		navigateFallback: PWA_KIT_OPTIONS.NAVIGATE_FALLBACK
+	},
+
+	// SvelteKit-specific options
+	kit: {
+		// Include version file for proper cache invalidation
+		includeVersionFile: true
+	}
+};

tools/ui/src/lib/constants/settings-keys.ts

@@ -31,6 +31,7 @@ export const SETTINGS_KEYS = {
 	SHOW_RAW_MODEL_NAMES: 'showRawModelNames',
 	SHOW_MODEL_QUANTIZATION: 'showModelQuantization',
 	SHOW_MODEL_TAGS: 'showModelTags',
+	SHOW_BUILD_VERSION: 'showBuildVersion',
 	SHOW_SYSTEM_MESSAGE: 'showSystemMessage',
 	// Sampling
 	TEMPERATURE: 'temperature',

tools/ui/src/lib/constants/settings-registry.ts

@@ -365,6 +365,14 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 					serverKey: SETTINGS_KEYS.ALWAYS_SHOW_AGENTIC_TURNS,
 					paramType: SyncableParameterType.BOOLEAN
 				}
+			},
+			{
+				key: SETTINGS_KEYS.SHOW_BUILD_VERSION,
+				label: 'Show build version information',
+				help: 'Display the current build version in the bottom-right corner of the interface.',
+				defaultValue: false,
+				type: SettingsFieldType.CHECKBOX,
+				section: SETTINGS_SECTION_SLUGS.DISPLAY
 			}
 		]
 	},

tools/ui/src/lib/constants/storage.ts

@@ -40,6 +40,9 @@ export const DEPRECATED_MCP_DEFAULT_ENABLED_LOCALSTORAGE_KEY = `${STORAGE_APP_NA
 /** @deprecated Use {@link USER_OVERRIDES_LOCALSTORAGE_KEY} instead */
 export const DEPRECATED_USER_OVERRIDES_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME_DEPRECATED}.userOverrides`;
 
+/** Build version stored in localStorage for non-PWA update detection */
+export const BUILD_VERSION_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.buildVersion`;
+
 /** Maps new keys to their deprecated fallback keys */
 export const NEW_TO_DEPRECATED_MAP: Record<string, string> = {
 	[ALWAYS_ALLOWED_TOOLS_LOCALSTORAGE_KEY]: DEPRECATED_ALWAYS_ALLOWED_TOOLS_LOCALSTORAGE_KEY,

tools/ui/src/lib/constants/ui.ts

@@ -5,7 +5,6 @@ import { ROUTES } from './routes';
 
 export const FORK_TREE_DEPTH_PADDING = 8;
 export const SYSTEM_MESSAGE_PLACEHOLDER = 'System message';
-export const APP_NAME = import.meta.env.VITE_PUBLIC_APP_NAME || 'llama-ui';
 
 export const ICON_STRIP_TRANSITION_DURATION = 150;
 export const ICON_STRIP_TRANSITION_DELAY_MULTIPLIER = 50;