ggml: support ggml_rope_set_offset on opencl, sycl, wgpu, hexagon (#27345 )

* ggml: support ggml_rope_set_offset on opencl, sycl, wgpu, hexagon * rm inplace optimization
ci: use shell script to check cmake pkg (#27414 )
2026-08-21 04:02:38 +02:00 · 2026-08-21 00:36:57 +02:00 · 2026-08-20 20:01:32 +00:00 · 2026-08-20 21:31:29 +03:00 · 2026-08-20 10:58:35 -07:00 · 2026-08-20 10:52:07 -07:00
504 changed files with 20698 additions and 35762 deletions
@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.36.3
-ARG IGC_VERSION_FULL=2_2.36.3+21719
-ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.33.0
-ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
-ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
    fi; \
    DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
    if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -1,20 +0,0 @@
-name: "Linux - Setup Vulkan SDK"
-description: "Setup Vulkan SDK for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "Vulkan SDK version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup Vulkan SDK
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://sdk.lunarg.com/sdk/download/${{ inputs.version }}/linux/vulkan_sdk.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1
@@ -6,8 +6,7 @@ inputs:
    required: true
  cuda_arch:
    description: "CUDA target architecture"
-    required: false
-    default: "x64"
+    required: true

 runs:
  using: "composite"
@@ -10,33 +10,6 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  ubuntu-24-vulkan-cache:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
-
  #ubuntu-24-spacemit-cache:
  #  runs-on: ubuntu-24.04

@@ -67,9 +40,9 @@ jobs:
    runs-on: ubuntu-24.04

    env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -96,8 +69,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -27,30 +27,26 @@ jobs:
          cmake --install build --prefix "$PREFIX" --config Release

          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)

-          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"
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"

-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")

-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
                  exit 1
-              }
-          }
-          puts "success."
-          EOF
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."

          cd examples/simple-cmake-pkg
          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
@@ -21,6 +21,7 @@ on:
    paths: [
      '.github/workflows/build-cpu.yml',
      '.github/workflows/build-cmake-pkg.yml',
+      'ggml/src/ggml-rpc/**',
      '**/CMakeLists.txt',
      '**/.cmake',
      '**/*.h',
@@ -118,28 +119,25 @@ jobs:
          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

  windows:
+    name: windows / ${{ matrix.build }}
    runs-on: windows-2025

    env:
      OPENBLAS_VERSION: 0.3.23
      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.357.0

    strategy:
      matrix:
        include:
          - build: 'x64-cpu-static'
            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
          - build: 'x64-openblas'
            arch: 'x64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'x64-vulkan'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
          - build: 'arm64'
            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'

    steps:
      - name: Clone
@@ -167,15 +165,6 @@ jobs:
          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll

-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'x64-vulkan' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
-          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
-          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
-          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
-
      - name: Install Ninja
        id: install_ninja
        run: |
@@ -22,6 +22,7 @@ env:

 jobs:
  cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    runs-on: windows-2022

    permissions:
@@ -29,7 +30,16 @@ 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
@@ -39,12 +49,13 @@ 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: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -54,26 +65,21 @@ jobs:
      - 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" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DGGML_RPC=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% -t ggml
-          cmake --build build --config Release
+          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 }}

  hip:
    runs-on: windows-2022
@@ -39,8 +39,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -81,7 +81,7 @@ jobs:
        # TODO: fix and re-enable the `test-llama-archs` test below
        run: |
          cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000

      - name: Test (GPU)
        id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
        run: |
          cd ${{ github.workspace }}
          export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000

  openvino-windows-2022:
    runs-on: windows-2022

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -166,4 +166,4 @@ jobs:
          call "%OPENVINO_ROOT%\setupvars.bat"

          cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
@@ -1,66 +0,0 @@
-name: CI (rpc)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      'ggml/src/ggml-rpc/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  ubuntu-24-rpc:
-    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libssl-dev ninja-build
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose
@@ -288,8 +288,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -93,19 +93,13 @@ jobs:
        run: |
          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"

-      - name: Use Vulkan SDK Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
+        id: setup
+        uses: ./.github/actions/unarchive-tar
        with:
+          url: https://sdk.lunarg.com/sdk/download/${{ env.VULKAN_SDK_VERSION }}/linux/vulkan_sdk.tar.xz
          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
+          strip: 1

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -133,3 +127,56 @@ jobs:
          # This is using llvmpipe and runs slower than other backends
          # test-backend-ops is too slow on llvmpipe, skip it
          ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.357.0
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-windows-2025-x64-vulkan
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_NATIVE=OFF `
+            -DLLAMA_BUILD_SERVER=ON `
+            -DGGML_RPC=ON `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_VULKAN=ON `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
@@ -44,6 +44,7 @@ jobs:
        uses: actions/checkout@v6
        with:
          fetch-depth: 0
+          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}

      - name: Determine source tag name
        id: srctag
@@ -393,6 +394,11 @@ jobs:
    name: Create shared tags from digests
    needs: [prepare_matrices, push_to_registry, create_tag]
    runs-on: ubuntu-24.04
+    permissions:
+      contents: read
+      packages: write
+      id-token: write
+      attestations: write
    strategy:
      fail-fast: false
      matrix:
@@ -427,6 +433,7 @@ jobs:
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Create tags from digests
+        id: create_tags
        shell: bash
        run: |
          set -euo pipefail
@@ -438,6 +445,7 @@ jobs:
          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
          BUILD_DATE="${{ steps.build_date.outputs.date }}"
          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
+          echo "image_repo=${IMAGE_REPO}" >> "$GITHUB_OUTPUT"
          TAGS="${{ matrix.config.tag }}"
          ARCHES="${{ matrix.config.arches }}"
          DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -504,6 +512,16 @@ jobs:

              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
+
+              if [[ "$tag_name" == "${TAGS%% *}" ]]; then
+                  local digest
+                  digest="$(docker buildx imagetools inspect "${merged_versioned_tag}" --format '{{.Manifest.Digest}}')"
+                  if [[ ! "$digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
+                      echo "Invalid digest for ${merged_versioned_tag}: ${digest}" >&2
+                      exit 1
+                  fi
+                  echo "${image_type}_digest=${digest}" >> "$GITHUB_OUTPUT"
+              fi
          }

          for tag in $TAGS; do
@@ -527,3 +545,24 @@ jobs:
          done
        env:
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Attest full image
+        if: ${{ matrix.config.full }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.full_digest }}
+
+      - name: Attest light image
+        if: ${{ matrix.config.light }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.light_digest }}
+
+      - name: Attest server image
+        if: ${{ matrix.config.server }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.server_digest }}
@@ -3,6 +3,11 @@ name: Make Release
 on:
  workflow_dispatch:
    inputs:
+      commit:
+        description: 'Commit SHA to release (empty = branch HEAD)'
+        required: false
+        default: ''
+        type: string
      dry_run:
        description: 'Dry run - validate without creating the tag'
        required: true
@@ -22,12 +27,17 @@ jobs:
    steps:
      - name: Checkout
        uses: actions/checkout@v6
+        with:
+          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
+          ref: ${{ inputs.commit != '' && inputs.commit || github.ref_name }}
+          fetch-depth: 0

      - 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 }}
+          RELEASE_BRANCH: ${{ github.ref_name }}

      - name: Create release tag
        if: ${{ github.event.inputs.dry_run == 'false' }}
@@ -39,8 +49,40 @@ jobs:
          git push origin "${VERSION}"
          echo "Created and pushed tag ${VERSION}"

+      - name: Generate release description
+        id: desc
+        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
+        env:
+          GITHUB_REPOSITORY: ${{ github.repository }}
+
+      - name: Create release
+        if: ${{ github.event.inputs.dry_run == 'false' }}
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        with:
+          tag_name: ${{ steps.checks.outputs.version }}
+          # TODO: remove the prerelease flag once the semantic versioning workflow is ready
+          # ref: https://github.com/ggml-org/ggml/discussions/1579
+          prerelease: true
+          body: |
+            > [!NOTE]
+            > Semantic versioning is still work in progress.
+            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+
+            ${{ steps.desc.outputs.nightly }}
+
+            ## ${{ steps.desc.outputs.changelog_title }}
+
+            ${{ steps.desc.outputs.changelog }}
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
-          echo "Dry run complete - all checks passed."
-          echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
+            echo "Dry run complete - all checks passed."
+            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+          else
+            echo "::error::Dry run found release check failures. A release tag would not be created."
+            exit 1
+          fi
@@ -446,8 +446,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -562,8 +562,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -681,6 +681,7 @@ jobs:
          name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip

  windows-cpu:
+    name: windows-cpu / ${{ matrix.arch }}
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

@@ -728,6 +729,7 @@ jobs:
            -DGGML_BACKEND_DL=ON ^
            -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
            -DGGML_OPENMP=ON ^
+            -DGGML_OPENMP_FETCH=ON ^
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release

@@ -739,7 +741,6 @@ jobs:
      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
          7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*

      - name: Upload artifacts
@@ -1439,7 +1440,9 @@ jobs:
      - name: xcodebuild for swift package
        id: xcodebuild
        run: |
-          ./build-xcframework.sh
+          # note: only macos and ios-device due to long build time
+          # ref: https://github.com/ggml-org/llama.cpp/pull/27252
+          ./build-xcframework.sh macos ios-device

      - name: Build Xcode project
        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
@@ -1567,6 +1570,8 @@ jobs:
    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
    permissions:
        contents: write # for creating release
+        id-token: write
+        attestations: write

    runs-on: ubuntu-slim

@@ -1575,14 +1580,14 @@ jobs:
      - windows
      - windows-cpu
      - windows-cuda
-      #- windows-sycl
+      - windows-sycl
      - windows-rocm
      - windows-openvino
      #- ubuntu-22-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
-      #- ubuntu-24-sycl
+      - ubuntu-24-sycl
      - android-arm64
      - macos-cpu
      - ios-xcode
@@ -1598,6 +1603,7 @@ jobs:
        uses: actions/checkout@v6
        with:
          fetch-depth: 0
+          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}

      - name: Determine tag name
        id: tag
@@ -1659,6 +1665,22 @@ jobs:
        run: |
          tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .

+      - name: Attest release artifacts
+        id: attest
+        uses: actions/attest@v4
+        with:
+          subject-path: 'release/*'
+
+      - name: Create and push git tag
+        run: |
+          TAG="${{ steps.tag.outputs.name }}"
+          if git rev-parse -q --verify "refs/tags/${TAG}" >/dev/null 2>&1; then
+            echo "Tag ${TAG} already exists, skipping creation"
+          else
+            git tag "${TAG}"
+            git push origin "${TAG}"
+          fi
+
      - name: Create release
        id: create_release
        uses: ggml-org/action-create-release@v1
@@ -1676,6 +1698,9 @@ jobs:
            **Website:**
            - <https://llama.app>

+            **Attestations:**
+            - <${{ steps.attest.outputs.attestation-url }}>
+
            **macOS/iOS:**
            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
            - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -2,12 +2,14 @@ You are a coding agent. Here are some very important rules that you must follow:

 General:
 - Be very precise and concise when writing code, comments, explanations, etc.
+- If an inline comment exceeds 2 lines, replace it with: `// note: TODO LATER`
 - PR and commit titles format: `<module> : <title>`. Lookup recents for examples
 - Don't try to build or run the code unless you are explicitly asked to do so
 - Use the `gh` CLI tool when querying PRs, issues, or other GitHub resources

 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
+- In `test-backend-ops.cpp`, do not mention specific backends (e.g. Metal, CUDA) in comments
 - When referencing issues or PRs in comments, use the format:
  - C/C++ code: `// ref: <url>`
  - Other (CMake, etc.): `# ref: <url>`
@@ -5,7 +5,7 @@ include(CheckIncludeFileCXX)
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
 set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 0)
+set(LLAMA_VERSION_PATCH 2)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -224,9 +224,10 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #

+add_subdirectory(vendor)
+
 if (LLAMA_BUILD_COMMON)
    add_subdirectory(common)
-    add_subdirectory(vendor/cpp-httplib)
 endif()

 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -7,10 +7,11 @@
 <b>LLM inference in C/C++</b>

 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
+[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
-[![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)
+[![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
+[![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](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) / [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)

@@ -119,7 +120,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 ## Acknowledgements

 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
+- [mackron/miniaudio](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [sheredom/subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
@@ -1,5 +1,8 @@
 #!/usr/bin/env bash
 #
+# usage: ./build-xcframework.sh [BUILD ...] (default: all builds)
+# builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device
+#
 # Options
 IOS_MIN_OS_VERSION=16.4
 MACOS_MIN_OS_VERSION=13.3
@@ -19,6 +22,43 @@ GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
 GGML_OPENMP=OFF

+# Max number of concurrent platform builds
+MAX_PARALLEL_BUILDS=1
+
+# Split the available cores between the concurrent builds (min 1)
+JOBS_PER_BUILD=$(( $(sysctl -n hw.logicalcpu) / MAX_PARALLEL_BUILDS ))
+if [[ "$JOBS_PER_BUILD" -lt 1 ]]; then
+    JOBS_PER_BUILD=1
+fi
+
+# echo "build_fn build_dir release_dir platform is_simulator min_os" for a build name
+build_spec() {
+    case "$1" in
+        ios-sim)      echo "build_ios_sim build-ios-sim Release-iphonesimulator ios true ${IOS_MIN_OS_VERSION}" ;;
+        ios-device)   echo "build_ios_device build-ios-device Release-iphoneos ios false ${IOS_MIN_OS_VERSION}" ;;
+        macos)        echo "build_macos build-macos Release macos false ${MACOS_MIN_OS_VERSION}" ;;
+        visionos)     echo "build_visionos build-visionos Release-xros visionos false ${VISIONOS_MIN_OS_VERSION}" ;;
+        visionos-sim) echo "build_visionos_sim build-visionos-sim Release-xrsimulator visionos true ${VISIONOS_MIN_OS_VERSION}" ;;
+        tvos-sim)     echo "build_tvos_sim build-tvos-sim Release-appletvsimulator tvos true ${TVOS_MIN_OS_VERSION}" ;;
+        tvos-device)  echo "build_tvos_device build-tvos-device Release-appletvos tvos false ${TVOS_MIN_OS_VERSION}" ;;
+        *)            return 1 ;;
+    esac
+}
+
+# Default: build everything
+if [[ $# -eq 0 ]]; then
+    BUILDS=(ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device)
+else
+    BUILDS=("$@")
+fi
+for b in "${BUILDS[@]}"; do
+    if ! build_spec "$b" >/dev/null; then
+        echo "Error: unknown build '$b'" >&2
+        echo "Valid builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device" >&2
+        exit 1
+    fi
+done
+
 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
 COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"

@@ -250,6 +290,7 @@ combine_static_libraries() {
        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
+        "${base_dir}/${build_dir}/vendor/hash/${release_dir}/libvendor-hash.a"
    )

    # Create temporary directory for processing
@@ -401,148 +442,189 @@ combine_static_libraries() {
    rm -rf "${temp_dir}"
 }

-echo "Building for iOS simulator..."
-cmake -B build-ios-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-    -DIOS=ON \
-    -DCMAKE_SYSTEM_NAME=iOS \
-    -DCMAKE_OSX_SYSROOT=iphonesimulator \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_ios_sim() {
+    echo "Building for iOS simulator..."
+    cmake -B build-ios-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+        -DIOS=ON \
+        -DCMAKE_SYSTEM_NAME=iOS \
+        -DCMAKE_OSX_SYSROOT=iphonesimulator \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-ios-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for iOS devices..."
-cmake -B build-ios-device -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=iOS \
-    -DCMAKE_OSX_SYSROOT=iphoneos \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_ios_device() {
+    echo "Building for iOS devices..."
+    cmake -B build-ios-device -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=iOS \
+        -DCMAKE_OSX_SYSROOT=iphoneos \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-ios-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for macOS..."
-cmake -B build-macos -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -S .
-cmake --build build-macos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_macos() {
+    echo "Building for macOS..."
+    cmake -B build-macos -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -S .
+    cmake --build build-macos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for visionOS..."
-cmake -B build-visionos -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DCMAKE_SYSTEM_NAME=visionOS \
-    -DCMAKE_OSX_SYSROOT=xros \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_visionos() {
+    echo "Building for visionOS..."
+    cmake -B build-visionos -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DCMAKE_SYSTEM_NAME=visionOS \
+        -DCMAKE_OSX_SYSROOT=xros \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DLLAMA_BUILD_SERVER=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-visionos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for visionOS simulator..."
-cmake -B build-visionos-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DCMAKE_SYSTEM_NAME=visionOS \
-    -DCMAKE_OSX_SYSROOT=xrsimulator \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_visionos_sim() {
+    echo "Building for visionOS simulator..."
+    cmake -B build-visionos-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DCMAKE_SYSTEM_NAME=visionOS \
+        -DCMAKE_OSX_SYSROOT=xrsimulator \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DLLAMA_BUILD_SERVER=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-visionos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
-echo "Building for tvOS simulator..."
-cmake -B build-tvos-sim -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=tvOS \
-    -DCMAKE_OSX_SYSROOT=appletvsimulator \
-    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-    -DGGML_METAL=ON \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_tvos_sim() {
+    echo "Building for tvOS simulator..."
+    cmake -B build-tvos-sim -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=tvOS \
+        -DCMAKE_OSX_SYSROOT=appletvsimulator \
+        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+        -DGGML_METAL=ON \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-tvos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}

-echo "Building for tvOS devices..."
-cmake -B build-tvos-device -G Xcode \
-    "${COMMON_CMAKE_ARGS[@]}" \
-    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=tvOS \
-    -DCMAKE_OSX_SYSROOT=appletvos \
-    -DCMAKE_OSX_ARCHITECTURES="arm64" \
-    -DGGML_METAL=ON \
-    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DMTMD_VIDEO=OFF \
-    -S .
-cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+build_tvos_device() {
+    echo "Building for tvOS devices..."
+    cmake -B build-tvos-device -G Xcode \
+        "${COMMON_CMAKE_ARGS[@]}" \
+        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+        -DCMAKE_SYSTEM_NAME=tvOS \
+        -DCMAKE_OSX_SYSROOT=appletvos \
+        -DCMAKE_OSX_ARCHITECTURES="arm64" \
+        -DGGML_METAL=ON \
+        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
+        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+        -DLLAMA_OPENSSL=OFF \
+        -DMTMD_VIDEO=OFF \
+        -S .
+    cmake --build build-tvos-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
+}
+
+run_builds_parallel() {
+    local -a pids=()
+    local -a names=()
+    local name i
+    for name in "$@"; do
+        # Wait for the oldest running build to free a slot
+        if [[ "${#pids[@]}" -ge "$MAX_PARALLEL_BUILDS" ]]; then
+            if ! wait "${pids[0]}"; then
+                echo "ERROR: build '${names[0]}' failed, log follows (${names[0]}.log):" >&2
+                kill "${pids[@]}" 2>/dev/null || true
+                cat "${names[0]}.log" >&2
+                exit 1
+            fi
+            pids=("${pids[@]:1}")
+            names=("${names[@]:1}")
+        fi
+        echo "Starting build: $name (log: ${name}.log, -j ${JOBS_PER_BUILD})"
+        "$name" > "${name}.log" 2>&1 &
+        pids+=("$!")
+        names+=("$name")
+    done
+    # Wait for the remaining builds
+    for i in "${!pids[@]}"; do
+        if ! wait "${pids[$i]}"; then
+            echo "ERROR: build '${names[$i]}' failed, log follows (${names[$i]}.log):" >&2
+            kill "${pids[@]}" 2>/dev/null || true
+            cat "${names[$i]}.log" >&2
+            exit 1
+        fi
+    done
+}
+
+BUILD_FNS=()
+for b in "${BUILDS[@]}"; do
+    read -r fn _ < <(build_spec "$b")
+    BUILD_FNS+=("$fn")
+done
+echo "Building: ${BUILDS[*]} (max ${MAX_PARALLEL_BUILDS} at a time, -j ${JOBS_PER_BUILD} each)..."
+run_builds_parallel "${BUILD_FNS[@]}"

 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."
-setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios"
-setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios"
-setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos"
-setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos"
-setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos"
-setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos"
-setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos"
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir _ platform _ min_os < <(build_spec "$b")
+    setup_framework_structure "$bdir" "$min_os" "$platform"
+done

 # Create dynamic libraries from static libraries
 echo "Creating dynamic libraries from static libraries..."
-combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true"
-combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false"
-combine_static_libraries "build-macos" "Release" "macos" "false"
-combine_static_libraries "build-visionos" "Release-xros" "visionos" "false"
-combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true"
-combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true"
-combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir rdir platform is_sim _ < <(build_spec "$b")
+    combine_static_libraries "$bdir" "$rdir" "$platform" "$is_sim"
+done

 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
+XCFW_ARGS=()
+for b in "${BUILDS[@]}"; do
+    read -r _ bdir _ _ _ _ < <(build_spec "$b")
+    XCFW_ARGS+=(-framework "$(pwd)/${bdir}/framework/llama.framework")
+    XCFW_ARGS+=(-debug-symbols "$(pwd)/${bdir}/dSYMs/llama.dSYM")
+done
 xcrun xcodebuild -create-xcframework \
-    -framework $(pwd)/build-ios-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-ios-device/framework/llama.framework \
-    -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-visionos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-visionos-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-visionos-sim/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-tvos-device/framework/llama.framework \
-    -debug-symbols $(pwd)/build-tvos-device/dSYMs/llama.dSYM \
-    -framework $(pwd)/build-tvos-sim/framework/llama.framework \
-    -debug-symbols $(pwd)/build-tvos-sim/dSYMs/llama.dSYM \
-    -output $(pwd)/build-apple/llama.xcframework
+    "${XCFW_ARGS[@]}" \
+    -output "$(pwd)/build-apple/llama.xcframework"
@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

    # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs"
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi

 ## helpers
@@ -8,6 +8,7 @@ set( CMAKE_CXX_COMPILER  clang++ )

 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
+set( CMAKE_ASM_COMPILER_TARGET ${target} )

 set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
@@ -126,7 +126,8 @@ set_target_properties(${TARGET} PROPERTIES
    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )

-target_include_directories(${TARGET} PUBLIC . ../vendor)
+target_include_directories(${TARGET} PUBLIC .)
+target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)

 if (LLAMA_SUBPROCESS)
@@ -2595,6 +2595,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.mmproj_use_gpu = value;
        }
    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
+    add_opt(common_arg(
+        // note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
+        {"-mmdev", "--mmproj-device"}, "DEVICE",
+        "device to use for multimodal projector (none = don't offload, default: auto)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.mmproj_use_gpu = false;
+                params.mmproj_device  = nullptr;
+                return;
+            }
+            auto devices = parse_device_list(value);
+            // parse_device_list pushes nullptr at back so devices is length 2 for single device.
+            if (devices.size() > 2) {
+                throw std::invalid_argument("only one device may be specified for mmproj");
+            }
+            params.mmproj_use_gpu = true;
+            params.mmproj_device  = devices.front();
+        }
+    ).set_examples(mmproj_examples).set_env("MTMD_BACKEND_DEVICE")); // no LLAMA_ARG_ prefix for backward compatibility reason
    add_opt(common_arg(
        {"--image", "--audio", "--video"}, "FILE",
        "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",
@@ -3362,7 +3382,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--tools"}, "TOOL1,TOOL2,...",
        "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
        "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, get_info\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_info\n"
        "note: for security reasons, this will limit --cors-origins to localhost by default",
        [](common_params & params, const std::string & value) {
            params.server_tools = parse_csv_row(value);
@@ -4658,6 +4678,12 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
        [](common_params &, int) { /* unused */ }
    ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());

+    args.push_back(common_arg(
+        {"dedup-cache-models"}, "0|1",
+        "in server router mode, hide a cached model from the model list when this preset resolves to the same model file",
+        [](common_params &, const std::string &) { /* unused */ }
+    ).set_env(COMMON_ARG_PRESET_DEDUP_CACHE_MODELS).set_preset_only());
+
    // args.push_back(common_arg(
    //     {"pin"},
    //     "in server router mode, do not unload this model if models_max is exceeded",
@@ -11,8 +11,9 @@
 #include <memory>

 // pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP    "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT       "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"

 //
 // CLI argument parsing
@@ -193,6 +193,14 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
              LOG_DBG(ANSI_ORANGE "[Patch: Laguna]\n" ANSI_RESET);
          }
      },
+      // Bailing V3
+      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
+          if (tmpl.src.find("Bailing V3 chat template") != std::string::npos) {
+              analysis.tools.arguments.value_suffix = trim_whitespace(analysis.tools.arguments.value_suffix);
+              analysis.tools.arguments.tolerate_intertag_whitespace = true;
+              LOG_DBG(ANSI_ORANGE "[Patch: Bailing V3]\n" ANSI_RESET);
+          }
+      },

    });

@@ -470,36 +470,80 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
    return msgs;
 }

+struct messages_inp_normalizer {
+    const jinja::caps & caps;
+
+    messages_inp_normalizer(const jinja::caps & c) : caps(c) {}
+
+    // handle supports_string_content / supports_typed_content
+    // if string=true and array=false, convert array to string
+    // if string=false and array=true, convert string to array
+    // if both are true, do nothing
+    json normalize(const json & messages) {
+        bool only_string = caps.supports_string_content && !caps.supports_typed_content;
+        bool only_typed  = !caps.supports_string_content && caps.supports_typed_content;
+        if ((!only_string && !only_typed) || !messages.is_array()) {
+            return messages;
+        }
+        json normalized = json::array();
+        for (const auto & msg : messages) {
+            json copy = msg;
+            auto it = copy.find("content");
+            if (it != copy.end()) {
+                if (only_typed && it->is_string()) {
+                    *it = json::array({
+                        json{
+                            {"type", "text"},
+                            {"text", it->get<std::string>()},
+                        }
+                    });
+                } else if (only_string && it->is_array()) {
+                    *it = concat_content_parts(*it);
+                }
+            }
+            normalized.push_back(std::move(copy));
+        }
+        return normalized;
+    }
+
+    // join parts with newline, do not add newline before or after media markers
+    static std::string concat_content_parts(const json & parts) {
+        std::string text;
+        bool last_was_media_marker = false;
+        for (const auto & part : parts) {
+            std::string type = part.value("type", "");
+            bool add_new_line = true;
+            if (type == "text") {
+                add_new_line = !last_was_media_marker && !text.empty();
+                last_was_media_marker = false;
+            } else if (type == "media_marker") {
+                add_new_line = false;
+                last_was_media_marker = true;
+            } else {
+                LOG_WRN("Ignoring content part type: %s\n", type.c_str());
+                continue;
+            }
+
+            if (add_new_line) {
+                text += '\n';
+            }
+
+            text += part.value("text", "");
+        }
+        return text;
+    }
+};
+
 static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
    if (!c.supports_string_content && !c.supports_typed_content) {
        LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
    }

-    bool only_string_accepted =  c.supports_string_content && !c.supports_typed_content;
-    bool only_typed_accepted  = !c.supports_string_content &&  c.supports_typed_content;
-
    json messages = json::array();
    for (const auto & msg : msgs) {
-        if (only_string_accepted) {
-            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
-            messages.push_back(jmsg);
-        } else if (only_typed_accepted) {
-            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
-            if (jmsg.at("content").is_string()) {
-                jmsg["content"] = json::array({
-                    json{
-                        {"type", "text"},
-                        {"text", jmsg.at("content").get<std::string>()},
-                    }
-                });
-            }
-            messages.push_back(jmsg);
-        } else {
-            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
-            messages.push_back(jmsg);
-        }
+        messages.push_back(msg.to_json_oaicompat(/* concat_typed_text= */ false));
    }
-    return messages;
+    return messages_inp_normalizer(c).normalize(messages);
 }

 // DEPRECATED: only used in tests
@@ -892,8 +936,11 @@ static std::string common_chat_template_direct_apply_impl(
    const std::optional<json> & additional_context = std::nullopt) {
    jinja::context ctx(tmpl.source());

+    // messages_override is already built for this template, do not touch its content parts
    nlohmann::ordered_json inp = nlohmann::ordered_json{
-        {"messages", messages_override.has_value() ? *messages_override : inputs.messages},
+        {"messages", messages_override.has_value()
+            ? *messages_override
+            : messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
        {"bos_token", tmpl.bos_token()},
        {"eos_token", tmpl.eos_token()},
        {"enable_thinking", inputs.enable_thinking},
@@ -957,14 +1004,12 @@ static std::string common_chat_template_generation_prompt_impl(
    const std::optional<json> & tools_override = std::nullopt,
    const std::optional<json> & additional_context = std::nullopt) {

-    auto adjusted_messages = messages_override ? *messages_override : inputs.messages;
-
    autoparser::generation_params params = inputs;
    params.add_generation_prompt = false;
    params.continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
-    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
+    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params, messages_override, tools_override, additional_context);
    params.add_generation_prompt = true;
-    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
+    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, messages_override, tools_override, additional_context);

    size_t prefix_len = 0;
    size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
@@ -1778,6 +1778,8 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    struct ggml_threadpool_params tpp =
            ggml_threadpool_params_from_cpu_params(params.cpuparams);

+    // each pool needs to match the respective n_threads exactly
+    // see: https://github.com/ggml-org/llama.cpp/pull/27138#issuecomment-5332307332
    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
        if (!threadpool_batch) {
@@ -581,9 +581,10 @@ struct common_params {

    // multimodal models (see tools/mtmd)
    struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;     // use GPU for multimodal model
-    bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
+    bool mmproj_use_gpu = true;                 // use GPU for multimodal model
+    ggml_backend_dev_t mmproj_device = nullptr; // GPU device to use for multimodal model
+    bool no_mmproj = false;                     // explicitly disable multimodal model
+    std::vector<std::string> image;             // path to image file(s) ; TODO: change the name to "media"
    int image_min_tokens = -1;
    int image_max_tokens = -1;
    int mtmd_batch_max_tokens = 1024;
@@ -989,6 +989,26 @@ std::vector<common_cached_model_info> common_list_cached_models() {
    return result;
 }

+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
+    auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
+
+    auto files = hf_cache::get_cached_files(repo);
+    if (files.empty()) {
+        return "";
+    }
+
+    if (!hf_file.empty()) {
+        for (const auto & f : files) {
+            if (f.path == hf_file) {
+                return f.local_path;
+            }
+        }
+        return "";
+    }
+
+    return find_best_model(files, tag).local_path;
+}
+
 bool common_download_remove(const std::string & hf_repo_with_tag) {
    namespace fs = std::filesystem;

@@ -85,6 +85,10 @@ std::vector<std::string> common_download_get_all_parts(const std::string & url);
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();

+// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
+// returns an empty string if the model is not present in the cache
+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
+
 // download single file from url to local path
 // returns status code or -1 on error
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
@@ -23,7 +23,7 @@ void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) {
    ctx.set_val("preserve_thinking",         mk_val<value_bool>(enabled));
    ctx.set_val("clear_thinking",            mk_val<value_bool>(!enabled));
    ctx.set_val("truncate_history_thinking", mk_val<value_bool>(!enabled));
-    ctx.set_val("drop_thinking",            mk_val<value_bool>(!enabled));
+    ctx.set_val("drop_thinking",             mk_val<value_bool>(!enabled));
 }

 void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort) {
@@ -117,6 +117,8 @@ caps caps_get(jinja::program & prog) {

    JJ_DEBUG("%s\n", ">>> Running capability check: typed content");

+    static const std::string content_marker = "STRING_MARKER";
+
    // case: typed content support
    caps_try_execute(
        prog,
@@ -125,22 +127,26 @@ caps caps_get(jinja::program & prog) {
            return json::array({
                {
                    {"role", "user"},
-                    {"content", "content"}
+                    {"content", content_marker}
                }
            });
        },
        nullptr, // ctx_fn
        nullptr, // tools_fn
-        [&](context &, bool success, value & messages, value &, const std::string &) {
+        [&](context &, bool success, value & messages, value &, const std::string & rendered) {
            auto & content = messages->at(0)->at("content");
            caps_print_stats(content, "messages[0].content");
-            if (has_op(content, "selectattr") || has_op(content, "array_access")) {
+            bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
+            if (used_as_array) {
                // accessed as an array
                result.supports_typed_content = true;
            }
            if (!success) {
                // failed to execute with content as string
                result.supports_string_content = false;
+            } else if (used_as_array && rendered.find(content_marker) == std::string::npos) {
+                // edge case: string may be accessed for checking, but does not appear in the output
+                result.supports_string_content = false;
            }
        }
    );
@@ -278,7 +278,9 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };

-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+static const int MAX_PATTERN_DEPTH = 100;
+
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};

 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -309,6 +311,32 @@ static std::string format_literal(const std::string & literal) {

 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }

+static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
+    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
+        return 0;
+    }
+    size_t n_hex = 0;
+    switch (pattern[pos + 1]) {
+        case 'x': n_hex = 2; break;
+        case 'u': n_hex = 4; break;
+        case 'U': n_hex = 8; break;
+        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
+            return 2;
+        default:
+            return 0;
+    }
+    if (pos + 2 + n_hex > pattern.length()) {
+        return 0;
+    }
+    for (size_t i = pos + 2; i < pos + 2 + n_hex; i++) {
+        char h = pattern[i];
+        if (!((h >= '0' && h <= '9') || (h >= 'a' && h <= 'f') || (h >= 'A' && h <= 'F'))) {
+            return 0;
+        }
+    }
+    return 2 + n_hex;
+}
+
 class common_schema_converter {
 private:
    friend class common_schema_info;
@@ -345,16 +373,42 @@ private:
        return string_join(rules, " | ");
    }

+    // thrown when the pattern is a valid regex with no grammar equivalent
+    struct unsupported_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
+    // thrown when the pattern is not a valid regex
+    struct invalid_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        if (!(pattern.front() == '^' && pattern.back() == '$')) {
-            _errors.push_back("Pattern must start with '^' and end with '$'");
+        auto rules_snapshot = _rules;
+        try {
+            return _pattern_to_rule(pattern, name);
+        } catch (const unsupported_pattern & err) {
+            // revert rules
+            _rules = std::move(rules_snapshot);
+            _warnings.push_back("pattern " + pattern + " is not supported (" + err.what() + "), accepting any string");
+            return _add_rule(name, _add_primitive("string", PRIMITIVE_RULES.at("string")));
+        } catch (const invalid_pattern & err) {
+            _rules = std::move(rules_snapshot);
+            _errors.push_back("Invalid pattern " + pattern + ": " + err.what());
            return "";
        }
+    }
+
+    std::string _pattern_to_rule(const std::string & pattern, const std::string & name) {
+        if (pattern.length() < 2 || pattern.front() != '^' || pattern.back() != '$') {
+            throw unsupported_pattern("not anchored with '^' and '$'");
+        }
        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
        std::unordered_map<std::string, std::string> sub_rule_ids;

        size_t i = 0;
        size_t length = sub_pattern.length();
+        int paren_depth = 0;

        using literal_or_rule = std::pair<std::string, bool>;
        auto to_rule = [&](const literal_or_rule & ls) {
@@ -363,7 +417,6 @@ private:
            return is_literal ? "\"" + s + "\"" : s;
        };
        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
-            size_t start = i;
            std::vector<literal_or_rule> seq;

            auto get_dot = [&]() {
@@ -420,43 +473,42 @@ private:
                        if (i + 1 < length && sub_pattern[i + 1] == ':') {
                            i += 2; // skip "?:" for non-capturing group, treat as regular group
                        } else {
-                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
-                            _warnings.push_back("Unsupported pattern syntax");
-                            // skip to matching ')' to avoid UB on empty seq
-                            int depth = 1;
-                            while (i < length && depth > 0) {
-                                if (sub_pattern[i] == '\\' && i + 1 < length) {
-                                    i += 2; // skip escaped character
-                                } else {
-                                    if (sub_pattern[i] == '(') depth++;
-                                    else if (sub_pattern[i] == ')') depth--;
-                                    i++;
-                                }
-                            }
-                            continue;
+                            // lookaround, named group, inline flags, ...
+                            throw unsupported_pattern("unsupported group syntax");
                        }
                    }
+                    paren_depth++;
+                    if (paren_depth > MAX_PATTERN_DEPTH) {
+                        throw unsupported_pattern("pattern nesting too deep");
+                    }
                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
                } else if (c == ')') {
                    i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
-                        _errors.push_back("Unbalanced parentheses");
+                    if (paren_depth == 0) {
+                        throw invalid_pattern("unbalanced parentheses");
                    }
+                    paren_depth--;
                    return join_seq();
+                } else if (c == '^' || c == '$') {
+                    throw unsupported_pattern("anchor inside the pattern");
                } else if (c == '[') {
                    std::string square_brackets = std::string(1, c);
                    i++;
                    while (i < length && sub_pattern[i] != ']') {
                        if (sub_pattern[i] == '\\') {
-                            square_brackets += sub_pattern.substr(i, 2);
-                            i += 2;
+                            auto escape_length = gbnf_escape_length(sub_pattern, i);
+                            if (escape_length == 0) {
+                                throw unsupported_pattern("unsupported escape in character class: " + sub_pattern.substr(i, 2));
+                            }
+                            square_brackets += sub_pattern.substr(i, escape_length);
+                            i += escape_length;
                        } else {
                            square_brackets += sub_pattern[i];
                            i++;
                        }
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced square brackets");
+                        throw invalid_pattern("unterminated character class");
                    }
                    square_brackets += ']';
                    i++;
@@ -465,6 +517,9 @@ private:
                    seq.emplace_back("|", false);
                    i++;
                } else if (c == '*' || c == '+' || c == '?') {
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
+                    }
                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
                    i++;
                } else if (c == '{') {
@@ -475,18 +530,19 @@ private:
                        i++;
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced curly brackets");
+                        throw unsupported_pattern("unterminated curly brackets");
                    }
                    curly_brackets += '}';
                    i++;
                    auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                    int min_times = 0;
                    int max_times = std::numeric_limits<int>::max();
+                    if (nums.size() != 1 && nums.size() != 2) {
+                        throw unsupported_pattern("wrong number of values in curly brackets");
+                    }
                    try {
                        if (nums.size() == 1) {
                            min_times = max_times = std::stoi(nums[0]);
-                        } else if (nums.size() != 2) {
-                            _errors.push_back("Wrong number of values in curly brackets");
                        } else {
                            if (!nums[0].empty()) {
                                min_times = std::stoi(nums[0]);
@@ -495,9 +551,11 @@ private:
                                max_times = std::stoi(nums[1]);
                            }
                        }
-                    } catch (const std::invalid_argument & e) {
-                        _errors.push_back("Invalid number in curly brackets");
-                        return std::make_pair("", false);
+                    } catch (const std::logic_error &) {
+                        throw unsupported_pattern("invalid number in curly brackets");
+                    }
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
                    }
                    auto &last = seq.back();
                    auto &sub = last.first;
@@ -523,15 +581,22 @@ private:
                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                    };
                    while (i < length) {
-                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                        if (sub_pattern[i] == '\\') {
+                            if (i == length - 1) {
+                                throw invalid_pattern("trailing backslash");
+                            }
                            char next = sub_pattern[i + 1];
                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
                                i++;
                                literal += sub_pattern[i];
                                i++;
                            } else {
-                                literal += sub_pattern.substr(i, 2);
-                                i += 2;
+                                auto escape_length = gbnf_escape_length(sub_pattern, i);
+                                if (escape_length == 0) {
+                                    throw unsupported_pattern("unsupported escape: " + sub_pattern.substr(i, 2));
+                                }
+                                literal += sub_pattern.substr(i, escape_length);
+                                i += escape_length;
                            }
                        } else if (sub_pattern[i] == '"') {
                            literal += "\\\"";
@@ -544,14 +609,21 @@ private:
                            break;
                        }
                    }
-                    if (!literal.empty()) {
-                        seq.emplace_back(literal, true);
+                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
+                        throw unsupported_pattern(std::string("unsupported character: ") + c);
                    }
+                    seq.emplace_back(literal, true);
                }
            }
            return join_seq();
        };
-        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
+
+        auto rule = to_rule(transform());
+        if (paren_depth != 0) {
+            throw invalid_pattern("unbalanced parentheses");
+        }
+
+        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
    }

    /*
@@ -926,6 +926,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
    const bool is_dspark;

+    // dspark speculators
+    bool sample_from_anchor = true;
+
    const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
    uint32_t        target_layer_ids_n = 0;

@@ -960,16 +963,20 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            if (llama_model_meta_val_str(model_dft, "dflash.block_size", buf, sizeof(buf)) >= 0) {
                block_size = std::atoi(buf);
            }
+            if (llama_model_meta_val_str(model_dft, "dflash.sample_from_anchor", buf, sizeof(buf)) >= 0) {
+                sample_from_anchor = std::strcmp(buf, "true") == 0;
+            }
        }
        mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));

        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
-        LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n);
+        LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u, sample_from_anchor=%s\n", __func__,
+                block_size, mask_token_id, target_layer_ids_n, sample_from_anchor ? "true" : "false");

        // DFlash input is [id_last, <mask> * (block_size-1)]: in-place denoising yields at most
-        // block_size-1 draft tokens, DSpark yield a full block_size draft tokens
-        const int32_t n_draft_max = is_dspark ? block_size : block_size - 1;
+        // block_size-1 draft tokens, anchor-first DSpark yields a full block_size draft tokens
+        const int32_t n_draft_max = is_dspark && sample_from_anchor ? block_size : block_size - 1;
        if (this->params.n_max > n_draft_max || this->params.n_min > n_draft_max) {
            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained block size %d -- clamping to %d\n",
                    __func__, this->params.n_max, this->params.n_min, block_size, n_draft_max);
@@ -1175,7 +1182,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

            const int32_t n_draft = params.n_max;

-            const int32_t n_block_tokens = n_draft + (is_dspark ? 0 : 1);
+            const int32_t n_block_tokens = n_draft + (is_dspark && sample_from_anchor ? 0 : 1);
            i_block_beg[seq_id] = batch.n_tokens;
            n_block    [seq_id] = n_block_tokens;
            for (int32_t i = 0; i < n_block_tokens; ++i) {
@@ -1208,11 +1215,11 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            auto & result = *dp.result;

            if (is_dspark) {
-                // DSpark predicts the next token from position 0 and optionally truncates
-                // at the first position below the confidence threshold.
+                // DSpark: read from the first draft slot, truncate below the confidence threshold
                const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
-
-                for (int32_t i = 0; i < n_block_tokens; ++i) {
+                // bonus-anchor drafts read the mask positions only, like DFlash
+                const int32_t i_draft_beg = sample_from_anchor ? 0 : 1;
+                for (int32_t i = i_draft_beg; i < n_block_tokens; ++i) {
                    const int32_t idx = beg + i;

                    if (conf && conf[(size_t) idx * n_embd_dec] < params.p_min) {
@@ -2642,6 +2649,10 @@ void common_speculative_draft(common_speculative * spec) {
        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) dparams.size(); ++seq_id) {
            auto & dp = dparams[seq_id];

+            if (!dp.drafting) {
+                continue;
+            }
+
            auto & result = *dp.result;

            // a new draft has been sampled
@@ -27,6 +27,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "BaichuanForCausalLM": "baichuan",
    "BailingMoeForCausalLM": "bailingmoe",
    "BailingMoeV2ForCausalLM": "bailingmoe",
+    "BailingMoeV3ForCausalLM": "bailingmoe3",
    "BambaForCausalLM": "granite",
    "BertForMaskedLM": "bert",
    "BertForSequenceClassification": "bert",
@@ -54,6 +55,9 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "DeepseekV32ForCausalLM": "deepseek",
    "DFlashDraftModel": "qwen",
    "Qwen3DSparkModel": "qwen",
+    "DSparkDraftModel": "qwen",
+    "DSparkSpeculator": "qwen",
+    "Lfm2DSparkDraftModel": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
    "DistilBertForMaskedLM": "bert",
@@ -106,6 +110,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
+    "GraniteSWAForCausalLM": "granite",
+    "GraniteMoeSWAForCausalLM": "granite",
    "Grok1ForCausalLM": "grok",
    "GrokForCausalLM": "grok",
    "GroveMoeForCausalLM": "grovemoe",
@@ -13,6 +13,7 @@ from .llama import LlamaModel


@ModelBase.register("AfmoeForCausalLM")
+@ModelBase.example("arcee-ai/Trinity-Large-Thinking")
 class AfmoeModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.AFMOE

@@ -16,6 +16,7 @@ from .llama import LlamaModel


@ModelBase.register("ArcticForCausalLM")
+@ModelBase.example("Snowflake/snowflake-arctic-instruct")
 class ArcticModel(TextModel):
    model_arch = gguf.MODEL_ARCH.ARCTIC

@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
+@ModelBase.example("baichuan-inc/Baichuan2-7B-Chat", "baichuan-inc/Baichuan-7B")
 class BaichuanModel(TextModel):
    model_arch = gguf.MODEL_ARCH.BAICHUAN

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("BailingMoeForCausalLM")
+@ModelBase.example("inclusionAI/Ling-lite")
 class BailingMoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.BAILINGMOE

@@ -108,6 +109,7 @@ class BailingMoeModel(TextModel):


@ModelBase.register("BailingMoeV2ForCausalLM")
+@ModelBase.example("inclusionAI/Ling-mini-2.0")
 class BailingMoeV2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.BAILINGMOE2

@@ -189,6 +191,7 @@ class BailingMoeV2Model(TextModel):


@ModelBase.register("SarvamMoEForCausalLM", "modeling_sarvam_moe.SarvamMoEForCausalLM")
+@ModelBase.example("sarvamai/sarvam-30b")
 class SarvamMoEModel(BailingMoeV2Model):
    model_arch = gguf.MODEL_ARCH.BAILINGMOE2
    # Sarvam-MoE shares the BailingMoeV2 architecture; only differences:
@@ -0,0 +1,193 @@
+from __future__ import annotations
+
+import re
+
+from typing import Callable, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf
+
+
+@ModelBase.register("BailingMoeV3ForCausalLM")
+@ModelBase.example("inclusionAI/Ling-3.0-tiny", "inclusionAI/Ling-3.0-flash")
+class BailingMoeV3Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE3
+    supports_mtp_export = True
+
+    _experts: list[dict[str, Tensor]] | None = None
+    _main_layers: int | None = None
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        nextn_layers = self.hparams.get("num_nextn_predict_layers", 0) or 0
+        if self.no_mtp:
+            nextn_layers = 0
+        self.block_count = self.hparams["num_hidden_layers"] + nextn_layers
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def is_full_attention(self, bid: int) -> bool:
+        n_layer = self.hparams["num_hidden_layers"]
+        layer_group_size = self.hparams["layer_group_size"]
+        return bid >= n_layer or (bid + 1) % layer_group_size == 0 or bid >= n_layer // layer_group_size * layer_group_size
+
+    def set_gguf_parameters(self):
+        if not self.hparams.get("no_kda_lora", False):
+            raise ValueError("BailingMoeV3 KDA LoRA projections are not supported")
+        if not self.hparams.get("kda_safe_gate", False):
+            raise ValueError("BailingMoeV3 non-safe KDA gates are not supported")
+        if self.hparams.get("gated_attention_proj_granularity_type") != "head_wise":
+            raise ValueError("BailingMoeV3 requires head-wise attention gates")
+
+        self.hparams["num_key_value_heads"] = 1
+        super().set_gguf_parameters()
+
+        n_head_kv = [1 if self.is_full_attention(il) else 0 for il in range(self.block_count)]
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_ssm_conv_kernel(self.hparams["short_conv_kernel_size"])
+        self.gguf_writer.add_kda_head_dim(self.hparams["head_dim"])
+        self.gguf_writer.add_kda_safe_gate(self.hparams["kda_safe_gate"])
+        self.gguf_writer.add_kda_gate_lower_bound(self.hparams["kda_lower_bound"])
+
+        kv_lora_rank = self.hparams["kv_lora_rank"]
+        qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+        qk_rope_head_dim = self.hparams["qk_rope_head_dim"]
+        if (q_lora_rank := self.hparams.get("q_lora_rank")) is not None:
+            self.gguf_writer.add_q_lora_rank(q_lora_rank)
+        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
+        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
+        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
+        self.gguf_writer.add_key_length_mla(qk_nope_head_dim + qk_rope_head_dim)
+        self.gguf_writer.add_value_length_mla(self.hparams["v_head_dim"])
+
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+
+        def clamp_limits(key: str) -> list[float] | None:
+            values = self.hparams.get(key)
+            if values is None:
+                return None
+            values = [0.0 if value is None else float(value) for value in values[:self.block_count]]
+            return values + [0.0] * (self.block_count - len(values))
+
+        if (values := clamp_limits("expert_swiglu_limit_list")) is not None:
+            self.gguf_writer.add_swiglu_clamp_exp(values)
+        if (values := clamp_limits("share_expert_swiglu_limit_list")) is not None:
+            self.gguf_writer.add_swiglu_clamp_shexp(values)
+
+        if not self.no_mtp and (nextn_layers := self.hparams.get("num_nextn_predict_layers", 0)):
+            self.gguf_writer.add_nextn_predict_layers(nextn_layers)
+
+    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"
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.endswith(".expert_bias"):
+            name += ".bias"
+
+        if cls._main_layers is None:
+            return super().filter_tensors((name, gen))
+
+        m = re.match(r"model\.layers\.(\d+)\.", name)
+        is_mtp = m is not None and int(m.group(1)) >= cls._main_layers
+
+        if is_mtp and cls.no_mtp:
+            return None
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.word_embeddings.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return super().filter_tensors((name, gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")) and data_torch.ndim in (2, 3):
+            d_inner = data_torch.shape[0]
+            d_conv = data_torch.shape[-1]
+            data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+
+        if name.endswith(".A_log"):
+            data_torch = torch.exp(data_torch).reshape(-1, 1)
+
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+
+        if name.endswith(".attention.f_proj.weight"):
+            assert bid is not None
+            if self.is_full_attention(bid):
+                raise ValueError(f"unexpected f_proj on full-attention layer {bid}")
+            name = self.format_tensor_name(gguf.MODEL_TENSOR.SSM_F_A, bid)
+
+        if name.endswith(".attention.g_proj.weight"):
+            assert bid is not None
+            tensor = gguf.MODEL_TENSOR.ATTN_GATE if self.is_full_attention(bid) else gguf.MODEL_TENSOR.SSM_G_A
+            name = self.format_tensor_name(tensor, bid)
+
+        if ".mlp.experts." in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+            if len(self._experts[bid]) >= n_experts * 3:
+                for weight_name in ("down_proj", "gate_proj", "up_proj"):
+                    tensors = []
+                    for expert_id in range(n_experts):
+                        expert_name = f"model.layers.{bid}.mlp.experts.{expert_id}.{weight_name}.weight"
+                        tensors.append(self._experts[bid].pop(expert_name))
+                    merged_name = f"model.layers.{bid}.mlp.experts.{weight_name}.weight"
+                    yield from super().modify_tensors(torch.stack(tensors, dim=0), merged_name, bid)
+            return
+
+        if name.endswith(".attention.kv_b_proj.weight"):
+            assert bid is not None
+            n_head = self.hparams["num_attention_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            assert data_torch.shape[0] == n_head * (v_head_dim + qk_nope_head_dim)
+            kv_b = data_torch.view(n_head, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            name_k = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K_B, bid)
+            name_v = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V_B, bid)
+            yield from super().modify_tensors(k_b.transpose(1, 2), name_k, bid)
+            yield from super().modify_tensors(v_b, name_v, bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            experts = [name for layer in self._experts for name in layer]
+            if experts:
+                raise ValueError(f"Unprocessed experts: {experts}")
@@ -1149,6 +1149,14 @@ class ModelBase:
            return modelcls
        return func

+    @classmethod
+    def example(cls, *hf_repos: str) -> Callable[[AnyModel], AnyModel]:
+        del hf_repos  # unused
+
+        def func(modelcls: AnyModel) -> AnyModel:
+            return modelcls
+        return func
+
    @classmethod
    def print_registered_models(cls):
        for model_type, model_classes in cls._model_classes.items():
@@ -15,6 +15,7 @@ from .base import ModelBase, SentencePieceTokenTypes, TextModel, gguf, logger


@ModelBase.register("BertModel", "BertForMaskedLM", "CamembertModel", "BertForSequenceClassification")
+@ModelBase.example("BAAI/bge-small-en-v1.5", "dangvantuan/sentence-camembert-base")
 class BertModel(TextModel):
    model_arch = gguf.MODEL_ARCH.BERT

@@ -240,6 +241,7 @@ class BertModel(TextModel):


@ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
+@ModelBase.example("distilbert/distilbert-base-uncased")
 class DistilBertModel(BertModel):
    model_arch = gguf.MODEL_ARCH.BERT

@@ -263,6 +265,7 @@ class DistilBertModel(BertModel):


@ModelBase.register("RobertaModel", "RobertaForSequenceClassification")
+@ModelBase.example("sentence-transformers/stsb-roberta-base")
 class RobertaModel(BertModel):
    model_arch = gguf.MODEL_ARCH.BERT

@@ -312,6 +315,7 @@ class RobertaModel(BertModel):


@ModelBase.register("NomicBertModel")
+@ModelBase.example("nomic-ai/nomic-embed-text-v1.5")
 class NomicBertModel(BertModel):
    model_arch = gguf.MODEL_ARCH.BERT

@@ -400,6 +404,7 @@ class NomicBertModel(BertModel):


@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
+@ModelBase.example("chandar-lab/NeoBERT")
 class NeoBert(BertModel):
    model_arch = gguf.MODEL_ARCH.NEO_BERT

@@ -431,6 +436,7 @@ class NeoBert(BertModel):


@ModelBase.register("EuroBertModel", "JinaEmbeddingsV5Model")
+@ModelBase.example("hf-tiny-v2/tiny-random-EuroBertModel", "jinaai/jina-embeddings-v5-text-nano")
 class EuroBertModel(TextModel):
    model_arch = gguf.MODEL_ARCH.EUROBERT

@@ -459,6 +465,7 @@ class EuroBertModel(TextModel):


@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
+@ModelBase.example("BAAI/bge-m3")
 class XLMRobertaModel(BertModel):
    model_arch = gguf.MODEL_ARCH.BERT
    _lora_files = {}
@@ -561,6 +568,7 @@ class XLMRobertaModel(BertModel):


@ModelBase.register("JinaBertModel", "JinaBertForMaskedLM")
+@ModelBase.example("jinaai/jina-embeddings-v2-base-en")
 class JinaBertV2Model(BertModel):
    model_arch = gguf.MODEL_ARCH.JINA_BERT_V2

@@ -588,6 +596,7 @@ class JinaBertV2Model(BertModel):


@ModelBase.register("ModernBertModel", "ModernBertForMaskedLM", "ModernBertForSequenceClassification")
+@ModelBase.example("answerdotai/ModernBERT-base")
 class ModernBertModel(BertModel):
    model_arch = gguf.MODEL_ARCH.MODERN_BERT

@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("BitnetForCausalLM", "BitNetForCausalLM")
+@ModelBase.example("microsoft/bitnet-b1.58-2B-4T")
 class BitnetModel(TextModel):
    model_arch = gguf.MODEL_ARCH.BITNET

@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("BloomForCausalLM", "BloomModel")
+@ModelBase.example("bigscience/bloom-560m")
 class BloomModel(TextModel):
    model_arch = gguf.MODEL_ARCH.BLOOM

@@ -12,6 +12,8 @@ from .llama import LlamaModel

@ModelBase.register("ChameleonForConditionalGeneration")
@ModelBase.register("ChameleonForCausalLM")  # obsolete
+# [TAG_HF_EXAMPLE_GATED] facebook/chameleon-7b is gated
+# [TAG_HF_EXAMPLE_MISSING]
 class ChameleonModel(TextModel):
    model_arch = gguf.MODEL_ARCH.CHAMELEON

@@ -9,6 +9,7 @@ from .base import ModelBase, SentencePieceTokenTypes, TextModel, gguf


@ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
+@ModelBase.example("THUDM/chatglm3-6b", "zai-org/glm-4-9b-chat-hf")
 class ChatGLMModel(TextModel):
    model_arch = gguf.MODEL_ARCH.CHATGLM

@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("CodeShellForCausalLM")
+@ModelBase.example("WisdomShell/CodeShell-7B")
 class CodeShellModel(TextModel):
    model_arch = gguf.MODEL_ARCH.CODESHELL

@@ -11,6 +11,7 @@ from .llama import LlamaModel


@ModelBase.register("CogVLMForCausalLM")
+@ModelBase.example("THUDM/cogvlm2-llama3-chat-19B", "THUDM/cogvlm-chat-hf")
 class CogVLMVisionModel(MmprojModel):

    def set_gguf_parameters(self):
@@ -29,5 +30,6 @@ class CogVLMVisionModel(MmprojModel):


@ModelBase.register("CogVLMForCausalLM")
+@ModelBase.example("THUDM/cogvlm2-llama3-chat-19B", "THUDM/cogvlm-chat-hf")
 class CogVLMModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.COGVLM
@@ -12,6 +12,8 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("CohereForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] CohereLabs/c4ai-command-r-v01 is gated
+# [TAG_HF_EXAMPLE_MISSING]
 class CommandR2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.COMMAND_R

@@ -30,6 +32,8 @@ class CommandR2Model(TextModel):


@ModelBase.register("Cohere2ForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] CohereLabs/c4ai-command-r7b-12-2024 is gated
+@ModelBase.example("hf-tiny-v2/tiny-random-Cohere2ForCausalLM")
 class Cohere2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.COHERE2

@@ -59,6 +63,7 @@ class Cohere2Model(TextModel):


@ModelBase.register("Cohere2MoeForCausalLM")
+@ModelBase.example("CohereLabs/North-Mini-Code-1.0")
 class Cohere2MoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.COHERE2MOE
    _n_main_layers: int | None = None
@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("DbrxForCausalLM")
+@ModelBase.example("alpindale/dbrx-instruct")
 class DbrxModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DBRX

@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("DeciLMForCausalLM")
+@ModelBase.example("nvidia/Llama-3_1-Nemotron-51B-Instruct", "Deci/DeciLM-7B")
 class DeciModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DECI

@@ -18,6 +18,7 @@ from .qwen import QwenModel


@ModelBase.register("DeepseekOCRForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-OCR")
 class DeepseekOCRVisionModel(MmprojModel):
    # HF dynamic_preprocess() max_num, which differs per model
    preproc_max_tiles = 9
@@ -100,11 +101,13 @@ class DeepseekOCRVisionModel(MmprojModel):


@ModelBase.register("UnlimitedOCRForCausalLM")
+@ModelBase.example("baidu/Unlimited-OCR")
 class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
    preproc_max_tiles = 32


@ModelBase.register("DeepseekOCR2ForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-OCR-2")
 class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
    preproc_max_tiles = 6

@@ -134,6 +137,7 @@ class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):


@ModelBase.register("DeepseekForCausalLM")
+@ModelBase.example("deepseek-ai/deepseek-moe-16b-chat")
 class DeepseekModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK

@@ -228,6 +232,7 @@ class DeepseekModel(TextModel):
    "YoutuForCausalLM",
    "YoutuVLForConditionalGeneration",
 )
+@ModelBase.example("deepseek-ai/DeepSeek-V2-Lite", "deepseek-ai/DeepSeek-V3")
 class DeepseekV2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK2

@@ -457,6 +462,7 @@ class DeepseekV2Model(TextModel):


@ModelBase.register("DeepseekV32ForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-V3.2-Exp")
 class DeepseekV32Model(DeepseekV2Model):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK32
    skip_mtp = False
@@ -517,6 +523,7 @@ class DeepseekV32Model(DeepseekV2Model):


@ModelBase.register("DeepseekV4ForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-V4-Flash-Base")
 class DeepseekV4Model(TextModel):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK4
    supports_mtp_export = True
@@ -911,6 +918,7 @@ class DeepseekV4Model(TextModel):


@ModelBase.register("DeepseekV4DSparkModel")
+@ModelBase.example("deepseek-ai/DeepSeek-V4-Flash-DSpark")
 class DeepseekV4DSparkModel(DeepseekV4Model):
    model_arch = gguf.MODEL_ARCH.DFLASH

@@ -11,6 +11,7 @@ from .qwen import Qwen2MoeModel


@ModelBase.register("Dots1ForCausalLM")
+@ModelBase.example("rednote-hilab/dots.llm1.inst")
 class Dots1Model(Qwen2MoeModel):
    model_arch = gguf.MODEL_ARCH.DOTS1

@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf


@ModelBase.register("DotsOCRForCausalLM")
+@ModelBase.example("rednote-hilab/dots.ocr")
 class DotsOCRVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("DreamModel")
+@ModelBase.example("Dream-org/Dream-v0-Instruct-7B")
 class DreamModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DREAM

@@ -15,6 +15,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf


@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")
+@ModelBase.example("baidu/ERNIE-4.5-0.3B-PT")
 class Ernie4_5Model(TextModel):
    model_arch = gguf.MODEL_ARCH.ERNIE4_5

@@ -73,6 +74,7 @@ class Ernie4_5Model(TextModel):


@ModelBase.register("Ernie4_5_MoeForCausalLM")
+@ModelBase.example("baidu/ERNIE-4.5-21B-A3B-PT")
 class Ernie4_5MoeModel(Ernie4_5Model):
    model_arch = gguf.MODEL_ARCH.ERNIE4_5_MOE
    _experts: list[dict[str, Tensor]] | None = None
@@ -156,11 +158,13 @@ class Ernie4_5MoeModel(Ernie4_5Model):


@ModelBase.register("PaddleOCRVLForConditionalGeneration")
+@ModelBase.example("PaddlePaddle/PaddleOCR-VL")
 class PaddleOCRModel(Ernie4_5Model):
    model_arch = gguf.MODEL_ARCH.PADDLEOCR


@ModelBase.register("PaddleOCRVisionModel")
+@ModelBase.example("PaddlePaddle/PaddleOCR-VL")
 class PaddleOCRVisionModel(MmprojModel):
    # PaddleOCR-VL uses a modified version of Siglip
    min_pixels: int = 0
@@ -15,6 +15,7 @@ from .qwenvl import Qwen2VLVisionModel


@ModelBase.register("ExaoneForCausalLM")
+@ModelBase.example("LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct")
 class ExaoneModel(TextModel):
    model_arch = gguf.MODEL_ARCH.EXAONE

@@ -60,6 +61,7 @@ class ExaoneModel(TextModel):


@ModelBase.register("Exaone4ForCausalLM")
+@ModelBase.example("LGAI-EXAONE/EXAONE-4.0-32B")
 class Exaone4Model(TextModel):
    model_arch = gguf.MODEL_ARCH.EXAONE4

@@ -126,6 +128,7 @@ class Exaone4Model(TextModel):
 # note: transformers >= 5.1 renamed the class to "ExaoneMoeForCausalLM" (lowercase 'e'),
 #       so accept both spellings - LG AI have updated the configs of already-released models
@ModelBase.register("ExaoneMoEForCausalLM", "ExaoneMoeForCausalLM")
+@ModelBase.example("LGAI-EXAONE/K-EXAONE-236B-A23B")
 class ExaoneMoEModel(Exaone4Model):
    model_arch = gguf.MODEL_ARCH.EXAONE_MOE

@@ -214,6 +217,7 @@ class ExaoneMoEModel(Exaone4Model):


@ModelBase.register("Exaone4_5_ForConditionalGeneration")
+@ModelBase.example("LGAI-EXAONE/EXAONE-4.5-33B")
 class Exaone4_5_TextModel(Exaone4Model):
    """Text tower of EXAONE 4.5; Tensors match EXAONE4"""

@@ -267,6 +271,7 @@ class Exaone4_5_TextModel(Exaone4Model):


@ModelBase.register("Exaone4_5_ForConditionalGeneration")
+@ModelBase.example("LGAI-EXAONE/EXAONE-4.5-33B")
 class Exaone4_5VisionModel(Qwen2VLVisionModel):
    """Vision tower for EXAONE 4.5; Qwen2-VL-style ViT (GQA) + patch merger"""

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("FalconForCausalLM", "RWForCausalLM")
+@ModelBase.example("tiiuae/falcon-7b")
 class FalconModel(TextModel):
    model_arch = gguf.MODEL_ARCH.FALCON

@@ -12,6 +12,7 @@ from .mamba import Mamba2Model


@ModelBase.register("FalconH1ForCausalLM")
+@ModelBase.example("tiiuae/Falcon-H1-0.5B-Base")
 class FalconH1Model(Mamba2Model):
    model_arch = gguf.MODEL_ARCH.FALCON_H1

@@ -14,6 +14,8 @@ from .base import MmprojModel, ModelBase, TextModel, gguf, logger


@ModelBase.register("GemmaForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-2b is gated
+@ModelBase.example("trl-internal-testing/tiny-GemmaForCausalLM")
 class GemmaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GEMMA

@@ -68,6 +70,8 @@ class GemmaModel(TextModel):


@ModelBase.register("Gemma2ForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-2-9b-it is gated
+@ModelBase.example("trl-internal-testing/tiny-Gemma2ForCausalLM")
 class Gemma2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.GEMMA2

@@ -118,6 +122,8 @@ class Gemma2Model(TextModel):


@ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-3-4b-it is gated
+@ModelBase.example("trl-internal-testing/tiny-Gemma3ForConditionalGeneration", "hf-tiny-v2/tiny-random-Gemma3ForCausalLM")
 class Gemma3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.GEMMA3

@@ -174,6 +180,8 @@ class Gemma3Model(TextModel):


@ModelBase.register("Gemma3TextModel")
+# [TAG_HF_EXAMPLE_GATED] google/embeddinggemma-300m is gated
+@ModelBase.example("hf-tiny-v2/tiny-random-Gemma3TextModel")
 class EmbeddingGemma(Gemma3Model):
    model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDING
    module_paths = []
@@ -248,6 +256,8 @@ class EmbeddingGemma(Gemma3Model):


@ModelBase.register("Gemma3ForConditionalGeneration")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-3-4b-it is gated
+@ModelBase.example("trl-internal-testing/tiny-Gemma3ForConditionalGeneration")
 class Gemma3VisionModel(MmprojModel):
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
@@ -352,6 +362,8 @@ class ConformerAudioModel(MmprojModel):


@ModelBase.register("Gemma3nForConditionalGeneration")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-3n-E2B-it is gated
+@ModelBase.example("hf-tiny-v2/tiny-random-Gemma3nForConditionalGeneration")
 class Gemma3nVisionAudioModel(ConformerAudioModel):
    has_audio_encoder = True
    has_vision_encoder = True
@@ -471,6 +483,8 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):


@ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration")
+# [TAG_HF_EXAMPLE_GATED] google/gemma-3n-E2B-it is gated
+@ModelBase.example("hf-tiny-v2/tiny-random-Gemma3nForConditionalGeneration")
 class Gemma3NModel(Gemma3Model):
    model_arch = gguf.MODEL_ARCH.GEMMA3N

@@ -615,6 +629,7 @@ class Gemma3NModel(Gemma3Model):


@ModelBase.register("Gemma4ForConditionalGeneration", "Gemma4ForCausalLM")
+@ModelBase.example("google/gemma-4-31B-it", "google/gemma-4-26B-A4B-it", "google/gemma-4-E2B-it")
 class Gemma4Model(Gemma3Model):
    model_arch = gguf.MODEL_ARCH.GEMMA4

@@ -795,6 +810,7 @@ class Gemma4Model(Gemma3Model):


@ModelBase.register("Gemma4UnifiedForConditionalGeneration")
+@ModelBase.example("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")
 class Gemma4UnifiedModel(Gemma4Model):
    model_arch = gguf.MODEL_ARCH.GEMMA4

@@ -815,6 +831,7 @@ class Gemma4UnifiedModel(Gemma4Model):


@ModelBase.register("Gemma4AssistantForCausalLM", "Gemma4UnifiedAssistantForCausalLM")
+@ModelBase.example("google/gemma-4-31B-it-assistant", "google/gemma-4-26B-A4B-it-assistant", "google/gemma-4-E2B-it-assistant")
 class Gemma4AssistantModel(Gemma4Model):
    model_arch = gguf.MODEL_ARCH.GEMMA4_ASSISTANT

@@ -835,6 +852,7 @@ class Gemma4AssistantModel(Gemma4Model):


@ModelBase.register("Gemma4ForConditionalGeneration")
+@ModelBase.example("google/gemma-4-31B-it", "google/gemma-4-26B-A4B-it", "google/gemma-4-E2B-it")
 class Gemma4VisionAudioModel(MmprojModel):
    has_audio_encoder = True
    has_vision_encoder = True
@@ -913,6 +931,7 @@ class Gemma4VisionAudioModel(MmprojModel):


@ModelBase.register("Gemma4UnifiedForConditionalGeneration")
+@ModelBase.example("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")
 class Gemma4UnifiedVisionAudioModel(Gemma4VisionAudioModel):
    has_audio_encoder = True
    has_vision_encoder = True
@@ -15,6 +15,7 @@ from .deepseek import DeepseekV2Model


@ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration")
+@ModelBase.example("zai-org/GLM-4-9B-0414")
 class Glm4Model(TextModel):
    model_arch = gguf.MODEL_ARCH.GLM4
    use_mrope = False
@@ -86,6 +87,7 @@ class Glm4Model(TextModel):


@ModelBase.register("GlmOcrForConditionalGeneration")
+@ModelBase.example("zai-org/GLM-OCR")
 class GlmOCRModel(Glm4Model):
    model_arch = gguf.MODEL_ARCH.GLM4
    use_mrope = False
@@ -107,6 +109,7 @@ class GlmOCRModel(Glm4Model):


@ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")
+@ModelBase.example("zai-org/GLM-4.5-Air")
 class Glm4MoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GLM4_MOE

@@ -204,6 +207,7 @@ class Glm4MoeModel(TextModel):


@ModelBase.register("Glm4MoeLiteForCausalLM")
+@ModelBase.example("zai-org/GLM-4.7-Flash")
 class Glm4MoeLiteModel(DeepseekV2Model):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
    skip_mtp = False
@@ -272,6 +276,7 @@ class Glm4MoeLiteModel(DeepseekV2Model):


@ModelBase.register("GlmMoeDsaForCausalLM")
+@ModelBase.example("zai-org/GLM-5.2")
 class GlmMoeDsaModel(DeepseekV2Model):
    model_arch = gguf.MODEL_ARCH.GLM_DSA
    skip_mtp = False
@@ -340,6 +345,7 @@ class GlmMoeDsaModel(DeepseekV2Model):


@ModelBase.register("SolarOpenForCausalLM")
+@ModelBase.example("upstage/Solar-Open-100B")
 class SolarOpenModel(Glm4MoeModel):
    model_arch = gguf.MODEL_ARCH.GLM4_MOE

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("GPT2LMHeadModel")
+@ModelBase.example("openai-community/gpt2")
 class GPT2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.GPT2

@@ -38,6 +39,7 @@ class GPT2Model(TextModel):


@ModelBase.register("RuGPT3XLForCausalLM")
+@ModelBase.example("evilfreelancer/ruGPT3XL")
 class RuGPT3XLModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GPT2

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("GptOssForCausalLM")
+@ModelBase.example("openai/gpt-oss-20b")
 class GptOssModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GPT_OSS

@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("GPTNeoXForCausalLM")
+@ModelBase.example("EleutherAI/pythia-70m")
 class GPTNeoXModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GPTNEOX

@@ -15,6 +15,7 @@ from .mamba import Mamba2Model


@ModelBase.register("GraniteForCausalLM")
+@ModelBase.example("ibm-granite/granite-3.3-2b-instruct")
 class GraniteModel(LlamaModel):
    """Conversion for IBM's GraniteForCausalLM"""
    model_arch = gguf.MODEL_ARCH.GRANITE
@@ -73,7 +74,110 @@ class GraniteModel(LlamaModel):
        return super().filter_tensors(item)


+@ModelBase.register("GraniteSWAForCausalLM")
+class GraniteSWAModel(GraniteModel):
+    """Conversion for IBM's GraniteSWAForCausalLM (interleaved sliding window attention)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if name.endswith("sinks"):
+            name += ".weight"
+
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        """GraniteSWA uses Granite parameters plus sliding window configuration."""
+        super().set_gguf_parameters()
+
+        # Add sliding_window from config
+        sliding_window = self.hparams.get("sliding_window", 128)
+        self.gguf_writer.add_sliding_window(sliding_window)
+        logger.info("gguf: (granite_swa) sliding_window = %s", sliding_window)
+
+        # Derive sliding_window_pattern from layer_types
+        if layer_types := self.hparams.get("layer_types"):
+            is_swa = [t == "sliding_attention" for t in layer_types]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern = %d SWA layers / %d total",
+                        sum(is_swa), len(is_swa))
+        else:
+            # Fall back to period-based pattern: i % 4 != 0
+            # This matches the transformers default pattern
+            n_layers = self.block_count
+            is_swa = [i % 4 != 0 for i in range(n_layers)]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern (inferred) = %d SWA layers / %d total",
+                        sum(is_swa), n_layers)
+
+        # Add rope_pattern from no_rope_layers
+        if no_rope_layers := self.hparams.get("no_rope_layers"):
+            # Convert 1/0 to bool (1 = use RoPE, 0 = NoPE)
+            rope_pattern = [bool(x) for x in no_rope_layers]
+            self.gguf_writer.add_rope_pattern(rope_pattern)
+            logger.info("gguf: (granite_swa) rope_pattern = %d RoPE layers / %d total",
+                        sum(rope_pattern), len(rope_pattern))
+
+
+@ModelBase.register("GraniteMoeSWAForCausalLM")
+class GraniteMoeSWAModel(GraniteSWAModel):
+    """Conversion for IBM's GraniteMoeSWAForCausalLM (unified dense + MoE with iSWA)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if shared_intermediate_size := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_intermediate_size)
+            logger.info("gguf: (granitemoewa) shared_intermediate_size = %s", shared_intermediate_size)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """Split merged MoE tensors (gate+up) following standard MoE pattern."""
+
+        # Handle expert FFN tensors (merged gate+up) - swash format: experts.gate_up_proj
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.experts.gate_up_proj"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, f"Merged FFN tensor size must be 2 * intermediate_size, got {data_torch.shape[-2]}"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle expert FFN down projection - swash format: experts.down_proj
+        if name.endswith("block_sparse_moe.experts.down_proj"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), bid)
+            return
+
+        # Handle expert FFN tensors (merged gate+up) - standard granite format: input_linear.weight
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle shared expert FFN tensors (if present) - kept fused since
+        # inference (build_ffn) supports a single ffn_up_shexp tensor with
+        # LLM_FFN_SWIGLU for the shared expert.
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams.get("shared_intermediate_size", self.hparams["intermediate_size"])
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
+            return
+
+        # Handle shared expert output (if present)
+        if name.endswith("shared_mlp.output_linear.weight"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, bid), bid)
+            return
+
+        # Pass through to parent for all other tensors (including sinks)
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
+@ModelBase.example("ibm-granite/granite-3.1-3b-a800m-instruct")
 class GraniteMoeModel(GraniteModel):
    """Conversion for IBM's GraniteMoeForCausalLM"""
    model_arch = gguf.MODEL_ARCH.GRANITE_MOE
@@ -124,6 +228,7 @@ class GraniteMoeModel(GraniteModel):


@ModelBase.register("GraniteSwitchForCausalLM")
+@ModelBase.example("ibm-granite/granite-switch-4.1-3b-preview")
 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)."""
@@ -284,6 +389,7 @@ class GraniteSwitchModel(GraniteMoeModel):


@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
+@ModelBase.example("ibm-granite/granite-4.0-h-tiny", "ibm-ai-platform/Bamba-9B-v2")
 class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
    """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
    layers and optionally uses MoE w/ a shared expert"""
@@ -426,6 +532,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):


@ModelBase.register("GraniteSpeechForConditionalGeneration")
+@ModelBase.example("ibm-granite/granite-speech-3.3-2b", "ibm-granite/granite-4.0-1b-speech")
 class GraniteSpeechMmprojModel(MmprojModel):
    has_vision_encoder = False
    has_audio_encoder = True
@@ -509,6 +616,7 @@ class GraniteSpeechMmprojModel(MmprojModel):


@ModelBase.register("GraniteSpeechPlusForConditionalGeneration")
+@ModelBase.example("ibm-granite/granite-speech-4.1-2b-plus")
 class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):
    """Conversion for GraniteSpeechPlus - extends GraniteSpeech with feature layer concatenation"""
    has_vision_encoder = False
@@ -537,6 +645,7 @@ class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):


@ModelBase.register("Granite4VisionForConditionalGeneration")
+@ModelBase.example("ibm-granite/granite-4.0-3b-vision")
 class Granite4VisionMmprojModel(MmprojModel):
    has_vision_encoder = True
    has_audio_encoder = False
@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")
+@ModelBase.example("keyfan/grok-1-hf")
 class GrokModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GROK

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
+@ModelBase.example("inclusionAI/GroveMoE-Inst")
 class GroveMoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GROVEMOE

@@ -17,6 +17,7 @@ from .qwen import QwenModel


@ModelBase.register("HunYuanMoEV1ForCausalLM")
+@ModelBase.example("tencent/Hunyuan-A13B-Instruct")
 class HunYuanMoEModel(TextModel):
    model_arch = gguf.MODEL_ARCH.HUNYUAN_MOE

@@ -154,6 +155,7 @@ class HunYuanMoEModel(TextModel):


@ModelBase.register("HunYuanDenseV1ForCausalLM")
+@ModelBase.example("tencent/Hunyuan-4B-Instruct")
 class HunYuanModel(TextModel):
    model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE

@@ -290,6 +292,7 @@ class HunYuanModel(TextModel):


@ModelBase.register("HunYuanVLForConditionalGeneration")
+@ModelBase.example("tencent/HunyuanOCR")
 class HunyuanVLVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -333,6 +336,7 @@ class HunyuanVLVisionModel(MmprojModel):


@ModelBase.register("HunYuanVLForConditionalGeneration")
+@ModelBase.example("tencent/HunyuanOCR")
 class HunyuanVLTextModel(HunYuanModel):
    model_arch = gguf.MODEL_ARCH.HUNYUAN_VL

@@ -365,6 +369,7 @@ class HunyuanVLTextModel(HunYuanModel):


@ModelBase.register("HYV3ForCausalLM")
+@ModelBase.example("tencent/Hy3")
 class HYV3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.HY_V3
    supports_mtp_export = True
@@ -14,6 +14,7 @@ from .llama import LlamaModel


@ModelBase.register("InternLM2ForCausalLM")
+@ModelBase.example("internlm/internlm2-chat-7b")
 class InternLM2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.INTERNLM2

@@ -170,6 +171,7 @@ class InternLM2Model(TextModel):


@ModelBase.register("InternLM3ForCausalLM")
+@ModelBase.example("internlm/internlm3-8b-instruct")
 class InternLM3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.LLAMA

@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf


@ModelBase.register("InternVisionModel")
+@ModelBase.example("OpenGVLab/InternVL3-2B", "OpenGVLab/InternVL2_5-1B")
 class InternVisionModel(MmprojModel):

    min_dynamic_tiles: int = 0
@@ -11,6 +11,8 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("Jais2ForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] inceptionai/Jais-2-8B-Chat is gated
+# [TAG_HF_EXAMPLE_MISSING]
 class Jais2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.JAIS2

@@ -22,6 +24,7 @@ class Jais2Model(TextModel):


@ModelBase.register("JAISLMHeadModel")
+@ModelBase.example("inceptionai/jais-family-590m")
 class JaisModel(TextModel):
    model_arch = gguf.MODEL_ARCH.JAIS

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("JambaForCausalLM")
+@ModelBase.example("ai21labs/Jamba-v0.1")
 class JambaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.JAMBA

@@ -11,6 +11,7 @@ from .llama import LlamaModel


@ModelBase.register("JanusForConditionalGeneration")
+@ModelBase.example("deepseek-community/Janus-Pro-1B")
 class JanusProModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA  # reuse Llama arch

@@ -34,6 +35,7 @@ class JanusProModel(LlamaModel):


@ModelBase.register("JanusForConditionalGeneration")
+@ModelBase.example("deepseek-community/Janus-Pro-1B")
 class JanusProVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -16,6 +16,7 @@ from .kimi_linear import KimiLinearModel


@ModelBase.register("KimiK3ForConditionalGeneration")
+@ModelBase.example("moonshotai/Kimi-K3")
 class KimiK3Model(TextModel):
    """
    Kimi-K3 text model (KimiLinearForCausalLM under a `language_model.` prefix).
@@ -13,6 +13,7 @@ from .qwen import QwenModel


@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
+@ModelBase.example("moonshotai/Kimi-Linear-48B-A3B-Instruct")
 class KimiLinearModel(TextModel):
    """Kimi-Linear model with hybrid MLA+KDA architecture"""
    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
@@ -11,6 +11,7 @@ from .base import MmprojModel, ModelBase, gguf


@ModelBase.register("KimiVLForConditionalGeneration")
+@ModelBase.example("moonshotai/Kimi-VL-A3B-Instruct")
 class KimiVLModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -52,6 +53,7 @@ class KimiVLModel(MmprojModel):


@ModelBase.register("KimiK25ForConditionalGeneration")
+@ModelBase.example("moonshotai/Kimi-K2.5")
 class KimiK25Model(MmprojModel):
    """Kimi-K2.5 with MoonViT3d vision encoder"""

@@ -155,6 +157,7 @@ class KimiK25Model(MmprojModel):


@ModelBase.register("Glm5vForConditionalGeneration")
+# [TAG_HF_EXAMPLE_MISSING]
 class Glm5vModel(KimiK25Model):
    """GLM-5.2-Vision MoonViT3d encoder and projector

@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("LagunaForCausalLM")
+@ModelBase.example("poolside/Laguna-XS.2", "poolside/Laguna-S-2.1")
 class LagunaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LAGUNA
    _experts: list[dict] | None = None
@@ -13,6 +13,7 @@ from .gemma import ConformerAudioModel


@ModelBase.register("Lfm2ForCausalLM", "LFM2ForCausalLM")
+@ModelBase.example("LiquidAI/LFM2-1.2B", "LiquidAI/LFM2.5-350M")
 class LFM2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.LFM2

@@ -65,6 +66,7 @@ class LFM2Model(TextModel):


@ModelBase.register("Lfm2Model", "Lfm2BidirectionalModel")
+@ModelBase.example("LiquidAI/LFM2.5-ColBERT-350M", "LiquidAI/LFM2.5-Embedding-350M")
 class LFM2ColBertModel(LFM2Model):
    model_arch = gguf.MODEL_ARCH.LFM2
    dense_tensor_name = "dense_2"
@@ -93,6 +95,7 @@ class LFM2ColBertModel(LFM2Model):


@ModelBase.register("Lfm2MoeForCausalLM")
+@ModelBase.example("LiquidAI/LFM2-8B-A1B")
 class LFM2MoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LFM2MOE

@@ -166,6 +169,7 @@ class LFM2MoeModel(TextModel):


@ModelBase.register("Lfm2VlForConditionalGeneration")
+@ModelBase.example("LiquidAI/LFM2-VL-450M")
 class LFM2VLModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -200,6 +204,7 @@ class LFM2VLModel(MmprojModel):


@ModelBase.register("Lfm2AudioForConditionalGeneration")
+@ModelBase.example("LiquidAI/LFM2.5-Audio-1.5B", "LiquidAI/LFM2-Audio-1.5B")
 class LFM2AudioModel(ConformerAudioModel):
    has_vision_encoder = False
    has_audio_encoder = True
@@ -238,6 +243,7 @@ class LFM2AudioModel(ConformerAudioModel):


@ModelBase.register("Lfm25AudioTokenizer")
+@ModelBase.example("LiquidAI/LFM2.5-Audio-1.5B")
 class LFM25AudioTokenizer(LFM2Model):
    model_arch = gguf.MODEL_ARCH.LFM2

@@ -11,6 +11,7 @@ from .llava import LlavaVisionModel


@ModelBase.register("LightOnOCRForConditionalGeneration")
+@ModelBase.example("lightonai/LightOnOCR-1B-1025")
 class LightOnOCRVisionModel(LlavaVisionModel):
    is_mistral_format = False
    use_break_tok = False
@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("LLaDAModelLM")
+@ModelBase.example("GSAI-ML/LLaDA-8B-Instruct")
 class LLaDAModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LLADA
    undo_permute = True
@@ -114,6 +115,7 @@ class LLaDAModel(TextModel):


@ModelBase.register("LLaDAMoEModel", "LLaDAMoEModelLM")
+@ModelBase.example("inclusionAI/LLaDA-MoE-7B-A1B-Instruct")
 class LLaDAMoEModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LLADA_MOE

@@ -28,6 +28,8 @@ from .base import ModelBase, TextModel, gguf, logger
    "Eagle3DraftModel",
    "IQuestCoderForCausalLM",
    "LlamaModel")
+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-3.2-1B-Instruct is gated
+@ModelBase.example("unsloth/Llama-3.2-1B-Instruct", "mistralai/Mistral-7B-Instruct-v0.3", "mistralai/Mixtral-8x7B-Instruct-v0.1")
 class LlamaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LLAMA
    undo_permute = True
@@ -359,6 +361,7 @@ class LlamaModel(TextModel):


@ModelBase.register("ArceeForCausalLM")
+@ModelBase.example("arcee-ai/AFM-4.5B")
 class ArceeModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.ARCEE

@@ -371,6 +374,8 @@ class ArceeModel(LlamaModel):
    "Llama4ForConditionalGeneration",
    "Llama4ForCausalLM",
 )
+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated
+@ModelBase.example("unsloth/Llama-4-Scout-17B-16E-Instruct")
 class Llama4Model(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA4
    undo_permute = False
@@ -412,16 +417,19 @@ class Llama4Model(LlamaModel):


@ModelBase.register("LlamaBidirectionalModel")
+@ModelBase.example("nvidia/llama-embed-nemotron-8b")
 class LlamaEmbedNemotronModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA_EMBED


@ModelBase.register("SmolLM3ForCausalLM")
+@ModelBase.example("HuggingFaceTB/SmolLM3-3B")
 class SmolLM3Model(LlamaModel):
    model_arch = gguf.MODEL_ARCH.SMOLLM3


@ModelBase.register("ApertusForCausalLM")
+@ModelBase.example("swiss-ai/Apertus-8B-Instruct-2509")
 class ApertusModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.APERTUS
    undo_permute = False
@@ -9,6 +9,8 @@ from .base import MmprojModel, ModelBase, gguf


@ModelBase.register("Llama4ForConditionalGeneration")
+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated
+@ModelBase.example("unsloth/Llama-4-Scout-17B-16E-Instruct")
 class Llama4VisionModel(MmprojModel):
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
@@ -16,6 +16,7 @@ from .llama import LlamaModel
    "LlavaForConditionalGeneration", # pixtral
    "Mistral3ForConditionalGeneration", # mistral small 3.1
 )
+@ModelBase.example("mistral-community/pixtral-12b", "mistralai/Mistral-Small-3.1-24B-Instruct-2503")
 class LlavaVisionModel(MmprojModel):
    img_break_tok_id = -1
    use_break_tok = True
@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("MaincoderForCausalLM")
+@ModelBase.example("Maincode/Maincoder-1B")
 class MaincoderModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MAINCODER

@@ -14,6 +14,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
+@ModelBase.example("state-spaces/mamba-130m-hf", "tiiuae/falcon-mamba-7b")
 class MambaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MAMBA

@@ -100,6 +101,7 @@ class MambaModel(TextModel):


@ModelBase.register("Mamba2ForCausalLM")
+@ModelBase.example("mistralai/Mamba-Codestral-7B-v0.1")
 class Mamba2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MAMBA2

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("MellumForCausalLM")
+@ModelBase.example("JetBrains/Mellum2-12B-A2.5B-Base")
 class MellumModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MELLUM

@@ -14,6 +14,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf


@ModelBase.register("MiMoV2FlashForCausalLM", "MiMoV2ForCausalLM")
+@ModelBase.example("XiaomiMiMo/MiMo-V2.5")
 class MimoV2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MIMO2

@@ -230,6 +231,7 @@ class MimoV2Model(TextModel):


@ModelBase.register("MiMoV2ForCausalLM")
+@ModelBase.example("XiaomiMiMo/MiMo-V2.5")
 class MiMoV2VisionAudioModel(MmprojModel):
    has_audio_encoder = True

@@ -14,6 +14,7 @@ from .qwen import Qwen3_5TextModel


@ModelBase.register("MiniCPMForCausalLM")
+@ModelBase.example("openbmb/MiniCPM-2B-sft-bf16")
 class MiniCPMModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MINICPM

@@ -61,6 +62,7 @@ class MiniCPMModel(TextModel):


@ModelBase.register("MiniCPM3ForCausalLM")
+@ModelBase.example("openbmb/MiniCPM3-4B")
 class MiniCPM3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MINICPM3

@@ -117,6 +119,7 @@ class MiniCPM3Model(TextModel):
 # the LM (text mode) and once as the mmproj (vision mode), mirroring the Qwen3-VL setup.

@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")
+@ModelBase.example("openbmb/MiniCPM-V-4_6")
 class MiniCPMV4_6TextModel(Qwen3_5TextModel):
    model_arch = gguf.MODEL_ARCH.QWEN35

@@ -134,6 +137,7 @@ class MiniCPMV4_6TextModel(Qwen3_5TextModel):


@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")
+@ModelBase.example("openbmb/MiniCPM-V-4_6")
 class MiniCPMV4_6VisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -12,6 +12,7 @@ from .base import ModelBase, TextModel, MmprojModel, gguf, logger

@ModelBase.register("MiniMaxText01ForCausalLM")
@ModelBase.register("MiniMaxM1ForCausalLM")
+@ModelBase.example("MiniMaxAI/MiniMax-Text-01", "MiniMaxAI/MiniMax-M1-40k")
 class MiniMaxText01Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MINIMAX01

@@ -119,6 +120,7 @@ class MiniMaxText01Model(TextModel):


@ModelBase.register("MiniMaxM2ForCausalLM")
+@ModelBase.example("MiniMaxAI/MiniMax-M2")
 class MiniMaxM2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MINIMAXM2
    _experts_cache: dict[int, dict[str, Tensor]] = {}
@@ -163,6 +165,7 @@ class MiniMaxM2Model(TextModel):


@ModelBase.register("MiniMaxM3SparseForCausalLM", "MiniMaxM3SparseForConditionalGeneration")
+@ModelBase.example("MiniMaxAI/MiniMax-M3")
 class MiniMaxM3Model(MiniMaxM2Model):
    model_arch = gguf.MODEL_ARCH.MINIMAXM3

@@ -203,6 +206,7 @@ class MiniMaxM3Model(MiniMaxM2Model):


@ModelBase.register("MiniMaxM3SparseForConditionalGeneration", "MiniMaxM3VLForConditionalGeneration")
+@ModelBase.example("MiniMaxAI/MiniMax-M3")
 class MiniMaxM3VisionModel(MmprojModel):
    @classmethod
    def filter_tensors(cls, item):
@@ -15,6 +15,7 @@ from .llama import LlamaModel
    "Mistral3ForConditionalGeneration",
    "Ministral3ForCausalLM",
 )
+@ModelBase.example("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM")
 class Mistral3Model(TextModel):
    class Ministral3Model(LlamaModel):
        model_arch = gguf.MODEL_ARCH.MISTRAL3
@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("MPTForCausalLM")
+@ModelBase.example("anas-awadalla/mpt-7b")
 class MPTModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MPT

@@ -24,6 +24,7 @@ def _unpermute_for_rope(tensor: "Tensor", n_heads: int) -> "Tensor":


@ModelBase.register("MuseGlimmerForConditionalGeneration")
+@ModelBase.example("meta-models/Muse-Glimmer-30B")
 class MuseGlimmerModel(TextModel):
    model_arch = gguf.MODEL_ARCH.MUSE_GLIMMER

@@ -78,6 +79,7 @@ class MuseGlimmerModel(TextModel):


@ModelBase.register("MuseGlimmerForConditionalGeneration")
+@ModelBase.example("meta-models/Muse-Glimmer-30B")
 class MuseGlimmerVisionModel(MmprojModel):
    def get_vision_config(self) -> dict[str, Any] | None:
        c = self.global_config.get("vision_config")
@@ -131,6 +133,7 @@ class MuseGlimmerVisionModel(MmprojModel):


@ModelBase.register("MuseGlimmerAssistantModel")
+@ModelBase.example("meta-models/Muse-Glimmer-30B-assistant")
 class MuseGlimmerAssistantModel(TextModel):
    model_arch = gguf.MODEL_ARCH.DFLASH

@@ -5,6 +5,7 @@ from .llama import LlamaModel


@ModelBase.register("NanbeigeForCausalLM")
+@ModelBase.example("Nanbeige/Nanbeige4.2-3B")
 class NanbeigeModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.NANBEIGE
    undo_permute = True
@@ -16,6 +16,7 @@ from .granite import GraniteHybridModel
    "NemotronH_Nano_VL_V2",
    "RADIOModel",
 )
+@ModelBase.example("nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16")
 class NemotronNanoV2VLModel(MmprojModel):
    # ViT-Huge architecture parameters for RADIO v2.5-h
    _vit_hidden_size = 1280
@@ -151,6 +152,7 @@ class NemotronNanoV2VLModel(MmprojModel):


@ModelBase.register("NemotronForCausalLM")
+@ModelBase.example("nvidia/Minitron-4B-Base")
 class NemotronModel(TextModel):
    model_arch = gguf.MODEL_ARCH.NEMOTRON

@@ -193,6 +195,7 @@ class NemotronModel(TextModel):


@ModelBase.register("NemotronHForCausalLM")
+@ModelBase.example("nvidia/Nemotron-H-8B-Base-8K")
 class NemotronHModel(GraniteHybridModel):
    """Hybrid mamba2/attention model from NVIDIA"""
    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
@@ -204,7 +207,9 @@ class NemotronHModel(GraniteHybridModel):
        # calling the parent __init__. This is because the parent constructor
        # uses self.model_arch to build the tensor name map, and all MoE-specific
        # mappings would be missed if it were called with the default non-MoE arch.
-        hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
        has_moe_params = (
            "num_experts_per_tok" in hparams
            or (isinstance(hparams.get("llm_config"), dict) and "num_experts_per_tok" in hparams["llm_config"])
@@ -212,8 +217,11 @@ class NemotronHModel(GraniteHybridModel):
        if has_moe_params:
            self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
            self.is_moe = True
+        layers_block_type = hparams.get("layers_block_type")
+        if layers_block_type is not None:
+            hparams["num_hidden_layers"] = len(layers_block_type)

-        super().__init__(*args, **kwargs)
+        super().__init__(*args, hparams=hparams, **kwargs)

        # Save the top-level head_dim for later
        self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))
@@ -14,6 +14,7 @@ from .llama import LlamaModel

@ModelBase.register("OlmoForCausalLM")
@ModelBase.register("OLMoForCausalLM")
+@ModelBase.example("allenai/OLMo-1.7-7B-hf")
 class OlmoModel(TextModel):
    model_arch = gguf.MODEL_ARCH.OLMO

@@ -39,12 +40,14 @@ class OlmoModel(TextModel):


@ModelBase.register("SeedOssForCausalLM")
+@ModelBase.example("ByteDance-Seed/Seed-OSS-36B-Instruct")
 class SeedOssModel(TextModel):
    model_arch = gguf.MODEL_ARCH.SEED_OSS


@ModelBase.register("Olmo2ForCausalLM")
@ModelBase.register("Olmo3ForCausalLM")
+@ModelBase.example("allenai/OLMo-2-1124-7B-Instruct", "allenai/Olmo-3-7B-Instruct")
 class Olmo2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.OLMO2

@@ -67,6 +70,7 @@ class Olmo2Model(TextModel):


@ModelBase.register("OlmoeForCausalLM")
+@ModelBase.example("allenai/OLMoE-1B-7B-0924")
 class OlmoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.OLMOE

@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("OpenELMForCausalLM")
+@ModelBase.example("apple/OpenELM-270M")
 class OpenELMModel(TextModel):
    model_arch = gguf.MODEL_ARCH.OPENELM

@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("OrionForCausalLM")
+@ModelBase.example("OrionStarAI/Orion-14B-Base")
 class OrionModel(TextModel):
    model_arch = gguf.MODEL_ARCH.ORION

@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger


@ModelBase.register("PanguEmbeddedForCausalLM")
+@ModelBase.example("FreedomIntelligence/openPangu-Embedded-7B-V1.1")
 class PanguEmbeddedModel(TextModel):
    model_arch = gguf.MODEL_ARCH.PANGU_EMBED

@@ -14,6 +14,7 @@ from .base import MmprojModel, ModelBase, SentencePieceTokenTypes, TextModel, gg


@ModelBase.register("PhiForCausalLM")
+@ModelBase.example("microsoft/phi-2")
 class Phi2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.PHI2

@@ -36,6 +37,7 @@ class Phi2Model(TextModel):


@ModelBase.register("Phi3ForCausalLM", "Phi4ForCausalLMV")
+@ModelBase.example("microsoft/Phi-3-mini-4k-instruct")
 class Phi3MiniModel(TextModel):
    model_arch = gguf.MODEL_ARCH.PHI3

@@ -210,6 +212,7 @@ class Phi3MiniModel(TextModel):


@ModelBase.register("Phi4ForCausalLMV")
+# [TAG_HF_EXAMPLE_MISSING]
 class Phi4VisionMmprojModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -336,6 +339,7 @@ class Phi4VisionMmprojModel(MmprojModel):


@ModelBase.register("PhiMoEForCausalLM")
+@ModelBase.example("microsoft/Phi-3.5-MoE-instruct")
 class PhiMoeModel(Phi3MiniModel):
    model_arch = gguf.MODEL_ARCH.PHIMOE

@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("PlamoForCausalLM")
+@ModelBase.example("pfnet/plamo-13b")
 class PlamoModel(TextModel):
    model_arch = gguf.MODEL_ARCH.PLAMO

@@ -58,6 +59,7 @@ class PlamoModel(TextModel):


@ModelBase.register("Plamo2ForCausalLM", "PLaMo2ForCausalLM")
+@ModelBase.example("pfnet/plamo-2-1b")
 class Plamo2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.PLAMO2

@@ -147,6 +149,8 @@ class Plamo2Model(TextModel):


@ModelBase.register("Plamo3ForCausalLM", "PLaMo3ForCausalLM")
+# [TAG_HF_EXAMPLE_GATED] pfnet/plamo-3-nict-2b-base is gated
+@ModelBase.example("midorin-Linux/plamo-3-12b-self-merged-base")
 class Plamo3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.PLAMO3

@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf


@ModelBase.register("PLMForCausalLM")
+@ModelBase.example("PLM-Team/PLM-1.8B-Instruct")
 class PLMModel(TextModel):
    model_arch = gguf.MODEL_ARCH.PLM

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