spec : auto-detect mtp draft model type (#27005 )

metal : add TQ2_0 support (#26980 )
* metal: add TQ2_0 support Add support for the GGML_TYPE_TQ2_0 (ternary, 2 bits per element) type in the Metal backend. Assisted-by: llama.cpp:DeepSeek-v4-Flash-0731 * cont : optimize mul_mv kernel - float ops over integer ops - precalculate sums - hoist coef out of the inner loop - contiguous y loads llama.cpp:DeepSeek-v4-Flash-0731
2026-08-13 16:26:55 +02:00 · 2026-08-13 13:39:16 +02:00 · 2026-08-13 14:33:53 +03:00 · 2026-08-13 12:34:27 +02:00 · 2026-08-13 12:07:28 +03:00 · 2026-08-13 12:04:46 +03:00
872 changed files with 25332 additions and 12275 deletions
@@ -57,7 +57,6 @@ COPY --from=web /app/tools/ui/dist tools/ui/dist
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DGGML_HIP=ON \
-        -DGGML_HIP_ROCWMMA_FATTN=ON \
        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
@@ -4,6 +4,10 @@ inputs:
  cuda_version:
    description: "CUDA toolkit version"
    required: true
+  cuda_arch:
+    description: "CUDA target architecture"
+    required: false
+    default: "x64"

 runs:
  using: "composite"
@@ -127,3 +131,26 @@ runs:
          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
          echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 13.4 for ARM64
+      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          choco install unzip -y
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cccl-windows-x86_64-13.3.4.1.2-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_crt-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_nvcc-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/libnvvm-windows-x86_64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_cudart-windows-arm64-13.4.46-archive.zip"
+          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/libcublas-windows-arm64-13.7.0.10-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.1.2-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.10-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
@@ -8,8 +8,26 @@ inputs:
 runs:
  using: "composite"
  steps:
-    - name: Setup ROCm
-      uses: ./.github/actions/install-exe
-      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
-        args: -install
+    - name: Install ROCm with Wheels
+      shell: pwsh
+      run: |
+        $ErrorActionPreference = "Stop"
+        write-host "Setting up Python virtual environment"
+
+        # Create the venv directly at the cache location to avoid relocation issues
+        New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
+        python -m venv C:\TheRock\build\.venv
+        & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+        write-host "Upgrading pip"
+        python -m pip install --upgrade pip
+
+        write-host "Installing ROCm wheels for multi-arch support"
+        # Install ROCm wheels for multi-arch support (this may take several minutes)
+        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
+
+        # Pre-expand the devel tree so it is included in the cache
+        write-host "Initializing ROCm devel tree"
+        rocm-sdk init
+        if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+        write-host "Completed ROCm wheel installation to C:\TheRock\build"
@@ -119,27 +119,27 @@ jobs:
          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
          version_full: ${{ env.OPENVINO_VERSION_FULL }}

-  windows-2022-rocm-cache:
-    runs-on: windows-2022
+  # windows-2022-rocm-cache:
+  #   runs-on: windows-2022

-    env:
-      # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
+  #   env:
+  #     # Make sure this is in sync with release.yml and build-cuda-windows.yml
+  #     ROCM_VERSION: "7.14.0"

-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v6

-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-rocm
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+  #     - name: Setup Cache
+  #       uses: actions/cache@v5
+  #       id: cache-rocm
+  #       with:
+  #         path: C:\TheRock\build
+  #         key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}

-      - name: Setup ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+  #     - name: Setup ROCm
+  #       if: steps.cache-rocm.outputs.cache-hit != 'true'
+  #       uses: ./.github/actions/windows-setup-rocm
+  #       with:
+  #         version: ${{ env.ROCM_VERSION }}
@@ -5,7 +5,7 @@ on:

 jobs:
  linux:
-    runs-on: [self-hosted, Linux, CPU]
+    runs-on: [self-hosted, Linux]
    steps:
      - uses: actions/checkout@v6
        with:
@@ -21,15 +21,21 @@ jobs:
                -DLLAMA_BUILD_TOOLS=OFF \
                -DLLAMA_BUILD_EXAMPLES=OFF \
                -DLLAMA_BUILD_APP=OFF \
+                -DLLAMA_BUILD_IS_DEV=OFF \
                -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release
+          cmake --build build --config Release -j $(nproc)
          cmake --install build --prefix "$PREFIX" --config Release

          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
          tclsh <<'EOF'
          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
          set build(number)  [string trim [exec git rev-list  --count HEAD]]
-          set build(version) "0.0.$build(number)"
+
+          set cmakelists [read [open "CMakeLists.txt" r]]
+          regexp {set\(LLAMA_VERSION_MAJOR\s+(\d+)\)} $cmakelists -> major
+          regexp {set\(LLAMA_VERSION_MINOR\s+(\d+)\)} $cmakelists -> minor
+          regexp {set\(LLAMA_VERSION_PATCH\s+(\d+)\)} $cmakelists -> patch
+          set build(version) "$major.$minor.$patch"

          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
@@ -48,4 +54,4 @@ jobs:

          cd examples/simple-cmake-pkg
          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build
+          cmake --build build -j $(nproc)
@@ -94,8 +94,10 @@ jobs:
        id: cmake_build
        run: |
          cmake -B build \
+            -DGGML_NATIVE=OFF \
            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DGGML_NATIVE=OFF
          time cmake --build build --config Release -j $(nproc)

      - name: Test
@@ -99,7 +99,6 @@ jobs:
        run: |
          cmake -B build -S . \
            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
            -DGPU_TARGETS="gfx1030" \
            -DGGML_HIP=ON
          cmake --build build --config Release -j $(nproc)
@@ -83,7 +83,7 @@ jobs:

    env:
      # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      ROCM_VERSION: "7.14.0"

    strategy:
      matrix:
@@ -97,36 +97,53 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      - name: Grab rocWMMA package
-        id: grab_rocwmma
-        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
-
-      - name: Use ROCm Installation Cache
-        uses: actions/cache@v5
-        id: cache-rocm
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+      # - name: Cache ROCm Installation
+      #   uses: actions/cache@v5
+      #   id: cache-rocm
+      #   with:
+      #     path: C:\TheRock\build
+      #     key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}

      - name: Setup ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        # if: steps.cache-rocm.outputs.cache-hit != 'true'
        uses: ./.github/actions/windows-setup-rocm
        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+          version: ${{ env.ROCM_VERSION }}
+
+      - name: Setup ROCm Environment
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH

      - name: Verify ROCm
        id: verify
        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
+          # Test the ROCm clang shipped in the installed wheel
+          & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -134,29 +151,27 @@ jobs:
          # TODO: this build does not match the build in release.yml, so we use a different cache key
          #       ideally, the builds should match, similar to the CUDA build above so that we would be able
          #       to populate the ccache for the release with manual runs of this workflow
-          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}

      - name: Build
        id: cmake_build
        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
            -DCMAKE_BUILD_TYPE=Release `
            -DLLAMA_BUILD_BORINGSSL=ON `
-            -DROCM_DIR="${env:HIP_PATH}" `
+            -DHIP_PATH="${env:HIP_PATH}" `
            -DGGML_HIP=ON `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGPU_TARGETS="gfx1100"  `
+            -DGPU_TARGETS="gfx1100" `
            -DGGML_RPC=ON
          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
@@ -15,6 +15,12 @@ on:
      '**/*.cpp'
    ]

+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-sanitize.yml'
+    ]
+
 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
  cancel-in-progress: true
@@ -28,19 +34,35 @@ env:

 jobs:
  ctest:
-    runs-on: [self-hosted, X64, CPU, Linux]
-
    continue-on-error: true

    strategy:
      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        include:
+        # thread and address doesn't run properly on some self hosted machines, so run it on Github instead
+        - sanitizer: ADDRESS
+          machine: ubuntu-24.04
+        - sanitizer: THREAD
+          machine: ubuntu-24.04
+        - sanitizer: UNDEFINED
+          machine: [self-hosted, X64, Linux]
+
+    runs-on: ${{ matrix.machine }}

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v6

+      # - name: ccache
+      #   uses: ggml-org/ccache-action@v1.2.21
+      #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
+      #   with:
+      #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04
+      #     variant: ccache
+      #     evict-old-files: 1d
+      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
      # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
      - name: Build (undefined)
        id: cmake_build_undefined
@@ -0,0 +1,46 @@
+name: Make Release
+
+on:
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: 'Dry run - validate without creating the tag'
+        required: true
+        type: boolean
+        default: true
+
+env:
+  GH_TOKEN: ${{ github.token }}
+
+permissions:
+  contents: write
+
+jobs:
+  make-release:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Run release checks
+        id: checks
+        run: bash scripts/make-release-checks.sh ${{ github.event.inputs.dry_run == 'true' && '--dry-run' || '' }}
+        env:
+          GITHUB_REPOSITORY: ${{ github.repository }}
+
+      - name: Create release tag
+        if: ${{ github.event.inputs.dry_run == 'false' }}
+        run: |
+          VERSION="${{ steps.checks.outputs.version }}"
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git tag -a "${VERSION}" -m "Release ${VERSION}"
+          git push origin "${VERSION}"
+          echo "Created and pushed tag ${VERSION}"
+
+      - name: Dry run summary
+        if: ${{ github.event.inputs.dry_run == 'true' }}
+        run: |
+          echo "Dry run complete - all checks passed."
+          echo "Would have created tag: ${{ steps.checks.outputs.version }}"
@@ -0,0 +1,23 @@
+name: Convert PR to draft
+
+on:
+  pull_request_target:
+    types: [labeled]
+
+permissions:
+  pull-requests: write
+  issues: write
+  contents: write # required for "gh pr ready" command, see https://github.com/cli/cli/issues/8910
+
+jobs:
+  convert-to-draft:
+    if: github.event.label.name == 'draft' && github.event.pull_request.draft == false
+    runs-on: ubuntu-slim
+    steps:
+      - name: Convert PR to draft
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_URL: ${{ github.event.pull_request.html_url }}
+        run: |
+          gh pr ready --undo "$PR_URL"
+          gh pr edit "$PR_URL" --remove-label draft
@@ -748,6 +748,135 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+  windows-rocm:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
+    runs-on: windows-2022
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+            build: x64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      # - name: Cache ROCm Installation
+      #   id: cache-rocm
+      #   uses: actions/cache@v5
+      #   with:
+      #     path: C:\TheRock\build
+      #     key: rocm-wheels-${{ matrix.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+
+      - name: Setup ROCm
+        # if: steps.cache-rocm.outputs.cache-hit != 'true'
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ matrix.ROCM_VERSION }}
+
+      - name: Setup ROCm Environment
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+          write-host "CMake path: $cmakePath"
+          write-host "Bin path: $binPath"
+
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_HIP=ON `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DHIP_PATH="${env:HIP_PATH}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
+      - name: Verify HIP backend was built
+        run: |
+          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
+          if (-not $hipDll) {
+            Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
+            Write-Host "Contents of build\bin:"
+            Get-ChildItem build\bin | Format-Table -AutoSize
+            exit 1
+          }
+          Write-Host "HIP backend artifact found:"
+          $hipDll | Format-Table FullName, Length -AutoSize
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Get ROCm short version
+        run: |
+          $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
+          echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
+
+      - name: Pack artifacts
+        run: |
+          cp "LICENSE" "build\bin\"
+          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
+          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
+
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -848,6 +977,7 @@ jobs:
          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip

  windows-cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

@@ -858,7 +988,16 @@ jobs:

    strategy:
      matrix:
-        cuda: ['12.4', '13.3']
+        include:
+          - cuda: '12.4'
+            arch: x64
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
+          - cuda: '13.3'
+            arch: x64
+            defines: ''
+          - cuda: '13.4'
+            arch: arm64
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -876,6 +1015,7 @@ jobs:
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -885,54 +1025,62 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Build
        id: cmake_build
        shell: cmd
        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
            -DGGML_BACKEND_DL=ON ^
            -DGGML_NATIVE=OFF ^
            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
+          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip .\build\bin\Release\ggml-cuda.dll

      - name: Upload artifacts
        uses: actions/upload-artifact@v6
        with:
-          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

-      - name: Copy and pack Cuda runtime
+      - name: Copy and pack Cuda runtime (x64)
+        if: ${{ matrix.arch == 'x64' }}
        run: |
          echo "Cuda install location: ${{ env.CUDA_PATH }}"
          $dst='.\build\bin\cudart\'
          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
          robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
+
+      - name: Copy and pack Cuda runtime (ARM64)
+        if: ${{ matrix.arch == 'arm64' }}
+        run: |
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{env.CUDA_PATH}}\bin\arm64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*

      - name: Upload Cuda runtime
        uses: actions/upload-artifact@v6
        with:
-          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

  windows-sycl:
    needs: [check-release]
@@ -1137,250 +1285,123 @@ jobs:
          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, get-version]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+#   ubuntu-22-rocm:
+#     needs: [check-release, get-version]
+#     if: ${{ needs.check-release.outputs.should_release == 'true' }}

-    runs-on: ubuntu-22.04
+#     runs-on: ubuntu-22.04

-    permissions:
-      actions: write
+#     permissions:
+#       actions: write

-    strategy:
-      matrix:
-        include:
-          - ROCM_VERSION: "7.2.1"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
-            build: 'x64'
+#     strategy:
+#       matrix:
+#         include:
+#           - ROCM_VERSION: "7.14.0"
+#             gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+#             build: 'x64'

-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
+#     steps:
+#       - name: Clone
+#         id: checkout
+#         uses: actions/checkout@v6
+#         with:
+#           fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+#       - name: Setup Node.js
+#         uses: actions/setup-node@v6
+#         with:
+#           node-version: "24"
+#           cache: "npm"
+#           cache-dependency-path: "tools/ui/package-lock.json"

-      - name: Free up disk space
-        uses: ggml-org/free-disk-space@v1.3.1
-        with:
-          tool-cache: true
+#       - name: Free up disk space
+#         uses: ggml-org/free-disk-space@v1.3.1
+#         with:
+#           tool-cache: true

-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+#       # - name: ccache
+#       #   uses: ggml-org/ccache-action@v1.2.21
+#       #   with:
+#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}

-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt install -y build-essential git cmake wget
+#       - name: Dependencies
+#         id: depends
+#         run: |
+#           sudo apt install -y build-essential git cmake wget

-      - name: Setup Legacy ROCm
-        if: matrix.ROCM_VERSION == '7.2.1'
-        id: legacy_env
-        run: |
-          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
-          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
-            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+#       - name: Setup TheRock with Wheels
+#         id: therock_env
+#         run: |
+#           # Create Python virtual environment
+#           python3 -m venv .venv
+#           source .venv/bin/activate

-          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
-          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
-          EOF
+#           # Install ROCm wheels for build
+#           # libraries = HIP runtime and CMake configs needed for linking
+#           # devel = compilers, headers, static libs
+#           python -m pip install --upgrade pip
+#           python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"

-          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
-          Package: *
-          Pin: release o=repo.radeon.com
-          Pin-Priority: 600
-          EOF
+#           # Get ROCm installation paths using the rocm-sdk CLI tool
+#           ROCM_PATH=$(rocm-sdk path --root)
+#           CMAKE_PATH=$(rocm-sdk path --cmake)
+#           BIN_PATH=$(rocm-sdk path --bin)
+#           echo "ROCM_PATH=$ROCM_PATH"
+#           echo "CMAKE_PATH=$CMAKE_PATH"
+#           echo "BIN_PATH=$BIN_PATH"

-          sudo apt update
-          sudo apt-get install -y libssl-dev rocm-hip-sdk
+#           # Set environment variables
+#           echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
+#           echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
+#           echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
+#           echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
+#           echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV

-      - name: Setup TheRock
-        if: matrix.ROCM_VERSION != '7.2.1'
-        id: therock_env
-        run: |
-          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
-          mkdir install
-          tar -xf *.tar.gz -C install
-          export ROCM_PATH=$(pwd)/install
-          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
-          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
-          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
+#           # Keep venv activated for subsequent steps
+#           echo "$(pwd)/.venv/bin" >> $GITHUB_PATH

-      - name: Build with native CMake HIP support
-        id: cmake_build
-        run: |
-          cmake -B build -S . \
-            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_NATIVE=OFF \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DGGML_CPU_ALL_VARIANTS=ON \
-            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
-            -DGGML_HIP=ON \
-            -DHIP_PLATFORM=amd \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j $(nproc)
+#       - name: Build with native CMake HIP support
+#         id: cmake_build
+#         run: |
+#           cmake -B build -S . \
+#             -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+#             -DCMAKE_BUILD_TYPE=Release \
+#             -DGGML_BACKEND_DL=ON \
+#             -DGGML_NATIVE=OFF \
+#             -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+#             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+#             -DGGML_CPU_ALL_VARIANTS=ON \
+#             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
+#             -DGGML_HIP=ON \
+#             -DHIP_PLATFORM=amd \
+#             -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
+#             ${{ env.CMAKE_ARGS }}
+#           cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+#       # - name: ccache-clear
+#       #   uses: ./.github/actions/ccache-clear
+#       #   with:
+#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}

-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+#       - name: Determine tag name
+#         id: tag
+#         uses: ./.github/actions/get-tag-name

-      - name: Get ROCm short version
-        run: echo "ROCM_VERSION_SHORT=$(echo '${{ matrix.ROCM_VERSION }}' | cut -d '.' -f 1,2)" >> $GITHUB_ENV
+#       - name: Get ROCm short version
+#         run: echo "ROCM_VERSION_SHORT=$(echo '${{ matrix.ROCM_VERSION }}' | cut -d '.' -f 1,2)" >> $GITHUB_ENV

-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .
+#       - name: Pack artifacts
+#         id: pack_artifacts
+#         run: |
+#           cp LICENSE ./build/bin/
+#           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .

-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
-          name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
-
-  windows-hip:
-    needs: [check-release, get-version]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: windows-2022
-
-    permissions:
-      actions: write
-
-    env:
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-
-    strategy:
-      matrix:
-        include:
-          - name: "radeon"
-            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
-      - name: Grab rocWMMA package
-        id: grab_rocwmma
-        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
-
-      - name: Cache ROCm Installation
-        id: cache-rocm
-        uses: actions/cache@v5
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-
-      - name: Install ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        id: depends
-        run: |
-          $ErrorActionPreference = "Stop"
-          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP SDK"
-          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $completed = $proc.WaitForExit(600000)
-          if (-not $completed) {
-              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
-              $proc.Kill()
-              exit 1
-          }
-          if ($proc.ExitCode -ne 0) {
-              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
-              exit 1
-          }
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        id: verify
-        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
-
-      - name: Build
-        id: cmake_build
-        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_BACKEND_DL=ON `
-            -DGGML_NATIVE=OFF `
-            -DGGML_CPU=OFF `
-            -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 ${env:NUMBER_OF_PROCESSORS}
-          md "build\bin\rocblas\library\"
-          md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
-          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
-          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+#       - name: Upload artifacts
+#         uses: actions/upload-artifact@v6
+#         with:
+#           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
+#           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz

  ios-xcode:
    needs: [check-release, get-version]
@@ -1555,9 +1576,9 @@ jobs:
      - windows-cpu
      - windows-cuda
      #- windows-sycl
-      - windows-hip
+      - windows-rocm
      - windows-openvino
-      - ubuntu-22-rocm
+      #- ubuntu-22-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
@@ -1667,7 +1688,7 @@ jobs:
            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
            - [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 (ROCm 7.14)[DISABLED](https://github.com/ggml-org/llama.cpp/pull/26969)
            - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ 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-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1681,10 +1702,11 @@ jobs:
            - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
            - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ 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/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
+            - [Windows arm64 (CUDA 13) (preview)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-arm64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-arm64.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 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ 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-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
+            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)

            **openEuler:**
            - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)
@@ -25,6 +25,12 @@ on:
      'tools/server/**.*'
    ]

+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/server-sanitize.yml'
+    ]
+
 env:
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
@@ -90,23 +96,27 @@ jobs:

      - name: Python setup
        id: setup_python
-        uses: actions/setup-python@v6
-        with:
-          python-version: '3.11'
-          pip-install: -r tools/server/tests/requirements.txt
+        uses: actions/setup-python@v7
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m venv .venv
+          .venv/bin/pip install -r tools/server/tests/requirements.txt

      - name: Tests
        id: server_integration_tests
        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
        run: |
+          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
+          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh
@@ -72,7 +72,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +99,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

  server-cuda:
    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
@@ -132,7 +132,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -141,7 +141,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -150,7 +150,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -159,7 +159,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

  server-kleidiai:
    runs-on: ah-ubuntu_22_04-c8g_8x
@@ -219,4 +219,4 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh
@@ -104,21 +104,21 @@ jobs:
        id: server_integration_tests
        run: |
          cd tools/server/tests
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
        run: |
          cd tools/server/tests
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh

      - name: Tests (Backend sampling)
        id: server_integration_tests_backend_sampling
        run: |
          cd tools/server/tests
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh

      - name: Slow tests (Backend sampling)
        id: server_integration_tests_slow_backend_sampling
@@ -126,7 +126,7 @@ jobs:
        run: |
          cd tools/server/tests
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh

  windows:
    runs-on: windows-2025
@@ -167,15 +167,17 @@ jobs:

      - name: Tests
        id: server_integration_tests
+        shell: bash
        run: |
          cd tools/server/tests
-          $env:PYTHONIOENCODING = ":replace"
-          pytest -v -x -m "not slow"
+          export PYTHONIOENCODING=":replace"
+          ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
+        shell: bash
        run: |
          cd tools/server/tests
-          $env:SLOW_TESTS = "1"
-          pytest -v -x
+          export SLOW_TESTS="1"
+          ./tests.sh
@@ -19,6 +19,8 @@ jobs:
        run: |
          cargo binstall komac@2.16.0 -y

+      # TODO: This should later be updated to publish releases instead of
+      # development release builds.
      - name: Find latest release
        id: find_latest_release
        uses: actions/github-script@v8
@@ -2,6 +2,26 @@ cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)

+### llama.cpp version
+set(LLAMA_VERSION_MAJOR 0)
+set(LLAMA_VERSION_MINOR 1)
+set(LLAMA_VERSION_PATCH 0)
+set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
+
+# whether this is a development/nightly build
+# set this to OFF when making a release from a release tag (vX.Y.Z)
+# ref: https://github.com/ggml-org/ggml/discussions/1579
+option(LLAMA_BUILD_IS_DEV "llama: dev build" ON)
+
+if (LLAMA_BUILD_IS_DEV)
+    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}-dev")
+else()
+    # TODO: check that the current commit is tagged correctly according to the version specified above
+    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}")
+endif()
+
+message(STATUS "llama.cpp version: ${LLAMA_VERSION}")
+
 #set(CMAKE_WARN_DEPRECATED YES)
 set(CMAKE_WARN_UNUSED_CLI YES)

@@ -24,9 +44,6 @@ if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
    set(LLAMA_STANDALONE ON)

    include(git-vars)
-
-    # configure project version
-    # TODO
 else()
    set(LLAMA_STANDALONE OFF)
 endif()
@@ -139,7 +156,6 @@ endif()
 if (NOT DEFINED LLAMA_BUILD_COMMIT)
    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
 endif()
-set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})

 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
@@ -275,12 +291,12 @@ configure_package_config_file(
              LLAMA_BIN_INSTALL_DIR )

 write_basic_package_version_file(
-        ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
-    VERSION ${LLAMA_INSTALL_VERSION}
+        ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
+        VERSION ${LLAMA_VERSION}
    COMPATIBILITY SameMajorVersion)

 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)

 configure_file(cmake/llama.pc.in
@@ -12,7 +12,7 @@
 [![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)

-[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev branches](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-features.md) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

 </div>

@@ -106,6 +106,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 - [XCFramework](docs/xcframework.md)
 - [Completions](docs/completions.md)
 - [Models](docs/models.md)
+- [Release process](docs/release.md)

 ## Contributing

@@ -1,5 +1,7 @@
 #include "build-info.h"

+#include "llama.h"
+
 #include <cstdio>
 #include <cstdlib>
 #include <string>
@@ -77,12 +79,12 @@ static const command cmds[] = {

 #undef UPDATE_HIDDEN

-static int version(int argc, char ** argv) {
-    printf("%s\n", llama_build_info());
+static int version(int /*argc*/, char ** /*argv*/) {
+    llama_print_build_info(llama_version());
    return 0;
 }

-static int licenses(int argc, char ** argv) {
+static int licenses(int /*argc*/, char ** /*argv*/) {
    for (int i = 0; LICENSES[i]; ++i) {
        printf("%s\n", LICENSES[i]);
    }
@@ -49,6 +49,14 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")

+# gpu-rocm self-hosted runner can't upload logs to blob; keep each run's logs in
+# their own dir keyed by the GitHub run id so an Actions run URL maps to its logs.
+if [ -n "${GG_BUILD_ROCM}" ] && [ -n "${GITHUB_RUN_ID}" ]; then
+    OUT="$OUT/run-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
+    mkdir -p "$OUT"
+    echo "ci results dir: $OUT"
+fi
+
 rm -f $OUT/*.log
 rm -f $OUT/*.exit
 rm -f $OUT/*.md
@@ -92,7 +100,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi

 if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON -DGGML_HIP_ROCWMMA_FATTN=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON"
    if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
        echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
        exit 1
@@ -0,0 +1,26 @@
+# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+if ( DEFINED CUDAToolkit_ROOT )
+    file( TO_CMAKE_PATH "${CUDAToolkit_ROOT}" CUDA_ROOT )
+elseif ( DEFINED ENV{CUDA_PATH} )
+    file( TO_CMAKE_PATH "$ENV{CUDA_PATH}" CUDA_ROOT )
+else()
+    message( FATAL_ERROR "Set CUDAToolkit_ROOT or CUDA_PATH to a Windows CUDA Toolkit with ARM64 target libraries" )
+endif()
+
+if ( DEFINED ENV{VCToolsInstallDir} )
+    file( TO_CMAKE_PATH "$ENV{VCToolsInstallDir}" MSVC_TOOLS_ROOT )
+    set( CMAKE_CUDA_HOST_COMPILER "${MSVC_TOOLS_ROOT}/bin/Hostx64/arm64/cl.exe" CACHE FILEPATH "" )
+endif()
+
+set( CMAKE_CUDA_COMPILER "${CUDA_ROOT}/bin/nvcc.exe" CACHE FILEPATH "" )
+set( CMAKE_CUDA_FLAGS_INIT "-target-dir=arm64" )
+
+# FindCUDAToolkit selects lib/x64 from the host architecture on Windows.
+set( CUDA_CUDART              "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cudart_LIBRARY      "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
+set( CUDA_cublas_LIBRARY      "${CUDA_ROOT}/lib/arm64/cublas.lib"   CACHE FILEPATH "" )
+set( CUDA_cublasLt_LIBRARY    "${CUDA_ROOT}/lib/arm64/cublasLt.lib" CACHE FILEPATH "" )
+set( CUDA_cuda_driver_LIBRARY "${CUDA_ROOT}/lib/arm64/cuda.lib"     CACHE FILEPATH "" )
@@ -1,4 +1,4 @@
-set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
+set(LLAMA_VERSION      @LLAMA_VERSION@)
 set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
@@ -5,6 +5,6 @@ includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@

 Name: llama
 Description: Port of Facebook's LLaMA model in C/C++
-Version: @LLAMA_INSTALL_VERSION@
+Version: @LLAMA_VERSION@
 Libs: -L${libdir} -lggml -lggml-base -lllama
 Cflags: -I${includedir}
@@ -121,8 +121,8 @@ add_library(${TARGET}
    )

 set_target_properties(${TARGET} PROPERTIES
-    VERSION ${LLAMA_INSTALL_VERSION}
-    SOVERSION 0
+    VERSION ${LLAMA_VERSION_BASE}
+    SOVERSION ${LLAMA_VERSION_MAJOR}
    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )

@@ -35,6 +35,7 @@
 #include <regex>
 #include <set>
 #include <string>
+#include <system_error>
 #include <thread> // for hardware_concurrency
 #include <vector>

@@ -560,6 +561,15 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        }
    }

+    // infer the speculative type from the draft GGUF metadata when none is requested
+    // note: reads only the first split - sharded drafts need an explicit --spec-type
+    if (spec_types_is_default(params) && !params.speculative.draft.mparams.path.empty()) {
+        const auto types_gguf = common_speculative_types_from_gguf(params.speculative.draft.mparams.path);
+        if (!types_gguf.empty()) {
+            params.speculative.types = types_gguf;
+        }
+    }
+
    // when a sidecar type is requested, the draft repo resolves to its sidecar instead of a full model
    const bool spec_sidecar_found = !plan_spec.mtp.local_path.empty() ||
                                    !plan_spec.dflash.local_path.empty() ||
@@ -704,12 +714,61 @@ void common_models_handler_apply(common_models_handler & handler, common_params
 // CLI argument parsing functions
 //

+// apply config files (if present), a later file overrides an earlier one:
+// 1. system-wide: /etc/llama.cpp/config.ini (%PROGRAMDATA%\llama.cpp\config.ini on windows)
+// 2. user-level: ${XDG_CONFIG_HOME:-~/.config}/llama.cpp/config.ini (%APPDATA%\llama.cpp\config.ini on windows)
+static void common_params_apply_system_config(common_params & params, llama_example ex) {
+    std::vector<std::string> paths;
+
+#if defined(_WIN32)
+    const std::string program_data = common_get_env("PROGRAMDATA");
+    if (!program_data.empty()) {
+        paths.push_back(program_data + "\\llama.cpp\\config.ini");
+    }
+#else
+    paths.push_back("/etc/llama.cpp/config.ini");
+#endif
+
+    try {
+        paths.push_back(fs_get_config_directory() + "config.ini");
+    } catch (const std::exception & e) {
+        LOG_DBG("cannot read user-level config file, skipping: %s\n", e.what());
+    }
+
+    std::vector<std::string> found;
+    for (const auto & path : paths) {
+        std::error_code ec;
+        if (std::filesystem::exists(path, ec)) {
+            found.push_back(path);
+        }
+    }
+    if (found.empty()) {
+        return;
+    }
+
+    common_preset_context ctx(ex);
+    ctx.ignore_unknown_keys = true; // the same config file is shared by all programs
+    for (const auto & path : found) {
+        LOG_INF("using config file: %s\n", path.c_str());
+        common_preset global;
+        common_presets presets = ctx.load_from_ini(path, global);
+        global.apply_to_params(params);
+        auto it = presets.find(COMMON_PRESET_DEFAULT_NAME);
+        if (it != presets.end()) {
+            it->second.apply_to_params(params);
+        }
+    }
+}
+
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
    common_params & params = ctx_arg.params;

    // setup log directly from params.verbosity: see tools/cli/cli.cpp
    common_log_set_verbosity_thold(params.verbosity);

+    // config file applies first, so env variables and CLI arguments override it
+    common_params_apply_system_config(params, ctx_arg.ex);
+
    std::unordered_map<std::string, std::pair<common_arg *, bool>> arg_to_options;
    for (auto & opt : ctx_arg.options) {
        for (const auto & arg : opt.args) {
@@ -1390,8 +1449,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--version"},
        "show version and build info",
        [](common_params &) {
-            fprintf(stderr, "version: %d (%s)\n", llama_build_number(), llama_commit());
-            fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
+            llama_print_build_info(llama_version());
            exit(0);
        }
    ));
@@ -2605,14 +2663,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_env("LLAMA_ARG_DIO"));
    add_opt(common_arg(
        {"-lm", "--load-mode"}, "MODE",
-        "model loading mode (default: mmap)\n"
+        "model loading mode (default: auto)\n"
+        "- auto: mmap, unless a device does not support it\n"
        "- none: no special loading mode\n"
        "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
        "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
        "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
        "- dio: use DirectIO if available\n",
        [](common_params & params, const std::string & value) {
-            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            /**/ if (value == "auto")       { params.load_mode = LLAMA_LOAD_MODE_AUTO;       }
+            else if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
            else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
            else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
            else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }
@@ -3308,6 +3368,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.server_tools = parse_csv_row(value);
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
+    add_opt(common_arg(
+        {"--tools-runtime"}, "OPTION",
+        "experimental: run tools in a separate runtime environment (default: none, use host environment)\n"
+        "available options:\n"
+        "  'docker:<image>', 'podman:<image>': spin up a new container and reuse it for all invocations, clean up on server exit\n"
+        "  'docker-container:<id>', 'podman-container:<id>': use an existing container by ID, won't stop on server exit\n"
+        "  'ssh:<target>': run tools on a remote POSIX host over SSH, key-based auth and a trusted host key are required\n",
+        [](common_params & params, const std::string & value) {
+            params.server_tools_runtime = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS_RUNTIME"));
    add_opt(common_arg(
        {"--mcp-servers-config"}, "PATH",
        "experimental: path to JSON file with MCP server definitions (Cursor-compatible format) - do not enable in untrusted environments (default: none)\n"
@@ -29,7 +29,7 @@ const char * llama_build_info(void) {
    return s.c_str();
 }

-void llama_print_build_info(void) {
-    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, llama_build_number(), llama_commit());
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, llama_compiler(), llama_build_target());
+void llama_print_build_info(const char * llama_version) {
+    fprintf(stderr, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
+    fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
 }
@@ -8,4 +8,4 @@ const char * llama_compiler(void);
 const char * llama_build_target(void);
 const char * llama_build_info(void);

-void llama_print_build_info(void);
+void llama_print_build_info(const char *);
@@ -1166,6 +1166,16 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
        data.prompt += data.generation_prompt;
    }

+    std::vector<std::string> tool_call_starts = { "<tool_call>" };
+
+    // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
+    // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
+    // constraining on <function which may occur in valid content generation, e.g. #include <functional>
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const std::string name = tool.at("function").at("name");
+        tool_call_starts.push_back("<function=" + name + ">");
+    });
+
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto generation_prompt = p.literal(GEN_PREFIX);

@@ -1238,7 +1248,7 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));

            return generation_prompt +
-                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
+                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
        }

        // Content only parser
@@ -1264,12 +1274,9 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
        });

        if (data.grammar_lazy) {
-            data.grammar_triggers = {
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
-                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
-                // the token with the function name e.g. "=read"
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
-            };
+            for (const auto & start : tool_call_starts) {
+                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
+            }
        }
    }

@@ -3086,6 +3093,153 @@ static common_chat_params common_chat_params_init_minicpm5(const common_chat_tem
    return data;
 }

+// An assistant turn is rendered as one or more messages, each
+// "<|start|>assistant to=<recipient><|message|>{content}{END}" where END is
+// <|eom|> (more messages follow) or <|eot|> (end of turn):
+//   - chain-of-thought: to=self, terminated by <|eom|>
+//   - final answer:     to=user, terminated by <|eot|>
+// The generation prompt is just "<|start|>assistant"; the model emits its own
+// " to=...<|message|>".
+static common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template &          tmpl,
+                                                               const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = "<|start|>assistant";
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    data.preserved_tokens = {
+        "<|start|>", "<|message|>", "<|eom|>", "<|eot|>",
+        // ATEM tool-call markup emitted on " to=<tool>" turns.
+        "<atem:function_calls>", "<atem:invoke", "<atem:parameter", "</atem:parameter>",
+        "</atem:invoke>", "</atem:function_calls>",
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>tool"      },
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|start|>assistant to=self<|message|>" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<|eom|><|start|>assistant to=user<|message|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    // Constrained grammar whenever tools are offered.
+    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto start = p.rule("start", p.literal("<|start|>assistant"));
+
+        if (!extract_reasoning && !include_grammar) {
+            return start + p.content(p.rest());
+        }
+
+        if (extract_reasoning) {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.reasoning(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        } else {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.content(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        }
+        auto analysis = p.ref("analysis");
+
+        auto recipient  = p.optional(p.literal(" to=user"));
+        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
+                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto string_value = p.ac(
+                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
+                "</atem:parameter>");
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto &      function = tool.at("function");
+                const std::string name     = function.at("name");
+                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto args = p.eps();
+                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
+                    auto schema_info = common_schema_info();
+                    schema_info.resolve_refs(params);
+
+                    auto arg_choice = p.choice();
+                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
+                        auto value_parser = p.eps();
+                        if (schema_info.resolves_to_string(prop_schema)) {
+                            value_parser = string_value;
+                        } else {
+                            value_parser = p.tool_arg_json_value(
+                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
+                                + p.tool_arg_close(p.literal("</atem:parameter>"));
+                        }
+
+                        auto arg_rule = p.tool_arg(
+                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
+                            value_parser);
+
+                        arg_choice |= arg_rule;
+                    }
+                    args = p.zero_or_more(arg_choice + p.space());
+                }
+
+                auto tool_parser = p.tool(
+                    p.tool_open(p.literal(" to=") + p.until("<|message|>") +
+                                p.literal("<|message|><atem:function_calls>") + p.space() +
+                                p.literal("<atem:invoke name=\"") + p.tool_name(p.literal(name)) + p.literal("\">") + p.space())
+                    << p.tool_args(args)
+                    << p.tool_close(p.literal("</atem:invoke>") + p.space() + p.literal("</atem:function_calls>")));
+
+                tool_choice |= p.rule("tool-" + name, tool_parser);
+            });
+
+            auto tool_calls = inputs.parallel_tool_calls
+                ? p.trigger_rule("tool-call", tool_choice + p.zero_or_more(p.literal("<|eom|>") + start + tool_choice))
+                : p.trigger_rule("tool-call", tool_choice);
+
+
+            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+                return p.zero_or_more(start + analysis) + start + tool_calls;
+            }
+            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
+            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
+        }
+
+        return p.zero_or_more(start + analysis) + start + final_msg;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
+        };
+    }
+
+    return data;
+}
+
 static json common_chat_extra_context() {
    json ctx = json::object();
    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
@@ -3114,6 +3268,12 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
        return common_chat_params_init_gpt_oss(tmpl, params);
    }

+    // Muse Glimmer format using " to=<recipient>" recipients and <|eom|>/<|eot|> message terminators.
+    if (src.find("<atem:function_calls>") != std::string::npos && src.find("<|eom|>") != std::string::npos) {
+        LOG_DBG("Using specialized template: Muse Glimmer\n");
+        return common_chat_params_init_muse_glimmer(tmpl, params);
+    }
+
    // Functionary v3.2 - uses recipient-based format with >>>recipient\n{content}
    // Detection: template has ">>>all" for content and ">>>" prefix for tool calls
    if (src.find(">>>all") != std::string::npos && src.find(">>>${recipient}") != std::string::npos) {
@@ -1019,20 +1019,21 @@ std::string fs_get_cache_directory() {
    std::string cache_directory = "";
    auto ensure_trailing_slash = [](std::string p) {
        // Make sure to add trailing slash
-        if (p.back() != DIRECTORY_SEPARATOR) {
+        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
            p += DIRECTORY_SEPARATOR;
        }
        return p;
    };
-    if (getenv("LLAMA_CACHE")) {
-        cache_directory = std::getenv("LLAMA_CACHE");
-    } else {
+    cache_directory = common_get_env("LLAMA_CACHE");
+    if (cache_directory.empty()) {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
        defined(__OpenBSD__) || defined(__NetBSD__)
-        if (std::getenv("XDG_CACHE_HOME")) {
-            cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else if (std::getenv("HOME")) {
-            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        const std::string xdg_cache_home = common_get_env("XDG_CACHE_HOME");
+        const std::string home           = common_get_env("HOME");
+        if (!xdg_cache_home.empty()) {
+            cache_directory = xdg_cache_home;
+        } else if (!home.empty()) {
+            cache_directory = home + "/.cache/";
        } else {
 #if defined(__linux__)
            /* no $HOME is defined, fallback to getpwuid */
@@ -1047,9 +1048,16 @@ std::string fs_get_cache_directory() {
 #endif /* defined(__linux__) */
        }
 #elif defined(__APPLE__)
-        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+        cache_directory = common_get_env("HOME");
+        if (cache_directory.empty()) {
+            throw std::runtime_error("Failed to find $HOME directory");
+        }
+        cache_directory += "/Library/Caches/";
 #elif defined(_WIN32)
-        cache_directory = std::getenv("LOCALAPPDATA");
+        cache_directory = common_get_env("LOCALAPPDATA");
+        if (cache_directory.empty()) {
+            throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
+        }
 #elif defined(__EMSCRIPTEN__)
        GGML_ABORT("not implemented on this platform");
 #else
@@ -1061,6 +1069,51 @@ std::string fs_get_cache_directory() {
    return ensure_trailing_slash(cache_directory);
 }

+std::string fs_get_config_directory() {
+    std::string config_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
+        defined(__OpenBSD__) || defined(__NetBSD__) || defined(__APPLE__)
+    const std::string xdg_config_home = common_get_env("XDG_CONFIG_HOME");
+    const std::string home            = common_get_env("HOME");
+    if (!xdg_config_home.empty()) {
+        config_directory = xdg_config_home;
+    } else if (!home.empty()) {
+        config_directory = home + "/.config/";
+    } else {
+#if defined(__linux__)
+        /* no $HOME is defined, fallback to getpwuid */
+        struct passwd *pw = getpwuid(getuid());
+        if ((!pw) || (!pw->pw_dir)) {
+            throw std::runtime_error("Failed to find $HOME directory");
+        }
+
+        config_directory = std::string(pw->pw_dir) + std::string("/.config/");
+#else
+        throw std::runtime_error("Failed to find $HOME directory");
+#endif
+    }
+#elif defined(_WIN32)
+    config_directory = common_get_env("APPDATA");
+    if (config_directory.empty()) {
+        throw std::runtime_error("Failed to find %APPDATA% directory");
+    }
+#elif defined(__EMSCRIPTEN__)
+    // caller decides what to do when there is no config directory
+    throw std::runtime_error("not implemented on this platform");
+#else
+#  error Unknown architecture
+#endif
+    config_directory = ensure_trailing_slash(config_directory);
+    config_directory += "llama.cpp";
+    return ensure_trailing_slash(config_directory);
+}
+
 std::string fs_get_cache_file(const std::string & filename) {
    GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
    std::string cache_directory = fs_get_cache_directory();
@@ -1639,6 +1692,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.n_seq_max         = params.n_parallel;
    cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
    cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
+    cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
    cparams.n_batch           = params.n_batch;
    cparams.n_ubatch          = params.n_ubatch;
    cparams.n_threads         = params.cpuparams.n_threads;
@@ -447,6 +447,7 @@ struct common_params {
    int32_t n_parallel            =     1; // number of parallel sequences to decode
    int32_t n_sequences           =     1; // number of sequences to decode
    int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
+    int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
    int32_t grp_attn_n            =     1; // group-attention factor
    int32_t grp_attn_w            =   512; // group-attention width
    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -472,7 +473,7 @@ struct common_params {
    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);

    enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
+    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model

    common_cpu_params cpuparams;
    common_cpu_params cpuparams_batch;
@@ -655,6 +656,7 @@ struct common_params {

    // enable built-in tools
    std::vector<std::string> server_tools;
+    std::string server_tools_runtime;

    // MCP server configs (Cursor-compatible JSON)
    std::string mcp_servers_config;   // path to JSON file with MCP server definitions
@@ -879,6 +881,7 @@ bool fs_is_directory(const std::string & path);

 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
+std::string fs_get_config_directory();

 struct common_file_info {
    std::string path;
@@ -116,6 +116,8 @@ static llama_sampler_i llama_sampler_llg_i = {
    /* .backend_accept    = */ NULL,
    /* .backend_apply     = */ NULL,
    /* .backend_set_input = */ NULL,
+    /* .backend_reset     = */ NULL,
+    /* .copy_state        = */ NULL,
 };

 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
@@ -570,23 +570,34 @@ struct parser_executor {
    }

    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
+        auto save = pos;
+
        ++pos; // consume '\'
        if (pos >= ctx.input.size()) {
            if (!ctx.is_lenient()) {
                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
            }
+            pos = save; // suppress unmatched '\'
            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
        }

        char c = ctx.input[pos];
+
        if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
            ++pos;
            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
-        } else if (c == 'u') {
-            return handle_unicode_escape(ctx, start, pos);
-        } else {
-            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
        }
+
+        if (c == 'u') {
+            auto result = handle_unicode_escape(ctx, start, pos);
+            if (result.need_more_input()) {
+                pos = save; // suppress incomplete sequence
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
+            }
+            return result;
+        }
+
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
    }

    static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {
@@ -322,6 +322,8 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
                    preset.options[opt] = value;
                }
                LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
+            } else if (ignore_unknown_keys) {
+                LOG_WRN("ignoring option '%s' from %s: not supported by this program\n", key.c_str(), path.c_str());
            } else {
                throw std::runtime_error(string_format(
                    "option '%s' not recognized in preset '%s'",
@@ -59,6 +59,10 @@ struct common_preset_context {
    bool filter_allowed_keys = false;
    std::set<std::string> allowed_keys;

+    // if true, options unknown to the current example are skipped instead of being an error
+    // used for config files shared by all binaries, where each binary only knows a subset of options
+    bool ignore_unknown_keys = false;
+
    // if only_remote_allowed is true, only accept whitelisted keys
    common_preset_context(llama_example ex);

@@ -217,6 +217,8 @@ static struct llama_sampler_i common_reasoning_budget_i = {
    /* .backend_accept    = */ nullptr,
    /* .backend_apply     = */ nullptr,
    /* .backend_set_input = */ nullptr,
+    /* .backend_reset     = */ nullptr,
+    /* .copy_state        = */ nullptr,
 };

 static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
@@ -518,6 +518,26 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
    };
 }

+void common_sampler_copy(const common_sampler * src, common_sampler * dst) {
+    if (!src || !dst || src == dst) {
+        return;
+    }
+
+    GGML_ASSERT((src->grmr == nullptr) == (dst->grmr == nullptr));
+    GGML_ASSERT((src->rbudget == nullptr) == (dst->rbudget == nullptr));
+
+    llama_sampler_copy(src->grmr,    dst->grmr);
+    llama_sampler_copy(src->rbudget, dst->rbudget);
+    llama_sampler_copy(src->chain,   dst->chain);
+
+    dst->params     = src->params;
+    dst->prev       = src->prev;
+    dst->cur        = src->cur;
+    dst->cur_p      = src->cur_p;
+    dst->cur_p.data = src->cur_p.data ? dst->cur.data() : nullptr; // re-point to dst's buffer
+    dst->t_total_us = src->t_total_us;
+}
+
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
    // TODO: measure grammar performance

@@ -47,6 +47,7 @@ void common_sampler_free(struct common_sampler * gsmpl);
 void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated);
 void                    common_sampler_reset (struct common_sampler * gsmpl);
 struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
+void                    common_sampler_copy  (const struct common_sampler * src, struct common_sampler * dst);

 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
@@ -2,6 +2,7 @@

 #include "common.h"
 #include "ggml.h"
+#include "ggml-cpp.h"
 #include "llama.h"
 #include "log.h"
 #include "ngram-cache.h"
@@ -171,12 +172,6 @@ struct common_speculative_impl {
    // (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
    virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
    virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}
-
-    // true if this implementation requires the target context to extract post-norm embeddings
-    virtual bool need_embd() const = 0;
-
-    // true if this implementation requires the target context to extract pre-norm embeddings
-    virtual bool need_embd_nextn() const { return false; }
 };

 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -193,6 +188,10 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        auto * ctx_dft = this->params.ctx_dft;
        auto * ctx_tgt = this->params.ctx_tgt;

+        if (!ctx_dft) {
+            throw std::runtime_error("draft-simple requires a draft context");
+        }
+
        SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n");
        SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min);
        SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n",
@@ -385,10 +384,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };


@@ -907,10 +902,6 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        pending_g_last[seq_id].resize(n_embd_dec);
        std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 // DFlash: block-diffusion drafting with a draft-side KV cache injection
@@ -922,6 +913,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

    std::vector<common_sampler_ptr> smpls;

+    // backend sampler chain per seq, attached to ctx_dft
+    std::vector<llama_sampler *> backend_chains;
+
    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
@@ -995,6 +989,22 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            s.reset(common_sampler_init(model_dft, sparams));
        }

+        // offload draft sampling to the backend
+        backend_chains.assign(n_seq, nullptr);
+        if (this->params.backend_sampling) {
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
+                llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
+
+                if (!llama_set_sampler(ctx_dft, seq_id, chain)) {
+                    SPC_WRN("backend offload failed for seq_id=%d; using CPU sampler\n", (int) seq_id);
+                    llama_sampler_free(chain);
+                    chain = nullptr;
+                }
+                backend_chains[seq_id] = chain;
+            }
+        }
+
        // 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);
@@ -1005,6 +1015,18 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    }

    ~common_speculative_impl_draft_dflash() override {
+        auto * ctx_dft = this->params.ctx_dft;
+        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) backend_chains.size(); ++seq_id) {
+            if (backend_chains[seq_id] == nullptr) {
+                continue;
+            }
+            if (ctx_dft) {
+                llama_set_sampler(ctx_dft, seq_id, nullptr);
+            }
+            llama_sampler_free(backend_chains[seq_id]);
+        }
+        backend_chains.clear();
+
        llama_batch_free(batch);
        llama_batch_free(batch_inject);
    }
@@ -1032,7 +1054,14 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            return true;
        }

-        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
+        // Target prefill may contain token IDs or multimodal embeddings. Both
+        // produce the target-layer features used to seed the draft KV cache, so
+        // skipping the embedding batches leaves a hole in the draft's cache and
+        // the next injection fails to initialize.
+        // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
+        const bool has_tokens     = batch_in.token != nullptr;
+        const bool has_embeddings = batch_in.embd  != nullptr;
+        if (has_tokens == has_embeddings) {
            return true;
        }

@@ -1240,10 +1269,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_draft_mtp : public common_speculative_impl {
@@ -1682,14 +1707,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
        const size_t row_bytes = (size_t) n_embd * sizeof(float);
        std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
    }
-
-    bool need_embd() const override {
-        return false;
-    }
-
-    bool need_embd_nextn() const override {
-        return true;
-    }
 };

 // state of self-speculation (simple implementation, not ngram-map)
@@ -1736,10 +1753,6 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
@@ -1794,10 +1807,6 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {

        common_ngram_map_accept(config[seq_id], n_accepted);
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_mod : public common_speculative_impl {
@@ -1973,10 +1982,6 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
            }
        }
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_cache : public common_speculative_impl {
@@ -2116,10 +2121,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative {
@@ -2227,6 +2228,43 @@ common_speculative_type common_speculative_type_from_name(const std::string & na
    return it->second;
 }

+std::vector<common_speculative_type> common_speculative_types_from_gguf(const std::string & path) {
+    struct gguf_init_params gguf_params = {
+        /* .no_alloc = */ true,
+        /* .ctx      = */ nullptr,
+    };
+
+    gguf_context_ptr gguf_ctx(gguf_init_from_file(path.c_str(), gguf_params));
+    if (!gguf_ctx) {
+        return {};
+    }
+
+    const int64_t arch_id = gguf_find_key(gguf_ctx.get(), "general.architecture");
+    if (arch_id < 0 || gguf_get_kv_type(gguf_ctx.get(), arch_id) != GGUF_TYPE_STRING) {
+        return {};
+    }
+
+    const std::string arch = gguf_get_val_str(gguf_ctx.get(), arch_id);
+    if (arch != "dflash") {
+        const uint32_t block_count = gguf_get_val_u32(gguf_ctx.get(), gguf_find_key(gguf_ctx.get(), (arch + ".block_count").c_str()));
+
+        if (gguf_find_tensor(gguf_ctx.get(), ("blk." + std::to_string(block_count - 1) + ".nextn.eh_proj.weight").c_str()) >= 0) {
+            return { COMMON_SPECULATIVE_TYPE_DRAFT_MTP };
+        }
+
+        return {};
+    }
+
+    // the Markov head distinguishes draft-dspark from draft-dflash
+    const auto type = gguf_find_tensor(gguf_ctx.get(), "markov_w1.weight") >= 0
+                    ? COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK
+                    : COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH;
+
+    SPC_INF("auto-detected speculative type '%s' from the draft model metadata\n", common_speculative_type_to_str(type).c_str());
+
+    return { type };
+}
+
 static uint32_t common_get_enabled_speculative_configs(const std::vector<common_speculative_type> & configs) {
    uint32_t result = 0;
    for (size_t i = 0; i < configs.size(); i++) {
@@ -2292,6 +2330,24 @@ common_params common_base_params_to_speculative(const common_params & params) {
    result.cache_type_k  = params_spec.cache_type_k;
    result.cache_type_v  = params_spec.cache_type_v;
    result.n_outputs_max = params.n_parallel;
+    result.n_outputs_max_per_seq = 1;
+
+    // dflash/dspark decode the whole noise block in a single pass and sample every block position on the backend
+    // TODO: refactor such properties to be announced by the speculative types
+    //       something like `struct common_speculative_type_props common_speculative_type_get_props(...);`
+    const bool has_block_draft = std::any_of(
+        params.speculative.types.begin(), params.speculative.types.end(),
+        [](common_speculative_type t) {
+            return t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
+        });
+    if (has_block_draft) {
+        // per-seq output positions: DFlash decodes anchor + n_max masks (n_max + 1); DSpark n_max -> +1 covers both
+        const int32_t per_seq = std::max(1, params_spec.n_max + 1);
+        result.n_outputs_max = params.n_parallel * per_seq;
+        if (params_spec.backend_sampling) {
+            result.n_outputs_max_per_seq = per_seq;
+        }
+    }

    return result;
 }
@@ -2314,7 +2370,6 @@ common_speculative_init_result::common_speculative_init_result(
    const bool spec_mtp = std::find(params.speculative.types.begin(),
                                    params.speculative.types.end(),
                                    COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-    GGML_ASSERT(has_draft || spec_mtp);

    auto mparams = common_model_params_to_llama(params);
    auto cparams = common_context_params_to_llama(params);
@@ -2377,6 +2432,17 @@ common_speculative_init_result_ptr common_speculative_init_from_params(common_pa
    return std::make_unique<common_speculative_init_result>(params, model_tgt, ctx_tgt);
 }

+common_speculative_output_limits common_speculative_get_output_limits(
+        int32_t n_batch, int32_t n_parallel, int32_t n_draft) {
+    const int64_t per_seq = 1 + (int64_t) std::max(0, n_draft);
+    const int64_t total   = (int64_t) n_parallel * per_seq;
+
+    return {
+        /* .total   = */ (int32_t) std::min<int64_t>(n_batch, total),
+        /* .per_seq = */ (int32_t) std::min<int64_t>(n_batch, per_seq),
+    };
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -2541,34 +2607,6 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
    return result;
 }

-bool common_speculative_need_embd(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
-bool common_speculative_need_embd_nextn(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd_nextn()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
 void common_speculative_draft(common_speculative * spec) {
    if (spec == nullptr) {
        return;
@@ -2653,7 +2691,10 @@ void common_speculative_draft(common_speculative * spec) {
 void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
    common_speculative_impl * impl = spec->impl_last[seq_id];

-    GGML_ASSERT(impl);
+    if (impl == nullptr) {
+        GGML_ASSERT(n_accepted == 0);
+        return;
+    }

    {
        common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
@@ -14,6 +14,9 @@ const char * common_speculative_all_types_str();
 // parse user provided types
 std::vector<enum common_speculative_type> common_speculative_types_from_names(const std::vector<std::string> & names);

+// infer the spec types from the GGUF metadata of a draft model; empty if unknown
+std::vector<enum common_speculative_type> common_speculative_types_from_gguf(const std::string & path);
+
 // convert string to type
 enum common_speculative_type common_speculative_type_from_name(const std::string & name);

@@ -25,6 +28,15 @@ int32_t common_speculative_n_max(const common_params_speculative * spec);

 common_params common_base_params_to_speculative(const common_params & params);

+struct common_speculative_output_limits {
+    int32_t total;
+    int32_t per_seq;
+};
+
+// return the output limits needed for speculative decoding
+common_speculative_output_limits common_speculative_get_output_limits(
+        int32_t n_batch, int32_t n_parallel, int32_t n_draft);
+
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);

 void common_speculative_free(common_speculative * spec);
@@ -58,12 +70,6 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);

-// true if any implementation requires target post-norm embeddings to be extracted
-bool common_speculative_need_embd(common_speculative * spec);
-
-// true if any implementation requires target nextn embeddings to be extracted
-bool common_speculative_need_embd_nextn(common_speculative * spec);
-
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);

@@ -103,6 +103,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteMoeForCausalLM": "granite",
    "GraniteMoeHybridForCausalLM": "granite",
    "GraniteMoeSharedForCausalLM": "granite",
+    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "Grok1ForCausalLM": "grok",
@@ -182,6 +183,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Olmo3ForCausalLM": "olmo",
    "OlmoForCausalLM": "olmo",
    "OlmoeForCausalLM": "olmo",
+    "MuseGlimmerAssistantModel": "muse_glimmer",
+    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "OpenELMForCausalLM": "openelm",
    "OrionForCausalLM": "orion",
    "PLMForCausalLM": "plm",
@@ -211,6 +214,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3MoeForCausalLM": "qwen",
    "Qwen3NextForCausalLM": "qwen",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -297,6 +301,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "MiniCPMV4_6ForConditionalGeneration": "minicpm",
    "Mistral3ForConditionalGeneration": "llava",
    "NemotronH_Nano_VL_V2": "nemotron",
+    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "PaddleOCRVisionModel": "ernie",
    "Phi4ForCausalLMV": "phi",
    "Qwen2AudioForConditionalGeneration": "ultravox",
@@ -306,6 +311,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Qwen2_5_VLForConditionalGeneration": "qwenvl",
    "Qwen3ASRForConditionalGeneration": "qwen3vl",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -58,6 +58,11 @@ logger = logging.getLogger("hf-to-gguf")
 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")


+# for checkpoints that ship no config.json, we will try to provide a synthetic one
+HparamsMatcher = Callable[[Path], bool]
+HparamsLoader = Callable[[Path], dict[str, Any]]
+
+
 class SentencePieceTokenTypes(IntEnum):
    NORMAL = 1
    UNKNOWN = 2
@@ -77,6 +82,7 @@ class ModelBase:
        ModelType.TEXT: {},
        ModelType.MMPROJ: {},
    }
+    _hparams_loaders: list[tuple[HparamsMatcher, HparamsLoader]] = []

    dir_model: Path
    ftype: gguf.LlamaFileType
@@ -823,7 +829,7 @@ class ModelBase:
            elif any(str(v.get("quant_algo")).endswith("NVFP4") for v in quant_layers.values() if isinstance(v, dict)):
                quant_algo = "NVFP4"

-        self._is_nvfp4 = quant_algo == "NVFP4"
+        self._is_nvfp4 = quant_algo in ("NVFP4", "W4A16_NVFP4")
        self._is_mxfp4 = quant_method == "mxfp4"

        # NVFP4 weights are repacked and written directly to gguf_writer.
@@ -1040,6 +1046,24 @@ class ModelBase:

        return part_names

+    @staticmethod
+    def load_hparams_guess(dir_model: Path) -> dict[str, Any] | None:
+        # some models ship no config.json, will try to guess them
+        from conversion import load_all_models
+        load_all_models()
+
+        for matcher, loader in ModelBase._hparams_loaders:
+            if matcher(dir_model):
+                return loader(dir_model)
+        return None
+
+    @classmethod
+    def register_hparams_loader(cls, matcher: HparamsMatcher) -> Callable[[HparamsLoader], HparamsLoader]:
+        def inner(loader: HparamsLoader) -> HparamsLoader:
+            cls._hparams_loaders.append((matcher, loader))
+            return loader
+        return inner
+
    @staticmethod
    def load_hparams(dir_model: Path, is_mistral_format: bool):
        if is_mistral_format:
@@ -1053,6 +1077,10 @@ class ModelBase:
            config = AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
        except Exception as e:
            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            if not (dir_model / "config.json").is_file():
+                config = ModelBase.load_hparams_guess(dir_model)
+                if config is not None:
+                    return config
            logger.warning("Trying to load config.json instead")
            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                config = json.load(f)
@@ -665,7 +665,18 @@ class Gemma4Model(Gemma3Model):
        swa_layers = [t == "sliding_attention" for t in self.hparams["layer_types"]]
        self.gguf_writer.add_sliding_window_pattern(swa_layers)

-        head_dim_full = self.hparams["global_head_dim"]
+        per_layer_config = self.hparams.get("per_layer_config")
+        layer_types = self.hparams.get("layer_types", [])
+        if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
+                        head_dim_full = layer_config["head_dim"]
+                        break
+
+        assert head_dim_full is not None
+
        head_dim_swa = self.hparams["head_dim"]
        # correct the head dim for global/swa layers
        self.gguf_writer.add_key_length(head_dim_full)
@@ -685,8 +696,14 @@ class Gemma4Model(Gemma3Model):
            n_ff_arr = [n_ff if il < first_kv_shared_layer_idx else n_ff * 2 for il in range(self.block_count)]
            self.gguf_writer.add_feed_forward_length(n_ff_arr)

-        # handle num_global_key_value_heads
-        num_key_value_heads_full = self.hparams.get("num_global_key_value_heads")
+        if (num_key_value_heads_full := self.hparams.get("num_global_key_value_heads")) is None and per_layer_config is not None:
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "num_key_value_heads" in layer_config:
+                        num_key_value_heads_full = layer_config["num_key_value_heads"]
+                        break
+
        num_key_value_heads_swa = self.hparams.get("num_key_value_heads")
        if num_key_value_heads_full is not None and num_key_value_heads_swa is not None:
            value_arr = [num_key_value_heads_swa if is_swa else num_key_value_heads_full for is_swa in swa_layers]
@@ -708,7 +725,19 @@ class Gemma4Model(Gemma3Model):
        # IMPORTANT: this ROPE_FREQS tensor is ONLY used by the full_attention layers
        rope_params_full = self.hparams["rope_parameters"]["full_attention"]
        assert rope_params_full["rope_type"] == "proportional"
-        head_dim_full = (self.hparams["global_head_dim"])
+
+        per_layer_config = self.hparams.get("per_layer_config")
+        if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
+            layer_types = self.hparams.get("layer_types", [])
+            for layer_idx, layer_config in per_layer_config.items():
+                layer_idx = int(layer_idx)
+                if layer_idx < len(layer_types):
+                    if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
+                        head_dim_full = layer_config["head_dim"]
+                        break
+
+        assert head_dim_full is not None
+
        partial_rotary_factor_full = rope_params_full["partial_rotary_factor"]
        n_rot_full = int(head_dim_full * partial_rotary_factor_full / 2)
        n_unrot_full = int(head_dim_full / 2) - n_rot_full
@@ -123,6 +123,166 @@ class GraniteMoeModel(GraniteModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("GraniteSwitchForCausalLM")
+class GraniteSwitchModel(GraniteMoeModel):
+    """Dense, all-attention Granite with N per-token embedded LoRA adapters, stacked
+    over the adapter dim with a zero adapter at slot 0 (N = num_adapters + 1)."""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWITCH
+
+    # permute q/k per-slice below (NORM-rope layout), not via the parent's auto-permute
+    undo_permute = False
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # the weightless switch reserves one cache slot: one fewer block than num_hidden_layers
+        self.block_count = self.block_count - 1
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self._n_adapters = int(self.hparams["num_adapters"])
+        self._max_lora_rank = int(self.hparams["max_lora_rank"])
+        self._n_slots = self._n_adapters + 1  # +1 for the zero slot at index 0
+
+        n_head = int(self.hparams["num_attention_heads"])
+        n_kv_head = int(self.hparams["num_key_value_heads"])
+        head_dim = (
+            self.hparams.get("projection_head_dim")
+            or self.hparams.get("head_dim")
+            or (self.hparams["hidden_size"] // n_head)
+        )
+        self._n_head = n_head
+        self._n_kv_head = n_kv_head
+        self._head_dim = int(head_dim)
+        self._q_size = n_head * self._head_dim
+        self._kv_size = n_kv_head * self._head_dim
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # dense: pin expert_used_count to 0 (config carries a leftover num_experts_per_tok)
+        if not self.hparams.get("num_local_experts"):
+            self.gguf_writer.add_expert_used_count(0)
+
+        self.gguf_writer.add_adapter_count(self._n_adapters)
+        self.gguf_writer.add_adapter_lora_rank(self._max_lora_rank)
+        self.gguf_writer.add_adapter_token_ids_activate(self.hparams["adapter_token_ids"])
+        self.gguf_writer.add_adapter_token_ids_substitute(self.hparams["adapter_substitute_token_ids"])
+        router_gain = float(self.hparams.get("control_token_gain", 15.0))
+        self.gguf_writer.add_adapter_router_gain(router_gain)
+        logger.info("gguf: (graniteswitch) num_adapters=%s max_lora_rank=%s n_slots=%s router_gain=%s", self._n_adapters, self._max_lora_rank, self._n_slots, router_gain)
+
+    def _lora_a(self, data: Tensor) -> Tensor:
+        # on-disk A: [n_adapters, 1, max_rank, in] -> [n_adapters+1, max_rank, in]
+        a = data.squeeze(1)
+        zero = torch.zeros_like(a[:1])
+        return torch.cat([zero, a], dim=0).contiguous()
+
+    def _lora_b(self, data: Tensor, permute_n_head: int | None = None) -> Tensor:
+        # on-disk B: [n_adapters, 1, out, max_rank] -> [n_adapters+1, out, max_rank]
+        b = data.squeeze(1)
+        if permute_n_head is not None:
+            # permute each adapter's B output rows to match the permuted q/k base
+            b = torch.stack([self.permute(b[i], permute_n_head, permute_n_head) for i in range(b.shape[0])], dim=0)
+        zero = torch.zeros_like(b[:1])
+        return torch.cat([zero, b], dim=0).contiguous()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+
+        # skip the weightless switch + control-token buffers (rebuilt at load time)
+        bare = name.split(".")[-1]
+        if (
+            name.startswith("model.switch.") or name.startswith("switch.")
+            or bare in ("adapter_token_ids", "control_to_substitute_lut")
+        ):
+            return
+
+        if "self_attn.qkv_proj" in name:
+            if name.endswith("base_layer.weight"):
+                # fused [q|k|v] rows: permute q/k row-blocks for ggml's NORM-rope layout
+                q, k, v = data_torch.split([self._q_size, self._kv_size, self._kv_size], dim=0)
+                q = self.permute(q, self._n_head, self._n_head)
+                k = self.permute(k, self._n_kv_head, self._n_kv_head)
+                fused = torch.cat([q, k, v], dim=0)
+                yield (self.format_tensor_name(T.ATTN_QKV, bid), fused)
+                return
+            if "lora_A_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.ATTN_Q, 1: T.ATTN_K, 2: T.ATTN_V}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if "lora_B_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key, ph = {
+                    0: (T.ATTN_Q, self._n_head),
+                    1: (T.ATTN_K, self._n_kv_head),
+                    2: (T.ATTN_V, None),
+                }[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch, ph))
+                return
+            raise ValueError(f"Unexpected qkv_proj tensor: {name}")
+
+        if "self_attn.o_proj" in name:
+            if name.endswith("base_layer.weight"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid), data_torch)
+                return
+            if name.endswith("lora_A"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if name.endswith("lora_B"):
+                yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected o_proj tensor: {name}")
+
+        if "shared_mlp.input_linear" in name:
+            ffn = self.hparams["shared_intermediate_size"]
+            if name.endswith("base_layer.weight"):
+                gate, up = data_torch.split([ffn, ffn], dim=0)
+                yield (self.format_tensor_name(T.FFN_GATE, bid), gate)
+                yield (self.format_tensor_name(T.FFN_UP, bid), up)
+                return
+            if "lora_A_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if "lora_B_slices." in name:
+                slot = int(name.rsplit(".", 1)[1])
+                key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+                yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected shared_mlp.input_linear tensor: {name}")
+
+        if "shared_mlp.output_linear" in name:
+            if name.endswith("base_layer.weight"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid), data_torch)
+                return
+            if name.endswith("lora_A"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_a"), self._lora_a(data_torch))
+                return
+            if name.endswith("lora_B"):
+                yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_b"), self._lora_b(data_torch))
+                return
+            raise ValueError(f"Unexpected shared_mlp.output_linear tensor: {name}")
+
+        if bid is not None and ".layers." in name and (
+            "input_layernorm" in name or "post_attention_layernorm" in name
+        ):
+            key = T.ATTN_NORM if "input_layernorm" in name else T.FFN_NORM
+            yield (self.format_tensor_name(key, bid), data_torch)
+            return
+
+        if name in ("model.embed_tokens.weight", "embed_tokens.weight"):
+            yield (self.format_tensor_name(T.TOKEN_EMBD), data_torch)
+            return
+        if name in ("model.norm.weight", "norm.weight"):
+            yield (self.format_tensor_name(T.OUTPUT_NORM), data_torch)
+            return
+        if name == "lm_head.weight":
+            return  # tied to token_embd
+
+        raise ValueError(f"graniteswitch: unhandled tensor {name!r} (bid={bid})")
+
+
@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
 class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
    """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
@@ -0,0 +1,179 @@
+from __future__ import annotations
+
+import json
+from typing import Any, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import MmprojModel, ModelBase, TextModel, gguf
+
+
+def _unpermute_for_rope(tensor: "Tensor", n_heads: int) -> "Tensor":
+    """Invert transformers' `_permute_for_rope`: HF stores Q/K in rotate_half layout,
+    llama.cpp consumes the interleaved (NORM) layout."""
+    if tensor.ndim == 2:
+        dim1, dim2 = tensor.shape
+        return tensor.view(n_heads, 2, dim1 // n_heads // 2, dim2).transpose(1, 2).reshape(dim1, dim2)
+    if tensor.ndim == 1:
+        (dim1,) = tensor.shape
+        return tensor.view(n_heads, 2, dim1 // n_heads // 2).transpose(1, 2).reshape(dim1)
+    raise ValueError(f"_unpermute_for_rope: unexpected shape {tuple(tensor.shape)}")
+
+
+@ModelBase.register("MuseGlimmerForConditionalGeneration")
+class MuseGlimmerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.MUSE_GLIMMER
+
+    def norm_shift(self, name: str) -> float:
+        # All four layer norms use 1, the final norm uses 0.
+        return 1.0 if name.endswith("layernorm.weight") else 0.0
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+        from transformers import AutoTokenizer
+        tok = AutoTokenizer.from_pretrained(self.dir_model)
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")
+        if isinstance(eot_id, int) and eot_id >= 0:
+            self.gguf_writer.add_eot_token_id(eot_id)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        self.gguf_writer.add_final_logit_softcapping(hparams["final_logit_softcapping"])
+        self.gguf_writer.add_logit_scale(hparams["output_multiplier"])
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        shift = self.norm_shift(name)
+        if shift != 0.0:
+            data_torch = data_torch + shift
+
+        # Invert transformers' `_permute_for_rope` on Q/K, we keep ggml's NORM (interleaved) rope
+        if ".self_attn.q_proj." in name:
+            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_attention_heads"]))
+        elif ".self_attn.k_proj." in name:
+            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_key_value_heads"]))
+
+        # Synthesize QK-norm weights to absorb qk_scale_factor.
+        # MuseGlimmer implementation: scaleless RMSNorm followed by qk_scale_factor..
+        if bid is not None and name.endswith(f"model.layers.{bid}.self_attn.q_proj.weight"):
+            head_dim = self.hparams["head_dim"]
+            q_scale = float(self.hparams["qk_scale_factor"])
+            yield (
+                self.map_tensor_name(f"model.layers.{bid}.self_attn.q_norm.weight"),
+                torch.full((head_dim,), q_scale, dtype=torch.float32),
+            )
+            yield (
+                self.map_tensor_name(f"model.layers.{bid}.self_attn.k_norm.weight"),
+                torch.ones((head_dim,), dtype=torch.float32),
+            )
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("MuseGlimmerForConditionalGeneration")
+class MuseGlimmerVisionModel(MmprojModel):
+    def get_vision_config(self) -> dict[str, Any] | None:
+        c = self.global_config.get("vision_config")
+        if not c:
+            return None
+        # MuseGlimmer actually uses dynamic size, initialize with nominal size
+        image_size = c["pos_emb_height"] * c["patch_size"] * c["merge_size"]
+        return {**c, "image_size": image_size}
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        c = self.hparams_vision  # enriched vision_config from get_vision_config()
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MUSE_GLIMMER)
+        self.gguf_writer.add_vision_attention_layernorm_eps(float(c["layer_norm_eps"]))
+        self.gguf_writer.add_vision_spatial_merge_size(int(c["merge_size"]))
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, gen = item
+        keep = ("model.vision_tower.", "model.vision_adapter.", "model.vision_projection.")
+        if not any(name.startswith(k) for k in keep):
+            return None
+        return super().filter_tensors((name, gen))
+
+    # 3-layer projector MLP
+    _MM_MLP_MAP = {
+        "model.vision_adapter.fc1": (gguf.MODEL_TENSOR.V_MMPROJ, 0),
+        "model.vision_adapter.fc2": (gguf.MODEL_TENSOR.V_MMPROJ, 1),
+        "model.vision_projection":  (gguf.MODEL_TENSOR.V_MMPROJ, 2),
+    }
+
+    def modify_tensors(self, data_torch, name, bid):
+        assert self.hparams_vision is not None
+        if ".attn.q_proj." in name or ".attn.k_proj." in name:
+            n_heads = int(self.hparams_vision["num_attention_heads"])
+            data_torch = _unpermute_for_rope(data_torch, n_heads)
+        # Lay out the pt=2 temporal slabs of the patch embedding as a conv2d for build_inp()
+        if name.endswith("patch_embedder.patch_embedding.weight"):
+            n_embd = data_torch.shape[0]
+            pt = int(self.hparams_vision["patch_temporal"])
+            ps = int(self.hparams_vision["patch_size"])
+            data_torch = data_torch.view(n_embd, pt, 3, ps, ps).sum(dim=1)  # (n_embd, 3, ps, ps)
+        stem, _, suffix = name.rpartition(".")
+        if stem in self._MM_MLP_MAP:
+            tensor_key, idx = self._MM_MLP_MAP[stem]
+            yield (self.format_tensor_name(tensor_key, bid=idx, suffix="." + suffix), data_torch)
+            return
+        yield (self.map_tensor_name(name), data_torch)
+
+
+@ModelBase.register("MuseGlimmerAssistantModel")
+class MuseGlimmerAssistantModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    def set_vocab(self):
+        if self.target_model_dir is None:
+            raise ValueError(
+                "MuseGlimmerAssistant (DFlash drafter) requires --target-model-dir pointing to the "
+                "target MuseGlimmer HF directory"
+            )
+
+        original_dir = self.dir_model
+        self.dir_model = self.target_model_dir
+
+        from . import get_model_class
+        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
+            target_arch = json.load(f)["architectures"][0]
+        target_cls = get_model_class(target_arch)
+        if target_cls is not type(self):
+            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
+        else:
+            super().set_vocab()
+
+        self.dir_model = original_dir
+
+        mask_token_id = self.hparams.get("mask_token_id")
+        if mask_token_id is not None:
+            self.gguf_writer.add_mask_token_id(int(mask_token_id))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        h = self.hparams
+
+        self.gguf_writer.add_block_size(int(h["block_size"]))
+
+        # dflash.target_layers[k] refers to the inputs going into the ith layer, which come from the (i-1)th layer's output.
+        # The transformers configuration refers to the outputs being recorded.
+        self.gguf_writer.add_target_layers([int(x) + 1 for x in h["target_layer_ids"]])
+
+        if h.get("sliding_window") and h.get("layer_types"):
+            self.gguf_writer.add_sliding_window(int(h["sliding_window"]))
+            self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in h["layer_types"]])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # DFlash defaults to NEOX (rotate_half) rope, matching transformers HF layout for Q/K, QK-norms
+        # no permutation needed.
+        yield (self.map_tensor_name(name), data_torch)
@@ -197,6 +197,7 @@ class NemotronHModel(GraniteHybridModel):
    """Hybrid mamba2/attention model from NVIDIA"""
    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
    is_moe: bool = False
+    supports_mtp_export = True

    def __init__(self, *args, **kwargs):
        # We have to determine the correct model architecture (MoE vs non-MoE) before
@@ -236,6 +237,25 @@ class NemotronHModel(GraniteHybridModel):
            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]

+        # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
+        self._mtp_bid: int | None = None
+        if self.is_moe and not self.no_mtp:
+            n_nextn = self.hparams.get("num_nextn_predict_layers", 0) or 0
+            if n_nextn > 0:
+                assert n_nextn == 1, (
+                    "NemotronH MTP conversion currently supports num_nextn_predict_layers == 1"
+                )
+                self._mtp_bid = self.block_count
+                self.block_count += 1
+                # The folded MTP block carries both an attention sub-layer and a
+                # MoE sub-layer, so register it as both so the per-layer metadata arrays cover it
+                self._attn_layers.append(self._mtp_bid)
+                self._mlp_layers.append(self._mtp_bid)
+                self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        if self.mtp_only and self._mtp_bid is None:
+            raise ValueError("--mtp was requested, but this model does not contain a supported MTP head")
+
    def get_attn_layers(self):
        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
        if pattern is None:
@@ -246,6 +266,44 @@ class NemotronHModel(GraniteHybridModel):

        return [i for i, val in enumerate(pattern) if val == "attention"]

+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.startswith("mtp."):
+            # --no-mtp: drop the MTP head entirely
+            if cls.no_mtp:
+                return None
+        elif cls.mtp_only:
+            # --mtp: export the MTP head plus the tensors it shares with the target model
+            # Include lm_head scale sidecars so NVFP4 packing sees them.
+            keep = name in (
+                "backbone.embeddings.weight",
+                "backbone.norm_f.weight",
+                "lm_head.weight",
+                "lm_head.weight_scale",
+                "lm_head.weight_scale_2",
+                "lm_head.weight_scale_inv",
+                "lm_head.input_scale",
+                "lm_head.input_global_scale",
+                "lm_head.weight_global_scale",
+                "lm_head.weight_packed",
+            )
+            if not keep:
+                return None
+        return super().filter_tensors((name, gen))
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()

@@ -284,6 +342,10 @@ class NemotronHModel(GraniteHybridModel):
            if (latent_size := self.hparams.get("moe_latent_size")) is not None:
                self.gguf_writer.add_moe_latent_size(latent_size)

+        # MTP head: number of trailing NextN blocks
+        if self._mtp_bid is not None:
+            self.gguf_writer.add_nextn_predict_layers(self.hparams["num_nextn_predict_layers"])
+
    def set_vocab(self):
        # The NemotronH config uses pattern characters (e.g. '-') that may not
        # be supported by the installed transformers version. AutoTokenizer
@@ -350,15 +412,24 @@ class NemotronHModel(GraniteHybridModel):
        if not self.is_moe:
            self.gguf_writer.add_add_bos_token(True)

-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if self.is_moe and bid is not None:
-            # Skip Multi-Token Prediction (MTP) tensors. These are used for
-            # for speculative decoding but we don't include them in this model
-            # conversion. See https://github.com/ggml-org/llama.cpp/pull/18886
-            if name.startswith("mtp."):
-                logger.info(f"gguf: Skipping MTP (Speculative) layer: {name}")
-                return
+    _MTP_SPECIAL_RENAMES = {
+        "mtp.layers.0.enorm.weight":           "model.layers.{bid}.enorm.weight",
+        "mtp.layers.0.hnorm.weight":           "model.layers.{bid}.hnorm.weight",
+        "mtp.layers.0.eh_proj.weight":         "model.layers.{bid}.eh_proj.weight",
+        "mtp.layers.1.norm.weight":            "model.layers.{bid}.post_attention_layernorm.weight",
+        "mtp.layers.1.final_layernorm.weight": "model.layers.{bid}.shared_head.norm.weight",
+    }

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        #   mtp.layers.0: NextN input fusion + attention
+        #   mtp.layers.1: MoE + final head norm
+        if self._mtp_bid is not None and name.startswith(("mtp.layers.0.", "mtp.layers.1.")):
+            suffix = name.split(".", 3)[3]
+            bid = self._mtp_bid
+            renamed = self._MTP_SPECIAL_RENAMES.get(name)
+            name = renamed.format(bid=bid) if renamed else f"backbone.layers.{bid}.{suffix}"
+
+        if self.is_moe and bid is not None:
            if name.endswith("mixer.gate.e_score_correction.bias"):
                yield from ModelBase.modify_tensors(self, data_torch, name, bid)
                return
@@ -0,0 +1,378 @@
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import Any, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, SentencePieceTokenTypes, TextModel, gguf, logger
+
+# Pocket TTS is a CALM: the backbone conditions a flow-matching decoder that generates one
+# continuous 32-d latent per frame. There is no codebook in this model.
+# The checkpoint ships no config.json, hparams come from _load_hparams() below.
+#
+# Tricks being used to support this model via existing llama.cpp code paths:
+# - bos_before_voice and bos_emb are learned input vectors, not tokens
+#   they are appended to the embedding table as extra tokens, to be looked up like any other row
+# - bos_emb lives in latent space, so input_linear is folded into it here
+# - the backbone has no lm_head, the embedding table is reused as output for the unused logits
+#
+# pipeline stage mapping:
+#   mimi encoder + speaker_proj --> mapped to normal mtmd audio encoder
+#   flow_lm.transformer         --> mapped to normal libllama text model (autoregressive)
+#   flow_lm.flow_net + out_eos  --> MTMD_GEN_PROCESS_TYPE_GEN_CODE
+#   mimi decoder                --> MTMD_GEN_PROCESS_TYPE_GEN_WAV
+
+# indices into mimi.encoder.model / mimi.decoder.model for stage i, see SEANetEncoder/SEANetDecoder
+_ENC_RES_IDX   = lambda i: 1 + 3 * i  # noqa: E731
+_ENC_SCALE_IDX = lambda i: 3 + 3 * i  # noqa: E731
+_DEC_SCALE_IDX = lambda i: 2 + 3 * i  # noqa: E731
+_DEC_RES_IDX   = lambda i: 3 + 3 * i  # noqa: E731
+
+_N_SEANET_STAGES = 3
+_SAMPLE_RATE = 24000
+
+
+def _tensor_shapes(dir_model: Path) -> dict[str, tuple[int, ...]]:
+    part_names = ModelBase.get_model_part_names(dir_model, "model", ".safetensors")
+    if len(part_names) != 1:
+        return {}
+    with gguf.utility.SafetensorsLocal(dir_model / part_names[0]) as part:
+        return {name: tuple(part[name].shape) for name in part.keys()}
+
+
+@ModelBase.register_hparams_loader(lambda dir_model: "flow_lm.bos_emb" in _tensor_shapes(dir_model))
+def _load_hparams(dir_model: Path) -> dict[str, Any]:
+    logger.info("gguf: detected pocket-tts checkpoint, deriving hparams from tensor shapes")
+    shapes = _tensor_shapes(dir_model)
+    n_vocab, n_embd = shapes["flow_lm.conditioner.embed.weight"]
+    n_layer = sum(1 for name in shapes if re.fullmatch(r"flow_lm\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_layer_a = sum(1 for name in shapes if re.fullmatch(r"mimi\.encoder_transformer\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_embd_a = shapes["mimi.encoder_transformer.transformer.layers.0.norm1.weight"][0]
+    return {
+        "architectures": ["PocketTTSModel"],
+        "model_type": "pockettts",
+        "num_hidden_layers": n_layer,
+        "hidden_size": n_embd,
+        "intermediate_size": shapes["flow_lm.transformer.layers.0.linear1.weight"][0],
+        # the transformer is fully causal with no context limit, this only bounds the KV cache
+        "max_position_embeddings": 4096,
+        # not in the checkpoint, but every released variant uses head_dim 64
+        "num_attention_heads": n_embd // 64,
+        # extra rows for the learned input vectors, see _embd_table()
+        "vocab_size": n_vocab + (2 if "flow_lm.bos_before_voice" in shapes else 1),
+        "rope_theta": 10000.0,
+        "layer_norm_eps": 1e-5,
+        "audio_config": {
+            "num_hidden_layers": n_layer_a,
+            "hidden_size": n_embd_a,
+            "intermediate_size": shapes["mimi.encoder_transformer.transformer.layers.0.linear1.weight"][0],
+            "num_attention_heads": n_embd_a // 64,
+        },
+    }
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.POCKETTTS
+
+    _LAYER_TENSOR_MAP = {
+        "norm1":               gguf.MODEL_TENSOR.ATTN_NORM,
+        "norm2":               gguf.MODEL_TENSOR.FFN_NORM,
+        "self_attn.out_proj":  gguf.MODEL_TENSOR.ATTN_OUT,
+        "linear1":             gguf.MODEL_TENSOR.FFN_UP,
+        "linear2":             gguf.MODEL_TENSOR.FFN_DOWN,
+    }
+
+    def set_vocab(self):
+        # this is a unigram sentencepiece model, llama.cpp's SPM tokenizer cannot do
+        # unigram segmentation, so use the UGM tokenizer instead
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        proto = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        proto.ParseFromString(open(self.dir_model / "tokenizer.model", "rb").read())
+        assert proto.trainer_spec.model_type == 1, "expected a unigram tokenizer"
+
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        # the last rows of the embedding table are not sentencepiece pieces
+        extra = self._extra_tokens()
+        for i, name in enumerate(extra):
+            tokens[len(tokens) - len(extra) + i] = name.encode("utf-8")
+            toktypes[len(tokens) - len(extra) + i] = SentencePieceTokenTypes.CONTROL
+            scores[len(tokens) - len(extra) + i] = -1000.0
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(proto.normalizer_spec.add_dummy_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(proto.normalizer_spec.remove_extra_whitespaces)
+        if proto.normalizer_spec.precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(proto.normalizer_spec.precompiled_charsmap)
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(False)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if not name.startswith("flow_lm."):
+            return  # mimi and the flow net go to the mmproj
+
+        if name == "flow_lm.conditioner.embed.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD), self._embd_table(data_torch))
+            return
+
+        if name.startswith("flow_lm.out_norm."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.transformer.layers."):
+            assert bid is not None
+            key_with_suffix = name.split(f"layers.{bid}.", 1)[1]
+            key, suffix = key_with_suffix.rsplit(".", 1)
+
+            if key == "self_attn.in_proj":
+                q, k, v = data_torch.chunk(3, dim=0)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+                return
+
+            tensor = self._LAYER_TENSOR_MAP.get(key)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix="." + suffix), data_torch)
+                return
+
+        return
+
+    def _extra_tokens(self) -> list[str]:
+        # the conditioner's padding row, then the learned vectors appended by _embd_table().
+        # bos_before_voice only exists when the pack sets insert_bos_before_voice
+        names = ["<|pad|>"]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            names.append("<|bos_before_voice|>")
+        names.append("<|audio_bos|>")
+        return names
+
+    def _embd_table(self, embed: Tensor) -> Tensor:
+        rows = [embed]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            rows.append(self.model_tensors["flow_lm.bos_before_voice"]().reshape(1, -1).to(embed.dtype))
+
+        # bos_emb is a latent, it only enters the backbone through input_linear
+        bos_emb = self.model_tensors["flow_lm.bos_emb"]()
+        input_linear = self.model_tensors["flow_lm.input_linear.weight"]()
+        audio_bos = torch.nn.functional.linear(bos_emb.float(), input_linear.float()).reshape(1, -1)
+        rows.append(audio_bos.to(embed.dtype))
+
+        return torch.cat(rows, dim=0)
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSMmprojModel(MmprojModel):
+    has_audio_encoder = True
+    has_vision_encoder = False
+
+    _MIMI_TFM_MAP = {
+        "norm1":              (gguf.MODEL_TENSOR.A_ENC_INPUT_NORM,  gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_NORM),
+        "norm2":              (gguf.MODEL_TENSOR.A_ENC_OUTPUT_NORM, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_NORM),
+        "self_attn.out_proj": (gguf.MODEL_TENSOR.A_ENC_OUTPUT,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_OUT),
+        "linear1":            (gguf.MODEL_TENSOR.A_ENC_FFN_UP,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_UP),
+        "linear2":            (gguf.MODEL_TENSOR.A_ENC_FFN_DOWN,    gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_DOWN),
+        "layer_scale_1.scale": (gguf.MODEL_TENSOR.A_ENC_ATTN_SCALE, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_SCALE),
+        "layer_scale_2.scale": (gguf.MODEL_TENSOR.A_ENC_FFN_SCALE_LS, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_SCALE),
+    }
+    _MIMI_TFM_QKV = (
+        (gguf.MODEL_TENSOR.A_ENC_ATTN_Q, gguf.MODEL_TENSOR.A_ENC_ATTN_K, gguf.MODEL_TENSOR.A_ENC_ATTN_V),
+        (gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V),
+    )
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        assert self.hparams_audio is not None
+
+        # voice-prompt encoder: mimi encoder + speaker_proj
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        # note: the 24kHz sample rate is hardcoded on the clip.cpp side, like the other audio models
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_SPKENC)
+        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+        # mimi convolves the waveform directly, it is passed around as a 1-row "mel"
+        self.gguf_writer.add_audio_num_mel_bins(1)
+
+        # generation: flow-matching decoder + mimi decoder
+        # the SEANet and flow net hparams are constant across the family, clip.cpp holds them
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_GEN)
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(1e-5)
+
+        self.gguf_writer.add_gen_audio_model_variant(self.dir_model.name)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del name, bid, n_dims
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if ".seanet." in new_name or new_name in ("a.downsample.conv.weight", "a.gen.wav.upsample.weight"):
+            return gguf.GGMLQuantizationType.F16
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # the block index of the mimi transformers is parsed here, not by the base class
+        T = gguf.MODEL_TENSOR
+
+        if name in ("flow_lm.bos_emb", "flow_lm.bos_before_voice", "flow_lm.conditioner.embed.weight"):
+            return  # folded into the backbone embedding table
+        if name.startswith("flow_lm.transformer.") or name.startswith("flow_lm.out_norm."):
+            return  # backbone
+
+        if name == "flow_lm.speaker_proj_weight":
+            yield (self.format_tensor_name(T.A_ENC_SPEAKER_PROJ), data_torch)
+            return
+        if name == "flow_lm.input_linear.weight":
+            yield (self.format_tensor_name(T.A_GEN_INPUT_LINEAR), data_torch)
+            return
+        if name == "flow_lm.emb_mean":
+            yield (self.format_tensor_name(T.A_GEN_EMB_MEAN, suffix=""), data_torch)
+            return
+        if name == "flow_lm.emb_std":
+            yield (self.format_tensor_name(T.A_GEN_EMB_STD, suffix=""), data_torch)
+            return
+        if name.startswith("flow_lm.out_eos."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(T.A_GEN_OUT_EOS, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.flow_net."):
+            yield from self._flow_net_tensor(name, data_torch)
+            return
+
+        if name == "mimi.downsample.conv.conv.weight":
+            yield (self.format_tensor_name(T.A_ENC_DOWNSAMPLE_CONV), data_torch)
+            return
+        if name == "mimi.upsample.convtr.convtr.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_UPSAMPLE), data_torch)
+            return
+        if name == "mimi.quantizer.output_proj.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_OUT), data_torch.squeeze(-1))
+            return
+
+        if "_transformer.transformer.layers." in name:
+            yield from self._mimi_tfm_tensor(name, data_torch)
+            return
+
+        if name.startswith("mimi.encoder.model.") or name.startswith("mimi.decoder.model."):
+            yield from self._seanet_tensor(name, data_torch)
+            return
+
+        return
+
+    def _flow_net_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        key = name.split("flow_lm.flow_net.", 1)[1]
+        suffix = "." + key.rsplit(".", 1)[1]
+
+        simple = {
+            "input_proj":                 T.A_GEN_FLOW_INPUT_PROJ,
+            "cond_embed":                 T.A_GEN_FLOW_COND_EMBD,
+            "final_layer.linear":         T.A_GEN_FLOW_FINAL_PROJ,
+            "final_layer.adaLN_modulation.1": T.A_GEN_FLOW_FINAL_ADA,
+        }
+        tensor = simple.get(key.rsplit(".", 1)[0])
+        if tensor is not None:
+            yield (self.format_tensor_name(tensor, suffix=suffix), data_torch)
+            return
+
+        if key.startswith("time_embed."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"time_embed.{bid}.", 1)[1]
+            time_map = {
+                "freqs":       (T.A_GEN_FLOW_TIME_FREQS, ""),
+                "mlp.0":       (T.A_GEN_FLOW_TIME_UP,    suffix),
+                "mlp.2":       (T.A_GEN_FLOW_TIME_DOWN,  suffix),
+                "mlp.3.alpha": (T.A_GEN_FLOW_TIME_NORM,  ""),
+            }
+            entry = time_map.get(rest) or time_map.get(rest.rsplit(".", 1)[0])
+            if entry is not None:
+                yield (self.format_tensor_name(entry[0], bid, suffix=entry[1]), data_torch)
+            return
+
+        if key.startswith("res_blocks."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"res_blocks.{bid}.", 1)[1].rsplit(".", 1)[0]
+            blk_map = {
+                "in_ln":                T.A_GEN_FLOW_BLK_NORM,
+                "mlp.0":                T.A_GEN_FLOW_BLK_UP,
+                "mlp.2":                T.A_GEN_FLOW_BLK_DOWN,
+                "adaLN_modulation.1":   T.A_GEN_FLOW_BLK_ADA,
+            }
+            tensor = blk_map.get(rest)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+            return
+
+    def _mimi_tfm_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        is_decoder = name.startswith("mimi.decoder_transformer.")
+        bid = int(name.split("_transformer.transformer.layers.", 1)[1].split(".")[0])
+        key_with_suffix = name.split(f".layers.{bid}.", 1)[1]
+
+        if key_with_suffix == "self_attn.in_proj.weight":
+            q, k, v = data_torch.chunk(3, dim=0)
+            names = self._MIMI_TFM_QKV[1 if is_decoder else 0]
+            for tensor, part in zip(names, (q, k, v)):
+                yield (self.format_tensor_name(tensor, bid), part)
+            return
+
+        key, suffix = key_with_suffix.rsplit(".", 1)
+        entry = self._MIMI_TFM_MAP.get(key) or self._MIMI_TFM_MAP.get(key_with_suffix)
+        if entry is None:
+            return
+        tensor = entry[1 if is_decoder else 0]
+        suffix = ".weight" if key_with_suffix.endswith(".scale") else "." + suffix
+        yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+
+    def _seanet_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        is_decoder = name.startswith("mimi.decoder.")
+        idx = int(name.split(".model.", 1)[1].split(".")[0])
+        suffix = "." + name.rsplit(".", 1)[1]
+
+        conv_in, conv_out, res1, res2, scale = (
+            (T.A_GEN_WAV_SEANET_CONV_IN, T.A_GEN_WAV_SEANET_CONV_OUT, T.A_GEN_WAV_SEANET_RES_CONV1,
+             T.A_GEN_WAV_SEANET_RES_CONV2, T.A_GEN_WAV_SEANET_SCALE_CONV)
+            if is_decoder else
+            (T.A_ENC_SEANET_CONV_IN, T.A_ENC_SEANET_CONV_OUT, T.A_ENC_SEANET_RES_CONV1,
+             T.A_ENC_SEANET_RES_CONV2, T.A_ENC_SEANET_SCALE_CONV)
+        )
+
+        if idx == 0:
+            yield (self.format_tensor_name(conv_in, suffix=suffix), data_torch)
+            return
+        if idx == 3 * _N_SEANET_STAGES + 2:
+            yield (self.format_tensor_name(conv_out, suffix=suffix), data_torch)
+            return
+
+        for stage in range(_N_SEANET_STAGES):
+            res_idx = _DEC_RES_IDX(stage) if is_decoder else _ENC_RES_IDX(stage)
+            scale_idx = _DEC_SCALE_IDX(stage) if is_decoder else _ENC_SCALE_IDX(stage)
+            if idx == scale_idx:
+                yield (self.format_tensor_name(scale, stage, suffix=suffix), data_torch)
+                return
+            if idx == res_idx:
+                # block.1 is the dilated conv, block.3 the pointwise one (0 and 2 are ELU)
+                inner = int(name.split(".block.", 1)[1].split(".")[0])
+                tensor = res1 if inner == 1 else res2
+                yield (self.format_tensor_name(tensor, stage, suffix=suffix), data_torch)
+                return
@@ -647,10 +647,13 @@ class DFlashModel(Qwen3Model):
        # own tokenizer logic, not the Qwen default).
        from . import get_model_class
        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
-            target_arch = json.load(f)["architectures"][0]
+            target_hparams = json.load(f)
+            target_arch = target_hparams["architectures"][0]
        target_cls = get_model_class(target_arch)

        if target_cls is not type(self):
+            if target_arch == "NemotronHForCausalLM":
+                setattr(self, "is_moe", "num_experts_per_tok" in target_hparams)
            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
        else:
            super().set_vocab()
@@ -688,6 +691,12 @@ class DFlashModel(Qwen3Model):
            name = "model." + name
        return super().filter_tensors((name, gen))

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "model.embed_tokens.weight" and not self.hparams.get("has_embed_tokens", True):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+

@ModelBase.register("Qwen3DSparkModel")
 class DSparkModel(DFlashModel):
@@ -804,6 +804,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
 | GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute (currently top-k MoE gating). |
+| GGML_SYCL_ENABLE_ESIMD | 0 or 1 (default)| Enable ESIMD kernels when available. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
 | GGML_SYCL_USM_SYSTEM | 0 (default) or 1 | Enable experimental support for [USM system allocations](https://github.khronos.org/SYCL_Reference/iface/usm_basic_concept.html#system-allocations) for large GPU buffers. This requires enough host memory for model weights and caches, an Intel Xe2+ GPU such as BMG or newer and supported on Linux only, with CONFIG_DRM_XE_GPUSVM enabled. |
@@ -4,7 +4,7 @@

 The INI preset feature, introduced in [PR#17859](https://github.com/ggml-org/llama.cpp/pull/17859), allows users to create reusable and shareable parameter configurations for llama.cpp.

-### Using Presets with the Server
+## Using Presets with the Server

 When running multiple models on the server (router mode), INI preset files can be used to configure model-specific parameters. Please refer to the [server documentation](../tools/server/README.md) for more details.

@@ -93,3 +93,18 @@ llama-server -hf user/repo:gpt-oss-120b-hf
 ```

 Please make sure to provide the correct `hf-repo` for each child preset. Otherwise, you may get error: `The specified tag is not a valid quantization scheme.`
+
+## System-level config
+
+The system-level config, added in PR [#26118](https://github.com/ggml-org/llama.cpp/pull/26118), allows sharing the same set of options among multiple tools and examples. Unlike the sections above, it is not limited to the server.
+
+These files are loaded on startup if present. A later file overrides an earlier one:
+1. System-wide: `/etc/llama.cpp/config.ini` (or `%PROGRAMDATA%\llama.cpp\config.ini` on Windows)
+2. User-level: `$XDG_CONFIG_HOME/llama.cpp/config.ini`, `~/.config/llama.cpp/config.ini` by default (or `%APPDATA%\llama.cpp\config.ini` on Windows)
+
+The config file is applied first, then its options are overridden by ENV variables, CLI arguments and model presets (in router mode).
+
+Note:
+- Only the `[*]` and default sections are used; options written before any section header belong to "default. Named sections are ignored
+- Tool-specific options can be specified, but will be ignored (with a warning) if the example doesn't support it<br/>Example: if you specify `port = 1234`, only `llama-server` will use it, other examples will ignore it
+- `model` or `hf-repo` are not recommended to be configured system-level, because it may introduce conflicts<br/>Example: a `hf-repo` in the config file still takes effect when you pass `-m` on the command line, so you may load a different model than expected
@@ -0,0 +1,49 @@
+# Release process
+
+llama.cpp uses [semantic versioning](https://semver.org) (`MAJOR.MINOR.PATCH`).
+
+## Version bump guidelines
+
+| Change type | Version component |
+|---|---|
+| Breaking change to the public C API (`include/llama.h`)         | `MAJOR` |
+| Backward-compatible features, model support, or API addition    | `MINOR` |
+| Bug fix with no API change                                      | `PATCH` |
+
+The version is set in the three variables at the top of the root `CMakeLists.txt`:
+
+```cmake
+set(LLAMA_VERSION_MAJOR 0)
+set(LLAMA_VERSION_MINOR 1)
+set(LLAMA_VERSION_PATCH 0)
+```
+
+_A version bump should be included in the PR that introduces the change, or in a
+dedicated bump commit merged before the release is cut._
+
+_TODO: add PR labels (`semver: patch`, `semver: minor`, `semver: major`) to help
+identify which PRs require a version bump before cutting a release._
+
+## Making a release
+
+Releases are created by running the [make-release](.github/workflows/make-release.yml)
+which is a manual workflow.
+
+The workflow creates an annotated git tag (e.g. `v0.1.0`) and pushes it to the
+remote. No GitHub Release object is created, the tag is the release artifact.
+
+## Building a release
+
+By default, `LLAMA_BUILD_IS_DEV=ON` which appends a `-dev` suffix to `LLAMA_VERSION`,
+marking the build as a nightly/development build. Distributors building from a
+release tag must pass `-DLLAMA_BUILD_IS_DEV=OFF` to produce a clean version string
+(e.g. `0.1.0` instead of `0.1.0-dev`).
+
+## How releases reach users
+Currently releases are not published to github releases, only nightly/development
+builds are available there. The way users can access releases are using the following
+channels:
+
+- **llama-install.sh**  — downloads pre-built binaries built from the release tag.
+- **Package managers**  — consume the git tag directly.
+- **Build from source** — users clone the repo and check out the tag.
@@ -202,6 +202,12 @@ Example Video:

 If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.

+### Backend Sampling
+
+Use `--backend-sampling` to run supported target-model samplers on the model backend. Draft-model sampling uses the backend by default and can be controlled with `--spec-draft-backend-sampling` and `--no-spec-draft-backend-sampling`.
+
+Unsupported samplers and device layouts fall back to CPU sampling. Tensor split mode does not support backend sampling. A fixed seed produces repeatable random draws, but stochastic CPU and backend sampling can still select different tokens because floating-point operations can differ between implementations and devices. Use greedy sampling when exact output matching is required.
+
 ### General Speculative Parameters

 ```
@@ -3,9 +3,11 @@
 #include "common.h"
 #include "ngram-cache.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "log.h"
 #include "llama.h"

+#include <algorithm>
 #include <clocale>
 #include <cstdint>
 #include <cstdio>
@@ -27,6 +29,10 @@ int main(int argc, char ** argv){
    // max. number of additional tokens to draft if match is found
    const int n_draft = params.speculative.draft.n_max;

+    const auto output_limits = common_speculative_get_output_limits(params.n_batch, params.n_parallel, n_draft);
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -1,6 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch
-torchvision
+torchvision; platform_machine != "s390x"
 transformers
 huggingface-hub
 accelerate
@@ -2,12 +2,15 @@

 import argparse
 import os
+import sys
 import importlib
 import torch
 import numpy as np

 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM
-from pathlib import Path
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+from utils.common import save_output_data

 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')

@@ -54,6 +57,7 @@ print(f"Model name: {model_name}")

 prompt = "Hello world today"
 input_ids = tokenizer(prompt, return_tensors="pt").input_ids  # ty: ignore[call-non-callable]
+token_ids = input_ids[0].cpu().tolist()
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")
 print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")  # ty: ignore[unresolved-attribute]
@@ -74,21 +78,8 @@ with torch.no_grad():
    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
    print(f"Number of tokens: {token_embeddings.shape[0]}")

-    # Save raw token embeddings
-    data_dir = Path("data")
-    data_dir.mkdir(exist_ok=True)
-    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
-    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
-
-    # Save all token embeddings as binary
    print(token_embeddings)
-    token_embeddings.astype(np.float32).tofile(bin_filename)
-
-    # Save as text for inspection
-    with open(txt_filename, "w") as f:
-        for i, embedding in enumerate(token_embeddings):
-            for j, val in enumerate(embedding):
-                f.write(f"{i} {j} {val:.6f}\n")
+    save_output_data(token_embeddings, token_ids, prompt, model_name, type_suffix="-embeddings")

    # Print embeddings per token in the requested format
    print("\nToken embeddings:")
@@ -110,5 +101,3 @@ with torch.no_grad():
    for i, token in enumerate(tokens):
        print(f"  Token {i}: {repr(token)}")

-    print(f"Saved bin logits to: {bin_filename}")
-    print(f"Saved txt logist to: {txt_filename}")
@@ -3,10 +3,47 @@
 Demonstration of basic greedy speculative decoding

 ```bash
+# spec-type draft-simple
 ./bin/llama-speculative-simple \
-    -m  ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
-    -md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
-    -f test.txt -c 0 -ngl 99 --color on \
-    --sampling-seq k --top-k 1 -fa on --temp 0.0 \
-    -ngld 99 --spec-draft-n-max 16 --spec-draft-n-draft-min 5 --draft-p-min 0.9
+        -hf  ggml-org/Qwen3-8B-Base-GGUF:Q8_0 \
+        -hfd ggml-org/Qwen3-0.6B-Base-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-simple --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3.6-27B-GGUF:Q8_0 \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp (with shared KV cache)
+# note: this model needs a <s> token at the start to somewhat work without the chat template
+./bin/llama-speculative-simple \
+        -hf ggml-org/Gemma-4-31B-it-GGUF:Q8_0 \
+        -p "<s>Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-eagle3
+./bin/llama-speculative-simple \
+        -hf ggml-org/gpt-oss-20b-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-eagle3 --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dflash
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dflash --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dspark
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dspark --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
 ```
@@ -5,6 +5,7 @@
 #include "log.h"
 #include "llama.h"

+#include <algorithm>
 #include <clocale>
 #include <cstdio>
 #include <cstring>
@@ -29,6 +30,11 @@ int main(int argc, char ** argv) {
        return 1;
    }

+    const auto output_limits = common_speculative_get_output_limits(
+            params.n_batch, params.n_parallel, common_speculative_n_max(&params.speculative));
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // init llama.cpp
    llama_backend_init();
    llama_numa_init(params.numa);
@@ -45,45 +51,23 @@ int main(int argc, char ** argv) {

    const llama_vocab * vocab = llama_model_get_vocab(model_tgt);

-    // load the draft model
-    llama_model_ptr model_dft;
-    llama_context_ptr ctx_dft;
+    // load the draft model (if any) - this also creates the MTP draft context when MTP speculation is enabled
+    common_speculative_init_result_ptr spec_init;

-    // TODO: simplify this logic
    {
-        const auto & params_spec = params.speculative.draft;
+        common_params params_dft = common_base_params_to_speculative(params);

-        auto params_dft = params;
-
-        params_dft.devices      = params_spec.devices;
-        params_dft.model        = params_spec.mparams;
-        params_dft.n_gpu_layers = params_spec.n_gpu_layers;
-
-        if (params_spec.cpuparams.n_threads > 0) {
-            params_dft.cpuparams.n_threads       = params.speculative.draft.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
-        }
-
-        params_dft.tensor_buft_overrides = params.speculative.draft.tensor_buft_overrides;
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-
-        model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
-        if (model_dft == nullptr) {
-            LOG_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
-            return 1;
-        }
-
-        auto cparams = common_context_params_to_llama(params_dft);
-        ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
+        spec_init = common_speculative_init_from_params(params_dft, model_tgt, ctx_tgt);

        params.speculative.draft.ctx_tgt = ctx_tgt;
-        params.speculative.draft.ctx_dft = ctx_dft.get();
+        params.speculative.draft.ctx_dft = spec_init->context();
    }

+    llama_context * ctx_dft = params.speculative.draft.ctx_dft;
+
    // check if the context supports partial sequence removal
-    const bool use_ckpt_tgt = (common_context_can_seq_rm(ctx_tgt)       == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
-    const bool use_ckpt_dft = (common_context_can_seq_rm(ctx_dft.get()) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
+    const bool use_ckpt_tgt = common_context_can_seq_rm(ctx_tgt) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+    const bool use_ckpt_dft = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;

    if (use_ckpt_tgt) {
        LOG_INF("speculative decoding will use checkpoints (context does not support partial sequence removal)\n");
@@ -129,9 +113,30 @@ int main(int argc, char ** argv) {
    // target model sampling context
    common_sampler_ptr smpl(common_sampler_init(model_tgt, params.sampling));

-    // eval the prompt
-    llama_decode(ctx_tgt,       llama_batch_get_one(inp.data(), inp.size() - 1));
-    llama_decode(ctx_dft.get(), llama_batch_get_one(inp.data(), inp.size() - 1));
+    // init the speculator
+    const auto & params_spec = params.speculative;
+
+    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
+
+    if (spec == nullptr) {
+        LOG_ERR("%s", "failed to initialize speculative decoding\n");
+        return 1;
+    }
+
+    // eval the prompt on the target and feed it to the speculative implementation(s)
+    {
+        llama_batch batch_prompt = llama_batch_init(inp.size(), 0, 1);
+        for (size_t i = 0; i < inp.size() - 1; ++i) {
+            common_batch_add(batch_prompt, inp[i], i, { seq_id }, false);
+        }
+
+        llama_decode(ctx_tgt, batch_prompt);
+
+        if (!common_speculative_process(spec, batch_prompt)) {
+            LOG_ERR("%s", "failed to process speculative prompt\n");
+            return 1;
+        }
+    }

    // note: keep the last token separate!
    llama_token id_last = inp.back();
@@ -142,18 +147,12 @@ int main(int argc, char ** argv) {

    int n_past = inp.size() - 1;

-    // init the speculator
-    const auto & params_spec = params.speculative;
-
-    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
-
    common_speculative_begin(spec, seq_id, prompt_tgt);

    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);

-    size_t n_draft = 0;
-
    llama_tokens draft;
+
    common_prompt_checkpoint ckpt;

    const auto t_enc_end = ggml_time_us();
@@ -175,13 +174,20 @@ int main(int argc, char ** argv) {
                    llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), seq_id));

            if (use_ckpt_dft) {
-                ckpt.update_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                ckpt.update_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
            }

+            // determine the max draft that fits the remaining context and generation budget
+            int n_draft_max = (int) llama_n_ctx(ctx_tgt) - n_past - 2;
+            if (params.n_predict >= 0) {
+                n_draft_max = std::min(n_draft_max, params.n_predict - n_predict - 1);
+            }
+            n_draft_max = std::max(n_draft_max, 0);
+
            // generate a new draft
            common_speculative_get_draft_params(spec, seq_id) = {
                /* .drafting   = */ true,
-                /* .n_max      = */ -1,
+                /* .n_max      = */ n_draft_max,
                /* .n_past     = */ n_past,
                /* .id_last    = */ id_last,
                /* .prompt     = */ &prompt_tgt,
@@ -189,9 +195,6 @@ int main(int argc, char ** argv) {
            };
            common_speculative_draft(spec);

-            // save the original draft size
-            n_draft = draft.size();
-
            // save a checkpoint of the target context before evaluating the draft
            // this allows us to restore the state if partial draft acceptance occurs
            if (!draft.empty()) {
@@ -200,10 +203,13 @@ int main(int argc, char ** argv) {
                }
            }

-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            // reset the draft context to the checkpoint before verification
+            if (ctx_dft) {
+                if (use_ckpt_dft) {
+                    ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                }

-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
            }
        } else {
            // we have a previous (partial) draft to reuse from checkpoint restoration
@@ -227,10 +233,10 @@ int main(int argc, char ** argv) {
            llama_decode(ctx_tgt, batch_tgt);
        }

-        // evaluate the same batch with the draft model
-        {
-            // TODO: extend to support MTP, Eagle, etc. See server code for reference
-            llama_decode(ctx_dft.get(), batch_tgt);
+        // feed the batch to the speculative implementation(s) - this drives the draft model, MTP, Eagle3, etc.
+        if (!common_speculative_process(spec, batch_tgt)) {
+            LOG_ERR("%s", "failed to process speculative batch\n");
+            break;
        }

        // only save the sampler sampler state if we use checkpoints
@@ -239,6 +245,9 @@ int main(int argc, char ** argv) {
            smpl_save.reset(common_sampler_clone(smpl.get()));
        }

+        // save the size of the draft being verified
+        const size_t n_draft = draft.size();
+
        // sample from the full target batch and return the accepted tokens based on the target sampler
        //
        // for each token to be accepted, the sampler would have to sample that same token
@@ -255,8 +264,8 @@ int main(int argc, char ** argv) {
        // check for partial draft acceptance:
        // if the context doesn't support partial sequence removal, restore the checkpoint
        // and make the accepted tokens the new partial draft for the next iteration
-        if (use_ckpt_tgt && ids.size() - 1 < draft.size()) {
-            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, draft.size());
+        if (use_ckpt_tgt && ids.size() - 1 < n_draft) {
+            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, n_draft);

            draft = std::move(ids);

@@ -266,10 +275,10 @@ int main(int argc, char ** argv) {
                llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, ckpt.pos_max + 1, -1);
            }

-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            if (ctx_dft) {
+                ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);

-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
            }

            prompt_tgt.resize(ckpt.n_tokens);
@@ -320,8 +329,11 @@ int main(int argc, char ** argv) {
        {
            LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);

-            llama_memory_seq_rm(llama_get_memory(ctx_tgt),       seq_id, n_past, -1);
-            llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, n_past, -1);
+
+            if (ctx_dft) {
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, n_past, -1);
+            }
        }

        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
@@ -347,6 +359,7 @@ int main(int argc, char ** argv) {

    LOG_INF("\n");
    LOG_INF("draft:\n\n");
+    common_speculative_print_stats(spec);

    LOG_INF("\n");
    LOG_INF("target:\n\n");
@@ -1,6 +1,7 @@
 #include "arg.h"
 #include "common.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "log.h"
 #include "llama.h"

@@ -57,6 +58,11 @@ int main(int argc, char ** argv) {
    // max number of parallel drafting sequences (i.e. tree branches)
    const int n_seq_dft = params.n_parallel;

+    const auto output_limits = common_speculative_get_output_limits(
+            params.n_batch, params.n_parallel, params.speculative.draft.n_max);
+    params.n_outputs_max = output_limits.total;
+    params.n_outputs_max_per_seq = output_limits.per_seq;
+
    // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
    const float p_draft_split = params.speculative.draft.p_split;

@@ -83,6 +89,8 @@ int main(int argc, char ** argv) {
    params.devices = params.speculative.draft.devices;
    params.model = params.speculative.draft.mparams;
    params.n_gpu_layers = params.speculative.draft.n_gpu_layers;
+    params.n_outputs_max = params.n_parallel;
+    params.n_outputs_max_per_seq = 1;
    if (params.speculative.draft.cpuparams.n_threads > 0) {
        params.cpuparams.n_threads = params.speculative.draft.cpuparams.n_threads;
    }
@@ -0,0 +1,3 @@
+llama-build-install
+install
+build
@@ -0,0 +1,13 @@
+cmake_minimum_required(VERSION 3.14)
+project(llama-simple)
+
+set(CMAKE_CXX_STANDARD 17)
+
+find_package(llama 0.1.0 REQUIRED)
+
+add_executable(test-cmake test-cmake.cpp)
+target_link_libraries(test-cmake PRIVATE llama)
+target_compile_definitions(test-cmake PRIVATE
+    LLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}
+    LLAMA_BUILD_COMMIT="${LLAMA_BUILD_COMMIT}"
+)
@@ -0,0 +1,36 @@
+## cmake-test
+
+This is just for manually testing/developing of a llama.cpp installation to
+enable troubleshooting issues and exploration. The idea is that this can be used
+after making changes to llama.cpp installation cmake configuration and then
+verify it locally.
+
+### Usage
+The following will configure, build, and install llama.cpp
+
+Configuring/build/install:
+```console
+./build-install.sh
+```
+The above command will create a directory named `install` in the current directory
+which will have the follwing files in its lib directory:
+```console
+(venv) $ ls install/lib/
+cmake                   libggml.so          libllama-common.so.0      libllama.so.0.1.0  llama.cpp
+libggml-base.so         libggml.so.0        libllama-common.so.0.1.0  libmtmd.so         pkgconfig
+libggml-base.so.0       libggml.so.0.19.0   libllama.so               libmtmd.so.0
+libggml-base.so.0.19.0  libllama-common.so  libllama.so.0             libmtmd.so.0.1.0
+```
+
+Build/run this project using the installation created above:
+```console
+(venv) $ ./build.sh
+-- Configuring done (0.0s)
+-- Generating done (0.0s)
+-- Build files have been written to: /home/danbev/work/ai/llama.cpp/examples/test-cmake/build
+[100%] Built target test-cmake
+[test-cmake] Using llama.cpp version 0.1.0-dev-b10335
+[test-cmake] Initializing backend...
+load_backend: loaded CPU backend from /home/danbev/work/ai/llama.cpp/examples/test-cmake/install/lib/llama.cpp/libggml-cpu-alderlake.so
+[test-cmake] Backend initialized.
+```
@@ -0,0 +1,19 @@
+#!/bin/bash
+
+set -e
+
+rm -rf llama-build-install install
+
+cmake --fresh -S ../../. -B llama-build-install -DCMAKE_BUILD_TYPE=Release \
+  -DBUILD_SHARED_LIBS=ON \
+  -DGGML_BACKEND_DL=ON \
+  -DGGML_CPU_ALL_VARIANTS=ON \
+  -DLLAMA_TESTS_INSTALL=OFF \
+  -DCMAKE_INSTALL_PREFIX="${PWD}/install" \
+  -DGGML_BACKEND_DIR="${PWD}/install/lib/llama.cpp" \
+  -DGGML_LIB_INSTALL_DIR="${PWD}/install/lib/llama.cpp" \
+  -DLLAMA_LIB_INSTALL_DIR="${PWD}/install/lib/llama.cpp" \
+  -DLLAMA_TOOLS_INSTALL=OFF
+
+cmake --build llama-build-install --parallel 12
+cmake --install llama-build-install
@@ -0,0 +1,7 @@
+#!/bin/bash
+
+set -e
+
+cmake -S . -B build -DCMAKE_PREFIX_PATH="${PWD}/install"
+cmake --build build
+LD_LIBRARY_PATH="${PWD}/install/lib/llama.cpp:${PWD}/install/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" ./build/test-cmake
@@ -0,0 +1,12 @@
+#include "llama.h"
+#include <cstdio>
+
+int main(void) {
+    printf("[test-cmake] version: %s, build: %d (%s)\n",
+           llama_version(), LLAMA_BUILD_NUMBER, LLAMA_BUILD_COMMIT);
+    printf("[test-cmake] Initializing backend...\n");
+    llama_backend_init();
+    printf("[test-cmake] Backend initialized.\n");
+    llama_backend_free();
+    return 0;
+}
@@ -402,7 +402,7 @@ configure_package_config_file(
              GGML_BIN_INSTALL_DIR)

 write_basic_package_version_file(
-        ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml-config-version.cmake
    VERSION ${GGML_INSTALL_VERSION}
    COMPATIBILITY SameMajorVersion)

@@ -414,7 +414,7 @@ message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
 message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")

 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/ggml-config-version.cmake
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)

 if (MSVC)
@@ -113,6 +113,7 @@ set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
 if(NOT TARGET ggml::ggml)
    find_package(Threads REQUIRED)

+    unset(GGML_LIBRARY CACHE)
    find_library(GGML_LIBRARY ggml
        REQUIRED
        HINTS ${GGML_LIB_DIR}
@@ -121,8 +122,10 @@ if(NOT TARGET ggml::ggml)
    add_library(ggml::ggml UNKNOWN IMPORTED)
    set_target_properties(ggml::ggml
        PROPERTIES
-            IMPORTED_LOCATION "${GGML_LIBRARY}")
+            IMPORTED_LOCATION "${GGML_LIBRARY}"
+            INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}")

+    unset(GGML_BASE_LIBRARY CACHE)
    find_library(GGML_BASE_LIBRARY ggml-base
        REQUIRED
        HINTS ${GGML_LIB_DIR}
@@ -132,6 +135,7 @@ if(NOT TARGET ggml::ggml)
    set_target_properties(ggml::ggml-base
        PROPERTIES
            IMPORTED_LOCATION "${GGML_BASE_LIBRARY}"
+            INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
            INTERFACE_LINK_LIBRARIES "${GGML_BASE_INTERFACE_LINK_LIBRARIES}")

    set(_ggml_all_targets "")
@@ -140,6 +144,7 @@ if(NOT TARGET ggml::ggml)
            string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
            string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)

+            unset(${_ggml_backend_pfx}_LIBRARY CACHE)
            find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
                REQUIRED
                HINTS ${GGML_LIB_DIR}
@@ -154,6 +154,8 @@ extern "C" {
        bool buffer_from_host_ptr;
        // event synchronization
        bool events;
+        // mmap is supported for loading
+        bool mmap_support;
    };

    // all the device properties
@@ -132,6 +132,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ false, // Not implemented.
        /* .buffer_from_host_ptr  = */ false, // Not implemented.
        /* .events                = */ false, // Not implemented.
+        /* .mmap_support          = */ true,
    };
    for (ggml_backend_dev_t simple_dev : meta_dev_ctx->simple_devs) {
        ggml_backend_dev_props tmp_props;
@@ -140,6 +141,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
        props->caps.host_buffer          = props->caps.host_buffer          && tmp_props.caps.host_buffer;
        props->caps.buffer_from_host_ptr = props->caps.buffer_from_host_ptr && tmp_props.caps.buffer_from_host_ptr;
        props->caps.events               = props->caps.events               && tmp_props.caps.events;
+        props->caps.mmap_support         = props->caps.mmap_support         && tmp_props.caps.mmap_support;
    }
 }

@@ -367,6 +367,7 @@ static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct gg
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ true,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -2815,6 +2815,7 @@ static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ true,
+        /* .mmap_support          = */ true,
    };
 }

@@ -1,90 +1,12 @@
 #include "ggml-backend-impl.h"
+#include "ggml-feats.h"

-#if defined(__aarch64__)
-
-#if defined(__linux__)
-#include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP_FPHP)
-#define HWCAP_FPHP (1 << 9)
-#endif
-
-#if !defined(HWCAP_ASIMDHP)
-#define HWCAP_ASIMDHP (1 << 10)
-#endif
-
-#if !defined(HWCAP_ASIMDDP)
-#define HWCAP_ASIMDDP (1 << 20)
-#endif
-
-#if !defined(HWCAP_SVE)
-#define HWCAP_SVE (1 << 22)
-#endif
-
-#if !defined(HWCAP2_SVE2)
-#define HWCAP2_SVE2 (1 << 1)
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
-#endif
-
-struct aarch64_features {
-    // has_neon not needed, aarch64 has NEON guaranteed
-    bool has_dotprod     = false;
-    bool has_fp16        = false;
-    bool has_sve         = false;
-    bool has_sve2        = false;
-    bool has_i8mm        = false;
-    bool has_sme         = false;
-    bool has_sme2        = false;
-
-    aarch64_features() {
-#if defined(__linux__)
-        uint32_t hwcap = getauxval(AT_HWCAP);
-        uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-        has_fp16    = !!(hwcap & HWCAP_FPHP) && !!(hwcap & HWCAP_ASIMDHP);
-        has_sve     = !!(hwcap & HWCAP_SVE);
-        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
-        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-        has_sme     = !!(hwcap2 & HWCAP2_SME);
-#elif defined(__APPLE__)
-        int oldp = 0;
-        size_t size = sizeof(oldp);
-
-        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
-            has_dotprod = static_cast<bool>(oldp);
-        }
-
-        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
-            has_i8mm = static_cast<bool>(oldp);
-        }
-
-        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
-            has_sme = static_cast<bool>(oldp);
-        }
-
-        if (sysctlbyname("hw.optional.arm.FEAT_SME2", &oldp, &size, NULL, 0) == 0) {
-            has_sme2 = static_cast<bool>(oldp);
-        }
-
-        // Apple apparently does not implement SVE yet
-#endif
-    }
-};
+#if defined(__aarch64__) || defined(_M_ARM64)

 static int ggml_backend_cpu_aarch64_score() {
    int score = 1;
-    aarch64_features af;
+    const ggml_feats_arch64_runtime_t af = ggml_feats_get_arch64_runtime();
+    GGML_UNUSED(af);

 #ifdef GGML_USE_DOTPROD
    if (!af.has_dotprod) { return 0; }
@@ -116,4 +38,4 @@ static int ggml_backend_cpu_aarch64_score() {

 GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)

-# endif // defined(__aarch64__)
+# endif // defined(__aarch64__) || defined(_M_ARM64)
@@ -2608,7 +2608,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
    return true;
 }

-#elif defined(__gnu_linux__)
+#elif defined(__linux__)
 // TODO: this may not work on BSD, to be verified

 static bool ggml_thread_apply_affinity(const bool * mask) {
@@ -397,6 +397,7 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ true,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -2,10 +2,12 @@
 // SPDX-License-Identifier: MIT
 //
 #include <arm_neon.h>
-#include <assert.h>
-#include <stdio.h>
+#include <cassert>
+#include <cstdio>
+#include <cstdlib>
 #include <atomic>
 #include <cfloat>
+#include <cctype>
 #include <algorithm>
 #include <cmath>
 #include <stdexcept>
@@ -17,25 +19,21 @@
 #include <cstddef>
 #include <cstdint>
 #include <fstream>
-#include <set>
+#include <map>
 #include <iostream>
 #include <climits>
+#include <charconv>
+#include <system_error>
 #if defined(__linux__)
 #include <asm/hwcap.h>
+#include <dirent.h>
 #include <sys/auxv.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <unistd.h>
-#ifndef HWCAP2_SME2
-#define HWCAP2_SME2 (1UL << 37)
-#endif
 #elif defined(__APPLE__)
-#include <string_view>
 #include <sys/sysctl.h>
 #include <sys/types.h>
-#elif defined(_WIN32)
-#include <windows.h>
-#include <excpt.h>
 #endif

 #include "kleidiai.h"
@@ -43,6 +41,7 @@
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
+#include "ggml-feats.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
 #include "traits.h"
@@ -64,8 +63,8 @@ struct ggml_kleidiai_context {
    ggml_kleidiai_kernels * kernels_q4;
    ggml_kleidiai_kernels * kernels_q8;
    ggml_kleidiai_kernels * kernels_f32;
-    int sme_thread_cap; // <= 0 means “SME disabled/unknown”;
-    int thread_hint;    // <= 0 means “no hint”
+    int sme_thread_cap; // <= 0 means "SME disabled/unknown"
+    int thread_hint;    // <= 0 means "no hint"
    int chunk_multiplier;
 } static ctx = { CPU_FEATURE_NONE, nullptr, nullptr, nullptr, 0, -1, 4 };

@@ -93,24 +92,117 @@ static const char* cpu_feature_to_string(cpu_feature f) {
    }
 }

+#if defined(__linux__) && defined(__aarch64__)
+static bool parse_cpu_dir_name(const char* name, size_t* cpu) {
+    if (strncmp(name, "cpu", 3) != 0 ||
+        name[3] < '0' || name[3] > '9') {
+        return false;
+    }
+
+    const char* first = name + 3;
+    const char* last = name + strlen(name);
+
+    size_t value = 0;
+    const auto [end, ec] = std::from_chars(first, last, value, 10);
+
+    if (ec != std::errc{} || end != last) {
+        return false;
+    }
+
+    *cpu = value;
+    return true;
+}
+
+static std::vector<size_t> detect_cpu_ids() {
+    std::vector<size_t> cpus;
+
+    DIR * dir = opendir("/sys/devices/system/cpu");
+    if (dir == nullptr) {
+        return cpus;
+    }
+
+    while (dirent * entry = readdir(dir)) {
+        size_t cpu = 0;
+        if (parse_cpu_dir_name(entry->d_name, &cpu)) {
+            cpus.push_back(cpu);
+        }
+    }
+    closedir(dir);
+
+    std::sort(cpus.begin(), cpus.end());
+    cpus.erase(std::unique(cpus.begin(), cpus.end()), cpus.end());
+    return cpus;
+}
+#endif
+
+#if defined(__APPLE__) && defined(__aarch64__)
+static bool apple_sme_counted_perf_level(std::string name) {
+    for (std::string::size_type i = 0; i < name.size(); ++i) {
+        name[i] = (char) std::tolower((unsigned char) name[i]);
+    }
+
+    // Conservative ceiling: only count perf-level names observed to provide full SME throughput.
+    // Future names should be calibrated here before they raise the automatic SME thread cap.
+    return name.find("super") != std::string::npos ||
+           name.find("performance") != std::string::npos;
+}
+#endif
+
+static void add_smcus_from_smidr(uint64_t smidr, size_t & num_private, std::map<uint32_t, size_t> & shared_counts) {
+    // Arm ARM: SMIDR_EL1. SH==0 is implementation-defined; keep the existing
+    // conservative policy and only treat zero affinity as private.
+    const uint32_t sh = (uint32_t)((smidr >> 13) & 0x3);
+    const uint32_t nsmc = (uint32_t)((smidr >> 56) & 0xF);
+    const size_t shared_count = nsmc == 0xF ? 1 : (size_t)nsmc + 1;
+    const uint32_t affinity = (uint32_t)(smidr & 0xFFFu);
+    const uint32_t affinity2 = (uint32_t)((smidr >> 32) & 0xFFFFFu);
+    const uint32_t id = (affinity2 << 12) | affinity;
+
+    if (nsmc == 0xF) {
+        GGML_LOG_WARN("kleidiai: NSMC detected as 0xF indicating reseved value, setting min safe shared SMCU count to 1");
+    }
+
+    switch (sh) {
+        case 2: // private SMCU
+            ++num_private;
+            break;
+        case 3: // shared SMCU
+            if (shared_counts[id] < shared_count) {
+                shared_counts[id] = shared_count;
+            }
+            break;
+        case 0:
+            if (id == 0) {
+                ++num_private;
+            } else if (shared_counts[id] < shared_count) {
+                shared_counts[id] = shared_count;
+            }
+            break;
+        default:
+            break;
+    }
+}
+
 static size_t detect_num_smcus() {
-    if (!ggml_cpu_has_sme()) {
+    const auto runtime_feat = ggml_feats_get_arch64_runtime();
+    if (!runtime_feat.has_sme) {
        return 0;
    }

 #if defined(__linux__) && defined(__aarch64__)
    // Linux/aarch64: Best-effort count of Streaming Mode Compute Units (SMCUs) via SMIDR_EL1 sysfs.
    size_t num_private = 0;
-    std::set<uint32_t> shared_ids;
+    std::map<uint32_t, size_t> shared_counts;

-    for (size_t cpu = 0;; ++cpu) {
+    const std::vector<size_t> cpus = detect_cpu_ids();
+    for (const size_t cpu : cpus) {
        const std::string path =
            "/sys/devices/system/cpu/cpu" + std::to_string(cpu) +
            "/regs/identification/smidr_el1";

        std::ifstream file(path);
        if (!file.is_open()) {
-            break;
+            continue;
        }

        uint64_t smidr = 0;
@@ -118,54 +210,69 @@ static size_t detect_num_smcus() {
            continue;
        }

-        // Arm ARM: SMIDR_EL1
-        const uint32_t sh = (uint32_t)((smidr >> 13) & 0x3);
-        // Build an "affinity-like" identifier for shared SMCUs.
-        // Keep the original packing logic, but isolate it here.
-        const uint32_t id = (uint32_t)((smidr & 0xFFFu) | ((smidr >> 20) & 0xFFFFF000u));
-
-        switch (sh) {
-            case 0b10: // private SMCU
-                ++num_private;
-                break;
-            case 0b11: // shared SMCU
-                shared_ids.emplace(id);
-                break;
-            case 0b00:
-                // Ambiguous / implementation-defined. Be conservative:
-                // treat id==0 as private, otherwise as shared.
-                if (id == 0) ++num_private;
-                else shared_ids.emplace(id);
-                break;
-            default:
-                break;
-        }
+        add_smcus_from_smidr(smidr, num_private, shared_counts);
    }

-    return num_private + shared_ids.size();
+    size_t total = num_private;
+    for (const auto & entry : shared_counts) {
+        total += entry.second;
+    }
+    return total;

 #elif defined(__APPLE__) && defined(__aarch64__)
-    // table for known M4 variants. Users can override via GGML_KLEIDIAI_SME=<n>.
-    char chip_name[256] = {};
-    size_t size = sizeof(chip_name);
+    int perf_levels = 0;
+    size_t size = sizeof(perf_levels);
+    if (sysctlbyname("hw.nperflevels", &perf_levels, &size, nullptr, 0) != 0 ||
+        size != sizeof(perf_levels) || perf_levels <= 0) {
+        return 0;
+    }

-    if (sysctlbyname("machdep.cpu.brand_string", chip_name, &size, nullptr, 0) == 0) {
-        const std::string brand(chip_name);
+    size_t units = 0;
+    for (int i = 0; i < perf_levels; ++i) {
+        char key[64] = {};
+        int physical_cpus = 0;
+        int cpus_per_l2 = 0;

-        struct ModelSMCU { const char *match; size_t smcus; };
-        static const ModelSMCU table[] = {
-            { "M4 Ultra", 2 },
-            { "M4 Max",   2 },
-            { "M4 Pro",   2 },
-            { "M4",       1 },
-        };
+        snprintf(key, sizeof(key), "hw.perflevel%d.physicalcpu", i);
+        size = sizeof(physical_cpus);
+        if (sysctlbyname(key, &physical_cpus, &size, nullptr, 0) != 0 ||
+            size != sizeof(physical_cpus) || physical_cpus <= 0) {
+            continue;
+        }

-        for (const auto &e : table) {
-            if (brand.find(e.match) != std::string::npos) {
-                return e.smcus;
-            }
+        snprintf(key, sizeof(key), "hw.perflevel%d.cpusperl2", i);
+        size = sizeof(cpus_per_l2);
+        if (sysctlbyname(key, &cpus_per_l2, &size, nullptr, 0) != 0 ||
+            size != sizeof(cpus_per_l2) || cpus_per_l2 <= 0) {
+            continue;
+        }
+
+        snprintf(key, sizeof(key), "hw.perflevel%d.name", i);
+        size = 0;
+        if (sysctlbyname(key, nullptr, &size, nullptr, 0) != 0 || size == 0) {
+            continue;
+        }
+
+        std::string name(size, '\0');
+        if (sysctlbyname(key, &name[0], &size, nullptr, 0) != 0) {
+            continue;
+        }
+        name.resize(size);
+        while (!name.empty() && name.back() == '\0') {
+            name.pop_back();
+        }
+
+        if (apple_sme_counted_perf_level(name)) {
+            units += (size_t) ((physical_cpus + cpus_per_l2 - 1) / cpus_per_l2);
        }
    }
+
+    return units;
+
+#elif defined(_WIN32) && (defined(_M_ARM64) || defined(__aarch64__))
+    // No verified Windows arm64 SMCU detection path yet. Return unknown and use
+    // GGML_KLEIDIAI_SME=N as a diagnostics/debug override for SME thread cap
+    // calibration until a detection mechanism is verified on real hardware.
    return 0;

 #else
@@ -198,15 +305,18 @@ static void init_kleidiai_context(void) {
    if (!initialized) {
        initialized = true;

+        // Optional diagnostics/debug overrides; production defaults come from runtime detection.
        const char *env_sme         = getenv("GGML_KLEIDIAI_SME");
        const char *env_threads     = getenv("GGML_TOTAL_THREADS");
        const char *env_chunk_mult  = getenv("GGML_KLEIDIAI_CHUNK_MULTIPLIER");

+        const auto runtime_feat = ggml_feats_get_arch64_runtime();
+
        size_t detected_smcus = 0;

-        ctx.features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
-                        (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                        ((ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
+        ctx.features  = (runtime_feat.has_dotprod  ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
+                        (runtime_feat.has_i8mm     ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
+                        (runtime_feat.sve_cnt == QK8_0 ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);

        if (env_threads) {
            bool ok = false;
@@ -224,54 +334,54 @@ static void init_kleidiai_context(void) {
            }
        }

-        // SME policy:
-        // - env unset => auto-detect SMCUs; enable SME only if detected > 0.
-        // - env=0     => force off.
-        // - env>0     => force N cores, if the binary was built with SME.
        int sme_cores = 0;
        bool sme_env_ok = false;
        bool sme_env_set = (env_sme != nullptr);

+        const bool has_supported_sme_family = runtime_feat.has_sme;
+        bool sme_cap_detected = false;
+
+        if (has_supported_sme_family) {
+            detected_smcus = detect_num_smcus();
+            sme_cap_detected = detected_smcus > 0;
+            // Some platforms expose SME without exposing a calibrated SMCU count.
+            // Use one SME thread as the conservative default; add platform SMCU detection to raise it.
+            sme_cores = sme_cap_detected ? (int)detected_smcus : 1;
+
+            if (!sme_env_set && !sme_cap_detected) {
+                GGML_LOG_INFO("kleidiai: SME detected; SMCU count unavailable, using conservative SME thread cap=1\n");
+            }
+        }
+
+        // Runtime-detect SME support and available SMCUs first. The detected SMCU
+        // count is used as the SME thread cap, and GGML_KLEIDIAI_SME can debug-override that:
+        //   - unset: use runtime detection.
+        //   - 0:     disable SME-family kernels.
+        //   - N > 0: use N as the SME thread cap, if an SME-family kernel is selectable.
        if (sme_env_set) {
            bool ok = false;
            int v = parse_uint_env(env_sme, "GGML_KLEIDIAI_SME", &ok);
            sme_env_ok = ok;

-            if (!ok) {
-                GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME set but parsing failed; falling back to runtime SME-core detection\n");
-                detected_smcus = detect_num_smcus();
-                sme_cores = detected_smcus > 0 ? (int)detected_smcus : 0;
-            } else if (v == 0) {
-                sme_cores = 0;
-            } else if (!ggml_cpu_has_sme()) {
-                GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME=%d but the binary was not built with SME; disabling SME\n", v);
-                sme_cores = 0;
+            if (ok) {
+                if (has_supported_sme_family) {
+                    sme_cores = v;
+                } else {
+                    if (v > 0) {
+                        GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME=%d but SME is not supported on this CPU; disabling SME-family kernels\n", v);
+                    }
+                    sme_cores = 0;
+                }
            } else {
-                sme_cores = v;
+                GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME set but parsing failed; using automatic SME thread cap\n");
            }
-        } else {
-            detected_smcus = detect_num_smcus();
-            sme_cores = detected_smcus > 0 ? (int)detected_smcus : 0;
        }

-        if (!sme_env_set && ggml_cpu_has_sme() && sme_cores == 0) {
-            GGML_LOG_WARN("kleidiai: runtime SME-core detection returned 0; falling back to NEON\n");
-        }
-
-        if (sme_cores > 0) {
+        if (sme_cores > 0 && has_supported_sme_family) {
            ctx.features |= CPU_FEATURE_SME;
-#if defined(__aarch64__) && defined(__linux__)
-            // ARM guarantees SME2 implies SME, so only check SME2 when SME is enabled.
-            if (getauxval(AT_HWCAP2) & HWCAP2_SME2) {
+            if (runtime_feat.has_sme2) {
                ctx.features |= CPU_FEATURE_SME2;
            }
-#elif defined(__aarch64__) && defined(__APPLE__)
-            int feat_sme2 = 0;
-            size_t size = sizeof(feat_sme2);
-            if (sysctlbyname("hw.optional.arm.FEAT_SME2", &feat_sme2, &size, NULL, 0) == 0 && feat_sme2) {
-                ctx.features |= CPU_FEATURE_SME2;
-            }
-#endif
        }

        // Kernel selection
@@ -297,16 +407,19 @@ static void init_kleidiai_context(void) {
            GGML_LOG_INFO("kleidiai: primary f32 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_f32->required_cpu));
        }

-        ctx.sme_thread_cap = (ctx.features & CPU_FEATURE_SME) ? sme_cores : 0;
+        const bool has_selected_sme_family_kernel =
+            (ctx.kernels_q4  && is_sme_family(ctx.kernels_q4->required_cpu)) ||
+            (ctx.kernels_q8  && is_sme_family(ctx.kernels_q8->required_cpu)) ||
+            (ctx.kernels_f32 && is_sme_family(ctx.kernels_f32->required_cpu));
+        ctx.sme_thread_cap = has_selected_sme_family_kernel ? sme_cores : 0;

-        if (ctx.features & CPU_FEATURE_SME) {
-            const bool has_sme2 = (ctx.features & CPU_FEATURE_SME2) != CPU_FEATURE_NONE;
+        if (has_selected_sme_family_kernel) {
            if (sme_env_set && sme_env_ok && sme_cores > 0) {
-                GGML_LOG_INFO("kleidiai: SME%s enabled (GGML_KLEIDIAI_SME=%d override)\n",
-                    has_sme2 ? "2" : "", sme_cores);
+                GGML_LOG_INFO("kleidiai: SME enabled (GGML_KLEIDIAI_SME=%d debug override)\n", sme_cores);
+            } else if (sme_cap_detected) {
+                GGML_LOG_INFO("kleidiai: SME enabled (runtime-detected SME thread cap=%d)\n", sme_cores);
            } else {
-                GGML_LOG_INFO("kleidiai: SME%s enabled (runtime-detected SME cores=%d)\n",
-                    has_sme2 ? "2" : "", sme_cores);
+                GGML_LOG_INFO("kleidiai: SME enabled (runtime SME detected, conservative thread cap=%d)\n", sme_cores);
            }
        } else {
            GGML_LOG_INFO("kleidiai: SME disabled\n");
@@ -467,7 +580,7 @@ static int kleidiai_collect_kernel_chain_common(
    }

    if (is_sme_family(primary->required_cpu)) {
-        const cpu_feature fallback_mask = static_cast<cpu_feature>(features & ~CPU_FEATURE_SME & ~CPU_FEATURE_SME2);
+        const cpu_feature fallback_mask = static_cast<cpu_feature>(features & ~(CPU_FEATURE_SME | CPU_FEATURE_SME2));
        if (fallback_mask != CPU_FEATURE_NONE) {
            ggml_kleidiai_kernels * fallback = select_fallback(fallback_mask);
            if (fallback && fallback != primary &&
@@ -1077,13 +1190,14 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        const int ith_total = params->ith;

        int sme_slot = -1;
+        int non_sme_slot = -1;
        for (int i = 0; i < runtime_count; ++i) {
            if (is_sme_family(runtime[i].kernels->required_cpu)) {
                sme_slot = i;
                break;
            }
        }
-        int non_sme_slot = -1;
+
        for (int i = 0; i < runtime_count; ++i) {
            if (!is_sme_family(runtime[i].kernels->required_cpu)) {
                non_sme_slot = i;
@@ -8941,7 +8941,7 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
            for (int tk = 0; tk < kv_tile; tk++) {
                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
                if (kv_type == GGML_TYPE_F16) {
-                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                    ggml_cpu_fp16_to_fp32((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
                } else {
                    memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
                }
@@ -195,6 +195,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -214,6 +215,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    case GGML_TYPE_Q4_K:
                    case GGML_TYPE_Q6_K:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q5_K:
                        //case GGML_TYPE_MXFP4:
@@ -1865,6 +1865,37 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    ggml_cuda_mul_mat_cublas(ctx, src0, src1, dst);
 }

+// returns true when ggml_cuda_mul_mat_id takes the fallback path that requires stream synchronization
+// [TAG_MUL_MAT_ID_CUDA_GRAPHS]
+static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int cc) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return true;
+    }
+
+    if (dst->ne[2] <= MMVQ_MAX_BATCH_SIZE) {
+        if (ggml_is_quantized(src0->type)) {
+            if (dst->ne[2] <= get_mmvq_mmid_max_batch(src0->type, cc)) {
+                return false;
+            }
+        } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+            return false;
+        }
+    }
+
+    if (ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[2], /*n_experts=*/src0->ne[2])) {
+        return false;
+    }
+
+    if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
+        return false;
+    }
+
+    return true;
+}
+
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
@@ -1907,7 +1938,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
    }

    // note: this path should not be reached when recording CUDA graphs, because it requires stream synchronization
-    // TODO: add asserts to verify this. should work with CUDA, HIP, etc.
+    GGML_ASSERT(ggml_cuda_mul_mat_id_needs_sync(dst, cc));
    cudaStream_t stream = ctx.stream();

    GGML_ASSERT(nb12 % nb11 == 0);
@@ -2522,10 +2553,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
        // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
        if (node->op == GGML_OP_MUL_MAT_ID) {
            const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
-            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
-                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+            if (ggml_cuda_mul_mat_id_needs_sync(node, cc)) {
+                // the mul_mat_id fallback path synchronizes the stream, so we cannot use CUDA graphs
                // ref: https://github.com/ggml-org/llama.cpp/pull/18958
                use_cuda_graph = false;
 #ifndef NDEBUG
@@ -2651,6 +2680,52 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
    return true;
 }

+static bool ggml_cuda_should_fuse_rms_norm_mul_rope(const ggml_tensor * rms_norm,
+                                                    const ggml_tensor * mul,
+                                                    const ggml_tensor * rope) {
+    if (rms_norm->op != GGML_OP_RMS_NORM || mul->op != GGML_OP_MUL || rope->op != GGML_OP_ROPE) {
+        return false;
+    }
+
+    if (rms_norm->src[0]->type != GGML_TYPE_F32 || rms_norm->type != GGML_TYPE_F32 ||
+        mul->src[0]->type != GGML_TYPE_F32 || mul->src[1]->type != GGML_TYPE_F32 ||
+        mul->type != GGML_TYPE_F32 || rope->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (rope->src[0] != mul) {
+        return false;
+    }
+
+    //if rms norm is the B operand, then we don't handle broadcast
+    if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
+        return false;
+    }
+
+    if (!ggml_are_same_shape(rms_norm, mul)) {
+        return false;
+    }
+
+    //rms_norm kernel assumes contiguous rows
+    if (!ggml_is_contiguous_rows(rms_norm->src[0]) ||
+        !ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+        return false;
+    }
+
+    // the fused kernel handles the norm/neox rope modes only
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    const int n_dims = ((const int32_t *) rope->op_params)[1];
+    if (n_dims % 2 != 0 || rope->src[0]->ne[0] % 2 != 0) {
+        return false;
+    }
+
+    return true;
+}
+
 // match gated_delta_net + the strided cpy that scatters its state snapshots into the cache
 // (slot i -> rollback group i, slot 0 newest), so the kernel can write them and skip the cpy.
 static int ggml_cuda_try_gdn_cache_fusion(
@@ -2980,6 +3055,36 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
        }
    }

+    std::initializer_list<enum ggml_op> rms_norm_mul_rope_ops          = { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE };
+    std::initializer_list<enum ggml_op> rms_norm_mul_rope_set_rows_ops = { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
+
+    if (is_equal(rms_norm_mul_rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
+        const ggml_tensor * rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor * mul      = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * rope     = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * view     = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 4];
+
+        if (ggml_check_edges(cgraph, node_idx, {{1, 0, 0}, {2, 0, 1}, {3, 0, 2}, {4, 0, 3}}) &&
+            ggml_cuda_should_fuse_rms_norm_mul_rope(rms_norm, mul, rope) &&
+            ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
+            int out_nodes[] = { node_idx + 4 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
+        }
+    }
+
+    if (is_equal(rms_norm_mul_rope_ops, ops) && ggml_can_fuse(cgraph, node_idx, ops)) {
+        const ggml_tensor * rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor * mul      = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * rope     = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_rms_norm_mul_rope(rms_norm, mul, rope)) {
+            int out_nodes[] = { node_idx + 2 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
+        }
+        return false;
+    }
+
    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };

    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
@@ -2988,7 +3093,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];

        if (ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
-            return true;
+            int out_nodes[] = { node_idx + 2 };
+            return ggml_cuda_check_fusion_memory_ranges(cgraph, node_idx, (int)ops.size(), out_nodes, 1);
        }
    }

@@ -3840,6 +3946,16 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
        return fused_node_count - 1;
    }

+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+        ggml_cuda_op_rms_norm_mul_rope_fused(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 4]);
+        return 4;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE }, {})) {
+        ggml_cuda_op_rms_norm_mul_rope_fused(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2], nullptr);
+        return 2;
+    }
+
    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD }, {})) {
        ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);
        return 2;
@@ -4714,6 +4830,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
        /* .host_buffer           = */ host_buffer,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ events,
+        /* .mmap_support          = */ props->type != GGML_BACKEND_DEVICE_TYPE_IGPU,
    };
 }

@@ -5098,7 +5215,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            return max_bias == 0.0f;
        }
        case GGML_OP_ROLL:
-            if(op->src[0]->type == GGML_TYPE_F32) {
+            if(op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0])) {
                return true;
            }
            return false;
@@ -670,3 +670,238 @@ void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rope, ggml_tensor * set_rows) {
    ggml_cuda_op_rope_impl<true>(ctx, rope, set_rows);
 }
+
+// fused RMS_NORM + MUL + ROPE (+ VIEW + SET_ROWS)
+// one block per row: block_reduce gives the norm scale, then each thread applies mul and rope to the elements it owns
+template <int block_size, bool has_ff, typename D>
+static __global__ void rms_norm_mul_rope_f32(
+        const float * x, D * dst, const int ncols,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s1, const int64_t s2, const int64_t s3,
+        const float eps,
+        const float * mul,
+        const int64_t mul_s01, const int64_t mul_s02, const int64_t mul_s03,
+        const uint3 mul_ncols_packed, const uint3 mul_nrows_packed,
+        const uint3 mul_nchannels_packed, const uint3 mul_nsamples_packed,
+        const int n_dims, const int32_t * pos,
+        const float freq_scale, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float theta_scale,
+        const float * freq_factors,
+        const int64_t * row_indices, const int set_rows_stride,
+        const bool is_neox) {
+    ggml_cuda_pdl_lc();
+    const int row     = blockIdx.x;
+    const int channel = blockIdx.y;
+    const int sample  = blockIdx.z;
+    const int tid     = threadIdx.x;
+
+    x += sample*s03 + channel*s02 + row*s01;
+
+    const uint32_t mul_row     = fastmodulo(row,     mul_nrows_packed);
+    const uint32_t mul_channel = fastmodulo(channel, mul_nchannels_packed);
+    const uint32_t mul_sample  = fastmodulo(sample,  mul_nsamples_packed);
+    mul += mul_sample*mul_s03 + mul_channel*mul_s02 + mul_row*mul_s01;
+
+    float tmp = 0.0f;
+
+    ggml_cuda_pdl_sync();
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[col];
+        tmp += xi * xi;
+    }
+
+    extern __shared__ float s_sum[];
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+
+    const float scale = rsqrtf(tmp/ncols + eps);
+
+    int64_t idst = sample*s3 + channel*s2 + row*s1;
+    if (set_rows_stride != 0) {
+        idst = row*s1 + row_indices[channel]*set_rows_stride;
+    }
+    dst += idst;
+
+    for (int i0 = 2*tid; i0 < ncols; i0 += 2*block_size) {
+        int ix0;
+        int ix1;
+        if (is_neox && i0 < n_dims) {
+            ix0 = i0/2;
+            ix1 = i0/2 + n_dims/2;
+        } else {
+            ix0 = i0 + 0;
+            ix1 = i0 + 1;
+        }
+
+        const float x0 = scale * x[ix0] * mul[fastmodulo(ix0, mul_ncols_packed)];
+        const float x1 = scale * x[ix1] * mul[fastmodulo(ix1, mul_ncols_packed)];
+
+        if (i0 >= n_dims) {
+            dst[ix0] = ggml_cuda_cast<D>(x0);
+            dst[ix1] = ggml_cuda_cast<D>(x1);
+            continue;
+        }
+
+        const float theta_base  = pos[channel]*powf(theta_scale, i0/2.0f);
+        const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+        float cos_theta;
+        float sin_theta;
+        rope_yarn<true>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+
+        dst[ix0] = ggml_cuda_cast<D>(x0*cos_theta - x1*sin_theta);
+        dst[ix1] = ggml_cuda_cast<D>(x0*sin_theta + x1*cos_theta);
+    }
+}
+
+template <typename D>
+static void rms_norm_mul_rope_cuda(
+        const float * x, D * dst,
+        const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s1, const int64_t s2, const int64_t s3,
+        const float eps,
+        const float * mul,
+        const int64_t mul_s01, const int64_t mul_s02, const int64_t mul_s03,
+        const uint32_t mul_ncols, const uint32_t mul_nrows,
+        const uint32_t mul_nchannels, const uint32_t mul_nsamples,
+        const int n_dims, const int32_t * pos,
+        const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims,
+        const float * freq_factors,
+        const int64_t * row_indices, const int set_rows_stride,
+        const bool is_neox, cudaStream_t stream) {
+    GGML_ASSERT(ncols % 2 == 0);
+
+    const dim3 blocks_num(nrows, nchannels, nsamples);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    const uint3 mul_ncols_packed     = init_fastdiv_values(mul_ncols);
+    const uint3 mul_nrows_packed     = init_fastdiv_values(mul_nrows);
+    const uint3 mul_nchannels_packed = init_fastdiv_values(mul_nchannels);
+    const uint3 mul_nsamples_packed  = init_fastdiv_values(mul_nsamples);
+
+    if (ncols < 1024) {
+        const dim3 block_dims(256, 1, 1);
+        const ggml_cuda_kernel_launch_params launch_params = {blocks_num, block_dims, 32*sizeof(float), stream};
+        if (freq_factors == nullptr) {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<256, false, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        } else {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<256, true, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        }
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        const ggml_cuda_kernel_launch_params launch_params = {blocks_num, block_dims, 32*sizeof(float), stream};
+        if (freq_factors == nullptr) {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<1024, false, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        } else {
+            ggml_cuda_kernel_launch(rms_norm_mul_rope_f32<1024, true, D>, launch_params,
+                x, dst, ncols, s01, s02, s03, s1, s2, s3, eps, mul, mul_s01, mul_s02, mul_s03,
+                mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+                freq_factors, row_indices, set_rows_stride, is_neox);
+        }
+    }
+}
+
+void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx,
+        ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows) {
+    const ggml_tensor * x = rms_norm->src[0];
+    const ggml_tensor * mul_src = mul->src[0] == rms_norm ? mul->src[1] : mul->src[0];
+
+    float eps = 0.0f;
+    memcpy(&eps, rms_norm->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    GGML_ASSERT(x->type == GGML_TYPE_F32);
+    GGML_ASSERT(mul_src->type == GGML_TYPE_F32);
+    GGML_ASSERT(rope->type == GGML_TYPE_F32);
+
+    void *          dst_d           = rope->data;
+    ggml_type       dst_type        = rope->type;
+    const int64_t * row_indices     = nullptr;
+    int             set_rows_stride = 0;
+
+    if (set_rows != nullptr) {
+        dst_d           = set_rows->data;
+        dst_type        = set_rows->type;
+        row_indices     = (const int64_t *) set_rows->src[1]->data;
+        set_rows_stride = set_rows->nb[1] / ggml_type_size(set_rows->type);
+    }
+
+    const int n_dims     = ((const int32_t *) rope->op_params)[1];
+    const int mode       = ((const int32_t *) rope->op_params)[2];
+    const int n_ctx_orig = ((const int32_t *) rope->op_params)[4];
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+
+    memcpy(&freq_base,   (const int32_t *) rope->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (const int32_t *) rope->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (const int32_t *) rope->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (const int32_t *) rope->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (const int32_t *) rope->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (const int32_t *) rope->op_params + 10, sizeof(float));
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    const int32_t * pos = (const int32_t *) rope->src[1]->data;
+
+    const float * freq_factors = rope->src[2] != nullptr ? (const float *) rope->src[2]->data : nullptr;
+
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+
+    const size_t ts0 = ggml_type_size(x->type);
+    GGML_ASSERT(x->nb[0] == ts0);
+    const int64_t s01 = x->nb[1] / ts0;
+    const int64_t s02 = x->nb[2] / ts0;
+    const int64_t s03 = x->nb[3] / ts0;
+
+    const size_t ts_mul = ggml_type_size(mul_src->type);
+    GGML_ASSERT(mul_src->nb[0] == ts_mul);
+    const int64_t mul_s01 = mul_src->nb[1] / ts_mul;
+    const int64_t mul_s02 = mul_src->nb[2] / ts_mul;
+    const int64_t mul_s03 = mul_src->nb[3] / ts_mul;
+
+    const size_t ts_dst = ggml_type_size(rope->type);
+    const int64_t s1 = rope->nb[1] / ts_dst;
+    const int64_t s2 = rope->nb[2] / ts_dst;
+    const int64_t s3 = rope->nb[3] / ts_dst;
+
+    cudaStream_t stream = ctx.stream();
+
+    if (dst_type == GGML_TYPE_F32) {
+        rms_norm_mul_rope_cuda((const float *) x->data, (float *) dst_d,
+            x->ne[0], x->ne[1], x->ne[2], x->ne[3], s01, s02, s03, s1, s2, s3, eps,
+            (const float *) mul_src->data, mul_s01, mul_s02, mul_s03,
+            mul_src->ne[0], mul_src->ne[1], mul_src->ne[2], mul_src->ne[3],
+            n_dims, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+            freq_factors, row_indices, set_rows_stride, is_neox, stream);
+    } else if (dst_type == GGML_TYPE_F16) {
+        rms_norm_mul_rope_cuda((const float *) x->data, (half *) dst_d,
+            x->ne[0], x->ne[1], x->ne[2], x->ne[3], s01, s02, s03, s1, s2, s3, eps,
+            (const float *) mul_src->data, mul_s01, mul_s02, mul_s03,
+            mul_src->ne[0], mul_src->ne[1], mul_src->ne[2], mul_src->ne[3],
+            n_dims, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+            freq_factors, row_indices, set_rows_stride, is_neox, stream);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
@@ -7,3 +7,5 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
+
+void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows);
@@ -141,6 +141,57 @@ static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, cons
    }
 }

+template <int rows_per_block>
+static __global__ void __launch_bounds__(WARP_SIZE * rows_per_block, 2)
+rwkv_wkv7_f32_t1_warp_row(const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
+    constexpr int head_size = CUDA_WKV_BLOCK_SIZE;
+    constexpr int half_head = head_size / 2;
+
+    const int lane = threadIdx.x;
+    const int row  = blockIdx.y * rows_per_block + threadIdx.y;
+    const int bid  = blockIdx.x;
+
+    const int batch_i = bid / H;
+    const int head_i  = bid % H;
+    const int state_size = C * head_size;
+    const int head_off = head_i * head_size;
+    const int t = batch_i * C + head_off + row;
+
+    __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
+
+    if (threadIdx.y == 0) {
+        _r[lane] = r[batch_i * C + head_off + lane];
+        _w[lane] = w[batch_i * C + head_off + lane];
+        _k[lane] = k[batch_i * C + head_off + lane];
+        _a[lane] = a[batch_i * C + head_off + lane];
+        _b[lane] = b[batch_i * C + head_off + lane];
+
+        _r[lane + half_head] = r[batch_i * C + head_off + lane + half_head];
+        _w[lane + half_head] = w[batch_i * C + head_off + lane + half_head];
+        _k[lane + half_head] = k[batch_i * C + head_off + lane + half_head];
+        _a[lane + half_head] = a[batch_i * C + head_off + lane + half_head];
+        _b[lane + half_head] = b[batch_i * C + head_off + lane + half_head];
+    }
+    __syncthreads();
+
+    const int64_t state_base = batch_i * state_size + head_i * head_size * head_size + row * head_size;
+    const float s0 = s[state_base + lane];
+    const float s1 = s[state_base + lane + half_head];
+    const float sa = warp_reduce_sum(_a[lane] * s0 + _a[lane + half_head] * s1);
+
+    const float vt  = v[t];
+    const float st0 = s0 * _w[lane]             + _k[lane]             * vt + sa * _b[lane];
+    const float st1 = s1 * _w[lane + half_head] + _k[lane + half_head] * vt + sa * _b[lane + half_head];
+    const float y   = warp_reduce_sum(st0 * _r[lane] + st1 * _r[lane + half_head]);
+
+    dst[T * C + state_base + lane]             = st0;
+    dst[T * C + state_base + lane + half_head] = st1;
+
+    if (lane == 0) {
+        dst[t] = y;
+    }
+}
+
 void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const float * k_d  = (const float *)dst->src[0]->data;
    const float * v_d  = (const float *)dst->src[1]->data;
@@ -191,7 +242,10 @@ void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
    GGML_ASSERT(C % H == 0);
    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);

-    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+    if (T / B == 1 && C / H == CUDA_WKV_BLOCK_SIZE) {
+        constexpr int rows_per_block = 4;
+        rwkv_wkv7_f32_t1_warp_row<rows_per_block><<<dim3(B * H, CUDA_WKV_BLOCK_SIZE / rows_per_block), dim3(WARP_SIZE, rows_per_block), 0, stream>>>(T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    } else if (C / H == CUDA_WKV_BLOCK_SIZE) {
        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
    } else {
        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
@@ -1646,6 +1646,7 @@ static void ggml_backend_et_device_get_props(ggml_backend_dev_t dev, struct ggml
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -0,0 +1,166 @@
+#pragma once
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#include <sys/prctl.h>
+
+#if !defined(HWCAP2_SVE2)
+#define HWCAP2_SVE2 (1ULL << 1)
+#endif
+
+#if !defined(HWCAP_FPHP)
+#define HWCAP_FPHP (1 << 9)
+#endif
+
+#if !defined(HWCAP_ASIMDHP)
+#define HWCAP_ASIMDHP (1 << 10)
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM (1ULL << 13)
+#endif
+
+#if !defined(HWCAP_ASIMDDP)
+#define HWCAP_ASIMDDP (1 << 20)
+#endif
+
+#if !defined(HWCAP_SVE)
+#define HWCAP_SVE (1 << 22)
+#endif
+
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1ULL << 23)
+#endif
+
+#if !defined(HWCAP2_SME2)
+#define HWCAP2_SME2 (1ULL << 37)
+#endif
+
+#if !defined(PR_SVE_GET_VL)
+#define PR_SVE_GET_VL 51
+#endif
+
+#if !defined(PR_SVE_VL_LEN_MASK)
+#define PR_SVE_VL_LEN_MASK 0xffff
+#endif
+
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#elif defined(_WIN32)
+#include <windows.h>
+
+#if !defined(PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE 43
+#endif
+
+#if !defined(PF_ARM_SVE_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_SVE_INSTRUCTIONS_AVAILABLE 46
+#endif
+
+#if !defined(PF_ARM_SVE2_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_SVE2_INSTRUCTIONS_AVAILABLE 47
+#endif
+
+#if !defined(PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE 66
+#endif
+
+#if !defined(PF_ARM_V82_FP16_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_V82_FP16_INSTRUCTIONS_AVAILABLE 67
+#endif
+
+#if !defined(PF_ARM_SME_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_SME_INSTRUCTIONS_AVAILABLE 70
+#endif
+
+#if !defined(PF_ARM_SME2_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_SME2_INSTRUCTIONS_AVAILABLE 71
+#endif
+
+#endif
+
+typedef struct ggml_feats_arch64_runtime {
+    bool has_dotprod;
+    bool has_fp16;
+    bool has_sve;
+    bool has_sve2;
+    bool has_i8mm;
+    bool has_sme;
+    bool has_sme2;
+    int sve_cnt;
+} ggml_feats_arch64_runtime_t;
+
+static inline ggml_feats_arch64_runtime_t ggml_feats_get_arch64_runtime(void) {
+    ggml_feats_arch64_runtime_t runtime_feat = {};
+
+#if defined(__linux__)
+    const unsigned long hwcap  = getauxval(AT_HWCAP);
+    const unsigned long hwcap2 = getauxval(AT_HWCAP2);
+
+    runtime_feat.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+    runtime_feat.has_fp16    = !!(hwcap & HWCAP_FPHP) && !!(hwcap & HWCAP_ASIMDHP);;
+    runtime_feat.has_sve     = !!(hwcap & HWCAP_SVE);
+    runtime_feat.has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
+    runtime_feat.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+    runtime_feat.has_sme     = !!(hwcap2 & HWCAP2_SME);
+    runtime_feat.has_sme2    = !!(hwcap2 & HWCAP2_SME2);
+
+    if (runtime_feat.has_sve) {
+        const int vl = prctl(PR_SVE_GET_VL);
+        if (vl >= 0) {
+            runtime_feat.sve_cnt = vl & PR_SVE_VL_LEN_MASK;
+        }
+    }
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+
+    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_dotprod = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_FP16", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_fp16 = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_SVE", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_sve = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_SVE2", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_sve2 = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_i8mm = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_sme = static_cast<bool>(oldp);
+    }
+
+    if (sysctlbyname("hw.optional.arm.FEAT_SME2", &oldp, &size, nullptr, 0) == 0) {
+        runtime_feat.has_sme2 = static_cast<bool>(oldp);
+    }
+
+    // Apple does not support userspace non-streaming SVE; keep SVE vector length unknown.
+    runtime_feat.sve_cnt = 0;
+#elif defined (_WIN32)
+    runtime_feat.has_dotprod = IsProcessorFeaturePresent(PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_fp16    = IsProcessorFeaturePresent(PF_ARM_V82_FP16_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_sve     = IsProcessorFeaturePresent(PF_ARM_SVE_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_sve2    = IsProcessorFeaturePresent(PF_ARM_SVE2_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_i8mm    = IsProcessorFeaturePresent(PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_sme     = IsProcessorFeaturePresent(PF_ARM_SME_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_sme2    = IsProcessorFeaturePresent(PF_ARM_SME2_INSTRUCTIONS_AVAILABLE) != 0;
+
+    // Windows exposes SVE feature presence, but not the runtime SVE vector length here.
+    runtime_feat.sve_cnt = 0;
+#endif
+
+    return runtime_feat;
+}
+
+#endif // defined(__aarch64__) || defined(_M_ARM64)
@@ -3930,6 +3930,7 @@ static void ggml_backend_hexagon_device_get_props(ggml_backend_dev_t dev, struct
        /* .host_buffer           = */ (bool) opt_hostbuf,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ false,
    };
 }

@@ -126,9 +126,6 @@ if (GGML_HIP_EXPORT_METRICS)
    set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
 endif()

-# Fast math for HIP, like CUDA's -use_fast_math. Not -ffast-math: that implies -ffinite-math-only, which breaks ggml's INFINITY masking and produces NaNs.
-set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -funsafe-math-optimizations")
-
 if (NOT GGML_CUDA_FA)
    add_compile_definitions(GGML_CUDA_NO_FA)
 endif()
@@ -953,6 +953,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
                nr0 = N_R0_IQ4_XS;
                smem = 32*sizeof(float);
            } break;
+        case GGML_TYPE_TQ2_0:
+            {
+                nsg = N_SG_TQ2_0;
+                nr0 = N_R0_TQ2_0;
+            } break;
        default:
            {
                GGML_LOG_ERROR("Asserting on type %d\n", (int) tsrc0);
@@ -1182,6 +1187,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m
                nr0 = N_R0_IQ4_XS;
                smem = 32*sizeof(float);
            } break;
+        case GGML_TYPE_TQ2_0:
+            {
+                nsg = N_SG_TQ2_0;
+                nr0 = N_R0_TQ2_0;
+            } break;
        default:
            {
                GGML_LOG_ERROR("Asserting on type %d\n", (int)op->src[2]->type);
@@ -1268,8 +1268,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_ARGSORT:
        case GGML_OP_TOP_K:
        case GGML_OP_ARANGE:
-        case GGML_OP_ROLL:
            return true;
+        case GGML_OP_ROLL:
+            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_FLASH_ATTN_EXT:
            // for new head sizes, add checks here
            if (op->src[0]->ne[0] != 32 &&
@@ -1406,6 +1407,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                           case GGML_TYPE_Q5_0:
                           case GGML_TYPE_Q5_1:
                           case GGML_TYPE_IQ4_NL:
+                           case GGML_TYPE_TQ2_0:
                           case GGML_TYPE_I32:
                                return true;
                           default:
@@ -1434,6 +1436,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_TQ2_0:
                        switch (op->type) {
                            case GGML_TYPE_F32:
                            case GGML_TYPE_F16:
@@ -1469,6 +1472,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_IQ4_NL:
+                    case GGML_TYPE_TQ2_0:
                        return true;
                    default:
                        return false;
@@ -87,6 +87,9 @@
 #define N_R0_IQ4_XS 2
 #define N_SG_IQ4_XS 2

+#define N_R0_TQ2_0 4
+#define N_SG_TQ2_0 2
+
 // function constants offsets
 #define FC_FLASH_ATTN_EXT_PAD          100
 #define FC_FLASH_ATTN_EXT_BLK          200
@@ -681,6 +681,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
        /* .host_buffer          = */ false,
        /* .buffer_from_host_ptr = */ true,
        /* .events               = */ true,
+        /* .mmap_support         = */ true,
    };
 }

@@ -468,6 +468,34 @@ void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
 }

+void quantize_tq2_0(device const float * src, device block_tq2_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK_K; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = (half) d;
+
+    for (int j = 0; j < QK_K/4; j += 32) {
+        for (int m = 0; m < 32; ++m) {
+            uint8_t q = 0;
+            for (int n = 0; n < 4; ++n) {
+                // -1, 0, 1 -> 0, 1, 2
+                int xi = (int)round(src[m + n*32] * id) + 1;
+                q += (uint8_t)((xi & 3) << (2*n));
+            }
+            dst.qs[j + m] = q;
+        }
+        src += 4*32;
+    }
+}
+
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
@@ -1021,6 +1049,25 @@ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4
    }
 }

+template <typename type4x4>
+void dequantize_tq2_0(device const block_tq2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    // 2 bits per element, 4 elements per byte, 128 elements per 32-byte group
+    const short base = il * 16;
+    for (int k = 0; k < 16; k++) {
+        const int i = base + k;
+        const int byte = ((i >> 7) & 1) * 32 + (i & 31);
+        const int l = (i >> 5) & 3;
+        reg_f[k/4][k%4] = d * (float)(((qs[byte] >> (2*l)) & 3) - 1);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
 enum ggml_sort_order {
    GGML_SORT_ORDER_ASC,
    GGML_SORT_ORDER_DESC,
@@ -8001,6 +8048,7 @@ template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_
 template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
 template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
 template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_cpy_f32_tq2_0")]]  kernel cpy_f_q_t kernel_cpy_f32_q<QK_K,   block_tq2_0,  quantize_tq2_0>;

 template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
 kernel void kernel_cpy_q_f32(
@@ -8048,6 +8096,8 @@ template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<
 template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
 template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;

+template [[host_name("kernel_cpy_tq2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
 template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
 template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
 template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
@@ -8056,6 +8106,8 @@ template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<
 template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
 template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;

+template [[host_name("kernel_cpy_tq2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
 template<typename T>
 kernel void kernel_concat(
        constant ggml_metal_kargs_concat & args,
@@ -9822,6 +9874,121 @@ kernel void kernel_mul_mv_mxfp4_f32(
    kernel_mul_mv_mxfp4_f32_impl<N_R0_MXFP4, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }

+template<int nr0, typename args_t>
+void kernel_mul_mv_tq2_0_f32_impl(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+    const short NSG = FC_mul_mv_nsg;
+
+    const int nb = args.ne00/QK_K;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * NSG + sgitg) * nr0;
+
+    const uint i12 = im%FC_mul_mv_ne12;
+    const uint i13 = im/FC_mul_mv_ne12;
+
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
+
+    device const float * y = (device const float *) (src1 + offset1);
+
+    device const block_tq2_0 * ax[nr0];
+    for (int row = 0; row < nr0; ++row) {
+        const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/FC_mul_mv_r2)*args.nb02 + (i13/FC_mul_mv_r3)*args.nb03;
+        ax[row] = (device const block_tq2_0 *) ((device char *) src0 + offset0);
+    }
+
+    float sumf[nr0] = {0.f};
+
+    // 8 threads per block, NBLOCK blocks per pass, 2 halves per block per pass
+    constexpr short NBLOCK = 4;
+
+    constexpr short NB = N_SIMDWIDTH/NBLOCK; // threads per block
+
+    const short blk = tiisg / NB;    // 0..NBLOCK-1, block handled by this thread
+    const short htg = tiisg % NB;    // 0..NB-1, thread within block (0..7)
+
+    // byte and y base offsets within the block (32 elements per thread, 4 per byte)
+    device const float4 * yb4 = (device const float4 *)(y + 4*htg + blk*QK_K);
+
+    // hoisted per-byte coefficients (from y) and total y-sum, shared across rows
+    // ref: https://github.com/ggml-org/llama.cpp/pull/26980
+    float4 coef[4];
+
+    for (int ib = blk; ib < nb; ib += NBLOCK) {
+        FOR_UNROLL (short h0 = 0; h0 < 2; ++h0) {
+            const float4 y0 = yb4[ 0 + 32*h0];
+            const float4 y1 = yb4[ 8 + 32*h0];
+            const float4 y2 = yb4[16 + 32*h0];
+            const float4 y3 = yb4[24 + 32*h0];
+
+            float sumy = 0.f;
+            FOR_UNROLL (short j = 0; j < 4; ++j) {
+                coef[j] = float4(
+                        y0[j],
+                        y1[j] - 4.0f*y0[j],
+                        y2[j] - 4.0f*y1[j],
+                        y3[j] - 4.0f*y2[j]);
+
+                sumy += (y0[j] + y1[j]) + (y2[j] + y3[j]);
+            }
+
+            FOR_UNROLL (short row = 0; row < nr0; ++row) {
+                device const block_tq2_0 & xb = ax[row][ib];
+                device const uchar * qs = xb.qs + 4*htg + 32*h0;
+
+                float sum = -sumy;
+                FOR_UNROLL (short j = 0; j < 4; ++j) {
+                    // express the 2-bit field shifts (v>>2, v>>4, v>>6) as float floor ops
+                    const float v = (float)qs[j];
+
+                    const float f0 = v;
+                    const float f1 = floor(v*0.25f);    // v>>2
+                    const float f2 = floor(v*0.0625);   // v>>4
+                    const float f3 = floor(v*0.015625); // v>>6
+
+                    sum += coef[j][0]*f0 + coef[j][1]*f1 + coef[j][2]*f2 + coef[j][3]*f3;
+                }
+
+                sumf[row] += xb.d * sum;
+            }
+        }
+
+        yb4 += QK_K * NBLOCK / 4;
+    }
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
+    for (int row = 0; row < nr0; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < args.ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_tq2_0_f32")]]
+kernel void kernel_mul_mv_tq2_0_f32(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_tq2_0_f32_impl<N_R0_TQ2_0, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows_q(
        constant ggml_metal_kargs_get_rows & args,
@@ -9915,6 +10082,38 @@ template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get
 template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+template [[host_name("kernel_get_rows_tq2_0")]]   kernel get_rows_q_t kernel_get_rows_q<block_tq2_0,   QK_NL, dequantize_tq2_0>;
+
+template<typename TS, typename TI, short QK, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + QK*ind, dst_row[ind]);
+    }
+}

 template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
 kernel void kernel_set_rows_q32(
@@ -10011,6 +10210,11 @@ template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t k
 template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
 template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;

+typedef decltype(kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>) set_rows_qK_t;
+
+template [[host_name("kernel_set_rows_f32_i64_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>;
+template [[host_name("kernel_set_rows_f32_i32_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int32_t, QK_K, block_tq2_0, quantize_tq2_0>;
+
 kernel void kernel_diag_f32(
        constant ggml_metal_kargs_diag & args,
        device   const char * src0,
@@ -10786,6 +10990,7 @@ template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_m
 template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;

 template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
@@ -10811,6 +11016,7 @@ template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_m
 template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;

 //
 // indirect matrix-matrix multiplication
@@ -10845,6 +11051,7 @@ template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_m
 template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;

 template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
@@ -10870,6 +11077,7 @@ template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_m
 template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;

 //
 // matrix-vector multiplication
@@ -11027,6 +11235,7 @@ template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl  <N_R0_IQ2_S>>>;
 template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl <N_R0_IQ4_NL>>>;
 template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl <N_R0_IQ4_XS>>>;
+template [[host_name("kernel_mul_mv_id_tq2_0_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_tq2_0_f32_impl  <N_R0_TQ2_0>>>;

 kernel void kernel_pool_2d_max_f32(
        constant    ggml_metal_kargs_pool_2d & args,
@@ -73,6 +73,7 @@ typedef const void * (*get_adreno_bin_kernel_func_t)(
 //------------------------------------------------------------------------------

 bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor);
+
 static bool ggml_cl_is_q4_0_soa(const ggml_tensor * tensor);
 static bool ggml_cl_is_q8_0_soa(const ggml_tensor * tensor);
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
@@ -4629,6 +4630,23 @@ static std::string ggml_opencl_fa_compile_opts(ggml_backend_opencl_context * bac
    if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E) {
        opts += " -D FA_C8_NO_SG_PIN";
    }
+    // Transposed K tile in local memory: the KV rows the QK loop walks together become
+    // adjacent, so a group of them is ONE 128-bit local read instead of several narrow
+    // ones. The QK loop is LDS-read-issue-bound (a wrong-math probe that kept every FMA/dp4a
+    // but removed the LDS reads ran the kernel ~40% faster), so this is worth up to +26% on
+    // fa=1 prefill. Output is bit-identical -- only the layout moves.
+    //
+    // DK <= 128 only. At DK=256 (gemma-3-4b) it measures 1-2% NEGATIVE and reproduces across
+    // rounds; padding the row stride does not recover it, so the cause is not a simple bank
+    // conflict and the wider tile does not want this layout.
+    //
+    // Default on within that gate; GGML_OPENCL_FA_K_LDS_T=0 restores the row-major tile.
+    {
+        const char * e = getenv("GGML_OPENCL_FA_K_LDS_T");
+        if ((e == nullptr || e[0] != '0') && cfg->dk <= 128) {
+            opts += " -D FA_K_LDS_T";
+        }
+    }
    return opts;
 }

@@ -4911,8 +4929,13 @@ static bool ggml_opencl_ensure_fa_variant(ggml_backend_opencl_context * backend_
        const int x = (e && e[0]) ? atoi(e) : 0;
        return (x == 8 || x == 16 || x == 32) ? x : 0;   // 0 = per-gen default
    }();
+    // X2E needs 16 to keep per-lane o_acc at 128B (the compiler spills the
+    // kernel-default width); X1E does not spill, but C=16 is still a measured
+    // +28-30% DK128-GQA4 decode win there (X1-85, kv 4096/8192), neutral on
+    // DK64 / GQA1 / quant-KV.
    const int fa_cl_c_gqa4 = fa_cl_c_env ? fa_cl_c_env
-        : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E ? 16 : 0);
+        : (backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E ||
+           backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E ? 16 : 0);
    const std::string opts_cl_c_gqa4 = fa_cl_c_gqa4
        ? " -D FA_CL_C=" + std::to_string(fa_cl_c_gqa4) : std::string();
    const std::string fa_cl_c_g8_val = std::to_string(fa_cl_c_gqa4 ? fa_cl_c_gqa4 * 2 : 16);
@@ -7058,6 +7081,19 @@ inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backen
    return ((elem_num < 128 * 1024 * 1024) && adreno_kernel && shape_ok);  // max element num: 2**27
 }

+inline bool enable_adreno_trans_weight_q5_K(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+    if (!use_adreno_kernels(backend_ctx, tensor)) {
+        return false;
+    }
+
+    const size_t elem_num = ggml_nelements(tensor);
+    const size_t q_img_width = elem_num / 8;
+    const size_t qh_img_width = elem_num / 16;
+
+    return q_img_width <= backend_ctx->image_max_buffer_size &&
+           qh_img_width <= backend_ctx->image_max_buffer_size;
+}
+
 static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
    // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
@@ -9237,7 +9273,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
        cl_kernel kernel = backend_ctx->kernel_convert_block_q5_K;
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {
            kernel = backend_ctx->kernel_convert_block_q5_K_noshuffle;
        }
 #else
@@ -9272,7 +9308,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,

        tensor->extra = extra;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {

            int M = tensor->ne[1];
            int K = tensor->ne[0];
@@ -10370,7 +10406,7 @@ 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)) {
+        if (enable_adreno_trans_weight_q5_K(backend_ctx, tensor)) {
            int M = tensor->ne[1];
            int K = tensor->ne[0];

@@ -10777,6 +10813,7 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ false,
    };
 }

@@ -18909,7 +18946,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
        }

        // q5_K x fp32
-        if (src0t == GGML_TYPE_Q5_K && src1t == GGML_TYPE_F32) {
+        if (src0t == GGML_TYPE_Q5_K && src1t == GGML_TYPE_F32 &&
+            enable_adreno_trans_weight_q5_K(backend_ctx, src0)) {
            ggml_cl_mul_mat_q5_K_f32_adreno(backend, src0, src1, dst);
            return;
        }
@@ -211,7 +211,30 @@ __kernel void FA_TILE_NAME(

    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

+#ifdef FA_K_LDS_T
+    // K tile transposed: [dk vec][kv row] instead of [kv row][dk vec].
+    //
+    // The QK loop walks 2 or 4 KV rows at a time against the same dk element. Row-major
+    // those are DK_VEC half4s apart, so each is its own 64-bit local read. Transposed they
+    // are adjacent, so a pair is one 128-bit read -- half the LDS issues for the same bytes,
+    // no extra registers, arithmetic untouched.
+    //
+    // This kernel looked like it should be FMA-bound (a half4 mad does ~4 ALU ops per LDS
+    // read, unlike the 1:1 of the dp4a loop), but it is NOT: a wrong-math probe that kept
+    // every FMA and removed the LDS reads ran it 38.6% faster (18.92 -> 11.62 ms/op).
+    // Explicitly 16-byte aligned: FA_LK_PAIR below reads two adjacent half4 as one float4,
+    // and the element type only obliges the compiler to align this array to 8. The indices
+    // are even so the offset is a multiple of 16, but the base has to be too, and relying
+    // on the compiler to over-align it is relying on luck.
+    __local KV_DATA_TYPE4 l_k[DK_VEC][BLOCK_N] __attribute__((aligned(16)));
+#define FA_LK(ROW, C) l_k[C][ROW]
+    // Two adjacent KV rows as one 128-bit local read (half4 pair == 16 B). j is even and
+    // BLOCK_N is even, so &l_k[c][j] is 16 B past a 16 B-aligned base.
+#define FA_LK_PAIR(C, J) as_half8(*(__local const float4 *)(&l_k[C][J]))
+#else
    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+#define FA_LK(ROW, C) l_k[ROW][C]
+#endif
    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

 #if N_SPLIT > 1 && !defined(HAS_SUBGROUP_SHUFFLE)
@@ -254,17 +277,17 @@ __kernel void FA_TILE_NAME(
 #ifdef FA_K_IMG
                if (use_kv_pad) {
                    const ulong k_row_offset = batch_idx * k_tile_nb3 + head_kv_idx * k_tile_nb2 + k_row_idx * k_nb1;
-                    l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
+                    FA_LK(row, col) = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
                } else {
                    const int k_row_px = batch_idx * k_pitch_px_batch + head_kv_idx * k_pitch_px_head + k_row_idx * k_pitch_px_row;
-                    l_k[row][col] = read_imageh(k_img, k_row_px + col);
+                    FA_LK(row, col) = read_imageh(k_img, k_row_px + col);
                }
 #else
                const ulong k_row_offset = batch_idx * k_tile_nb3 + head_kv_idx * k_tile_nb2 + k_row_idx * k_nb1;
-                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
+                FA_LK(row, col) = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
 #endif
            } else {
-                l_k[row][col] = (KV_DATA_TYPE4)(0.0h);
+                FA_LK(row, col) = (KV_DATA_TYPE4)(0.0h);
            }
        }
        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
@@ -292,8 +315,15 @@ __kernel void FA_TILE_NAME(
                FA_UNROLL
                for (int k = 0; k < SPLIT_DK_VEC; k++) {
                    const ACC_TYPE4 qk = q_priv[k];
+#if defined(FA_K_LDS_T)
+                    // 2 KV rows adjacent in the transposed tile: one 128-bit local read.
+                    const half8 kk = FA_LK_PAIR(dk_off + k, j);
+                    ACC_TYPE4 dot0 = qk * CONVERT_KV_ACC4(kk.lo);
+                    ACC_TYPE4 dot1 = qk * CONVERT_KV_ACC4(kk.hi);
+#else
                    ACC_TYPE4 dot0 = qk * CONVERT_KV_ACC4(l_k[j  ][dk_off + k]);
                    ACC_TYPE4 dot1 = qk * CONVERT_KV_ACC4(l_k[j+1][dk_off + k]);
+#endif
                    partial0 += dot0.s0 + dot0.s1 + dot0.s2 + dot0.s3;
                    partial1 += dot1.s0 + dot1.s1 + dot1.s2 + dot1.s3;
                }
@@ -359,7 +389,7 @@ __kernel void FA_TILE_NAME(
            ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
            FA_UNROLL
            for (int k = 0; k < SPLIT_DK_VEC; k++) {
-                dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][dk_off + k]), dot_acc);
+                dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(FA_LK(j, dk_off + k)), dot_acc);
            }
            local_partial[j][tid] =
                dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3;
@@ -452,10 +482,21 @@ __kernel void FA_TILE_NAME(
                FA_UNROLL
                for (int k = 0; k < DK_VEC; k++) {
                    const ACC_TYPE4 qk = q_priv[k];
+#if defined(FA_K_LDS_T)
+                    // 4 KV rows adjacent in the transposed tile: two 128-bit local reads
+                    // instead of four 64-bit ones.
+                    const half8 kk01 = FA_LK_PAIR(k, j);
+                    const half8 kk23 = FA_LK_PAIR(k, j + 2);
+                    dot_acc0 = mad(qk, CONVERT_KV_ACC4(kk01.lo), dot_acc0);
+                    dot_acc1 = mad(qk, CONVERT_KV_ACC4(kk01.hi), dot_acc1);
+                    dot_acc2 = mad(qk, CONVERT_KV_ACC4(kk23.lo), dot_acc2);
+                    dot_acc3 = mad(qk, CONVERT_KV_ACC4(kk23.hi), dot_acc3);
+#else
                    dot_acc0 = mad(qk, CONVERT_KV_ACC4(l_k[j][k]),   dot_acc0);
                    dot_acc1 = mad(qk, CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
                    dot_acc2 = mad(qk, CONVERT_KV_ACC4(l_k[j+2][k]), dot_acc2);
                    dot_acc3 = mad(qk, CONVERT_KV_ACC4(l_k[j+3][k]), dot_acc3);
+#endif
                }
                ACC_TYPE s0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
                ACC_TYPE s1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
@@ -1631,8 +1631,25 @@ __kernel void flash_attn_f32_q4_0(
    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

 #ifdef FA_HAVE_INT_DOT
+// Accessors so the staging code is layout-agnostic.
+#ifdef FA_K_LDS_T
+#define FA_K_PACKED(ROW, IDX) l_k_packed[IDX][ROW]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[BLK][ROW]
+#else
+#define FA_K_PACKED(ROW, IDX) l_k_packed[ROW][IDX]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[ROW][BLK]
+#endif
+
+#ifdef FA_K_LDS_T
+    // K tile transposed: the 4 KV rows the QK loop walks together become adjacent, so each
+    // (block, group) step is ONE 128-bit local read instead of four 32-bit ones. The QK
+    // loop is LDS-read-issue-bound.
+    __local uint  l_k_packed[DK_Q4_BLOCKS_PREFILL * 8][BLOCK_N];
+    __local float l_k_scale [DK_Q4_BLOCKS_PREFILL][BLOCK_N];
+#else
    __local uint  l_k_packed[BLOCK_N][DK_Q4_BLOCKS_PREFILL * 8];
    __local float l_k_scale [BLOCK_N][DK_Q4_BLOCKS_PREFILL];
+#endif
 #else
    __local half4 l_k[BLOCK_N][DK_VEC];
 #endif
@@ -1660,17 +1677,17 @@ __kernel void flash_attn_f32_q4_0(
                    const global char * blk_ptr = k_base + k_row_off + blk * Q4_0_BLOCK_SIZE;
                    const float df = (float) vload_half(0, (const global half *) blk_ptr);
                    const global uchar * qs = (const global uchar *)(blk_ptr + 2);
-                    l_k_scale[row][blk] = df;
+                    FA_K_SCALE(row, blk) = df;
                    uint k_packed[8];
                    pack_q4_0_nibbles(qs, k_packed);
                    #pragma unroll
                    for (int j = 0; j < 8; ++j) {
-                        l_k_packed[row][blk * 8 + j] = k_packed[j];
+                        FA_K_PACKED(row, blk * 8 + j) = k_packed[j];
                    }
                } else {
-                    l_k_scale[row][blk] = 0.0f;
+                    FA_K_SCALE(row, blk) = 0.0f;
                    #pragma unroll
-                    for (int j = 0; j < 8; ++j) l_k_packed[row][blk * 8 + j] = 0u;
+                    for (int j = 0; j < 8; ++j) FA_K_PACKED(row, blk * 8 + j) = 0u;
                }
            }
 #else
@@ -1760,6 +1777,19 @@ __kernel void flash_attn_f32_q4_0(
                for (int b_local = 0; b_local < SPLIT_DK_Q4_BLOCKS; ++b_local) {
                    const int b = k_blk_base + b_local;
                    int sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+#ifdef FA_K_LDS_T
+                    // 4 KV rows are adjacent in the transposed tile: one 128-bit local
+                    // read per (block, group) instead of four 32-bit ones.
+                    #pragma unroll
+                    for (int g = 0; g < 8; ++g) {
+                        const uint  qp  = q_packed_pf[b_local * 8 + g];
+                        const uint4 kq4 = vload4(0, &l_k_packed[b * 8 + g][j]);
+                        sum0 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s0, sum0);
+                        sum1 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s1, sum1);
+                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s2, sum2);
+                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s3, sum3);
+                    }
+#else
                    #pragma unroll
                    for (int g = 0; g < 8; ++g) {
                        const uint qp = q_packed_pf[b_local * 8 + g];
@@ -1768,12 +1798,21 @@ __kernel void flash_attn_f32_q4_0(
                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+2][b * 8 + g], sum2);
                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+3][b * 8 + g], sum3);
                    }
+#endif
                    const float qd    = q_d_pf[b_local];
                    const int   q_sum = q_sum_pf[b_local];
+#ifdef FA_K_LDS_T
+                    const float4 ks4 = vload4(0, &l_k_scale[b][j]);
+                    s0 += (float)(sum0 - 8 * q_sum) * qd * ks4.s0;
+                    s1 += (float)(sum1 - 8 * q_sum) * qd * ks4.s1;
+                    s2 += (float)(sum2 - 8 * q_sum) * qd * ks4.s2;
+                    s3 += (float)(sum3 - 8 * q_sum) * qd * ks4.s3;
+#else
                    s0 += (float)(sum0 - 8 * q_sum) * qd * l_k_scale[j  ][b];
                    s1 += (float)(sum1 - 8 * q_sum) * qd * l_k_scale[j+1][b];
                    s2 += (float)(sum2 - 8 * q_sum) * qd * l_k_scale[j+2][b];
                    s3 += (float)(sum3 - 8 * q_sum) * qd * l_k_scale[j+3][b];
+#endif
                }
 #else
                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
@@ -1393,8 +1393,31 @@ __kernel void flash_attn_f32_q8_0(
    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);

 #ifdef FA_HAVE_INT_DOT
+// Accessors so the staging code is layout-agnostic.
+#ifdef FA_K_LDS_T
+#define FA_K_PACKED(ROW, IDX) l_k_packed[IDX][ROW]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[BLK][ROW]
+#else
+#define FA_K_PACKED(ROW, IDX) l_k_packed[ROW][IDX]
+#define FA_K_SCALE(ROW, BLK)  l_k_scale[ROW][BLK]
+#endif
+
+#ifdef FA_K_LDS_T
+    // K tile transposed: [block*8 + g][kv row] instead of [kv row][block*8 + g].
+    //
+    // The QK loop walks 4 KV rows at a time against the same (b, g), so in the original
+    // layout those 4 values are BLOCK_N*8 uints apart and cost 4 separate 32-bit local
+    // reads. Transposed they are adjacent, so they are one 128-bit read -- 4x fewer LDS
+    // issues for the same bytes and no extra registers. That matters because the QK loop
+    // is LDS-read-issue-bound: a wrong-math probe that kept every dp4a but cut the LDS
+    // reads ran the whole kernel 41% faster (18.51 -> 10.91 ms/op), and deleting QK
+    // outright only reached 10.88 -- i.e. essentially ALL of QK's cost is these reads.
+    __local uint  l_k_packed[DK_Q8_BLOCKS_PREFILL * 8][BLOCK_N];
+    __local float l_k_scale [DK_Q8_BLOCKS_PREFILL][BLOCK_N];
+#else
    __local uint  l_k_packed[BLOCK_N][DK_Q8_BLOCKS_PREFILL * 8];
    __local float l_k_scale [BLOCK_N][DK_Q8_BLOCKS_PREFILL];
+#endif
 #else
    __local half4 l_k[BLOCK_N][DK_VEC];
 #endif
@@ -1427,7 +1450,7 @@ __kernel void flash_attn_f32_q8_0(
                    const global char * blk_ptr = k_base + k_row_off + blk * Q8_0_BLOCK_SIZE;
                    const float df = (float) vload_half(0, (const global half *) blk_ptr);
                    const global uchar * qs = (const global uchar *)(blk_ptr + 2);
-                    l_k_scale[row][blk] = df;
+                    FA_K_SCALE(row, blk) = df;
                    #pragma unroll
                    for (int j = 0; j < 8; ++j) {
                        uint k_packed =
@@ -1435,12 +1458,12 @@ __kernel void flash_attn_f32_q8_0(
                             ((uint) qs[j*4 + 1]) <<  8 |
                             ((uint) qs[j*4 + 2]) << 16 |
                             ((uint) qs[j*4 + 3]) << 24;
-                        l_k_packed[row][blk * 8 + j] = k_packed;
+                        FA_K_PACKED(row, blk * 8 + j) = k_packed;
                    }
                } else {
-                    l_k_scale[row][blk] = 0.0f;
+                    FA_K_SCALE(row, blk) = 0.0f;
                    #pragma unroll
-                    for (int j = 0; j < 8; ++j) l_k_packed[row][blk * 8 + j] = 0u;
+                    for (int j = 0; j < 8; ++j) FA_K_PACKED(row, blk * 8 + j) = 0u;
                }
            }
 #else
@@ -1556,6 +1579,19 @@ __kernel void flash_attn_f32_q8_0(
                for (int b_local = 0; b_local < SPLIT_DK_Q8_BLOCKS; ++b_local) {
                    const int b = k_blk_base + b_local;
                    int sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+#if defined(FA_K_LDS_T)
+                    // The 4 KV rows are adjacent in the transposed tile, so each (b, g)
+                    // step is ONE 128-bit local read instead of four 32-bit ones.
+                    #pragma unroll
+                    for (int g = 0; g < 8; ++g) {
+                        const uint qp = q_packed_pf[b_local * 8 + g];
+                        const uint4 kq4 = vload4(0, &l_k_packed[b * 8 + g][j]);
+                        sum0 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s0, sum0);
+                        sum1 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s1, sum1);
+                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s2, sum2);
+                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, kq4.s3, sum3);
+                    }
+#else
                    #pragma unroll
                    for (int g = 0; g < 8; ++g) {
                        const uint qp = q_packed_pf[b_local * 8 + g];
@@ -1564,11 +1600,20 @@ __kernel void flash_attn_f32_q8_0(
                        sum2 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+2][b * 8 + g], sum2);
                        sum3 = dot_acc_sat_4x8packed_ss_int(qp, l_k_packed[j+3][b * 8 + g], sum3);
                    }
+#endif
                    const float qd = q_d_pf[b_local];
+#ifdef FA_K_LDS_T
+                    const float4 ks4 = vload4(0, &l_k_scale[b][j]);
+                    s0 += (float)sum0 * qd * ks4.s0;
+                    s1 += (float)sum1 * qd * ks4.s1;
+                    s2 += (float)sum2 * qd * ks4.s2;
+                    s3 += (float)sum3 * qd * ks4.s3;
+#else
                    s0 += (float)sum0 * qd * l_k_scale[j  ][b];
                    s1 += (float)sum1 * qd * l_k_scale[j+1][b];
                    s2 += (float)sum2 * qd * l_k_scale[j+2][b];
                    s3 += (float)sum3 * qd * l_k_scale[j+3][b];
+#endif
                }
 #else
                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
@@ -763,6 +763,7 @@ static void ggml_backend_openvino_device_get_props(ggml_backend_dev_t dev, ggml_
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -1881,6 +1881,7 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
        /* .events                = */ false,
+        /* .mmap_support          = */ true,
    };
 }

@@ -61,6 +61,7 @@ void ggml_sycl_host_free(void* ptr);
 extern int g_ggml_sycl_debug;
 extern int g_ggml_sycl_enable_optimize;
 extern int g_ggml_sycl_enable_fusion;
+extern int g_ggml_sycl_enable_esimd;
 extern int g_ggml_sycl_prioritize_dmmv;
 extern int g_ggml_sycl_enable_flash_attention;
 extern int g_ggml_sycl_dev2dev_memcpy;
@@ -184,8 +184,8 @@ void concat_impl_sycl(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
            const size_t size0 = ggml_nbytes(src0);
            const size_t size1 = ggml_nbytes(src1);

-            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0).wait()));
-            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d + size0 / type_size, src1_d, size1).wait()));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d + size0 / type_size, src1_d, size1)));
        }
    } else {
        concat_T_sycl_non_cont<T>(stream, (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
@@ -196,6 +196,270 @@ void concat_impl_sycl(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
    }
 }

+static void concat_impl_q4_0_sycl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0   = dst->src[0];
+    const ggml_tensor * src1   = dst->src[1];
+    queue_ptr           stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(src1->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(dst->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(src0->ne[0] % QK4_0 == 0);
+    GGML_ASSERT(src1->ne[0] % QK4_0 == 0);
+    GGML_ASSERT(dst->ne[0] % QK4_0 == 0);
+
+    const int ne00_blk = src0->ne[0] / QK4_0;
+    const int ne0_blk  = dst->ne[0] / QK4_0;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const block_q4_0 * src0_d = (const block_q4_0 *) src0->data;
+        const block_q4_0 * src1_d = (const block_q4_0 *) src1->data;
+        block_q4_0 * dst_d        = (block_q4_0 *) dst->data;
+        const size_t type_size    = sizeof(block_q4_0);
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_T_sycl<block_q4_0>(
+                    src0_d + i3 * (src0->nb[3] / type_size),
+                    src1_d + i3 * (src1->nb[3] / type_size),
+                    dst_d + i3 * (dst->nb[3] / type_size),
+                    ne00_blk, src0->ne[1], src0->ne[2], ne0_blk,
+                    dst->ne[1], dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy((char *) dst_d + size0, src1_d, size1)));
+        }
+    } else {
+        concat_T_sycl_non_cont<block_q4_0>(
+            stream, (const char *) src0->data, (const char *) src1->data,
+            (char *) dst->data,
+            ne00_blk, src0->ne[1], src0->ne[2], src0->ne[3],
+            src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src1->ne[0] / QK4_0, src1->ne[1], src1->ne[2], src1->ne[3],
+            src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+            ne0_blk, dst->ne[1], dst->ne[2], dst->ne[3],
+            dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+    }
+}
+
+static void concat_impl_q4_1_sycl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0   = dst->src[0];
+    const ggml_tensor * src1   = dst->src[1];
+    queue_ptr           stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q4_1);
+    GGML_ASSERT(src1->type == GGML_TYPE_Q4_1);
+    GGML_ASSERT(dst->type == GGML_TYPE_Q4_1);
+    GGML_ASSERT(src0->ne[0] % QK4_1 == 0);
+    GGML_ASSERT(src1->ne[0] % QK4_1 == 0);
+    GGML_ASSERT(dst->ne[0] % QK4_1 == 0);
+
+    const int ne00_blk = src0->ne[0] / QK4_1;
+    const int ne0_blk  = dst->ne[0] / QK4_1;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const block_q4_1 * src0_d = (const block_q4_1 *) src0->data;
+        const block_q4_1 * src1_d = (const block_q4_1 *) src1->data;
+        block_q4_1 * dst_d        = (block_q4_1 *) dst->data;
+        const size_t type_size    = sizeof(block_q4_1);
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_T_sycl<block_q4_1>(
+                    src0_d + i3 * (src0->nb[3] / type_size),
+                    src1_d + i3 * (src1->nb[3] / type_size),
+                    dst_d + i3 * (dst->nb[3] / type_size),
+                    ne00_blk, src0->ne[1], src0->ne[2], ne0_blk,
+                    dst->ne[1], dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy((char *) dst_d + size0, src1_d, size1)));
+        }
+    } else {
+        concat_T_sycl_non_cont<block_q4_1>(
+            stream, (const char *) src0->data, (const char *) src1->data,
+            (char *) dst->data,
+            ne00_blk, src0->ne[1], src0->ne[2], src0->ne[3],
+            src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src1->ne[0] / QK4_1, src1->ne[1], src1->ne[2], src1->ne[3],
+            src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+            ne0_blk, dst->ne[1], dst->ne[2], dst->ne[3],
+            dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+    }
+}
+
+static void concat_impl_q5_0_sycl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0   = dst->src[0];
+    const ggml_tensor * src1   = dst->src[1];
+    queue_ptr           stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q5_0);
+    GGML_ASSERT(src1->type == GGML_TYPE_Q5_0);
+    GGML_ASSERT(dst->type == GGML_TYPE_Q5_0);
+    GGML_ASSERT(src0->ne[0] % QK5_0 == 0);
+    GGML_ASSERT(src1->ne[0] % QK5_0 == 0);
+    GGML_ASSERT(dst->ne[0] % QK5_0 == 0);
+
+    const int ne00_blk = src0->ne[0] / QK5_0;
+    const int ne0_blk  = dst->ne[0] / QK5_0;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const block_q5_0 * src0_d = (const block_q5_0 *) src0->data;
+        const block_q5_0 * src1_d = (const block_q5_0 *) src1->data;
+        block_q5_0 * dst_d        = (block_q5_0 *) dst->data;
+        const size_t type_size    = sizeof(block_q5_0);
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_T_sycl<block_q5_0>(
+                    src0_d + i3 * (src0->nb[3] / type_size),
+                    src1_d + i3 * (src1->nb[3] / type_size),
+                    dst_d + i3 * (dst->nb[3] / type_size),
+                    ne00_blk, src0->ne[1], src0->ne[2], ne0_blk,
+                    dst->ne[1], dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy((char *) dst_d + size0, src1_d, size1)));
+        }
+    } else {
+        concat_T_sycl_non_cont<block_q5_0>(
+            stream, (const char *) src0->data, (const char *) src1->data,
+            (char *) dst->data,
+            ne00_blk, src0->ne[1], src0->ne[2], src0->ne[3],
+            src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src1->ne[0] / QK5_0, src1->ne[1], src1->ne[2], src1->ne[3],
+            src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+            ne0_blk, dst->ne[1], dst->ne[2], dst->ne[3],
+            dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+    }
+}
+
+static void concat_impl_q5_1_sycl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0   = dst->src[0];
+    const ggml_tensor * src1   = dst->src[1];
+    queue_ptr           stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q5_1);
+    GGML_ASSERT(src1->type == GGML_TYPE_Q5_1);
+    GGML_ASSERT(dst->type == GGML_TYPE_Q5_1);
+    GGML_ASSERT(src0->ne[0] % QK5_1 == 0);
+    GGML_ASSERT(src1->ne[0] % QK5_1 == 0);
+    GGML_ASSERT(dst->ne[0] % QK5_1 == 0);
+
+    const int ne00_blk = src0->ne[0] / QK5_1;
+    const int ne0_blk  = dst->ne[0] / QK5_1;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const block_q5_1 * src0_d = (const block_q5_1 *) src0->data;
+        const block_q5_1 * src1_d = (const block_q5_1 *) src1->data;
+        block_q5_1 * dst_d        = (block_q5_1 *) dst->data;
+        const size_t type_size    = sizeof(block_q5_1);
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_T_sycl<block_q5_1>(
+                    src0_d + i3 * (src0->nb[3] / type_size),
+                    src1_d + i3 * (src1->nb[3] / type_size),
+                    dst_d + i3 * (dst->nb[3] / type_size),
+                    ne00_blk, src0->ne[1], src0->ne[2], ne0_blk,
+                    dst->ne[1], dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy((char *) dst_d + size0, src1_d, size1)));
+        }
+    } else {
+        concat_T_sycl_non_cont<block_q5_1>(
+            stream, (const char *) src0->data, (const char *) src1->data,
+            (char *) dst->data,
+            ne00_blk, src0->ne[1], src0->ne[2], src0->ne[3],
+            src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src1->ne[0] / QK5_1, src1->ne[1], src1->ne[2], src1->ne[3],
+            src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+            ne0_blk, dst->ne[1], dst->ne[2], dst->ne[3],
+            dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+    }
+}
+
+static void concat_impl_q8_0_sycl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0   = dst->src[0];
+    const ggml_tensor * src1   = dst->src[1];
+    queue_ptr           stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(src1->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(dst->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(src0->ne[0] % QK8_0 == 0);
+    GGML_ASSERT(src1->ne[0] % QK8_0 == 0);
+    GGML_ASSERT(dst->ne[0] % QK8_0 == 0);
+
+    const int ne00_blk = src0->ne[0] / QK8_0;
+    const int ne0_blk  = dst->ne[0] / QK8_0;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const block_q8_0 * src0_d = (const block_q8_0 *) src0->data;
+        const block_q8_0 * src1_d = (const block_q8_0 *) src1->data;
+        block_q8_0 * dst_d        = (block_q8_0 *) dst->data;
+        const size_t type_size    = sizeof(block_q8_0);
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_T_sycl<block_q8_0>(
+                    src0_d + i3 * (src0->nb[3] / type_size),
+                    src1_d + i3 * (src1->nb[3] / type_size),
+                    dst_d + i3 * (dst->nb[3] / type_size),
+                    ne00_blk, src0->ne[1], src0->ne[2], ne0_blk,
+                    dst->ne[1], dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0)));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy((char *) dst_d + size0, src1_d, size1)));
+        }
+    } else {
+        concat_T_sycl_non_cont<block_q8_0>(
+            stream, (const char *) src0->data, (const char *) src1->data,
+            (char *) dst->data,
+            ne00_blk, src0->ne[1], src0->ne[2], src0->ne[3],
+            src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+            src1->ne[0] / QK8_0, src1->ne[1], src1->ne[2], src1->ne[3],
+            src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+            ne0_blk, dst->ne[1], dst->ne[2], dst->ne[3],
+            dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+    }
+}
+
 void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {

    switch (dst->type) {
@@ -222,6 +486,21 @@ void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
    case GGML_TYPE_I8:
        concat_impl_sycl<int8_t>(ctx, dst);
        break;
+    case GGML_TYPE_Q4_0:
+        concat_impl_q4_0_sycl(ctx, dst);
+        break;
+    case GGML_TYPE_Q4_1:
+        concat_impl_q4_1_sycl(ctx, dst);
+        break;
+    case GGML_TYPE_Q5_0:
+        concat_impl_q5_0_sycl(ctx, dst);
+        break;
+    case GGML_TYPE_Q5_1:
+        concat_impl_q5_1_sycl(ctx, dst);
+        break;
+    case GGML_TYPE_Q8_0:
+        concat_impl_q8_0_sycl(ctx, dst);
+        break;
    default:
        fprintf(stderr, "%s: unsupported types: dst: %s\n", __func__, ggml_type_name(dst->type));
        GGML_ASSERT(false);
@@ -8,6 +8,9 @@
        #include <sycl/ext/oneapi/bfloat16.hpp>
        #define GGML_SYCL_DMMV_HAS_BF16
    #endif
+    #include <sycl/ext/intel/esimd.hpp>
+    #include "esimd.hpp"
+    #define GGML_SYCL_DMMV_HAS_ESIMD
 #endif

 static void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
@@ -1864,6 +1867,113 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
        });
 }

+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+using ggml_sycl_esimd::GGML_SYCL_DMMV_ESIMD_WG_SIZE;
+
+// generic reordered dequantize-matvec: each work-group owns a pair of
+// consecutive output rows and updates one 32-wide accumulator per row
+template <ggml_type T>
+ESIMD_INLINE void dequantize_mul_mat_vec_reorder_esimd(
+        const void * vx, const float * y, float * dst,
+        const int ncols, const int nrows,
+        sycl::local_accessor<float, 1> lmem,
+        const sycl::nd_item<1> & it) {
+    using namespace sycl::ext::intel::esimd;
+    using traits = ggml_sycl_esimd::esimd_reorder_q_traits<T>;
+
+    const int    num_blocks_per_row = ncols / QK_K;
+    const size_t nb = (size_t) nrows * num_blocks_per_row;
+    const auto   ps = traits::make_ptrs(vx, nb);
+
+    const int  tid      = it.get_local_id(0);
+    const int  row_pair = it.get_group(0);
+    const int  row0     = row_pair * 2; // two consecutive output rows
+    const bool has_row1 = row0 + 1 < nrows;
+
+    // one 32-wide accumulator per output row (small footprint, no spill)
+    simd<float, 32> acc0 = 0.0f;
+    simd<float, 32> acc1 = 0.0f;
+
+    for (int ib = tid; ib < num_blocks_per_row; ib += GGML_SYCL_DMMV_ESIMD_WG_SIZE) {
+        simd<float, 256> y_vec = block_load<float, 256>(y + (size_t) ib * QK_K);
+
+        const size_t bi0 = (size_t) (row0 + 0) * num_blocks_per_row + ib;
+        const size_t bi1 = (size_t) (row0 + 1) * num_blocks_per_row + ib;
+
+        traits::mac_pair(ps, bi0, ps, bi1, has_row1, y_vec, acc0, acc1);
+    }
+
+    lmem[tid * 2 + 0] = reduce<float>(acc0, std::plus<>{});
+    lmem[tid * 2 + 1] = reduce<float>(acc1, std::plus<>{});
+    it.barrier(sycl::access::fence_space::local_space);
+
+    if (tid == 0) {
+        float sum0 = 0.0f;
+        float sum1 = 0.0f;
+        for (int p = 0; p < GGML_SYCL_DMMV_ESIMD_WG_SIZE; ++p) {
+            sum0 += lmem[p * 2 + 0];
+            sum1 += lmem[p * 2 + 1];
+        }
+        dst[row0 + 0] = sum0;
+        if (has_row1) {
+            dst[row0 + 1] = sum1;
+        }
+    }
+}
+
+static void dequantize_mul_mat_vec_q3_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q3_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
+static void dequantize_mul_mat_vec_q4_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q4_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
+static void dequantize_mul_mat_vec_q6_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q6_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
+#endif // GGML_SYCL_DMMV_HAS_ESIMD
+
 static void dequantize_mul_mat_vec_q4_K_sycl_reorder(const void *vx, const float *y,
                                                     float *dst, const int ncols,
                                                     const int nrows,
@@ -1992,7 +2102,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
        case GGML_TYPE_Q3_K:
            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q3_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q3_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q3_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
            } else {
                dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
            }
@@ -2000,7 +2118,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
        case GGML_TYPE_Q4_K:
            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q4_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q4_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q4_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
            } else {
                dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
            }
@@ -2016,7 +2142,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
        case GGML_TYPE_Q6_K:
            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q6_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q6_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q6_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
            } else {
                dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
            }
@@ -448,6 +448,47 @@ static void unary_gated_op_generic_kernel(
    }
 }

+// Fused UNARY + MUL. Unlike the gated ops above, `x` and `g` are separate tensors of the
+// same shape; `o0`/`o1` are their row strides in elements, so a half-view needs no repack.
+// `dst` is contiguous and indexed flat. Math is done in f32, as the CPU and CUDA references do.
+template<typename T, typename F>
+static void unary_mul_flat_kernel(const T * x, const T * g, T * dst, const int64_t k, const sycl::nd_item<1> &item_ct1, F op) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = (T) (op((float) x[i]) * (float) g[i]);
+    }
+}
+
+template<typename T, typename F>
+static void unary_mul_strided_kernel(const T * x, const T * g, T * dst, const int64_t k, const sycl::uint3 n_fd, const int64_t o0, const int64_t o1, const sycl::nd_item<1> &item_ct1, F op) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
+        const int64_t j0 = rc.x() * o0 + rc.y();
+        const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
+        dst[i] = (T) (op((float) x[j0]) * (float) g[j1]);
+    }
+}
+
+template<typename T, typename F>
+static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, queue_ptr main_stream, F op) {
+    const size_t            num_blocks = ceil_div((size_t) k, (size_t) SYCL_GLU_BLOCK_SIZE);
+    const sycl::nd_range<1> range(num_blocks * sycl::range<1>(SYCL_GLU_BLOCK_SIZE), sycl::range<1>(SYCL_GLU_BLOCK_SIZE));
+
+    // o0 == o1 == n makes (i/n)*o0 + (i%n) == i, so the strided kernel degenerates to the flat one
+    if (o0 == n && o1 == n) {
+        main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+            unary_mul_flat_kernel(x, g, dst, k, item_ct1, op);
+        });
+        return;
+    }
+
+    // 32-bit fastdiv, exact only below 2^31; ggml_sycl_can_fuse() already declined past that
+    GGML_ASSERT(k < ((int64_t) 1 << 31));
+    const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
+    main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+        unary_mul_strided_kernel(x, g, dst, k, n_fd, o0, o1, item_ct1, op);
+    });
+}
+
 namespace ggml_sycl_detail {
 static void acc_f32_sycl(const char *x, const char *y, float *dst,
                         const int64_t n_elements,
@@ -991,6 +1032,52 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
    });
 }

+// dst = op(unary_node->src[0]) * other, written straight to the MUL output, saving the
+// standalone unary launch. Preconditions come from ggml_sycl_can_fuse(); re-asserted here.
+void ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
+    scope_op_debug_print scope_dbg_print(__func__, mul_node, /*num_src=*/2);
+
+    const ggml_tensor * x = unary_node->src[0];
+    const ggml_tensor * g = (mul_node->src[0] == unary_node) ? mul_node->src[1] : mul_node->src[0];
+
+    // g is picked by elimination; ggml_can_fuse()'s single-use rule rules out MUL(unary, unary)
+    GGML_ASSERT(g != unary_node);
+    GGML_ASSERT(x->type == g->type && x->type == mul_node->type);
+    GGML_ASSERT(ggml_are_same_shape(x, g) && ggml_are_same_shape(x, mul_node));
+    GGML_ASSERT(ggml_is_contiguous_1(x) && ggml_is_contiguous_1(g));
+    // dst is indexed flat
+    GGML_ASSERT(ggml_is_contiguous(mul_node));
+
+    queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int64_t k = ggml_nelements(mul_node);
+    const int64_t n = mul_node->ne[0];
+
+    const auto dispatch_type = [&](auto op) {
+        switch (mul_node->type) {
+            case GGML_TYPE_F32:
+                unary_mul_sycl((const float *) x->data, (const float *) g->data, (float *) mul_node->data,
+                               k, n, x->nb[1] / sizeof(float), g->nb[1] / sizeof(float), main_stream, op);
+                break;
+            case GGML_TYPE_F16:
+                unary_mul_sycl((const sycl::half *) x->data, (const sycl::half *) g->data, (sycl::half *) mul_node->data,
+                               k, n, x->nb[1] / sizeof(sycl::half), g->nb[1] / sizeof(sycl::half), main_stream, op);
+                break;
+            default:
+                GGML_ABORT("fused unary+mul: unsupported type %s", ggml_type_name(mul_node->type));
+        }
+    };
+
+    switch (ggml_get_unary_op(unary_node)) {
+        case GGML_UNARY_OP_SILU:     dispatch_type([](float v) { return op_silu(v); });     break;
+        case GGML_UNARY_OP_SIGMOID:  dispatch_type([](float v) { return op_sigmoid(v); });  break;
+        case GGML_UNARY_OP_SOFTPLUS: dispatch_type([](float v) { return op_softplus(v); }); break;
+        default:
+            GGML_ABORT("fused unary+mul: unsupported unary op %s", ggml_unary_op_name(ggml_get_unary_op(unary_node)));
+    }
+}
+
 __dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
    x = sycl::fmin(x, limit);
    g = sycl::fmax(sycl::fmin(g, limit), -limit);
@@ -95,4 +95,7 @@ void ggml_sycl_trunc(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

 void ggml_sycl_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

+// fused UNARY(silu|sigmoid|softplus) + MUL; see ggml_sycl_can_fuse() for the accepted shapes
+void ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);
+
 #endif // GGML_SYCL_ELEMENTWISE_HPP
@@ -0,0 +1,392 @@
+//
+// MIT license
+// Copyright (C) 2026 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_ESIMD_HPP
+#define GGML_SYCL_ESIMD_HPP
+
+#include <sycl/ext/intel/esimd.hpp>
+
+#include "common.hpp"
+
+namespace ggml_sycl_esimd {
+
+constexpr int GGML_SYCL_DMMV_ESIMD_WG_SIZE = 4;
+
+//
+// Shared ESIMD building blocks for the reordered K-quant dequantize-matvec
+// kernels.
+//
+// The reordered K-quant ESIMD matvec kernels share one skeleton: per super-block,
+// load a 256-float activation slice, load one weight block, dequantize it into 8
+// chunks of 32 and MAC each chunk against the matching activation slice, then
+// reduce and run a lane-0 epilogue.
+//
+// Each K-quant kernel emits exactly 8 chunks of 32 mapping to activation slices
+// 0..7, so the per-block work is captured by esimd_reorder_q_traits<T>::mac_pair,
+// which dequantizes two weight blocks and MACs both against a shared activation
+// vector with the two FMA chains interleaved (co-scheduled to hide FMA latency).
+// The "pair" is the (row0,row1) row pair owned by one work-group, so the
+// layout+dequant is written once per quant type here.
+//
+
+template <ggml_type T> struct esimd_reorder_q_traits;
+
+// build a 32-lane vector whose low 16 lanes are `lo` and high 16 are `hi`
+// (a super-chunk splits into two 16-wide halves with distinct scale/min codes).
+static ESIMD_INLINE sycl::ext::intel::esimd::simd<float, 32> splat_lo_hi(float lo, float hi) {
+    using namespace sycl::ext::intel::esimd;
+    simd<float, 32> v;
+    v.select<16, 1>(0)  = lo;
+    v.select<16, 1>(16) = hi;
+    return v;
+}
+
+// unpack one block of Q4_K/Q5_K scale/min codes (get_scale_min_k4 layout) into 8
+// float scales (dall * sc) and 8 float mins (-dmin * m); the min carries the
+// negation so the dequant epilogue adds.
+static ESIMD_INLINE void unpack_scale_min_k4(
+        sycl::ext::intel::esimd::simd<uint8_t, 12> scales, float dall, float dmin,
+        sycl::ext::intel::esimd::simd<float, 8> & scale_f,
+        sycl::ext::intel::esimd::simd<float, 8> & min_f) {
+    using namespace sycl::ext::intel::esimd;
+    simd<uint8_t, 8> sc = 0;
+    simd<uint8_t, 8> m  = 0;
+    simd<uint8_t, 4> scale_lo = scales.select<4, 1>(0);
+    simd<uint8_t, 4> min_lo   = scales.select<4, 1>(4);
+    simd<uint8_t, 4> hi_bits  = scales.select<4, 1>(8);
+    sc.select<4, 1>(0) = scale_lo & simd<uint8_t, 4>(0x3F);
+    sc.select<4, 1>(4) = (hi_bits & simd<uint8_t, 4>(0x0F)) |
+                         ((scale_lo >> simd<uint8_t, 4>(6)) << simd<uint8_t, 4>(4));
+    m.select<4, 1>(0)  = min_lo & simd<uint8_t, 4>(0x3F);
+    m.select<4, 1>(4)  = (hi_bits >> simd<uint8_t, 4>(4)) |
+                         ((min_lo >> simd<uint8_t, 4>(6)) << simd<uint8_t, 4>(4));
+    scale_f = convert<float>(sc) * dall;
+    min_f   = convert<float>(m) * (-dmin);
+}
+
+// ---------------------------------------------------------------------------
+// Q3_K, SOA reorder layout produced by reorder_qw_q3_k:
+//   [qs: nb*(QK_K/4)] [hmask: nb*(QK_K/8)] [scales: nb*12] [d: nb*sizeof(half)]
+// with nb = nrows*num_blocks_per_row. Single super-block scale d, no dmin.
+//
+// 3 bits per weight: 2 low bits in qs, 1 high bit in hmask. The 8 output chunks
+// of 32 (matching dequantize_row_q3_K) map to super-chunk s (0..7): byte base
+// 32*(s/4) into the 64-byte qs array, bit shift 2*(s%4); the low 16 lanes use
+// scale code 2s, the high 16 use 2s+1. hmask is a 32-byte array (like Q5_K's
+// qh) where chunk s uses bit s of the same 32 bytes, but INVERTED: the value is
+// (q & 3) - (hmask_bit_set ? 0 : 4), i.e. (q & 3) + 4*bit - 4.
+//
+// The 16 6-bit scale codes are packed into 12 bytes (get_scale_min layout for
+// Q3_K): low nibbles from bytes 0..7, high 2 bits from bytes 8..11 shifted by
+// 0/2/4/6; the dequant scale is d * (code - 32).
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q3_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    hmask;
+        const uint8_t *    scales;
+        const sycl::half * d;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * hmask  = qs + nb * (QK_K / 4);
+        const uint8_t * scales = hmask + nb * (QK_K / 8);
+        const sycl::half * d   = (const sycl::half *) (scales + nb * 12);
+        return { qs, hmask, scales, d };
+    }
+
+    // unpack the 12 packed bytes into 16 6-bit scale codes (dequantize_row_q3_K
+    // aux layout), returned as float scale = d * (code - 32).
+    // done with wide (8/16-lane) ops rather than four 4-lane groups.
+    static ESIMD_INLINE sycl::ext::intel::esimd::simd<float, 16> unpack_scales(
+            sycl::ext::intel::esimd::simd<uint8_t, 12> in, float d) {
+        using namespace sycl::ext::intel::esimd;
+
+        // low 6-bit part: codes 0..7 = low nibble of bytes 0..7,
+        //                 codes 8..15 = high nibble of bytes 0..7
+        simd<uint8_t, 8>  lo8 = in.select<8, 1>(0);
+        simd<uint8_t, 16> code;
+        code.select<8, 1>(0) = lo8 & simd<uint8_t, 8>(0x0F);
+        code.select<8, 1>(8) = lo8 >> simd<uint8_t, 8>(4);
+
+        // high 2-bit part: bytes 8..11 replicated 4x, group g (0..3) shifted 2*g
+        simd<uint8_t, 16> hib;
+        hib.select<4, 1>(0)  = in.select<4, 1>(8);
+        hib.select<4, 1>(4)  = in.select<4, 1>(8);
+        hib.select<4, 1>(8)  = in.select<4, 1>(8);
+        hib.select<4, 1>(12) = in.select<4, 1>(8);
+        simd<uint8_t, 16> hshift;
+        hshift.select<4, 1>(0)  = 0;
+        hshift.select<4, 1>(4)  = 2;
+        hshift.select<4, 1>(8)  = 4;
+        hshift.select<4, 1>(12) = 6;
+        hib = (hib >> hshift) & simd<uint8_t, 16>(0x03);
+
+        code = code | (hib << simd<uint8_t, 16>(4));
+        return (convert<float>(code) - 32.0f) * d;
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 64> qs_a     = block_load<uint8_t, 64>(pa.qs + bia * (QK_K / 4));
+        simd<uint8_t, 64> qs_b     = 0;
+        simd<uint8_t, 32> hmask_a  = block_load<uint8_t, 32>(pa.hmask + bia * (QK_K / 8));
+        simd<uint8_t, 32> hmask_b  = 0;
+        simd<uint8_t, 12> scales_a = block_load<uint8_t, 12>(pa.scales + bia * 12);
+        simd<uint8_t, 12> scales_b = 0;
+
+        const float d_a = (float) pa.d[bia];
+        float d_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 64>(pb.qs + bib * (QK_K / 4));
+            hmask_b  = block_load<uint8_t, 32>(pb.hmask + bib * (QK_K / 8));
+            scales_b = block_load<uint8_t, 12>(pb.scales + bib * 12);
+            d_b = (float) pb.d[bib];
+        }
+
+        simd<float, 16> scale_f_a = unpack_scales(scales_a, d_a);
+        simd<float, 16> scale_f_b = unpack_scales(scales_b, d_b);
+
+#pragma unroll
+        for (int s = 0; s < 8; ++s) {
+            const int     byte_base = 32 * (s / 4);
+            const uint8_t shift     = (uint8_t) (2 * (s % 4));
+            simd<float, 32> y_s = y_vec.select<32, 1>(s * 32);
+
+            // 2 low bits from qs, high bit from hmask (bit s of the same 32 bytes);
+            // value = (q & 3) + 4*bit - 4  (inverted hmask: subtract 4 when bit clear).
+            // merge in the integer domain: q3 = (q & 3) | (bit << 2) in {0..7},
+            // then a single convert + subtract yields q3 - 4 (one convert, not two)
+            simd<uint16_t, 32> q3_a = convert<uint16_t>(
+                    (qs_a.select<32, 1>(byte_base) >> shift) & simd<uint8_t, 32>(3));
+            q3_a |= convert<uint16_t>(
+                    ((hmask_a >> simd<uint8_t, 32>((uint8_t) s)) & simd<uint8_t, 32>(1)) << simd<uint8_t, 32>(2));
+            simd<uint16_t, 32> q3_b = convert<uint16_t>(
+                    (qs_b.select<32, 1>(byte_base) >> shift) & simd<uint8_t, 32>(3));
+            q3_b |= convert<uint16_t>(
+                    ((hmask_b >> simd<uint8_t, 32>((uint8_t) s)) & simd<uint8_t, 32>(1)) << simd<uint8_t, 32>(2));
+
+            simd<float, 32> qf_a = convert<float>(q3_a) - 4.0f;
+            simd<float, 32> qf_b = convert<float>(q3_b) - 4.0f;
+
+            const float scale_a_lo = scale_f_a[2 * s + 0];
+            const float scale_a_hi = scale_f_a[2 * s + 1];
+            const float scale_b_lo = scale_f_b[2 * s + 0];
+            const float scale_b_hi = scale_f_b[2 * s + 1];
+
+            simd<float, 32> scale_vec_a = splat_lo_hi(scale_a_lo, scale_a_hi);
+            simd<float, 32> scale_vec_b = splat_lo_hi(scale_b_lo, scale_b_hi);
+
+            simd<float, 32> deq_a = qf_a * scale_vec_a;
+            simd<float, 32> deq_b = qf_b * scale_vec_b;
+
+            acc_a += y_s * deq_a;
+            acc_b += y_s * deq_b;
+        }
+    }
+};
+
+// ---------------------------------------------------------------------------
+// Q4_K, SOA reorder layout produced by reorder_qw_q4_k:
+//   [qs: nb*(QK_K/2)] [scales: nb*K_SCALE_SIZE] [dm: nb*sizeof(half2)]
+// with nb = nrows*num_blocks_per_row.
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q4_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    scales;
+        const sycl::half * dm;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * scales = qs + nb * (QK_K / 2);
+        const sycl::half * dm  = (const sycl::half *) (scales + nb * K_SCALE_SIZE);
+        return { qs, scales, dm };
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 128> qs_a     = block_load<uint8_t, 128>(pa.qs + bia * (QK_K / 2));
+        simd<uint8_t, 128> qs_b     = 0;
+        simd<uint8_t, 12>  scales_a = block_load<uint8_t, 12>(pa.scales + bia * K_SCALE_SIZE);
+        simd<uint8_t, 12>  scales_b = 0;
+
+        const float dall_a = (float) pa.dm[bia * 2 + 0];
+        const float dmin_a = (float) pa.dm[bia * 2 + 1];
+        float dall_b = 0.0f;
+        float dmin_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 128>(pb.qs + bib * (QK_K / 2));
+            scales_b = block_load<uint8_t, 12>(pb.scales + bib * K_SCALE_SIZE);
+            dall_b = (float) pb.dm[bib * 2 + 0];
+            dmin_b = (float) pb.dm[bib * 2 + 1];
+        }
+
+        simd<float, 8> scale_f_a, min_f_a, scale_f_b, min_f_b;
+        unpack_scale_min_k4(scales_a, dall_a, dmin_a, scale_f_a, min_f_a);
+        unpack_scale_min_k4(scales_b, dall_b, dmin_b, scale_f_b, min_f_b);
+
+        simd<uint8_t, 128> qs_lo_a = qs_a & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_a = qs_a >> simd<uint8_t, 128>(4);
+        simd<uint8_t, 128> qs_lo_b = qs_b & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_b = qs_b >> simd<uint8_t, 128>(4);
+
+#pragma unroll
+        for (int sb = 0; sb < 8; sb += 2) {
+            const int q_offset = sb * 16;
+            simd<float, 32> y_lo = y_vec.select<32, 1>(sb * 32);
+            simd<float, 32> y_hi = y_vec.select<32, 1>((sb + 1) * 32);
+
+            const float scale_a_lo = scale_f_a[sb];
+            const float scale_a_hi = scale_f_a[sb + 1];
+            const float min_a_lo   = min_f_a[sb];
+            const float min_a_hi   = min_f_a[sb + 1];
+            const float scale_b_lo = scale_f_b[sb];
+            const float scale_b_hi = scale_f_b[sb + 1];
+            const float min_b_lo   = min_f_b[sb];
+            const float min_b_hi   = min_f_b[sb + 1];
+
+            simd<uint8_t, 32> qa_lo = qs_lo_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qa_hi = qs_hi_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_lo = qs_lo_b.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_hi = qs_hi_b.select<32, 1>(q_offset);
+
+            simd<float, 32> deq_a_lo = convert<float>(qa_lo) * scale_a_lo + min_a_lo;
+            simd<float, 32> deq_a_hi = convert<float>(qa_hi) * scale_a_hi + min_a_hi;
+            simd<float, 32> deq_b_lo = convert<float>(qb_lo) * scale_b_lo + min_b_lo;
+            simd<float, 32> deq_b_hi = convert<float>(qb_hi) * scale_b_hi + min_b_hi;
+
+            acc_a += y_lo * deq_a_lo;
+            acc_b += y_lo * deq_b_lo;
+            acc_a += y_hi * deq_a_hi;
+            acc_b += y_hi * deq_b_hi;
+        }
+    }
+};
+
+// ---------------------------------------------------------------------------
+// Q6_K, SOA reorder layout:
+//   [ql: nb*(QK_K/2)] [qh: nb*(QK_K/4)] [scales(int8): nb*(QK_K/16)] [d: nb*half]
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q6_K> {
+    struct ptrs {
+        const uint8_t *    ql;
+        const uint8_t *    qh;
+        const int8_t *     scales;
+        const sycl::half * d;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t *    ql     = (const uint8_t *) vx;
+        const uint8_t *    qh     = ql + nb * (QK_K / 2);
+        const int8_t *     scales = (const int8_t *) (qh + nb * (QK_K / 4));
+        const sycl::half * d      = (const sycl::half *) (scales + nb * (QK_K / 16));
+        return { ql, qh, scales, d };
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 128> ql_a     = block_load<uint8_t, 128>(pa.ql + bia * (QK_K / 2));
+        simd<uint8_t, 128> ql_b     = 0;
+        simd<uint8_t, 64>  qh_a     = block_load<uint8_t, 64>(pa.qh + bia * (QK_K / 4));
+        simd<uint8_t, 64>  qh_b     = 0;
+        simd<int8_t, 16>   scales_a = block_load<int8_t, 16>(pa.scales + bia * (QK_K / 16));
+        simd<int8_t, 16>   scales_b = 0;
+
+        const float d_a = (float) pa.d[bia];
+        float d_b = 0.0f;
+        if (has_b) {
+            ql_b     = block_load<uint8_t, 128>(pb.ql + bib * (QK_K / 2));
+            qh_b     = block_load<uint8_t, 64>(pb.qh + bib * (QK_K / 4));
+            scales_b = block_load<int8_t, 16>(pb.scales + bib * (QK_K / 16));
+            d_b = (float) pb.d[bib];
+        }
+
+        simd<float, 16> sc_a = convert<float>(scales_a);
+        simd<float, 16> sc_b = convert<float>(scales_b);
+
+#pragma unroll
+        for (int im = 0; im < 2; ++im) {
+            simd<uint8_t, 32> ql_lo_a   = ql_a.select<32, 1>(64 * im);
+            simd<uint8_t, 32> ql_hi_a   = ql_a.select<32, 1>(64 * im + 32);
+            simd<uint8_t, 32> qh_bits_a = qh_a.select<32, 1>(32 * im);
+            simd<uint8_t, 32> ql_lo_b   = ql_b.select<32, 1>(64 * im);
+            simd<uint8_t, 32> ql_hi_b   = ql_b.select<32, 1>(64 * im + 32);
+            simd<uint8_t, 32> qh_bits_b = qh_b.select<32, 1>(32 * im);
+
+            // reconstruct each 32-wide 6-bit group (matches dequantize_row_q6_K)
+#pragma unroll
+            for (int g = 0; g < 4; ++g) {
+                simd<float, 32> y_g = y_vec.select<32, 1>(32 * (4 * im + g));
+
+                const float scale_a_lo = sc_a[8 * im + 2 * g + 0] * d_a;
+                const float scale_a_hi = sc_a[8 * im + 2 * g + 1] * d_a;
+                const float scale_b_lo = sc_b[8 * im + 2 * g + 0] * d_b;
+                const float scale_b_hi = sc_b[8 * im + 2 * g + 1] * d_b;
+
+                simd<float, 32> scale_vec_a = splat_lo_hi(scale_a_lo, scale_a_hi);
+                simd<float, 32> scale_vec_b = splat_lo_hi(scale_b_lo, scale_b_hi);
+
+                simd<uint8_t, 32> qa;
+                simd<uint8_t, 32> qb;
+                switch (g) {
+                    case 0:
+                        qa = (ql_lo_a & simd<uint8_t, 32>(0x0F)) | ((qh_bits_a & simd<uint8_t, 32>(0x03)) << simd<uint8_t, 32>(4));
+                        qb = (ql_lo_b & simd<uint8_t, 32>(0x0F)) | ((qh_bits_b & simd<uint8_t, 32>(0x03)) << simd<uint8_t, 32>(4));
+                        break;
+                    case 1:
+                        qa = (ql_hi_a & simd<uint8_t, 32>(0x0F)) | ((qh_bits_a & simd<uint8_t, 32>(0x0C)) << simd<uint8_t, 32>(2));
+                        qb = (ql_hi_b & simd<uint8_t, 32>(0x0F)) | ((qh_bits_b & simd<uint8_t, 32>(0x0C)) << simd<uint8_t, 32>(2));
+                        break;
+                    case 2:
+                        qa = (ql_lo_a >> simd<uint8_t, 32>(4)) | (qh_bits_a & simd<uint8_t, 32>(0x30));
+                        qb = (ql_lo_b >> simd<uint8_t, 32>(4)) | (qh_bits_b & simd<uint8_t, 32>(0x30));
+                        break;
+                    default:
+                        qa = (ql_hi_a >> simd<uint8_t, 32>(4)) | ((qh_bits_a & simd<uint8_t, 32>(0xC0)) >> simd<uint8_t, 32>(2));
+                        qb = (ql_hi_b >> simd<uint8_t, 32>(4)) | ((qh_bits_b & simd<uint8_t, 32>(0xC0)) >> simd<uint8_t, 32>(2));
+                        break;
+                }
+
+                simd<float, 32> deq_a = (convert<float>(qa) - 32.0f) * scale_vec_a;
+                simd<float, 32> deq_b = (convert<float>(qb) - 32.0f) * scale_vec_b;
+
+                acc_a += y_g * deq_a;
+                acc_b += y_g * deq_b;
+            }
+        }
+    }
+};
+
+} // namespace ggml_sycl_esimd
+
+#endif // GGML_SYCL_ESIMD_HPP
--- a/Show More
+++ b/Show More