server: (router) do not evict busy models

.devops/openvino.Dockerfile

@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.3
-ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
+ARG OPENVINO_VERSION_MAJOR=2026.2.1
+ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-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 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 IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-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
+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
 
 # 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 "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
     fi; \
     mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
     apt-get update; \

.devops/rocm.Dockerfile

@@ -57,6 +57,7 @@ COPY --from=web /app/tools/ui/dist tools/ui/dist
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
     cmake -S . -B build \
         -DGGML_HIP=ON \
+        -DGGML_HIP_ROCWMMA_FATTN=ON \
         -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
         -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
         -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \

.github/actions/ccache-clear/action.yml

@@ -1,88 +1,22 @@
-# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
-description: "Delete GitHub Actions caches matching a key prefix, oldest first"
+description: "Delete all GitHub Actions caches matching a key prefix"
 inputs:
   key:
     description: "Cache key prefix to match and delete"
     required: true
-  older:
-    description: "Only delete caches created more than this long ago (e.g. 90m, 1h, 1d). By default all matching caches are deleted"
-    required: false
-    default: ""
-  min:
-    description: "Stop deleting if fewer than this many caches would remain (e.g. 1). By default there is no minimum"
-    required: false
-    default: "0"
-  dry-run:
-    description: "Only print the caches that would be deleted, without deleting them"
-    required: false
-    default: "false"
 
 runs:
   using: "composite"
   steps:
     - name: Clear caches
       shell: bash
-      env:
-        CLEAR_KEY: ${{ inputs.key }}
-        CLEAR_OLDER: ${{ inputs.older }}
-        CLEAR_MIN: ${{ inputs.min }}
-        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        # Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
-        to_seconds() {
-          local val="$1"
-          [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
-          local num="${val%?}" unit="${val: -1}" mult
-          [[ "$num" =~ ^[0-9]+$ ]] || return 1
-          case "$unit" in
-            s) mult=1 ;;
-            m) mult=60 ;;
-            h) mult=3600 ;;
-            d) mult=86400 ;;
-            *) return 1 ;;
-          esac
-          echo $((num * mult))
-        }
-
-        [[ "$CLEAR_MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $CLEAR_MIN" >&2; exit 1; }
-        [[ "$CLEAR_DRY_RUN" =~ ^(true|false)$ ]] || { echo "Invalid dry-run value: $CLEAR_DRY_RUN" >&2; exit 1; }
-
-        CACHES=$(gh cache list --key "ccache-$CLEAR_KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' 2>/dev/null | LC_ALL=C sort)
+        CACHES=$(gh cache list --key "ccache-${{ inputs.key }}" --json id,key --jq '.[] | "\(.id) \(.key)"' 2>/dev/null)
         if [ -z "$CACHES" ]; then
-          echo "No caches found with key prefix: $CLEAR_KEY"
+          echo "No caches found with key prefix: ${{ inputs.key }}"
           exit 0
         fi
-
-        TOTAL=$(( $(wc -l <<< "$CACHES") ))
-
-        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
-        done <<< "$CACHES"
-
-        CUTOFF=""
-        if [ -n "$CLEAR_OLDER" ]; then
-          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
-          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
-        fi
-
-        # Caches are sorted oldest first
-        DELETED=0
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
-            echo "Rest are not older than $CLEAR_OLDER, stopping"
-            break
-          fi
-          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
-            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
-            break
-          fi
-          if [ "$CLEAR_DRY_RUN" = "true" ]; then
-            echo "Would delete cache: $ID ($KEY)"
-          else
-            echo "Deleting cache: $ID ($KEY)"
-            gh cache delete "$ID"
-          fi
-          DELETED=$((DELETED + 1))
+        while read -r id key; do
+          echo "Deleting cache: $id ($key)"
+          gh cache delete "$id"
         done <<< "$CACHES"

.github/actions/linux-setup-vulkan/action.yml

@@ -0,0 +1,20 @@
+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

.github/actions/windows-setup-cuda/action.yml

@@ -4,9 +4,6 @@ inputs:
   cuda_version:
     description: "CUDA toolkit version"
     required: true
-  cuda_arch:
-    description: "CUDA target architecture"
-    required: true
 
 runs:
   using: "composite"
@@ -130,26 +127,3 @@ runs:
           echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
           echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-    - name: Install Cuda Toolkit 13.4 for ARM64
-      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          choco install unzip -y
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cccl-windows-x86_64-13.3.4.1.2-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_crt-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_nvcc-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/libnvvm-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_cudart-windows-arm64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/libcublas-windows-arm64-13.7.0.10-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.1.2-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.10-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

.github/actions/windows-setup-rocm/action.yml

@@ -8,26 +8,8 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Install ROCm with Wheels
-      shell: pwsh
-      run: |
-        $ErrorActionPreference = "Stop"
-        write-host "Setting up Python virtual environment"
-
-        # Create the venv directly at the cache location to avoid relocation issues
-        New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
-        python -m venv C:\TheRock\build\.venv
-        & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-        write-host "Upgrading pip"
-        python -m pip install --upgrade pip
-
-        write-host "Installing ROCm wheels for multi-arch support"
-        # Install ROCm wheels for multi-arch support (this may take several minutes)
-        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
-
-        # Pre-expand the devel tree so it is included in the cache
-        write-host "Initializing ROCm devel tree"
-        rocm-sdk init
-        if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-        write-host "Completed ROCm wheel installation to C:\TheRock\build"
+    - name: Setup ROCm
+      uses: ./.github/actions/install-exe
+      with:
+        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
+        args: -install

.github/workflows/build-cache.yml

@@ -10,6 +10,33 @@ 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
 
@@ -40,9 +67,9 @@ jobs:
     runs-on: ubuntu-24.04
 
     env:
-      # 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"
+      # 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"
 
     steps:
       - name: Clone
@@ -69,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build.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"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone
@@ -92,27 +119,27 @@ jobs:
           version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
           version_full: ${{ env.OPENVINO_VERSION_FULL }}
 
-  # windows-2022-rocm-cache:
-  #   runs-on: windows-2022
+  windows-2022-rocm-cache:
+    runs-on: windows-2022
 
-  #   env:
-  #     # Make sure this is in sync with release.yml and build-cuda-windows.yml
-  #     ROCM_VERSION: "7.14.0"
+    env:
+      # Make sure this is in sync with build.yml
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
-  #   steps:
-  #     - name: Clone
-  #       id: checkout
-  #       uses: actions/checkout@v6
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
 
-  #     - name: Setup Cache
-  #       uses: actions/cache@v5
-  #       id: cache-rocm
-  #       with:
-  #         path: C:\TheRock\build
-  #         key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+      - name: Setup Cache
+        uses: actions/cache@v5
+        id: cache-rocm
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
-  #     - name: Setup ROCm
-  #       if: steps.cache-rocm.outputs.cache-hit != 'true'
-  #       uses: ./.github/actions/windows-setup-rocm
-  #       with:
-  #         version: ${{ env.ROCM_VERSION }}
+      - name: Setup ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}

.github/workflows/build-cmake-pkg.yml

@@ -5,7 +5,7 @@ on:
 
 jobs:
   linux:
-    runs-on: [self-hosted, Linux]
+    runs-on: [self-hosted, Linux, CPU]
     steps:
       - uses: actions/checkout@v6
         with:
@@ -21,33 +21,31 @@ jobs:
                 -DLLAMA_BUILD_TOOLS=OFF \
                 -DLLAMA_BUILD_EXAMPLES=OFF \
                 -DLLAMA_BUILD_APP=OFF \
-                -DLLAMA_BUILD_IS_DEV=OFF \
                 -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release -j $(nproc)
+          cmake --build build --config Release
           cmake --install build --prefix "$PREFIX" --config Release
 
           export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          build_commit=$(git rev-parse --short HEAD | xargs)
-          build_number=$(git rev-list --count HEAD | xargs)
+          tclsh <<'EOF'
+          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
+          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          set build(version) "0.0.$build(number)"
 
-          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\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")
-
-          for check in "${checks[@]}"; do
-              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
-                  echo "Checking llama-config.cmake version... \"$check\" failed!"
+          puts -nonewline "Checking llama-config.cmake version... "
+          foreach check $checks {
+              if {![regexp -expanded -- $check $llamaconfig]} {
+                  puts "\"$check\" failed!"
                   exit 1
-              fi
-          done
-
-          echo "Checking llama-config.cmake version... success."
+              }
+          }
+          puts "success."
+          EOF
 
           cd examples/simple-cmake-pkg
           cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build -j $(nproc)
+          cmake --build build

.github/workflows/build-cpu.yml

@@ -21,7 +21,6 @@ on:
     paths: [
       '.github/workflows/build-cpu.yml',
       '.github/workflows/build-cmake-pkg.yml',
-      'ggml/src/ggml-rpc/**',
       '**/CMakeLists.txt',
       '**/.cmake',
       '**/*.h',
@@ -95,7 +94,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DGGML_NATIVE=OFF \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_RPC=ON
           time cmake --build build --config Release -j $(nproc)
@@ -117,38 +115,29 @@ jobs:
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
-      # note: real deletion only on push to master (same condition as the ccache save),
-      #       dry-run otherwise (the token is read-only on PRs from forks)
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-${{ matrix.os }}
-          older: 1h
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   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 -DGGML_OPENMP_FETCH=ON -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 -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 -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
 
     steps:
       - name: Clone
@@ -176,6 +165,15 @@ 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: |

.github/workflows/build-cuda-ubuntu.yml

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

.github/workflows/build-cuda-windows.yml

@@ -22,7 +22,6 @@ env:
 
 jobs:
   cuda:
-    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     runs-on: windows-2022
 
     permissions:
@@ -30,16 +29,7 @@ jobs:
 
     strategy:
       matrix:
-        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'
+        cuda: ['12.4', '13.3']
 
     steps:
       - name: Clone
@@ -49,13 +39,12 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-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
@@ -65,21 +54,26 @@ 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" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DGGML_BACKEND_DL=ON ^
+            -DLLAMA_BUILD_SERVER=ON ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
             -DGGML_NATIVE=OFF ^
-            -DGGML_CPU=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=ON ^
             -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
+            -DGGML_RPC=ON ^
+            -DGGML_CUDA_CUB_3DOT2=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
+          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
+          cmake --build build --config Release
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
 
   hip:
     runs-on: windows-2022
@@ -89,7 +83,7 @@ jobs:
 
     env:
       # Make sure this is in sync with build-cache.yml
-      ROCM_VERSION: "7.14.0"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     strategy:
       matrix:
@@ -103,53 +97,36 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # - name: Cache ROCm Installation
-      #   uses: actions/cache@v5
-      #   id: cache-rocm
-      #   with:
-      #     path: C:\TheRock\build
-      #     key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
+        run: |
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
+          7z x rocwmma.deb
+          7z x data.tar
+
+      - name: Use ROCm Installation Cache
+        uses: actions/cache@v5
+        id: cache-rocm
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
       - name: Setup ROCm
-        # if: steps.cache-rocm.outputs.cache-hit != 'true'
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         uses: ./.github/actions/windows-setup-rocm
         with:
-          version: ${{ env.ROCM_VERSION }}
-
-      - name: Setup ROCm Environment
-        run: |
-          $ErrorActionPreference = "Stop"
-
-          # Activate venv from cache or fresh install
-          & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-          # Expand the devel tree (idempotent; no-op if already done during install)
-          rocm-sdk init
-          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          $rocmPath = (rocm-sdk path --root)
-          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
-          $rocmPath = $rocmPath.Trim()
-          $cmakePath = (rocm-sdk path --cmake).Trim()
-          $binPath = (rocm-sdk path --bin).Trim()
-          write-host "ROCm root: $rocmPath"
-
-          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
-          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
-          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
-          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
-          echo "$binPath" >> $env:GITHUB_PATH
-
-          # Keep venv in PATH for subsequent steps
-          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
 
       - name: Verify ROCm
         id: verify
         run: |
-          # Test the ROCm clang shipped in the installed wheel
-          & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -157,27 +134,29 @@ jobs:
           # TODO: this build does not match the build in release.yml, so we use a different cache key
           #       ideally, the builds should match, similar to the CUDA build above so that we would be able
           #       to populate the ccache for the release with manual runs of this workflow
-          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
 
       - name: Build
         id: cmake_build
         run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
           cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
-            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DLLAMA_BUILD_BORINGSSL=ON `
-            -DHIP_PATH="${env:HIP_PATH}" `
+            -DROCM_DIR="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
-            -DGPU_TARGETS="gfx1100" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGPU_TARGETS="gfx1100"  `
             -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}

.github/workflows/build-openvino.yml

@@ -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: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     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|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --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|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --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: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     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|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000

.github/workflows/build-rpc.yml

@@ -0,0 +1,66 @@
+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

.github/workflows/build-sanitize.yml

@@ -15,12 +15,6 @@ on:
       '**/*.cpp'
     ]
 
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-sanitize.yml'
-    ]
-
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
@@ -34,35 +28,19 @@ env:
 
 jobs:
   ctest:
+    runs-on: [self-hosted, X64, CPU, Linux]
+
     continue-on-error: true
 
     strategy:
       matrix:
-        include:
-        # thread and address doesn't run properly on some self hosted machines, so run it on Github instead
-        - sanitizer: ADDRESS
-          machine: ubuntu-24.04
-        - sanitizer: THREAD
-          machine: ubuntu-24.04
-        - sanitizer: UNDEFINED
-          machine: [self-hosted, X64, Linux]
-
-    runs-on: ${{ matrix.machine }}
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
-      #   with:
-      #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04
-      #     variant: ccache
-      #     evict-old-files: 1d
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
       # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
       - name: Build (undefined)
         id: cmake_build_undefined

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

@@ -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: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/build-vulkan.yml

@@ -93,13 +93,19 @@ jobs:
         run: |
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
-      - name: Setup Vulkan SDK
-        id: setup
-        uses: ./.github/actions/unarchive-tar
+      - name: Use Vulkan SDK Cache
+        uses: actions/cache@v5
+        id: cache-sdk
         with:
-          url: https://sdk.lunarg.com/sdk/download/${{ env.VULKAN_SDK_VERSION }}/linux/vulkan_sdk.tar.xz
           path: ./vulkan_sdk
-          strip: 1
+          key: cache-gha-vulkan-sdk-${{ 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 }}
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -127,56 +133,3 @@ 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

.github/workflows/docker.yml

@@ -44,7 +44,6 @@ jobs:
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
-          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
 
       - name: Determine source tag name
         id: srctag
@@ -394,11 +393,6 @@ 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:
@@ -433,7 +427,6 @@ jobs:
           password: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Create tags from digests
-        id: create_tags
         shell: bash
         run: |
           set -euo pipefail
@@ -445,7 +438,6 @@ 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"
@@ -512,16 +504,6 @@ 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 $TAGS; do
@@ -545,24 +527,3 @@ jobs:
           done
         env:
           GITHUB_REPOSITORY_OWNER: '${{ 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 }}

.github/workflows/make-release.yml

@@ -1,126 +0,0 @@
-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
-        type: boolean
-        default: true
-
-env:
-  GH_TOKEN: ${{ github.token }}
-
-permissions:
-  contents: write
-
-jobs:
-  make-release:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        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: ${{ github.event.inputs.dry_run == 'false' }}
-        run: |
-          VERSION="${{ steps.checks.outputs.version }}"
-          git config user.name "github-actions[bot]"
-          git config user.email "github-actions[bot]@users.noreply.github.com"
-          git tag -a "${VERSION}" -m "Release ${VERSION}"
-          git push origin "${VERSION}"
-          echo "Created and pushed tag ${VERSION}"
-
-      - name: Generate release description
-        id: desc
-        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
-        env:
-          GITHUB_REPOSITORY: ${{ github.repository }}
-
-      - name: Create nightly-tag.txt
-        id: nightly_tag_file
-        run: |
-          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
-          if [[ -z "${NIGHTLY_TAG}" ]]; then
-            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
-            echo "create=false" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          echo "${NIGHTLY_TAG}" > nightly-tag.txt
-          echo "create=true" >> "$GITHUB_OUTPUT"
-          echo "nightly-tag.txt:"
-          cat nightly-tag.txt
-
-      - name: Create release
-        id: 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 }}
-          prerelease: false
-          # TODO: enrich the body of the release with more information
-          body: |
-            ## Overview
-
-            New version has been released.
-
-            ${{ steps.desc.outputs.nightly }}
-
-            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
-
-            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
-
-            ## ${{ steps.desc.outputs.changelog_title }}
-
-            ${{ steps.desc.outputs.changelog }}
-
-      - name: Upload nightly-tag.txt
-        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
-        uses: actions/github-script@v8
-        with:
-          github-token: ${{secrets.GITHUB_TOKEN}}
-          script: |
-            const fs = require('fs');
-            const release_id = '${{ steps.create_release.outputs.id }}';
-            console.log('uploadReleaseAsset', 'nightly-tag.txt');
-            await github.rest.repos.uploadReleaseAsset({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              release_id: release_id,
-              name: 'nightly-tag.txt',
-              data: await fs.readFileSync('./nightly-tag.txt')
-            });
-
-      - name: Dry run summary
-        if: ${{ github.event.inputs.dry_run == 'true' }}
-        run: |
-          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
-            echo "Dry run complete - all checks passed."
-            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
-            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
-              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
-            fi
-          else
-            echo "::error::Dry run found release check failures. A release tag would not be created."
-            exit 1
-          fi

.github/workflows/pr-draft-label.yml

@@ -1,23 +0,0 @@
-name: Convert PR to draft
-
-on:
-  pull_request_target:
-    types: [labeled]
-
-permissions:
-  pull-requests: write
-  issues: write
-  contents: write # required for "gh pr ready" command, see https://github.com/cli/cli/issues/8910
-
-jobs:
-  convert-to-draft:
-    if: github.event.label.name == 'draft' && github.event.pull_request.draft == false
-    runs-on: ubuntu-slim
-    steps:
-      - name: Convert PR to draft
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_URL: ${{ github.event.pull_request.html_url }}
-        run: |
-          gh pr ready --undo "$PR_URL"
-          gh pr edit "$PR_URL" --remove-label draft

.github/workflows/release.yml

@@ -145,6 +145,11 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -161,11 +166,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
           name: llama-bin-macos-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
   ubuntu-cpu:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,6 +231,12 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -247,12 +253,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
   ubuntu-vulkan:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,6 +318,11 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -334,11 +339,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
   android-arm64:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -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: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -512,6 +512,11 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build/ReleaseOV --config Release --parallel
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -546,11 +551,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
           name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
   windows-openvino:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -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: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -637,6 +637,11 @@ jobs:
 
           cmake --build build\ReleaseOV --config Release -- /m
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -675,13 +680,7 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
           name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
   windows-cpu:
-    name: windows-cpu / ${{ matrix.arch }}
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
@@ -729,13 +728,18 @@ 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
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
       - name: Pack artifacts
         id: pack_artifacts
         run: |
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ 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
@@ -744,141 +748,6 @@ jobs:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
-  windows-rocm:
-    needs: [check-release]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: windows-2022
-
-    strategy:
-      matrix:
-        include:
-          - ROCM_VERSION: "7.14.0"
-            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
-            build: x64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-          max-size: "1G"
-
-      # - name: Cache ROCm Installation
-      #   id: cache-rocm
-      #   uses: actions/cache@v5
-      #   with:
-      #     path: C:\TheRock\build
-      #     key: rocm-wheels-${{ matrix.ROCM_VERSION }}-multi-arch-${{ runner.os }}
-
-      - name: Setup ROCm
-        # if: steps.cache-rocm.outputs.cache-hit != 'true'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ matrix.ROCM_VERSION }}
-
-      - name: Setup ROCm Environment
-        run: |
-          $ErrorActionPreference = "Stop"
-
-          # Activate venv from cache or fresh install
-          & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-          # Expand the devel tree (idempotent; no-op if already done during install)
-          rocm-sdk init
-          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          $rocmPath = (rocm-sdk path --root)
-          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
-          $rocmPath = $rocmPath.Trim()
-          $cmakePath = (rocm-sdk path --cmake).Trim()
-          $binPath = (rocm-sdk path --bin).Trim()
-          write-host "ROCm root: $rocmPath"
-          write-host "CMake path: $cmakePath"
-          write-host "Bin path: $binPath"
-
-          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
-          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
-          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
-          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
-          echo "$binPath" >> $env:GITHUB_PATH
-
-          # Keep venv in PATH for subsequent steps
-          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
-
-      - name: Build
-        run: |
-          mkdir build
-          cd build
-          cmake .. `
-            -G "Unix Makefiles" `
-            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_BACKEND_DL=ON `
-            -DGGML_NATIVE=OFF `
-            -DGGML_CPU=ON `
-            -DGGML_CPU_ALL_VARIANTS=ON `
-            -DGGML_HIP=ON `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
-            -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
-            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DHIP_PATH="${env:HIP_PATH}" `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
-          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Verify HIP backend was built
-        run: |
-          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
-          if (-not $hipDll) {
-            Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
-            Write-Host "Contents of build\bin:"
-            Get-ChildItem build\bin | Format-Table -AutoSize
-            exit 1
-          }
-          Write-Host "HIP backend artifact found:"
-          $hipDll | Format-Table FullName, Length -AutoSize
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Get ROCm short version
-        run: |
-          $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
-          echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
-
-      - name: Pack artifacts
-        run: |
-          cp "LICENSE" "build\bin\"
-          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
-          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
   windows:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -979,7 +848,6 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
@@ -990,16 +858,7 @@ jobs:
 
     strategy:
       matrix:
-        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'
+        cuda: ['12.4', '13.3']
 
     steps:
       - name: Clone
@@ -1017,7 +876,6 @@ jobs:
         uses: ./.github/actions/windows-setup-cuda
         with:
           cuda_version: ${{ matrix.cuda }}
-          cuda_arch: ${{ matrix.arch }}
 
       - name: Install Ninja
         id: install_ninja
@@ -1027,62 +885,54 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
 
       - name: Build
         id: cmake_build
         shell: cmd
         # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DGGML_CUDA_CUB_3DOT2=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip .\build\bin\Release\ggml-cuda.dll
+          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
         with:
-          path: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-          name: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
-      - name: Copy and pack Cuda runtime (x64)
-        if: ${{ matrix.arch == 'x64' }}
+      - name: Copy and pack Cuda runtime
         run: |
           echo "Cuda install location: ${{ env.CUDA_PATH }}"
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
-
-      - name: Copy and pack Cuda runtime (ARM64)
-        if: ${{ matrix.arch == 'arm64' }}
-        run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
-          $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin\arm64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         uses: actions/upload-artifact@v6
         with:
-          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
     needs: [check-release]
@@ -1143,6 +993,11 @@ jobs:
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
       - name: Build the release package
         id: pack_artifacts
         run: |
@@ -1189,11 +1044,6 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
   ubuntu-24-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1266,6 +1116,11 @@ jobs:
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -1282,11 +1137,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
           name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
   ubuntu-22-rocm:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1299,8 +1149,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+          - ROCM_VERSION: "7.2.1"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
             build: 'x64'
 
     steps:
@@ -1325,54 +1175,45 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-          max-size: "1G"
-
-      - name: Tune ccache for reinstalled ROCm toolchain
-        run: |
-          # ROCm is pip-installed fresh each run, so the clang binary's mtime
-          # changes every time. With the default compiler_check=mtime that
-          # invalidates the cache; hash compiler contents instead so warm
-          # builds hit.
-          ccache --set-config=compiler_check=content
-          ccache --set-config=sloppiness=time_macros,include_file_mtime,include_file_ctime
+          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
 
       - name: Dependencies
         id: depends
         run: |
           sudo apt install -y build-essential git cmake wget
 
-      - name: Setup TheRock with Wheels
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2.1'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk
+
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2.1'
         id: therock_env
         run: |
-          # Create Python virtual environment
-          python3 -m venv .venv
-          source .venv/bin/activate
-
-          # Install ROCm wheels for build
-          # libraries = HIP runtime and CMake configs needed for linking
-          # devel = compilers, headers, static libs
-          python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          ROCM_PATH=$(rocm-sdk path --root)
-          CMAKE_PATH=$(rocm-sdk path --cmake)
-          BIN_PATH=$(rocm-sdk path --bin)
-          echo "ROCM_PATH=$ROCM_PATH"
-          echo "CMAKE_PATH=$CMAKE_PATH"
-          echo "BIN_PATH=$BIN_PATH"
-
-          # Set environment variables
-          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
-          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
-          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
-          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
-
-          # Keep venv activated for subsequent steps
-          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
 
       - name: Build with native CMake HIP support
         id: cmake_build
@@ -1388,10 +1229,16 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DGGML_HIP=ON \
             -DHIP_PLATFORM=amd \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -1411,10 +1258,129 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 
+  windows-hip:
+    needs: [check-release, get-version]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
+    runs-on: windows-2022
+
+    permissions:
+      actions: write
+
+    env:
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
+
+    strategy:
+      matrix:
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
+        run: |
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
+          7z x rocwmma.deb
+          7z x data.tar
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v5
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGGML_HIP=ON `
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          md "build\bin\hipblaslt\library"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
+
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode:
     needs: [check-release, get-version]
@@ -1452,9 +1418,7 @@ jobs:
       - name: xcodebuild for swift package
         id: xcodebuild
         run: |
-          # 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
+          ./build-xcframework.sh
 
       - 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
@@ -1582,8 +1546,6 @@ 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
 
@@ -1592,14 +1554,14 @@ jobs:
       - windows
       - windows-cpu
       - windows-cuda
-      - windows-sycl
-      - windows-rocm
+      #- windows-sycl
+      - windows-hip
       - windows-openvino
       - ubuntu-22-rocm
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
-      - ubuntu-24-sycl
+      #- ubuntu-24-sycl
       - android-arm64
       - macos-cpu
       - ios-xcode
@@ -1615,7 +1577,6 @@ jobs:
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
-          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
 
       - name: Determine tag name
         id: tag
@@ -1677,22 +1638,6 @@ 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
@@ -1700,7 +1645,6 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
-          prerelease: true
           body: |
             <details open>
 
@@ -1711,9 +1655,6 @@ 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-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
             - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -1726,7 +1667,7 @@ jobs:
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
             - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1740,11 +1681,10 @@ jobs:
             - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
             - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
             - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.3-x64.zip) - [CUDA 13.3 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
-            - [Windows arm64 (CUDA 13) (preview)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-arm64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-arm64.zip)
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
             - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
             - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)
+            - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
 
             **openEuler:**
             - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)

.github/workflows/server-sanitize.yml

@@ -25,12 +25,6 @@ on:
       'tools/server/**.*'
     ]
 
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/server-sanitize.yml'
-    ]
-
 env:
   LLAMA_ARG_LOG_COLORS: 1
   LLAMA_ARG_LOG_PREFIX: 1
@@ -96,27 +90,23 @@ jobs:
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v7
-
-      - name: Install Python dependencies
-        run: |
-          python3 -m venv .venv
-          .venv/bin/pip install -r tools/server/tests/requirements.txt
+        uses: actions/setup-python@v6
+        with:
+          python-version: '3.11'
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests
         if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
         run: |
-          source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
-          source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x

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

@@ -72,7 +72,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +99,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
   server-cuda:
     runs-on: [self-hosted, llama-server, Linux, NVIDIA]
@@ -132,7 +132,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -141,7 +141,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -150,7 +150,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -159,7 +159,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
   server-kleidiai:
     runs-on: ah-ubuntu_22_04-c8g_8x
@@ -219,4 +219,4 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"

.github/workflows/server.yml

@@ -104,21 +104,21 @@ jobs:
         id: server_integration_tests
         run: |
           cd tools/server/tests
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
         run: |
           cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x
 
       - name: Tests (Backend sampling)
         id: server_integration_tests_backend_sampling
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests (Backend sampling)
         id: server_integration_tests_slow_backend_sampling
@@ -126,7 +126,7 @@ jobs:
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x
 
   windows:
     runs-on: windows-2025
@@ -167,17 +167,15 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
-        shell: bash
         run: |
           cd tools/server/tests
-          export PYTHONIOENCODING=":replace"
-          ./tests.sh
+          $env:PYTHONIOENCODING = ":replace"
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
-        shell: bash
         run: |
           cd tools/server/tests
-          export SLOW_TESTS="1"
-          ./tests.sh
+          $env:SLOW_TESTS = "1"
+          pytest -v -x

.github/workflows/winget.yml

@@ -19,8 +19,6 @@ jobs:
         run: |
           cargo binstall komac@2.16.0 -y
 
-      # TODO: This should later be updated to publish releases instead of
-      # development release builds.
       - name: Find latest release
         id: find_latest_release
         uses: actions/github-script@v8

.pi/gg/SYSTEM.md

@@ -2,14 +2,12 @@ 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>`
@@ -17,7 +15,6 @@ Coding:
 Pull requests (PRs):
 - New branch names are prefixed with "gg/"
 - Before opening a pull request, ask the user to confirm the description
-- Don't explicitly wrap lines in the PR description (each paragraph and bullet is a single line)
 - When creating a pull request, look for the repository's PR template and follow it
 - For the AI usage disclosure section, write "YES. pi:llama.cpp/[MODEL]"
 - Ask the user to tell you what model was used and write it in place of [MODEL]

AGENTS.md

@@ -84,7 +84,6 @@ These points are extremely important - failing to follow them won't necessarily
 Common mistakes that AI agents usually make:
 - Write comments first then write code: this usually leads to extensive redundant comments. Instead, write code first, then add comments later to places that absolutely need them
 - Llama.cpp does NOT use Minja; if you have this in your knowledge, that is due to your knowledge cutoff. Llama.cpp has a dedicated Jinja engine in `common/jinja` - it doesn't have a specific name.
-- Do NOT add a new file in `tests/*` without maintainers' approval. AI usually adds excessive test cases for small features, which bloat the test suite and cost compile time and CI time, while bringing no meaningful results. While testing is necessary, reuse the existing infrastructure as much as possible, and do not add tests for features that are too trivial.
 
 ### Prohibited Actions
 

CMakeLists.txt

@@ -2,26 +2,6 @@ cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
-### llama.cpp version
-set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 2)
-set(LLAMA_VERSION_PATCH 0)
-set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
-
-# whether this is a development/nightly build
-# set this to OFF when making a release from a release tag (vX.Y.Z)
-# ref: https://github.com/ggml-org/ggml/discussions/1579
-option(LLAMA_BUILD_IS_DEV "llama: dev build" ON)
-
-if (LLAMA_BUILD_IS_DEV)
-    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}-dev")
-else()
-    # TODO: check that the current commit is tagged correctly according to the version specified above
-    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}")
-endif()
-
-message(STATUS "llama.cpp version: ${LLAMA_VERSION}")
-
 #set(CMAKE_WARN_DEPRECATED YES)
 set(CMAKE_WARN_UNUSED_CLI YES)
 
@@ -44,6 +24,9 @@ if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     include(git-vars)
+
+    # configure project version
+    # TODO
 else()
     set(LLAMA_STANDALONE OFF)
 endif()
@@ -156,6 +139,7 @@ endif()
 if (NOT DEFINED LLAMA_BUILD_COMMIT)
     set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
 endif()
+set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
 
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
@@ -224,10 +208,9 @@ 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)
@@ -292,12 +275,12 @@ configure_package_config_file(
               LLAMA_BIN_INSTALL_DIR )
 
 write_basic_package_version_file(
-        ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
-        VERSION ${LLAMA_VERSION}
+        ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
+    VERSION ${LLAMA_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
 
 configure_file(cmake/llama.pc.in

README.md

@@ -7,13 +7,12 @@
 <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?filter=v*&color=brightgreen)](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&filter=b*&color=orange)](https://github.com/ggml-org/llama.cpp/releases?q=b)
-[![Server](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/server.yml?label=Server)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.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)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](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)
 
-[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)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev branches](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-features.md) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 
 </div>
 
@@ -107,7 +106,6 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 - [XCFramework](docs/xcframework.md)
 - [Completions](docs/completions.md)
 - [Models](docs/models.md)
-- [Release process](docs/release.md)
 
 ## Contributing
 
@@ -120,7 +118,7 @@ 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
-- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [stb-image](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
-- [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
+- [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

app/llama.cpp

@@ -1,7 +1,5 @@
 #include "build-info.h"
 
-#include "llama.h"
-
 #include <cstdio>
 #include <cstdlib>
 #include <string>
@@ -79,12 +77,12 @@ static const command cmds[] = {
 
 #undef UPDATE_HIDDEN
 
-static int version(int /*argc*/, char ** /*argv*/) {
-    llama_print_build_info(llama_version());
+static int version(int argc, char ** argv) {
+    printf("%s\n", llama_build_info());
     return 0;
 }
 
-static int licenses(int /*argc*/, char ** /*argv*/) {
+static int licenses(int argc, char ** argv) {
     for (int i = 0; LICENSES[i]; ++i) {
         printf("%s\n", LICENSES[i]);
     }

build-xcframework.sh

@@ -1,8 +1,5 @@
 #!/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
@@ -22,43 +19,6 @@ 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"
 
@@ -290,7 +250,6 @@ 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
@@ -442,189 +401,148 @@ combine_static_libraries() {
     rm -rf "${temp_dir}"
 }
 
-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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for iOS simulator..."
+cmake -B build-ios-sim -G Xcode \
+    "${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_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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for iOS devices..."
+cmake -B build-ios-device -G Xcode \
+    "${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_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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for macOS..."
+cmake -B build-macos -G Xcode \
+    "${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_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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for visionOS..."
+cmake -B build-visionos -G Xcode \
+    "${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_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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for visionOS simulator..."
+cmake -B build-visionos-sim -G Xcode \
+    "${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
 
 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
-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 "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for tvOS simulator..."
+cmake -B build-tvos-sim -G Xcode \
+    "${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_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[@]}"
+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
 
 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."
-for b in "${BUILDS[@]}"; do
-    read -r _ bdir _ platform _ min_os < <(build_spec "$b")
-    setup_framework_structure "$bdir" "$min_os" "$platform"
-done
+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"
 
 # Create dynamic libraries from static libraries
 echo "Creating dynamic libraries from static libraries..."
-for b in "${BUILDS[@]}"; do
-    read -r _ bdir rdir platform is_sim _ < <(build_spec "$b")
-    combine_static_libraries "$bdir" "$rdir" "$platform" "$is_sim"
-done
+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"
 
 # 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 \
-    "${XCFW_ARGS[@]}" \
-    -output "$(pwd)/build-apple/llama.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

ci/run.sh

@@ -49,14 +49,6 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-# gpu-rocm self-hosted runner can't upload logs to blob; keep each run's logs in
-# their own dir keyed by the GitHub run id so an Actions run URL maps to its logs.
-if [ -n "${GG_BUILD_ROCM}" ] && [ -n "${GITHUB_RUN_ID}" ]; then
-    OUT="$OUT/run-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
-    mkdir -p "$OUT"
-    echo "ci results dir: $OUT"
-fi
-
 rm -f $OUT/*.log
 rm -f $OUT/*.exit
 rm -f $OUT/*.md
@@ -100,7 +92,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON -DGGML_HIP_ROCWMMA_FATTN=ON"
     if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
         echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
         exit 1
@@ -190,7 +182,7 @@ 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|test-recurrent-state-rollback-nemotron-h"
+    CTEST_EXTRA="-E test-llama-archs"
 fi
 
 ## helpers

cmake/arm64-windows-llvm.cmake

@@ -8,7 +8,6 @@ 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" )

cmake/arm64-windows-msvc-cuda.cmake

@@ -1,26 +0,0 @@
-# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR arm64 )
-
-if ( DEFINED CUDAToolkit_ROOT )
-    file( TO_CMAKE_PATH "${CUDAToolkit_ROOT}" CUDA_ROOT )
-elseif ( DEFINED ENV{CUDA_PATH} )
-    file( TO_CMAKE_PATH "$ENV{CUDA_PATH}" CUDA_ROOT )
-else()
-    message( FATAL_ERROR "Set CUDAToolkit_ROOT or CUDA_PATH to a Windows CUDA Toolkit with ARM64 target libraries" )
-endif()
-
-if ( DEFINED ENV{VCToolsInstallDir} )
-    file( TO_CMAKE_PATH "$ENV{VCToolsInstallDir}" MSVC_TOOLS_ROOT )
-    set( CMAKE_CUDA_HOST_COMPILER "${MSVC_TOOLS_ROOT}/bin/Hostx64/arm64/cl.exe" CACHE FILEPATH "" )
-endif()
-
-set( CMAKE_CUDA_COMPILER "${CUDA_ROOT}/bin/nvcc.exe" CACHE FILEPATH "" )
-set( CMAKE_CUDA_FLAGS_INIT "-target-dir=arm64" )
-
-# FindCUDAToolkit selects lib/x64 from the host architecture on Windows.
-set( CUDA_CUDART              "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
-set( CUDA_cudart_LIBRARY      "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
-set( CUDA_cublas_LIBRARY      "${CUDA_ROOT}/lib/arm64/cublas.lib"   CACHE FILEPATH "" )
-set( CUDA_cublasLt_LIBRARY    "${CUDA_ROOT}/lib/arm64/cublasLt.lib" CACHE FILEPATH "" )
-set( CUDA_cuda_driver_LIBRARY "${CUDA_ROOT}/lib/arm64/cuda.lib"     CACHE FILEPATH "" )

cmake/llama-config.cmake.in

@@ -1,4 +1,4 @@
-set(LLAMA_VERSION      @LLAMA_VERSION@)
+set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
 set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)

cmake/llama.pc.in

@@ -5,6 +5,6 @@ includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@
 
 Name: llama
 Description: Port of Facebook's LLaMA model in C/C++
-Version: @LLAMA_VERSION@
+Version: @LLAMA_INSTALL_VERSION@
 Libs: -L${libdir} -lggml -lggml-base -lllama
 Cflags: -I${includedir}

common/CMakeLists.txt

@@ -81,8 +81,6 @@ add_library(${TARGET}
     imatrix-loader.cpp
     imatrix-loader.h
     json-schema-to-grammar.cpp
-    json.cpp
-    json.h
     llguidance.cpp
     log.cpp
     log.h
@@ -123,13 +121,12 @@ add_library(${TARGET}
     )
 
 set_target_properties(${TARGET} PROPERTIES
-    VERSION ${LLAMA_VERSION_BASE}
-    SOVERSION ${LLAMA_VERSION_MAJOR}
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
     MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )
 
-target_include_directories(${TARGET} PUBLIC .)
-target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
+target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
 if (LLAMA_SUBPROCESS)

common/arg.cpp

@@ -5,7 +5,6 @@
 #include "common.h"
 #include "download.h"
 #include "json-schema-to-grammar.h"
-#include "json.h"
 #include "llama.h"
 #include "log.h"
 #include "sampling.h"
@@ -22,6 +21,9 @@
 #include <shellapi.h>
 #endif
 
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
@@ -30,11 +32,9 @@
 #include <filesystem>
 #include <fstream>
 #include <list>
-#include <numeric>
 #include <regex>
 #include <set>
 #include <string>
-#include <system_error>
 #include <thread> // for hardware_concurrency
 #include <vector>
 
@@ -54,7 +54,7 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 using namespace common_arg_utils;
 
 static std::initializer_list<enum llama_example> mmproj_examples = {
@@ -560,15 +560,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
         }
     }
 
-    // infer the speculative type from the draft GGUF metadata when none is requested
-    // note: reads only the first split - sharded drafts need an explicit --spec-type
-    if (spec_types_is_default(params) && !params.speculative.draft.mparams.path.empty()) {
-        const auto types_gguf = common_speculative_types_from_gguf(params.speculative.draft.mparams.path);
-        if (!types_gguf.empty()) {
-            params.speculative.types = types_gguf;
-        }
-    }
-
     // when a sidecar type is requested, the draft repo resolves to its sidecar instead of a full model
     const bool spec_sidecar_found = !plan_spec.mtp.local_path.empty() ||
                                     !plan_spec.dflash.local_path.empty() ||
@@ -713,61 +704,12 @@ void common_models_handler_apply(common_models_handler & handler, common_params
 // CLI argument parsing functions
 //
 
-// apply config files (if present), a later file overrides an earlier one:
-// 1. system-wide: /etc/llama.cpp/config.ini (%PROGRAMDATA%\llama.cpp\config.ini on windows)
-// 2. user-level: ${XDG_CONFIG_HOME:-~/.config}/llama.cpp/config.ini (%APPDATA%\llama.cpp\config.ini on windows)
-static void common_params_apply_system_config(common_params & params, llama_example ex) {
-    std::vector<std::string> paths;
-
-#if defined(_WIN32)
-    const std::string program_data = common_get_env("PROGRAMDATA");
-    if (!program_data.empty()) {
-        paths.push_back(program_data + "\\llama.cpp\\config.ini");
-    }
-#else
-    paths.push_back("/etc/llama.cpp/config.ini");
-#endif
-
-    try {
-        paths.push_back(fs_get_config_directory() + "config.ini");
-    } catch (const std::exception & e) {
-        LOG_DBG("cannot read user-level config file, skipping: %s\n", e.what());
-    }
-
-    std::vector<std::string> found;
-    for (const auto & path : paths) {
-        std::error_code ec;
-        if (std::filesystem::exists(path, ec)) {
-            found.push_back(path);
-        }
-    }
-    if (found.empty()) {
-        return;
-    }
-
-    common_preset_context ctx(ex);
-    ctx.ignore_unknown_keys = true; // the same config file is shared by all programs
-    for (const auto & path : found) {
-        LOG_INF("using config file: %s\n", path.c_str());
-        common_preset global;
-        common_presets presets = ctx.load_from_ini(path, global);
-        global.apply_to_params(params);
-        auto it = presets.find(COMMON_PRESET_DEFAULT_NAME);
-        if (it != presets.end()) {
-            it->second.apply_to_params(params);
-        }
-    }
-}
-
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
     common_params & params = ctx_arg.params;
 
     // setup log directly from params.verbosity: see tools/cli/cli.cpp
     common_log_set_verbosity_thold(params.verbosity);
 
-    // config file applies first, so env variables and CLI arguments override it
-    common_params_apply_system_config(params, ctx_arg.ex);
-
     std::unordered_map<std::string, std::pair<common_arg *, bool>> arg_to_options;
     for (auto & opt : ctx_arg.options) {
         for (const auto & arg : opt.args) {
@@ -1448,7 +1390,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--version"},
         "show version and build info",
         [](common_params &) {
-            llama_print_build_info(llama_version());
+            fprintf(stderr, "version: %d (%s)\n", llama_build_number(), llama_commit());
+            fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
             exit(0);
         }
     ));
@@ -1897,7 +1840,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.conversation_mode = value ? COMMON_CONVERSATION_MODE_ENABLED : COMMON_CONVERSATION_MODE_DISABLED;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-st", "--single-turn"},
         "run conversation for a single turn only, then exit when done\n"
@@ -2594,26 +2537,6 @@ 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",
@@ -2682,16 +2605,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_DIO"));
     add_opt(common_arg(
         {"-lm", "--load-mode"}, "MODE",
-        "model loading mode (default: auto)\n"
-        "- auto: mmap, unless a device does not support it\n"
+        "model loading mode (default: mmap)\n"
         "- none: no special loading mode\n"
         "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
         "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
         "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
         "- dio: use DirectIO if available\n",
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "auto")       { params.load_mode = LLAMA_LOAD_MODE_AUTO;       }
-            else if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
             else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
             else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
             else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }
@@ -3381,23 +3302,12 @@ 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_info\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, 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);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
-    add_opt(common_arg(
-        {"--tools-runtime"}, "OPTION",
-        "experimental: run tools in a separate runtime environment (default: none, use host environment)\n"
-        "available options:\n"
-        "  'docker:<image>', 'podman:<image>': spin up a new container and reuse it for all invocations, clean up on server exit\n"
-        "  'docker-container:<id>', 'podman-container:<id>': use an existing container by ID, won't stop on server exit\n"
-        "  'ssh:<target>': run tools on a remote POSIX host over SSH, key-based auth and a trusted host key are required\n",
-        [](common_params & params, const std::string & value) {
-            params.server_tools_runtime = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS_RUNTIME"));
     add_opt(common_arg(
         {"--mcp-servers-config"}, "PATH",
         "experimental: path to JSON file with MCP server definitions (Cursor-compatible format) - do not enable in untrusted environments (default: none)\n"
@@ -3665,18 +3575,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING"));
-    add_opt(common_arg(
-        {"--reasoning-effort"}, "LEVEL",
-        "reasoning effort level given to the chat template: 'default' to keep the template default,\n"
-        "or a level such as 'minimal', 'low', 'medium', 'high', 'xhigh' or 'max' (default: default)",
-        [](common_params & params, const std::string & value) {
-            if (value == "default") {
-                params.default_template_kwargs.erase("reasoning_effort");
-            } else {
-                params.default_template_kwargs["reasoning_effort"] = json(value).dump();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_EFFORT"));
     add_opt(common_arg(
         {"--reasoning-budget"}, "N",
         "token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)",
@@ -4096,9 +3994,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--spec-draft-n-max"}, "N",
         string_format("number of tokens to draft for speculative decoding (default: %d)", params.speculative.draft.n_max),
         [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("invalid value");
-            }
             params.speculative.draft.n_max = value;
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_MAX"));
@@ -4677,12 +4572,6 @@ 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",

common/arg.h

@@ -11,9 +11,8 @@
 #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_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
 
 //
 // CLI argument parsing

common/build-info.cpp.in

@@ -29,7 +29,7 @@ const char * llama_build_info(void) {
     return s.c_str();
 }
 
-void llama_print_build_info(const char * llama_version) {
-    fprintf(stderr, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
-    fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
+void llama_print_build_info(void) {
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, llama_build_number(), llama_commit());
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, llama_compiler(), llama_build_target());
 }

common/build-info.h

@@ -8,4 +8,4 @@ const char * llama_compiler(void);
 const char * llama_build_target(void);
 const char * llama_build_info(void);
 
-void llama_print_build_info(const char *);
+void llama_print_build_info(void);

common/chat-auto-parser-generator.cpp

@@ -5,12 +5,13 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <stdexcept>
 #include <string>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 // Helper to iterate over tools/functions
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
@@ -390,7 +391,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
 
         std::set<std::string> required;
         if (params.contains("required")) {
-            required = params.at("required").get<std::set<std::string>>();
+            params.at("required").get_to(required);
         }
 
         auto schema_info = common_schema_info();

common/chat-auto-parser-helpers.cpp

@@ -4,11 +4,14 @@
 #include "chat-peg-parser.h"
 #include "chat.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <cctype>
 #include <numeric>
 
+using json = nlohmann::ordered_json;
+
 std::string trim_whitespace(const std::string & str) {
     size_t start = 0;
     while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {

common/chat-auto-parser.h

@@ -4,7 +4,7 @@
 #include "common.h"
 #include "jinja/caps.h"
 #include "peg-parser.h"
-#include "json.h"
+#include "nlohmann/json.hpp"
 
 #include <chrono>
 #include <optional>
@@ -12,7 +12,7 @@
 #include <utility>
 #include <vector>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 class common_chat_peg_builder;
 

common/chat-diff-analyzer.cpp

@@ -4,11 +4,11 @@
 #include "chat.h"
 #include "common.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <algorithm>
 #include <cctype>
-#include <numeric>
 #include <ostream>
 #include <sstream>
 
@@ -17,7 +17,7 @@
 #define ANSI_ORANGE "\033[1m\x1b[38;5;214m"
 #define ANSI_RED    "\033[1m\x1b[38;5;196m"
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 namespace autoparser {
 
@@ -193,14 +193,6 @@ 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);
-          }
-      },
 
     });
 
@@ -929,7 +921,7 @@ void analyze_tools::analyze_tool_call_format_json_native(const std::string & cle
     int  json_end       = clean_haystack.find_last_of('}');
     std::string cut     = clean_haystack.substr(json_start, json_end - json_start + 1);
     json call_struct    = json::parse(cut);
-    auto register_field = [&](const std::string & prefix, const common_json_entry & subel) {
+    auto register_field = [&](const std::string & prefix, const nlohmann::detail::iteration_proxy_value<json::iterator> & subel) {
         if (subel.value().is_string() && std::string(subel.value()).find("call0000") != std::string::npos) {
             format.id_field = !prefix.empty() ? prefix + "." + subel.key() : subel.key();
         } else if (subel.value().is_string() && std::string(subel.value()) == fun_name_needle) {

common/chat-peg-parser.cpp

@@ -4,10 +4,12 @@
 #include "ggml.h"
 #include "peg-parser.h"
 
+#include <nlohmann/json.hpp>
+
 #include <cstdint>
 #include <functional>
 
-using ordered_json = common_json;
+using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
@@ -592,7 +594,9 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
 
                 // Full argument: name="value" or name=value
                 auto arg_rule = tool_arg(
-                    tool_arg_open(tool_arg_name(arg_name_parser) + literal("=")) +
+                    tool_arg_open(eps()) +
+                    tool_arg_name(arg_name_parser) +
+                    literal("=") +
                     arg_value_parser +
                     tool_arg_close(eps())
                 );

common/chat-peg-parser.h

@@ -128,7 +128,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const common_json &   tools,
+                                          const nlohmann::ordered_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -143,13 +143,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const common_json &             tools,
+                                                 const nlohmann::ordered_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const common_json & tools,
+    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
                                               bool                           parallel_tool_calls,
                                               bool                           allow_json_literals);
 
@@ -158,19 +158,19 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser python_or_json_value();
 
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const common_json & tools,
+    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
                                                        const std::string &            args_key,
                                                        const std::string &            effective_args_key,
                                                        const std::string &            call_id_key,
                                                        const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const common_json & tools,
+    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
                                                    const std::string &            effective_name_key,
                                                    const std::string &            effective_args_key,
                                                    const std::string &            call_id_key,
                                                    const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const common_json &   tools,
+    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,

common/chat.cpp

@@ -6,7 +6,6 @@
 #include "common.h"
 #include "ggml.h"
 #include "json-schema-to-grammar.h"
-#include "json.h"
 #include "log.h"
 
 #include "jinja/value.h"
@@ -14,13 +13,14 @@
 #include "jinja/caps.h"
 #include "peg-parser.h"
 
+#include "nlohmann/json.hpp"
+
 #include <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <ctime>
 #include <exception>
 #include <functional>
-#include <iomanip>
 #include <map>
 
 #include <optional>
@@ -30,7 +30,7 @@
 #include <utility>
 #include <vector>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto               time       = std::chrono::system_clock::to_time_t(now);
@@ -48,7 +48,7 @@ static json safe_args_parse(const std::string & to_parse) {
     }
     try {
         return json::parse(stripped);
-    } catch (const common_json_error & e) {
+    } catch (json::exception & e) {
         return stripped;
     }
 }
@@ -470,80 +470,36 @@ 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;
-            if (copy.contains("content")) {
-                json & it = copy.at("content");
-                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) {
-        messages.push_back(msg.to_json_oaicompat(/* concat_typed_text= */ false));
+        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);
+        }
     }
-    return messages_inp_normalizer(c).normalize(messages);
+    return messages;
 }
 
 // DEPRECATED: only used in tests
@@ -608,7 +564,7 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-common_chat_continuation common_chat_continuation_parse(const common_json & value) {
+common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
     if (value.is_boolean() && value.get<bool>()) {
         return COMMON_CHAT_CONTINUATION_AUTO;
     }
@@ -920,7 +876,7 @@ static void foreach_parameter(const json &
     const auto &          props = params.at("properties");
     std::set<std::string> required;
     if (params.contains("required") && params.at("required").is_array()) {
-        required = params.at("required").get<std::set<std::string>>();
+        params.at("required").get_to(required);
     }
     for (const auto & [name, prop] : props.items()) {
         bool is_required = (required.find(name) != required.end());
@@ -936,11 +892,8 @@ 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
-    json inp = json{
-        {"messages", messages_override.has_value()
-            ? *messages_override
-            : messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
+    nlohmann::ordered_json inp = nlohmann::ordered_json{
+        {"messages", messages_override.has_value() ? *messages_override : inputs.messages},
         {"bos_token", tmpl.bos_token()},
         {"eos_token", tmpl.eos_token()},
         {"enable_thinking", inputs.enable_thinking},
@@ -967,10 +920,6 @@ static std::string common_chat_template_direct_apply_impl(
         bool enabled = inp["preserve_reasoning"].get<bool>();
         jinja::caps_apply_preserve_reasoning(ctx, enabled);
     }
-    if (inp.contains("reasoning_effort") && inp["reasoning_effort"].is_string() && !inp["reasoning_effort"].empty()) {
-        std::string reasoning_effort = inp["reasoning_effort"].get<std::string>();
-        jinja::caps_apply_reasoning_effort(ctx, reasoning_effort);
-    }
 
     jinja::global_from_json(ctx, inp, inputs.mark_input);
 
@@ -1004,12 +953,14 @@ 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, messages_override, tools_override, additional_context);
+    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
     params.add_generation_prompt = true;
-    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, messages_override, tools_override, additional_context);
+    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
 
     size_t prefix_len = 0;
     size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
@@ -1058,7 +1009,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
                 });
             } else if (msg.at("content").is_array()) {
                 auto blocks = msg.at("content");
-                content.insert(blocks);
+                content.insert(content.end(), blocks.begin(), blocks.end());
             }
         }
 
@@ -1215,16 +1166,6 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
         data.prompt += data.generation_prompt;
     }
 
-    std::vector<std::string> tool_call_starts = { "<tool_call>" };
-
-    // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
-    // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
-    // constraining on <function which may occur in valid content generation, e.g. #include <functional>
-    foreach_function(inputs.tools, [&](const json & tool) {
-        const std::string name = tool.at("function").at("name");
-        tool_call_starts.push_back("<function=" + name + ">");
-    });
-
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto generation_prompt = p.literal(GEN_PREFIX);
 
@@ -1297,7 +1238,7 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
             auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
 
             return generation_prompt +
-                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
+                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
         }
 
         // Content only parser
@@ -1323,9 +1264,12 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
         });
 
         if (data.grammar_lazy) {
-            for (const auto & start : tool_call_starts) {
-                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
-            }
+            data.grammar_triggers = {
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
+                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
+                // the token with the function name e.g. "=read"
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
+            };
         }
     }
 
@@ -2238,7 +2182,7 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
 
                 std::set<std::string> required;
                 if (params.contains("required")) {
-                    required = params.at("required").get<std::set<std::string>>();
+                    params.at("required").get_to(required);
                 }
 
                 auto schema_info = common_schema_info();
@@ -2370,179 +2314,6 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     return data;
 }
 
-// Kimi K3 - XTML tagged format, built by open_tag/close_tag macros:
-//   open_tag(t, attrs) = <|open|>t k="v"...<|sep|>   close_tag(t) = <|close|>t<|sep|>
-//   assistant := [think] [response] [tools] close_tag(message) <|end_of_msg|>
-// the generation prompt already opens the think (or response) section, so the
-// section opener is optional here - same as Kimi K2 Thinking
-static common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &          tmpl,
-                                                          const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    const std::string SEP         = "<|sep|>";
-    const std::string MSG_START   = "<|open|>message role=\"assistant\"<|sep|>";
-    const std::string THINK_START = "<|open|>think<|sep|>";
-    const std::string THINK_END   = "<|close|>think<|sep|>";
-    const std::string RESP_START  = "<|open|>response<|sep|>";
-    const std::string RESP_END    = "<|close|>response<|sep|>";
-    const std::string TOOLS_START = "<|open|>tools<|sep|>";
-    const std::string TOOLS_END   = "<|close|>tools<|sep|>";
-    const std::string CALL_START  = "<|open|>call tool=\"";
-    const std::string CALL_END    = "<|close|>call<|sep|>";
-    const std::string ARG_START   = "<|open|>argument key=\"";
-    const std::string ARG_END     = "<|close|>argument<|sep|>";
-    const std::string MSG_END     = "<|close|>message<|sep|>";
-    const std::string EOM_TOKEN   = "<|end_of_msg|>";
-
-    // only the markers are special tokens. tag names ("think", "response", ...) are
-    // normal tokens and must not be preserved, or prose with those words is broken
-    data.preserved_tokens = {
-        "<|open|>",
-        "<|close|>",
-        "<|sep|>",
-        "<|end_of_msg|>",
-    };
-
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = { THINK_END };
-
-    // per-role message-start delimiters. user/assistant messages only have the role
-    // attribute, so the full opener is used. system and tool messages have more
-    // attributes, so those delimiters stop after the closing quote of the role
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|open|>message role=\"assistant\"<|sep|>" },
-        { COMMON_CHAT_ROLE_USER,      "<|open|>message role=\"user\"<|sep|>"      },
-        { COMMON_CHAT_ROLE_TOOL,      "<|open|>message role=\"tool\""             },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|open|>message role=\"system\""           },
-    };
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = MSG_START + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + RESP_START + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto end = p.end();
-
-        auto start = p.optional(p.literal(MSG_START));
-
-        // the think section is always consumed, even with reasoning extraction off:
-        // the generation prompt ends with open_tag('think'), so it is always present.
-        // reasoning stops at its own closer, or at the response opener if the model
-        // skips the closer
-        auto think_body = extract_reasoning ? p.reasoning(p.until_one_of({ THINK_END, RESP_START })) :
-                                              p.content(p.until_one_of({ THINK_END, RESP_START }));
-
-        auto reasoning = p.optional(p.optional(p.literal(THINK_START)) + think_body +
-                                    p.optional(p.literal(THINK_END)));
-
-        // content runs to the response closer, or to the next section if truncated
-        auto response = p.optional(p.literal(RESP_START)) +
-                        p.content(p.until_one_of({ RESP_END, TOOLS_START, MSG_END })) +
-                        p.optional(p.literal(RESP_END));
-
-        // the EOG token after the message closer reaches the parser as text,
-        // so it must be consumed or the parse stays incomplete
-        auto trailer = p.optional(p.literal(MSG_END)) + p.optional(p.literal(EOM_TOKEN));
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return start + reasoning + response + trailer + end;
-        }
-
-        auto tool_choices = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            const json   schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-            // arguments come one tag per key, with the JSON type in a type="..."
-            // attribute. the type is taken from the tool schema instead, as it tells
-            // us if the value is JSON or a literal string
-            auto args = p.eps();
-            if (schema.contains("properties") && !schema.at("properties").empty()) {
-                auto arg_choices = p.choice();
-                for (const auto & prop : schema.at("properties").items()) {
-                    const std::string & key = prop.key();
-
-                    std::string type = "string";
-                    if (prop.value().is_object() && prop.value().contains("type") &&
-                        prop.value().at("type").is_string()) {
-                        type = prop.value().at("type").get<std::string>();
-                    }
-
-                    auto value = type == "string" ? p.tool_arg_string_value(p.until(ARG_END)) :
-                                                    p.tool_arg_value(p.until(ARG_END));
-
-                    // skip the trailing type="..." attribute: anything up to <|sep|>
-                    arg_choices |= p.rule("kimi-k3-arg-" + name + "-" + key,
-                                          p.tool_arg(p.tool_arg_open(p.literal(ARG_START)) +
-                                                     p.tool_arg_name(p.literal(key)) + p.literal("\"") +
-                                                     p.until(SEP) + p.literal(SEP) + value +
-                                                     p.tool_arg_close(p.literal(ARG_END))));
-                }
-                args = p.zero_or_more(arg_choices);
-            }
-
-            // skip the trailing index="N" attribute the same way
-            auto call = p.tool(p.tool_open(p.literal(CALL_START) + p.tool_name(p.literal(name)) + p.literal("\"") +
-                                           p.until(SEP) + p.literal(SEP)) +
-                               p.tool_args(args) + p.tool_close(p.literal(CALL_END)));
-
-            tool_choices |= p.rule("kimi-k3-tool-" + name, call);
-        });
-
-        // all calls go inside one tools section, then the message is closed. the
-        // message closer is part of the trigger rule, or else the lazy grammar
-        // rejects it once tool calls have started
-        auto tools_section =
-            p.trigger_rule("kimi-k3-tool-call", p.literal(TOOLS_START) + p.one_or_more(tool_choices) +
-                                                    p.literal(TOOLS_END) + p.optional(p.literal(MSG_END)) +
-                                                    p.optional(p.literal(EOM_TOKEN)));
-
-        auto tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? tools_section :
-                                                                              p.optional(tools_section);
-
-        return start + reasoning + response + tools + trailer + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                if (function.contains("parameters")) {
-                    auto schema = function.at("parameters");
-                    builder.resolve_refs(schema);
-                }
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOLS_START },
-        };
-    }
-
-    return data;
-}
-
 // Cohere2 MoE (a.k.a. "North Code") parser.
 //
 // The assistant turn is fully marker-wrapped:
@@ -2860,7 +2631,7 @@ static common_chat_params common_chat_params_init_minimax_m3(const common_chat_t
 
                 std::set<std::string> required;
                 if (schema.contains("required")) {
-                    required = schema.at("required").get<std::set<std::string>>();
+                    schema.at("required").get_to(required);
                 }
 
                 std::vector<common_peg_parser> required_elements;
@@ -2972,10 +2743,10 @@ static void system_message_not_supported(json & messages) {
             auto & second_msg = messages[1];
             second_msg["content"] = first_msg.at("content").get<std::string>()
                 + "\n" + second_msg.at("content").get<std::string>();
-            messages.erase(0);
+            messages.erase(messages.begin());
         } else {
             LOG_WRN("Removing system prompt due to template not supporting system role\n");
-            messages.erase(0);
+            messages.erase(messages.begin());
         }
     }
 }
@@ -3315,153 +3086,6 @@ static common_chat_params common_chat_params_init_minicpm5(const common_chat_tem
     return data;
 }
 
-// An assistant turn is rendered as one or more messages, each
-// "<|start|>assistant to=<recipient><|message|>{content}{END}" where END is
-// <|eom|> (more messages follow) or <|eot|> (end of turn):
-//   - chain-of-thought: to=self, terminated by <|eom|>
-//   - final answer:     to=user, terminated by <|eot|>
-// The generation prompt is just "<|start|>assistant"; the model emits its own
-// " to=...<|message|>".
-static common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template &          tmpl,
-                                                               const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = "<|start|>assistant";
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    data.preserved_tokens = {
-        "<|start|>", "<|message|>", "<|eom|>", "<|eot|>",
-        // ATEM tool-call markup emitted on " to=<tool>" turns.
-        "<atem:function_calls>", "<atem:invoke", "<atem:parameter", "</atem:parameter>",
-        "</atem:invoke>", "</atem:function_calls>",
-    };
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
-        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
-        { COMMON_CHAT_ROLE_TOOL,      "<|start|>tool"      },
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|start|>assistant to=self<|message|>" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<|eom|><|start|>assistant to=user<|message|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
-    // Constrained grammar whenever tools are offered.
-    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start = p.rule("start", p.literal("<|start|>assistant"));
-
-        if (!extract_reasoning && !include_grammar) {
-            return start + p.content(p.rest());
-        }
-
-        if (extract_reasoning) {
-            p.rule("analysis", p.literal(" to=self<|message|>") + p.reasoning(p.until("<|eom|>")) + p.literal("<|eom|>"));
-        } else {
-            p.rule("analysis", p.literal(" to=self<|message|>") + p.content(p.until("<|eom|>")) + p.literal("<|eom|>"));
-        }
-        auto analysis = p.ref("analysis");
-
-        auto recipient  = p.optional(p.literal(" to=user"));
-        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
-                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto string_value = p.ac(
-                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
-                "</atem:parameter>");
-
-            auto tool_choice = p.choice();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto &      function = tool.at("function");
-                const std::string name     = function.at("name");
-                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-                auto args = p.eps();
-                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
-                    auto schema_info = common_schema_info();
-                    schema_info.resolve_refs(params);
-
-                    auto arg_choice = p.choice();
-                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
-                        auto value_parser = p.eps();
-                        if (schema_info.resolves_to_string(prop_schema)) {
-                            value_parser = string_value;
-                        } else {
-                            value_parser = p.tool_arg_json_value(
-                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
-                                + p.tool_arg_close(p.literal("</atem:parameter>"));
-                        }
-
-                        auto arg_rule = p.tool_arg(
-                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
-                            value_parser);
-
-                        arg_choice |= arg_rule;
-                    }
-                    args = p.zero_or_more(arg_choice + p.space());
-                }
-
-                auto tool_parser = p.tool(
-                    p.tool_open(p.literal(" to=") + p.until("<|message|>") +
-                                p.literal("<|message|><atem:function_calls>") + p.space() +
-                                p.literal("<atem:invoke name=\"") + p.tool_name(p.literal(name)) + p.literal("\">") + p.space())
-                    << p.tool_args(args)
-                    << p.tool_close(p.literal("</atem:invoke>") + p.space() + p.literal("</atem:function_calls>")));
-
-                tool_choice |= p.rule("tool-" + name, tool_parser);
-            });
-
-            auto tool_calls = inputs.parallel_tool_calls
-                ? p.trigger_rule("tool-call", tool_choice + p.zero_or_more(p.literal("<|eom|>") + start + tool_choice))
-                : p.trigger_rule("tool-call", tool_choice);
-
-
-            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return p.zero_or_more(start + analysis) + start + tool_calls;
-            }
-            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
-            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
-        }
-
-        return p.zero_or_more(start + analysis) + start + final_msg;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
-        };
-    }
-
-    return data;
-}
-
 static json common_chat_extra_context() {
     json ctx = json::object();
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
@@ -3490,12 +3114,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_gpt_oss(tmpl, params);
     }
 
-    // Muse Glimmer format using " to=<recipient>" recipients and <|eom|>/<|eot|> message terminators.
-    if (src.find("<atem:function_calls>") != std::string::npos && src.find("<|eom|>") != std::string::npos) {
-        LOG_DBG("Using specialized template: Muse Glimmer\n");
-        return common_chat_params_init_muse_glimmer(tmpl, params);
-    }
-
     // Functionary v3.2 - uses recipient-based format with >>>recipient\n{content}
     // Detection: template has ">>>all" for content and ">>>" prefix for tool calls
     if (src.find(">>>all") != std::string::npos && src.find(">>>${recipient}") != std::string::npos) {
@@ -3511,13 +3129,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
-    // Kimi K3 - the <|open|>/<|close|>/<|end_of_msg|> markers are unique to it
-    if (src.find("<|open|>") != std::string::npos && src.find("<|close|>") != std::string::npos &&
-        src.find("<|end_of_msg|>") != std::string::npos) {
-        LOG_DBG("Using specialized template: Kimi K3\n");
-        return common_chat_params_init_kimi_k3(tmpl, params);
-    }
-
     // Cohere2 MoE / North Code - marker-wrapped format with <|START_TEXT|> content and
     // <|START_ACTION|> JSON tool calls. <|START_TEXT|> is unique to this template (the older
     // Command-R templates use <|START_RESPONSE|>).

common/chat.h

@@ -8,7 +8,7 @@
 #include "jinja/runtime.h"
 #include "jinja/caps.h"
 
-#include "json.h"
+#include "nlohmann/json_fwd.hpp"
 
 #include <chrono>
 #include <functional>
@@ -17,6 +17,7 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
+using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 
@@ -86,7 +87,7 @@ struct common_chat_msg {
     std::string                               tool_name;
     std::string                               tool_call_id;
 
-    common_json to_json_oaicompat(bool concat_typed_text = false) const;
+    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     std::string render_content(const std::string & delimiter = "\n\n") const;
 
@@ -210,7 +211,7 @@ struct common_chat_msg_delimiters {
     // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
     common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
 
-    common_json to_json() const;
+    nlohmann::ordered_json to_json() const;
 };
 
 struct common_chat_tool {
@@ -349,16 +350,16 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const common_json & messages);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const common_json & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
 
-common_chat_continuation common_chat_continuation_parse(const common_json & value);
+common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
 
 // DEPRECATED: only used in tests
-common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
@@ -385,4 +386,4 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-common_chat_msg_delimiters common_chat_msg_delimiters_parse(const common_json & delimiters);
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);

common/common.cpp

@@ -1019,21 +1019,20 @@ std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
         // Make sure to add trailing slash
-        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
+        if (p.back() != DIRECTORY_SEPARATOR) {
             p += DIRECTORY_SEPARATOR;
         }
         return p;
     };
-    cache_directory = common_get_env("LLAMA_CACHE");
-    if (cache_directory.empty()) {
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
         defined(__OpenBSD__) || defined(__NetBSD__)
-        const std::string xdg_cache_home = common_get_env("XDG_CACHE_HOME");
-        const std::string home           = common_get_env("HOME");
-        if (!xdg_cache_home.empty()) {
-            cache_directory = xdg_cache_home;
-        } else if (!home.empty()) {
-            cache_directory = home + "/.cache/";
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else if (std::getenv("HOME")) {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
         } else {
 #if defined(__linux__)
             /* no $HOME is defined, fallback to getpwuid */
@@ -1048,16 +1047,9 @@ std::string fs_get_cache_directory() {
 #endif /* defined(__linux__) */
         }
 #elif defined(__APPLE__)
-        cache_directory = common_get_env("HOME");
-        if (cache_directory.empty()) {
-            throw std::runtime_error("Failed to find $HOME directory");
-        }
-        cache_directory += "/Library/Caches/";
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
-        cache_directory = common_get_env("LOCALAPPDATA");
-        if (cache_directory.empty()) {
-            throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
-        }
+        cache_directory = std::getenv("LOCALAPPDATA");
 #elif defined(__EMSCRIPTEN__)
         GGML_ABORT("not implemented on this platform");
 #else
@@ -1069,51 +1061,6 @@ std::string fs_get_cache_directory() {
     return ensure_trailing_slash(cache_directory);
 }
 
-std::string fs_get_config_directory() {
-    std::string config_directory = "";
-    auto ensure_trailing_slash = [](std::string p) {
-        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
-            p += DIRECTORY_SEPARATOR;
-        }
-        return p;
-    };
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
-        defined(__OpenBSD__) || defined(__NetBSD__) || defined(__APPLE__)
-    const std::string xdg_config_home = common_get_env("XDG_CONFIG_HOME");
-    const std::string home            = common_get_env("HOME");
-    if (!xdg_config_home.empty()) {
-        config_directory = xdg_config_home;
-    } else if (!home.empty()) {
-        config_directory = home + "/.config/";
-    } else {
-#if defined(__linux__)
-        /* no $HOME is defined, fallback to getpwuid */
-        struct passwd *pw = getpwuid(getuid());
-        if ((!pw) || (!pw->pw_dir)) {
-            throw std::runtime_error("Failed to find $HOME directory");
-        }
-
-        config_directory = std::string(pw->pw_dir) + std::string("/.config/");
-#else
-        throw std::runtime_error("Failed to find $HOME directory");
-#endif
-    }
-#elif defined(_WIN32)
-    config_directory = common_get_env("APPDATA");
-    if (config_directory.empty()) {
-        throw std::runtime_error("Failed to find %APPDATA% directory");
-    }
-#elif defined(__EMSCRIPTEN__)
-    // caller decides what to do when there is no config directory
-    throw std::runtime_error("not implemented on this platform");
-#else
-#  error Unknown architecture
-#endif
-    config_directory = ensure_trailing_slash(config_directory);
-    config_directory += "llama.cpp";
-    return ensure_trailing_slash(config_directory);
-}
-
 std::string fs_get_cache_file(const std::string & filename) {
     GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
     std::string cache_directory = fs_get_cache_directory();
@@ -1275,8 +1222,6 @@ struct common_init_result::impl {
 
     // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
 
-    common_threadpools threadpools;
-
     llama_model_ptr   model;
     llama_context_ptr context;
 
@@ -1294,34 +1239,11 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     if (params.fit_params) {
         COM_TRC("%s", "fitting params to device memory ...\n");
         COM_TRC("%s", "(for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n");
-
-        // the draft context is created from the same base params and follows the main context, fit both together
-        const bool has_draft = params.speculative.has_dft();
-        const bool spec_mtp  = std::find(params.speculative.types.begin(), params.speculative.types.end(),
-            COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-
-        common_params params_dft = common_base_params_to_speculative(params);
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-        auto cparams_dft = common_context_params_to_llama(params_dft);
-        if (spec_mtp) {
-            cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-        }
-        cparams_dft.n_rs_seq = 0;
-
-        const common_fit_extra_model extra = {
-            /*.path_model   =*/ params_dft.model.path.c_str(),
-            /*.mparams      =*/ &mparams_dft,
-            /*.cparams      =*/ &cparams_dft,
-            /*.shares_model =*/ !has_draft, // an MTP context runs on the weights of the main model
-        };
-
         common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
-            has_draft || spec_mtp ? &extra : nullptr,
             params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1401,10 +1323,6 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     }
 
     pimpl->context.reset(lctx);
-
-    set_process_priority(params.cpuparams.priority);
-
-    pimpl->threadpools.init(lctx, params);
 }
 
 llama_model * common_init_result::model() {
@@ -1721,7 +1639,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_seq_max         = params.n_parallel;
     cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
     cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
-    cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
@@ -1753,10 +1670,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     return cparams;
 }
 
-//
-// Threadpool utils
-//
-
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params) {
     struct ggml_threadpool_params tpp;
 
@@ -1773,58 +1686,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const commo
     return tpp;
 }
 
-common_threadpools::~common_threadpools() {
-    if (!free_fn) {
-        return;
-    }
-    free_fn(threadpool);
-    free_fn(threadpool_batch);
-}
-
-void common_threadpools::init(llama_context * ctx, const common_params & params) {
-    GGML_ASSERT(!threadpool);
-    GGML_ASSERT(!threadpool_batch);
-
-    COM_INF("llama threadpool init, n_threads = %d\n", (int) params.cpuparams.n_threads);
-
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (!cpu_dev) {
-        COM_WRN("%s", "no CPU backend found\n");
-        return;
-    }
-    auto * reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
-    free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
-
-    struct ggml_threadpool_params tpp_batch =
-            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
-    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) {
-            COM_WRN("batch threadpool create failed : n_threads %d\n", tpp_batch.n_threads);
-            return;
-        }
-
-        // start the non-batch threadpool in the paused state
-        tpp.paused = true;
-    }
-
-    threadpool = ggml_threadpool_new_fn(&tpp);
-    if (!threadpool) {
-        COM_WRN("threadpool create failed : n_threads %d\n", tpp.n_threads);
-        free_fn(threadpool_batch);
-        threadpool_batch = nullptr;
-        return;
-    }
-
-    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
-}
-
 //
 // Batch utils
 //

common/common.h

@@ -447,7 +447,6 @@ struct common_params {
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
     int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
-    int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -473,7 +472,7 @@ struct common_params {
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
+    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
 
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;
@@ -581,10 +580,9 @@ struct common_params {
 
     // multimodal models (see tools/mtmd)
     struct common_params_model mmproj;
-    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"
+    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"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
     int mtmd_batch_max_tokens = 1024;
@@ -657,7 +655,6 @@ struct common_params {
 
     // enable built-in tools
     std::vector<std::string> server_tools;
-    std::string server_tools_runtime;
 
     // MCP server configs (Cursor-compatible JSON)
     std::string mcp_servers_config;   // path to JSON file with MCP server definitions
@@ -882,7 +879,6 @@ bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
-std::string fs_get_config_directory();
 
 struct common_file_info {
     std::string path;
@@ -930,8 +926,9 @@ using common_init_result_ptr = std::unique_ptr<common_init_result>;
 
 common_init_result_ptr common_init_from_params(common_params & params, bool model_only = false);
 
-struct llama_model_params   common_model_params_to_llama  (      common_params & params);
-struct llama_context_params common_context_params_to_llama(const common_params & params);
+struct llama_model_params     common_model_params_to_llama  (      common_params & params);
+struct llama_context_params   common_context_params_to_llama(const common_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
@@ -942,28 +939,6 @@ std::string common_get_model_endpoint();
 // for testing purposes
 char * common_get_model_or_exit(int, char*[]);
 
-//
-// Threadpool utils
-//
-
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
-
-struct common_threadpools {
-    common_threadpools() = default;
-    ~common_threadpools();
-
-    common_threadpools(const common_threadpools &) = delete;
-    common_threadpools & operator=(const common_threadpools &) = delete;
-
-    void init(llama_context * ctx, const common_params & params);
-
-private:
-    ggml_threadpool * threadpool       = nullptr;
-    ggml_threadpool * threadpool_batch = nullptr;
-
-    decltype(ggml_threadpool_free) * free_fn = nullptr;
-};
-
 //
 // Context utils
 //

common/download.cpp

@@ -5,7 +5,9 @@
 #include "log.h"
 #include "download.h"
 #include "hf-cache.h"
-#include "json.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
 
 #include <algorithm>
 #include <filesystem>
@@ -42,6 +44,8 @@
 #include <unistd.h>
 #endif
 
+using json = nlohmann::ordered_json;
+
 //
 // downloader
 //
@@ -852,8 +856,8 @@ static std::string common_docker_get_token(const std::string & repo) {
         throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
     }
 
-    std::string response_str(res.second.begin(), res.second.end());
-    common_json response = common_json::parse(response_str);
+    std::string            response_str(res.second.begin(), res.second.end());
+    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
 
     if (!response.contains("token")) {
         throw std::runtime_error("Docker registry token response missing 'token' field");
@@ -915,9 +919,9 @@ std::string common_docker_resolve_model(const std::string & docker) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
         }
 
-        std::string manifest_str(manifest_res.second.begin(), manifest_res.second.end());
-        common_json manifest = common_json::parse(manifest_str);
-        std::string gguf_digest;  // Find the GGUF layer
+        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
+        std::string            gguf_digest;  // Find the GGUF layer
         if (manifest.contains("layers")) {
             for (const auto & layer : manifest["layers"]) {
                 if (layer.contains("mediaType")) {
@@ -985,26 +989,6 @@ 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;
 

common/download.h

@@ -85,10 +85,6 @@ 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)

common/fit.cpp

@@ -178,7 +178,7 @@ common_device_memory_data_vec common_get_device_memory_data(
 static void common_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t * margins_s, uint32_t n_ctx_min, const common_fit_extra_model * extra, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
         throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
     }
@@ -191,92 +191,10 @@ static void common_params_fit_impl(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
-    // with non-unified kv, we need to take into account n_streams
-    // for example, if memory can hold more than model's trained context size, we must extend the n_ctx to hold enough n_streams
-    const uint32_t n_streams  = cparams->kv_unified ? 1 : std::max<uint32_t>(1, cparams->n_seq_max);
-    const bool     n_ctx_auto = cparams->n_ctx == 0;
-
-    dmds_t   dmds_extra;       // memory of the extra model, laid out on the devices of the main model
-    uint32_t n_ctx_extra = 0;  // context that memory was measured at
-
-    // the extra model competes for the same memory as the main model, add it to every measurement
-    // its memory is measured again whenever the context it follows changes
-    auto add_extra_memory = [&](dmds_t & dmds) {
-        if (extra == nullptr) {
-            return;
-        }
-
-        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
-            std::vector<ggml_backend_dev_t> devs_extra;
-            uint32_t ngl_extra = 0;
-            uint32_t nct_extra = 0;
-            uint32_t nex_extra = 0;
-
-            extra->cparams->n_ctx = cparams->n_ctx;
-
-            LOG_TRC("%s: getting device memory data for the extra model at a context size of %" PRIu32 ":\n",
-                __func__, cparams->n_ctx);
-
-            dmds_t measured;
-            try {
-                measured = common_get_device_memory_data_impl(
-                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
-            } catch (const std::runtime_error & e) {
-                // the extra model is optional, fit the main model alone rather than giving up
-                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
-                dmds_extra = dmds_t(devs.size() + 1);
-                n_ctx_extra = cparams->n_ctx;
-                return;
-            }
-
-            dmds_extra = dmds_t(devs.size() + 1);
-            dmds_extra.back().mb = measured.back().mb;
-            for (size_t je = 0; je < devs_extra.size(); je++) {
-                for (size_t id = 0; id < devs.size(); id++) {
-                    if (devs_extra[je] == devs[id]) {
-                        dmds_extra[id].mb.model   += measured[je].mb.model;
-                        dmds_extra[id].mb.context += measured[je].mb.context;
-                        dmds_extra[id].mb.compute += measured[je].mb.compute;
-                        break;
-                    }
-                }
-            }
-            if (extra->shares_model) {
-                for (llama_device_memory_data & dmd : dmds_extra) {
-                    dmd.mb.model = 0;
-                }
-            }
-
-            n_ctx_extra = cparams->n_ctx;
-        }
-
-        for (size_t id = 0; id < dmds.size(); id++) {
-            dmds[id].mb.model   += dmds_extra[id].mb.model;
-            dmds[id].mb.context += dmds_extra[id].mb.context;
-            dmds[id].mb.compute += dmds_extra[id].mb.compute;
-        }
-    };
-
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-
-    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
-    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
-    const uint32_t n_ctx_min_total = (uint32_t) std::min<uint64_t>(uint64_t(n_ctx_min) * n_streams, UINT32_MAX);
-
-    // llama_context would use only hp_nct in total for n_ctx == 0, resolve the context before measuring anything else:
-    if (n_ctx_auto) {
-        cparams->n_ctx = n_ctx_max;
-        if (n_streams > 1) {
-            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
-                __func__, n_ctx_max, n_streams);
-            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        }
-    }
-    add_extra_memory(dmds_full);
-
+    const dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
     const size_t nd = devs.size(); // number of devices
 
     std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
@@ -389,8 +307,8 @@ static void common_params_fit_impl(
                     "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
                     __func__, -global_surplus/MiB);
             }
-            if (n_ctx_auto) {
-                if (n_ctx_max > n_ctx_min_total) {
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
                     int64_t sum_used_target = sum_free;
                     if (nd == 0) {
                         sum_used_target -= margins[0];
@@ -410,9 +328,8 @@ static void common_params_fit_impl(
                     }
 
                     int64_t sum_projected_used_min_ctx = 0;
-                    cparams->n_ctx = n_ctx_min_total;
-                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-                    add_extra_memory(dmds_min_ctx);
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
                     if (nd == 0) {
                         sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
                     } else {
@@ -422,16 +339,14 @@ static void common_params_fit_impl(
                     }
                     if (sum_used_target > sum_projected_used_min_ctx) {
                         // linear interpolation between minimum and maximum context size:
-                        cparams->n_ctx += (n_ctx_max - n_ctx_min_total) * (sum_used_target - sum_projected_used_min_ctx)
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
                             / (sum_projected_used - sum_projected_used_min_ctx);
-                        // round down context for CUDA backend, keep it divisible by the number of streams:
-                        const uint32_t align = 256 * n_streams;
-                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % align, n_ctx_min_total);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
 
-                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (n_ctx_max - n_ctx_min_total);
-                        const int64_t memory_reduction = (n_ctx_max - cparams->n_ctx) * bytes_per_ctx;
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd <= 1) {
                             LOG_TRC("%s: entire model can be fit by reducing context\n", __func__);
                             return;
@@ -440,14 +355,14 @@ static void common_params_fit_impl(
                     } else {
                         const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     if (n_ctx_min == UINT32_MAX) {
-                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, n_ctx_max);
+                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
                     } else {
                         LOG_TRC("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                            __func__, n_ctx_max, n_ctx_min_total);
+                            __func__, hp_nct, n_ctx_min);
                     }
                 }
             } else {
@@ -592,9 +507,8 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        dmds_t dmd_nl = common_get_device_memory_data_impl(
+        const dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmd_nl);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
         for (size_t id = 0; id < nd; id++) {
@@ -621,9 +535,8 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmds_cpu_moe);
 
         for (size_t id = 0; id < nd; id++) {
             global_surplus_cpu_moe += dmds_cpu_moe[id].free;
@@ -883,12 +796,11 @@ enum common_params_fit_status common_fit_params(
         llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t * margins,
         uint32_t n_ctx_min,
-        const common_fit_extra_model * extra,
         ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
     common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, extra, log_level);
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LOG_TRC("%s: successfully fit params to free device memory\n", __func__);
     } catch (const common_params_fit_exception & e) {
         LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());

common/fit.h

@@ -11,16 +11,6 @@ enum common_params_fit_status {
     COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
 };
 
-// a second model that shares the devices of the main model, e.g. a draft model
-//   - its context follows the context of the main model, so its memory is measured again whenever that context changes
-//   - shares_model tells the fit that the weights are already counted in the main model, as for an MTP context
-struct common_fit_extra_model {
-    const char * path_model;
-    llama_model_params * mparams;
-    llama_context_params * cparams;
-    bool shares_model;
-};
-
 // fits mparams and cparams to free device memory (assumes system memory is unlimited)
 //   - returns true if the parameters could be successfully modified to fit device memory
 //   - this function is NOT thread safe because it modifies the global llama logger state
@@ -34,7 +24,6 @@ common_params_fit_status common_fit_params(
    llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
                              size_t * margins,               // margins of memory to leave per device in bytes
                            uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
-      const common_fit_extra_model * extra,                  // model to fit alongside the main one, nullptr if there is none
                      ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout

common/hf-cache.cpp

@@ -4,7 +4,9 @@
 #include "common.h"
 #include "log.h"
 #include "http.h"
-#include "json.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
 
 #include <filesystem>
 #include <fstream>
@@ -13,6 +15,8 @@
 #include <string_view>
 #include <stdexcept>
 
+namespace nl = nlohmann;
+
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
@@ -191,8 +195,8 @@ static void safe_write_file(const fs::path & path, const std::string & data) {
     }
 }
 
-static common_json api_get(const std::string & url,
-                           const std::string & token) {
+static nl::json api_get(const std::string & url,
+                        const std::string & token) {
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
@@ -210,10 +214,10 @@ static common_json api_get(const std::string & url,
         auto body = res->body;
 
         if (res->status == 200) {
-            return common_json::parse(res->body);
+            return nl::json::parse(res->body);
         }
         try {
-            body = common_json::parse(res->body)["error"].get<std::string>();
+            body = nl::json::parse(res->body)["error"].get<std::string>();
         } catch (...) { }
 
         throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
@@ -276,7 +280,7 @@ static std::string get_repo_commit(const std::string & repo_id,
         safe_write_file(refs_path / name, commit);
         return commit;
 
-    } catch (const common_json_error & e) {
+    } catch (const nl::json::exception & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
@@ -354,7 +358,7 @@ hf_files get_repo_files(const std::string & repo_id,
 
             files.push_back(file);
         }
-    } catch (const common_json_error & e) {
+    } catch (const nl::json::exception & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());

common/imatrix-loader.cpp

@@ -102,8 +102,7 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) {
     const int64_t chunk_count_key = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT);
     const int64_t chunk_size_key  = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE);
 
-    if (datasets_key != -1 && gguf_get_kv_type(ctx_gguf, datasets_key) == GGUF_TYPE_ARRAY &&
-        gguf_get_arr_type(ctx_gguf, datasets_key) == GGUF_TYPE_STRING) {
+    if (datasets_key != -1 && gguf_get_arr_type(ctx_gguf, datasets_key) == GGUF_TYPE_STRING) {
         const int64_t n = gguf_get_arr_n(ctx_gguf, datasets_key);
         imatrix.datasets.reserve(imatrix.datasets.size() + n);
         for (int64_t i = 0; i < n; ++i) {
@@ -144,13 +143,6 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) {
             return false;
         }
 
-        if (in_sum2->type != GGML_TYPE_F32 || counts->type != GGML_TYPE_F32) {
-            LOG_ERR("%s: sums and counts for %s must be F32\n", __func__, name.c_str());
-            gguf_free(ctx_gguf);
-            ggml_free(ctx);
-            return false;
-        }
-
         auto & e = imatrix.entries[name];
 
         const int64_t nval    = ggml_nelements(in_sum2);

common/jinja/README.md

@@ -7,7 +7,7 @@ The implementation can be found in the `common/jinja` directory.
 ## Key Features
 
 - Input marking: security against special token injection
-- Decoupled from the JSON library: `common_json` is only used for JSON-to-internal type translation and is completely optional
+- Decoupled from `nlohmann::json`: this dependency is only used for JSON-to-internal type translation and is completely optional
 - Minimal primitive types: int, float, bool, string, array, object, none, undefined
 - Detailed logging: allow source tracing on error
 - Clean architecture: workarounds are applied to input data before entering the runtime (see `common/chat.cpp`)

common/jinja/caps.cpp

@@ -4,32 +4,26 @@
 
 // note: the json dependency is only for defining input in a convenient way
 // we can remove it in the future when we figure out a better way to define inputs using jinja::value
-#include "json.h"
+#include <nlohmann/json.hpp>
 
 #include <functional>
 #include <sstream>
 
 #define FILENAME "jinja-caps"
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 namespace jinja {
 
 using caps_json_fn = std::function<json()>;
 using caps_ctx_fn = std::function<void(context &)>;
-using caps_analyze_fn = std::function<void(context &, bool, value &, value &, const std::string &)>;
+using caps_analyze_fn = std::function<void(bool, value &, value &, const std::string &)>;
 
 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));
-}
-
-void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort) {
-    value var = mk_val<value_string>(effort); // bind to the same value for stats
-    ctx.set_val("reasoning_effort",   var);
-    ctx.set_val("reasoning_strength", var);
+    ctx.set_val("drop_thinking",            mk_val<value_bool>(!enabled));
 }
 
 static void caps_try_execute(jinja::program & prog,
@@ -68,7 +62,7 @@ static void caps_try_execute(jinja::program & prog,
         // ignore exceptions during capability analysis
     }
 
-    analyze_fn(ctx, success, messages, tools, result);
+    analyze_fn(success, messages, tools, result);
 }
 
 // for debugging only
@@ -93,7 +87,6 @@ std::map<std::string, bool> caps::to_map() const {
         {"supports_parallel_tool_calls", supports_parallel_tool_calls},
         {"supports_system_role", supports_system_role},
         {"supports_preserve_reasoning", supports_preserve_reasoning},
-        {"supports_reasoning_effort", supports_reasoning_effort},
         {"supports_object_arguments", supports_object_arguments},
     };
 }
@@ -117,8 +110,6 @@ 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,
@@ -127,26 +118,22 @@ caps caps_get(jinja::program & prog) {
             return json::array({
                 {
                     {"role", "user"},
-                    {"content", content_marker}
+                    {"content", "content"}
                 }
             });
         },
         nullptr, // ctx_fn
         nullptr, // tools_fn
-        [&](context &, bool success, value & messages, value &, const std::string & rendered) {
+        [&](bool success, value & messages, value &, const std::string &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
-            bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
-            if (used_as_array) {
+            if (has_op(content, "selectattr") || has_op(content, "array_access")) {
                 // 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;
             }
         }
     );
@@ -171,7 +158,7 @@ caps caps_get(jinja::program & prog) {
         },
         nullptr, // ctx_fn
         nullptr, // tools_fn
-        [&](context &, bool, value & messages, value &, const std::string &) {
+        [&](bool, value & messages, value &, const std::string &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
             if (!content->stats.used) {
@@ -247,7 +234,7 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        [&](context &, bool success, value & messages, value & tools, const std::string &) {
+        [&](bool success, value & messages, value & tools, const std::string &) {
             if (!success) {
                 return; // Nothing can be inferred
             }
@@ -340,7 +327,7 @@ caps caps_get(jinja::program & prog) {
                     },
                 });
             },
-            [&](context &, bool success, value & messages, value & tools, const std::string &) {
+            [&](bool success, value & messages, value & tools, const std::string &) {
                 if (!success) {
                     result.supports_tool_calls = false;
                     result.supports_tools = false;
@@ -442,7 +429,7 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        [&](context &, bool success, value & messages, value &, const std::string &) {
+        [&](bool success, value & messages, value &, const std::string &) {
             if (!success) {
                 result.supports_parallel_tool_calls = false;
                 return;
@@ -499,7 +486,7 @@ caps caps_get(jinja::program & prog) {
             caps_apply_preserve_reasoning(ctx, true);
         },
         nullptr, // tools_fn
-        [&](context &, bool, value &, value &, const std::string & output) {
+        [&](bool, value &, value &, const std::string & output) {
             // note: we cannot use stats here because the reasoning_content may be used for "if" condition test, but not actually outputted in the final result
             if (output.find(reasoning_placeholder) != std::string::npos) {
                 result.supports_preserve_reasoning = true;
@@ -507,32 +494,6 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
-    JJ_DEBUG("%s\n", ">>> Running capability check: reasoning effort");
-
-    // case: reasoning effort level
-    caps_try_execute(
-        prog,
-        [&]() {
-            // messages
-            return json::array({
-                {
-                    {"role", "user"},
-                    {"content", "User message"}
-                },
-            });
-        },
-        [&](context & ctx) {
-            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
-            caps_apply_reasoning_effort(ctx, "low");
-        },
-        nullptr, // tools_fn
-        [&](context & ctx, bool, value &, value &, const std::string &) {
-            value effort = ctx.get_val("reasoning_effort");
-            caps_print_stats(effort, "reasoning_effort");
-            result.supports_reasoning_effort = effort->stats.used;
-        }
-    );
-
     JJ_DEBUG("%s\n", result.to_string().c_str());
 
     return result;

common/jinja/caps.h

@@ -16,9 +16,6 @@ struct caps {
     // supports preserve reasoning trace in the full history, not just the last assistant message
     bool supports_preserve_reasoning = false;
 
-    // supports reasoning effort levels
-    bool supports_reasoning_effort = false;
-
     // one of the 2 content capabilities must be true
     bool supports_string_content = true;
     bool supports_typed_content = false;
@@ -35,6 +32,5 @@ struct caps {
 caps caps_get(jinja::program & prog);
 
 void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled);
-void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort);
 
 } // namespace jinja

common/jinja/runtime.cpp

@@ -263,7 +263,7 @@ value binary_expression::execute_impl(context & ctx) {
             return res;
         }
         for (int64_t i = 0; i < repeat; ++i) {
-            res->val_str.append(str);
+            res->val_str = res->val_str.append(str);
         }
         return res;
     }

common/jinja/runtime.h

@@ -763,22 +763,14 @@ struct runtime {
         gather_string_parts_recursive(val, parts);
         // join consecutive parts with the same type
         auto & p = parts->val_str.parts;
-        if (p.empty()) {
-            return parts;
-        }
-        size_t w = 0;
-        for (size_t r = 1; r < p.size(); r++) {
-            if (p[w].is_input == p[r].is_input) {
-                p[w].val += p[r].val;
+        for (size_t i = 1; i < p.size(); ) {
+            if (p[i].is_input == p[i - 1].is_input) {
+                p[i - 1].val += p[i].val;
+                p.erase(p.begin() + i);
             } else {
-                w++;
-                if (w != r) {
-                    // the guard is needed, self-move leaves the string in an unspecified state
-                    p[w] = std::move(p[r]);
-                }
+                i++;
             }
         }
-        p.resize(w + 1);
         return parts;
     }
 

common/jinja/string.cpp

@@ -103,7 +103,7 @@ void string::mark_input_based_on(const string & other) {
     }
 }
 
-string & string::append(const string & other) {
+string string::append(const string & other) {
     for (const auto & part : other.parts) {
         parts.push_back(part);
     }

common/jinja/string.h

@@ -47,7 +47,7 @@ struct string {
     // mark this string as input if other has ALL parts as input
     void mark_input_based_on(const string & other);
 
-    string & append(const string & other);
+    string append(const string & other);
 
     // in-place transformations
 

common/jinja/value.cpp

@@ -3,7 +3,7 @@
 #include "value.h"
 
 // for converting from JSON to jinja values
-#include "json.h"
+#include <nlohmann/json.hpp>
 
 #include <sstream>
 #include <string>
@@ -1355,7 +1355,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const common_json & j, bool mark_input) {
+static value from_json(const nlohmann::ordered_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1452,7 +1452,7 @@ bool value_compare(const value & a, const value & b, value_compare_op op) {
 }
 
 template<>
-void global_from_json(context & ctx, const common_json & json_obj, bool mark_input) {
+void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bool mark_input) {
     // printf("global_from_json: %s\n" , json_obj.dump(2).c_str());
     if (json_obj.is_null() || !json_obj.is_object()) {
         throw std::runtime_error("global_from_json: input JSON value must be an object");

common/jinja/value.h

@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be common_json
+// Note: T_JSON can be nlohmann::ordered_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 

common/json-schema-to-grammar.cpp

@@ -1,8 +1,9 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
-#include <limits>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -11,7 +12,7 @@
 #include <unordered_set>
 #include <vector>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
@@ -277,9 +278,7 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
     {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
-static const int MAX_PATTERN_DEPTH = 100;
-
-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -310,32 +309,6 @@ 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;
@@ -372,42 +345,16 @@ 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) {
-        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());
+        if (!(pattern.front() == '^' && pattern.back() == '$')) {
+            _errors.push_back("Pattern must start with '^' and end with '$'");
             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) {
@@ -416,6 +363,7 @@ 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 = [&]() {
@@ -472,42 +420,43 @@ private:
                         if (i + 1 < length && sub_pattern[i + 1] == ':') {
                             i += 2; // skip "?:" for non-capturing group, treat as regular group
                         } else {
-                            // lookaround, named group, inline flags, ...
-                            throw unsupported_pattern("unsupported group syntax");
+                            // 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;
                         }
                     }
-                    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 (paren_depth == 0) {
-                        throw invalid_pattern("unbalanced parentheses");
+                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
+                        _errors.push_back("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] == '\\') {
-                            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;
+                            square_brackets += sub_pattern.substr(i, 2);
+                            i += 2;
                         } else {
                             square_brackets += sub_pattern[i];
                             i++;
                         }
                     }
                     if (i >= length) {
-                        throw invalid_pattern("unterminated character class");
+                        _errors.push_back("Unbalanced square brackets");
                     }
                     square_brackets += ']';
                     i++;
@@ -516,9 +465,6 @@ 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 == '{') {
@@ -529,19 +475,18 @@ private:
                         i++;
                     }
                     if (i >= length) {
-                        throw unsupported_pattern("unterminated curly brackets");
+                        _errors.push_back("Unbalanced 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]);
@@ -550,11 +495,9 @@ private:
                                 max_times = std::stoi(nums[1]);
                             }
                         }
-                    } catch (const std::logic_error &) {
-                        throw unsupported_pattern("invalid number in curly brackets");
-                    }
-                    if (seq.empty()) {
-                        throw invalid_pattern("nothing to repeat");
+                    } catch (const std::invalid_argument & e) {
+                        _errors.push_back("Invalid number in curly brackets");
+                        return std::make_pair("", false);
                     }
                     auto &last = seq.back();
                     auto &sub = last.first;
@@ -580,22 +523,15 @@ private:
                         return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                     };
                     while (i < length) {
-                        if (sub_pattern[i] == '\\') {
-                            if (i == length - 1) {
-                                throw invalid_pattern("trailing backslash");
-                            }
+                        if (sub_pattern[i] == '\\' && i < length - 1) {
                             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 {
-                                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;
+                                literal += sub_pattern.substr(i, 2);
+                                i += 2;
                             }
                         } else if (sub_pattern[i] == '"') {
                             literal += "\\\"";
@@ -608,21 +544,14 @@ private:
                             break;
                         }
                     }
-                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
-                        throw unsupported_pattern(std::string("unsupported character: ") + c);
+                    if (!literal.empty()) {
+                        seq.emplace_back(literal, true);
                     }
-                    seq.emplace_back(literal, true);
                 }
             }
             return join_seq();
         };
-
-        auto rule = to_rule(transform());
-        if (paren_depth != 0) {
-            throw invalid_pattern("unbalanced parentheses");
-        }
-
-        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
+        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
     }
 
     /*
@@ -916,11 +845,7 @@ public:
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
         }
         if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
-            std::vector<json> alt_schemas;
-            for (const auto & alt : alts) {
-                alt_schemas.push_back(alt);
-            }
+            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
         }
         if (schema_type.is_array()) {
@@ -1114,7 +1039,7 @@ common_schema_info::~common_schema_info() = default;
 common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
 common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
 
-void common_schema_info::resolve_refs(common_json & schema) {
+void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
     impl_->resolve_refs(schema, "");
 }
 
@@ -1122,7 +1047,7 @@ void common_schema_info::resolve_refs(common_json & schema) {
 // Some models emit raw string values rather than JSON-encoded strings for string parameters.
 // If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
 // true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const common_json & schema) {
+bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
     std::unordered_set<std::string> visited_refs;
 
     std::function<bool(const json &)> check = [&](const json & s) -> bool {
@@ -1230,7 +1155,7 @@ bool common_schema_info::resolves_to_string(const common_json & schema) {
     return check(schema);
 }
 
-std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
+std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
         return "%llguidance {}\nstart: %json " + schema.dump();
@@ -1251,10 +1176,10 @@ std::string build_grammar(const std::function<void(const common_grammar_builder
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
+        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
             return converter.visit(schema, name == "root" ? "" : name);
         },
-        /* .resolve_refs = */ [&](common_json & schema) {
+        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
             converter.resolve_refs(schema, "");
         }
     };

common/json-schema-to-grammar.h

@@ -1,12 +1,12 @@
 #pragma once
 
-#include "json.h"
+#include <nlohmann/json_fwd.hpp>
 
 #include <functional>
 #include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const common_json & schema,
+std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);
 
 class common_schema_converter;
@@ -24,14 +24,14 @@ class common_schema_info {
     common_schema_info(common_schema_info &&) noexcept;
     common_schema_info & operator=(common_schema_info &&) noexcept;
 
-    void resolve_refs(common_json & schema);
-    bool resolves_to_string(const common_json & schema);
+    void resolve_refs(nlohmann::ordered_json & schema);
+    bool resolves_to_string(const nlohmann::ordered_json & schema);
 };
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const common_json &)> add_schema;
-    std::function<void(common_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
+    std::function<void(nlohmann::ordered_json &)> resolve_refs;
 };
 
 struct common_grammar_options {

common/json.cpp

@@ -1,437 +0,0 @@
-#include "json.h"
-
-#include "ggml.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
-
-#include <iterator>
-#include <new>
-#include <set>
-#include <unordered_map>
-#include <vector>
-
-using nlohmann::ordered_json;
-
-// a common_json is the backing value, so any value of a tree can be used as a common_json
-static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
-static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
-
-// runs fn and gives every error of the backing library as a common_json_error
-template <typename F>
-static decltype(auto) guard(F && fn) {
-    try {
-        return fn();
-    } catch (const ordered_json::exception & e) {
-        throw common_json_error(e.what());
-    }
-}
-
-static ordered_json & as_json(common_json * self) {
-    return *reinterpret_cast<ordered_json *>(self);
-}
-
-static const ordered_json & as_json(const common_json * self) {
-    return *reinterpret_cast<const ordered_json *>(self);
-}
-
-static common_json & as_common(ordered_json & json) {
-    return *reinterpret_cast<common_json *>(&json);
-}
-
-static const common_json & as_common(const ordered_json & json) {
-    return *reinterpret_cast<const common_json *>(&json);
-}
-
-static ordered_json to_json(const common_json_value & val) {
-    switch (val.type) {
-        case common_json_value::VAL_NULL:   return nullptr;
-        case common_json_value::VAL_BOOL:   return val.val_bool;
-        case common_json_value::VAL_INT:    return val.val_int;
-        case common_json_value::VAL_UINT:   return val.val_uint;
-        case common_json_value::VAL_DOUBLE: return val.val_double;
-        case common_json_value::VAL_STRING: return val.val_string;
-        case common_json_value::VAL_JSON:
-            // one owner means no one else can see this tree, so it is safe to move it out
-            // note: this makes a value single use, same as the json_ref of the backing library
-            if (val.val_json.use_count() == 1) {
-                return std::move(as_json(val.val_json.get()));
-            }
-            return as_json(val.val_json.get());
-    }
-
-    return nullptr;
-}
-
-common_json_value::common_json_value(const char * val) {
-    if (val) {
-        type       = VAL_STRING;
-        val_string = val;
-    } else {
-        type = VAL_NULL;
-    }
-}
-
-common_json_value::common_json_value(const common_json & val) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
-
-common_json_value::common_json_value(common_json && val) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
-
-template <typename T>
-common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::array();
-
-    for (const auto & val : vals) {
-        out.push_back(val);
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
-
-COMMON_JSON_SET(int)
-COMMON_JSON_SET(std::string)
-
-#undef COMMON_JSON_SET
-
-template <typename T>
-common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
-
-COMMON_JSON_MAP(bool)
-COMMON_JSON_MAP(std::string)
-
-#undef COMMON_JSON_MAP
-
-template <typename T>
-common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
-
-COMMON_JSON_UMAP(size_t)
-
-#undef COMMON_JSON_UMAP
-
-template <typename T>
-common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::array();
-
-    for (const auto & val : vals) {
-        out.push_back(val);
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a vector value is usable only for the types below
-// note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
-
-COMMON_JSON_VEC(int)
-COMMON_JSON_VEC(unsigned char)
-COMMON_JSON_VEC(unsigned int)
-COMMON_JSON_VEC(long)
-COMMON_JSON_VEC(unsigned long)
-COMMON_JSON_VEC(long long)
-COMMON_JSON_VEC(unsigned long long)
-COMMON_JSON_VEC(float)
-COMMON_JSON_VEC(double)
-COMMON_JSON_VEC(std::string)
-COMMON_JSON_VEC(std::vector<float>)
-COMMON_JSON_VEC(common_json)
-
-#undef COMMON_JSON_VEC
-
-common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
-
-// null, same as the backing library
-// operator[] turns it into an object, push_back() into an array
-common_json::common_json() {
-    new (storage) ordered_json();
-}
-
-common_json::common_json(const common_json & other) {
-    new (storage) ordered_json(as_json(&other));
-}
-
-common_json::common_json(common_json && other) noexcept {
-    new (storage) ordered_json(std::move(as_json(&other)));
-}
-
-common_json::common_json(std::initializer_list<common_json_item> items) {
-    new (storage) ordered_json(ordered_json::object());
-
-    for (const auto & item : items) {
-        set(item);
-    }
-}
-
-common_json::common_json(const common_json_value & val) {
-    new (storage) ordered_json(to_json(val));
-}
-
-common_json::common_json(std::nullptr_t) {
-    new (storage) ordered_json(nullptr);
-}
-
-common_json & common_json::operator=(common_json other) noexcept {
-    as_json(this).swap(as_json(&other));
-
-    return *this;
-}
-
-common_json::~common_json() {
-    as_json(this).~basic_json();
-}
-
-common_json common_json::parse(const std::string & text) {
-    try {
-        // the assignment moves the parsed tree in, it does not copy
-        common_json out;
-        as_json(&out) = ordered_json::parse(text);
-        return out;
-    } catch (const std::exception & e) {
-        throw common_json_error(e.what());
-    }
-}
-
-common_json common_json::parse_no_throw(const std::string & text) {
-    common_json out;
-    as_json(&out) = ordered_json::parse(text, nullptr, false);
-    return out;
-}
-
-bool common_json::is_discarded() const {
-    return as_json(this).is_discarded();
-}
-
-common_json common_json::array() {
-    common_json out;
-    as_json(&out) = ordered_json::array();
-    return out;
-}
-
-common_json common_json::array(std::initializer_list<common_json_value> vals) {
-    common_json out;
-    ordered_json & arr = as_json(&out);
-    arr = ordered_json::array();
-
-    for (const auto & val : vals) {
-        arr.push_back(to_json(val));
-    }
-
-    return out;
-}
-
-common_json common_json::object() {
-    common_json out;
-    as_json(&out) = ordered_json::object();
-    return out;
-}
-
-common_json common_json::object(std::initializer_list<common_json_item> items) {
-    return common_json(items);
-}
-
-common_json common_json::make(const common_json_value & val) {
-    return common_json(val);
-}
-
-bool common_json::is_null()           const { return as_json(this).is_null(); }
-bool common_json::is_object()         const { return as_json(this).is_object(); }
-bool common_json::is_array()          const { return as_json(this).is_array(); }
-bool common_json::is_string()         const { return as_json(this).is_string(); }
-bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
-bool common_json::is_number()         const { return as_json(this).is_number(); }
-bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
-bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
-
-bool   common_json::empty() const { return as_json(this).empty(); }
-size_t common_json::size()  const { return as_json(this).size(); }
-
-bool common_json::contains(const std::string & key) const {
-    return as_json(this).contains(key);
-}
-
-bool common_json::operator==(const common_json_value & val) const {
-    // compare a tree in place, to_json() would copy it
-    if (val.type == common_json_value::VAL_JSON) {
-        return as_json(this) == as_json(val.val_json.get());
-    }
-    return as_json(this) == to_json(val);
-}
-
-bool common_json::operator!=(const common_json_value & val) const {
-    return !(*this == val);
-}
-
-common_json       & common_json::at(const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this).at(key)); }); }
-const common_json & common_json::at(const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
-common_json       & common_json::at(size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this).at(idx)); }); }
-const common_json & common_json::at(size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
-
-common_json       & common_json::operator[](const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this)[key]); }); }
-const common_json & common_json::operator[](const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
-common_json       & common_json::operator[](size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this)[idx]); }); }
-const common_json & common_json::operator[](size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
-
-common_json       & common_json::front()       { return as_common(as_json(this).front()); }
-const common_json & common_json::front() const { return as_common(as_json(this).front()); }
-common_json       & common_json::back()        { return as_common(as_json(this).back()); }
-const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
-
-void common_json::clear() {
-    as_json(this).clear();
-}
-
-void common_json::erase(const std::string & key) {
-    guard([&] { as_json(this).erase(key); });
-}
-
-void common_json::erase(size_t idx) {
-    guard([&] { as_json(this).erase(idx); });
-}
-
-void common_json::assign(const common_json_value & val) {
-    as_json(this) = to_json(val);
-}
-
-void common_json::set(const common_json_item & item) {
-    guard([&] { as_json(this)[item.key] = to_json(item.val); });
-}
-
-void common_json::push_back(const common_json_value & val) {
-    guard([&] { as_json(this).push_back(to_json(val)); });
-}
-
-void common_json::push_back(std::initializer_list<common_json_item> items) {
-    common_json val(items);
-
-    guard([&] { as_json(this).push_back(std::move(as_json(&val))); });
-}
-
-size_t common_json::count(const std::string & key) const {
-    return as_json(this).count(key);
-}
-
-void common_json::insert(const common_json & vals) {
-    guard([&] {
-        ordered_json & self = as_json(this);
-
-        self.insert(self.end(), as_json(&vals).begin(), as_json(&vals).end());
-    });
-}
-
-std::string common_json::dump(int indent) const {
-    return guard([&] { return as_json(this).dump(indent); });
-}
-
-std::string common_json::dump_safe(int indent) const {
-    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
-}
-
-// an array is indexed directly, an object needs a walk from the start
-common_json & common_json::iterator::operator*() const {
-    return guard([&]() -> common_json & {
-        ordered_json & j = as_json(node);
-
-        if (j.is_object()) {
-            return as_common(std::next(j.begin(), idx).value());
-        }
-        if (j.is_array()) {
-            return as_common(j[idx]);
-        }
-
-        // a plain value gives itself once, same as the backing library
-        return *node;
-    });
-}
-
-std::string common_json::iterator::key() const {
-    return guard([&] { return std::next(as_json(node).begin(), idx).key(); });
-}
-
-common_json::iterator common_json::begin() const {
-    return iterator(const_cast<common_json *>(this), 0);
-}
-
-common_json::iterator common_json::end() const {
-    return iterator(const_cast<common_json *>(this), size());
-}
-
-// the keys follow the backing library: the index for an array, "" for a plain value
-common_json::items_view::entry common_json::items_view::iterator::operator*() const {
-    return guard([&]() -> entry {
-        ordered_json & j = as_json(node);
-
-        if (j.is_object()) {
-            auto it = std::next(j.begin(), idx);
-
-            return { it.key(), as_common(it.value()) };
-        }
-        if (j.is_array()) {
-            return { std::to_string(idx), as_common(j[idx]) };
-        }
-
-        return { std::string(), *node };
-    });
-}
-
-common_json::items_view common_json::items() const {
-    return items_view(const_cast<common_json *>(this), size());
-}
-
-template <typename T> T common_json::get() const {
-    return guard([&] { return as_json(this).get<T>(); });
-}
-
-// the backing library cannot build a common_json, so this one is just a copy
-template <> common_json common_json::get<common_json>() const {
-    return *this;
-}
-
-// get<T>() is usable only for the types below
-
-#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
-
-COMMON_JSON_GET(bool)
-COMMON_JSON_GET(int)
-COMMON_JSON_GET(unsigned int)
-COMMON_JSON_GET(long)
-COMMON_JSON_GET(unsigned long)
-COMMON_JSON_GET(long long)
-COMMON_JSON_GET(unsigned long long)
-COMMON_JSON_GET(float)
-COMMON_JSON_GET(double)
-COMMON_JSON_GET(std::string)
-COMMON_JSON_GET(std::vector<float>)
-COMMON_JSON_GET(std::vector<std::string>)
-COMMON_JSON_GET(std::set<std::string>)
-COMMON_JSON_GET(std::vector<int>)
-COMMON_JSON_GET(std::vector<size_t>)
-COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
-
-#undef COMMON_JSON_GET

common/json.h

@@ -1,354 +0,0 @@
-#pragma once
-
-#include <cstddef>
-#include <cstdint>
-#include <initializer_list>
-#include <iterator>
-#include <map>
-#include <memory>
-#include <set>
-#include <stdexcept>
-#include <string>
-#include <string_view>
-#include <type_traits>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-// common_json, a thin wrapper around vendor json library
-// the underlay library is pimpl, we are using nlohmann::json for now
-//
-// many features of the library are deliberately left out, to keep this interface small and generic and to keep compile time down
-//
-// some main differences compared to nlohmann::json :
-// - object keys keep the order in which they are added
-// - errors are always throw as common_json_error
-// - obj.push_back({key, val}) is intentionally unsupported to avoid confusion with push_back on a vector; write it as obj[key] = val for clarity
-// - a braced pair in value position does not build, e.g. {"key", {"a", "b"}}; write array({"a", "b"}) where nlohmann made an array
-//
-// in doubt, search the code base for an existing usage example; do not add anything to this header unless absolutely necessary
-
-class common_json;
-
-// common_json_value holds a list of these, and each of them holds a value, so one must come first
-struct common_json_item;
-
-struct common_json_error : std::runtime_error {
-    using std::runtime_error::runtime_error;
-};
-
-// one value, tagged so that this header stays free of the backing library
-// note: a value that holds a tree is single use, the second use gives null
-struct common_json_value {
-    enum value_type {
-        VAL_NULL,
-        VAL_BOOL,
-        VAL_INT,
-        VAL_UINT,
-        VAL_DOUBLE,
-        VAL_STRING,
-        VAL_JSON,
-    };
-
-    value_type type = VAL_NULL;
-
-    union {
-        bool     val_bool;
-        int64_t  val_int;
-        uint64_t val_uint = 0;
-        double   val_double;
-    };
-
-    std::string                  val_string;
-    std::shared_ptr<common_json> val_json;
-
-    common_json_value(std::nullptr_t = nullptr) : type(VAL_NULL) {}
-    common_json_value(bool val) : type(VAL_BOOL), val_bool(val) {}
-    common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
-    // without this a string_view lands on the common_json ctor below and recurses
-    common_json_value(std::string_view val) : type(VAL_STRING), val_string(val) {}
-    common_json_value(const char * val);
-    common_json_value(const common_json & val);
-    common_json_value(common_json && val);
-    // only for the types instantiated in json.cpp, the rest fails at link time
-    template <typename T> common_json_value(const std::vector<T> & vals);
-    // a set becomes an array, in the set's own order
-    template <typename T> common_json_value(const std::set<T> & vals);
-    // a map becomes an object, keyed in the map's own order
-    template <typename T> common_json_value(const std::map<std::string, T> & vals);
-    template <typename T> common_json_value(const std::unordered_map<std::string, T> & vals);
-
-    // nested object, e.g. {"fn", {{"name", "x"}}}
-    // note: a nested pair {"a", "b"} does not build, use common_json::array({"a", "b"}) for an array
-    common_json_value(std::initializer_list<common_json_item> items);
-
-    template <typename T, typename std::enable_if<std::is_integral<T>::value && !std::is_same<T, bool>::value, int>::type = 0>
-    common_json_value(T val) : type(std::is_signed<T>::value ? VAL_INT : VAL_UINT) {
-        if (std::is_signed<T>::value) {
-            val_int = (int64_t) val;
-        } else {
-            val_uint = (uint64_t) val;
-        }
-    }
-
-    template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
-    common_json_value(T val) : type(VAL_DOUBLE), val_double((double) val) {}
-};
-
-struct common_json_item {
-    std::string       key;
-    common_json_value val;
-
-    template <typename T>
-    common_json_item(std::string key, T && val) :
-        key(std::move(key)), val(std::forward<T>(val)) {}
-
-    // a braced list cannot deduce T, so it needs its own overload
-    common_json_item(std::string key, std::initializer_list<common_json_item> items) :
-        key(std::move(key)), val(items) {}
-};
-
-// the types common_json_value holds on its own
-// anything else reaches its common_json ctor and recurses forever
-template <typename T> struct common_json_is_value : std::integral_constant<bool,
-    std::is_arithmetic<T>::value ||
-    std::is_same<T, std::nullptr_t>::value ||
-    std::is_same<T, std::string>::value ||
-    std::is_same<T, std::string_view>::value ||
-    std::is_same<T, char *>::value ||
-    std::is_same<T, const char *>::value ||
-    std::is_same<T, common_json>::value> {};
-
-template <typename T, typename A>
-struct common_json_is_value<std::vector<T, A>> : std::true_type {};
-
-template <typename T, typename C, typename A>
-struct common_json_is_value<std::set<T, C, A>> : std::true_type {};
-
-template <typename V, typename C, typename A>
-struct common_json_is_value<std::map<std::string, V, C, A>> : std::true_type {};
-
-template <typename V, typename H, typename E, typename A>
-struct common_json_is_value<std::unordered_map<std::string, V, H, E, A>> : std::true_type {};
-
-class common_json {
-  public:
-    common_json();
-    common_json(const common_json & other);
-    common_json(common_json && other) noexcept;
-    common_json(std::initializer_list<common_json_item> items);
-    common_json(const common_json_value & val);
-
-    // direct, a value would need two conversions in a row
-    common_json(std::nullptr_t);
-
-    // one step, so that "abc" or a vector can go straight into a common_json
-    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value &&
-                                                  !std::is_same<typename std::decay<T>::type, common_json_value>::value, int>::type = 0>
-    common_json(T && val) : common_json(common_json_value(std::forward<T>(val))) {
-        static_assert(common_json_is_value<typename std::decay<T>::type>::value,
-                      "no common_json_value ctor holds this type, add one instead of letting it recurse");
-    }
-
-    // by value, same as the backing library
-    // the right side is copied before the left side can invalidate it, e.g. msg["a"] = msg.at("b")
-    common_json & operator=(common_json other) noexcept;
-
-    ~common_json();
-
-    // throws common_json_error if the text is not valid JSON
-    static common_json parse(const std::string & text);
-
-    // gives a discarded value instead of throwing, check it with is_discarded()
-    static common_json parse_no_throw(const std::string & text);
-
-    bool is_discarded() const;
-
-    static common_json array();
-    static common_json array(std::initializer_list<common_json_value> vals);
-    static common_json object();
-    static common_json object(std::initializer_list<common_json_item> items);
-
-    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
-    static common_json make(const common_json_value & val);
-
-    bool is_null()    const;
-    bool is_object()  const;
-    bool is_array()   const;
-    bool is_string()  const;
-    bool is_boolean() const;
-    bool is_number()  const;
-    bool is_number_integer() const;
-    bool is_number_float()   const;
-
-    bool   empty() const;
-    size_t size()  const;
-
-    bool contains(const std::string & key) const;
-
-    bool operator==(const common_json_value & val) const;
-    bool operator!=(const common_json_value & val) const;
-
-    // at() throws common_json_error if the key is missing, operator[] adds a null value instead
-    // note: a const operator[] cannot add, it throws like at()
-    common_json       & at(const std::string & key);
-    const common_json & at(const std::string & key) const;
-    common_json       & at(size_t idx);
-    const common_json & at(size_t idx) const;
-
-    common_json       & operator[](const std::string & key);
-    const common_json & operator[](const std::string & key) const;
-    common_json       & operator[](const char * key)       { return (*this)[std::string(key)]; }
-    const common_json & operator[](const char * key) const { return (*this)[std::string(key)]; }
-    common_json       & operator[](int idx)       { return (*this)[to_idx(idx)]; }
-    const common_json & operator[](int idx) const { return (*this)[to_idx(idx)]; }
-    common_json       & operator[](size_t idx);
-    const common_json & operator[](size_t idx) const;
-
-    common_json       & front();
-    const common_json & front() const;
-    common_json       & back();
-    const common_json & back()  const;
-
-    void clear();
-
-    void erase(const std::string & key);
-    void erase(size_t idx);
-
-    // only for the types instantiated in json.cpp, the rest fails at link time
-    template <typename T> T get() const;
-
-    // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
-    // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
-    // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const { return get<std::string>(); }
-
-    template <typename T>
-    T value(const std::string & key, T def) const {
-        return contains(key) ? at(key).get<T>() : def;
-    }
-
-    std::string value(const std::string & key, const char * def) const {
-        return contains(key) ? at(key).get<std::string>() : std::string(def);
-    }
-
-    // a JSON default needs no get<T>(), it is already the right type
-    common_json value(const std::string & key, const common_json & def) const {
-        return contains(key) ? at(key) : def;
-    }
-
-    void assign(const common_json_value & val);
-    void set(const common_json_item & item);
-    void push_back(const common_json_value & val);
-
-    // appends one object, e.g. push_back({{"a", 1}})
-    void push_back(std::initializer_list<common_json_item> items);
-
-    // 1 if the key is there, 0 if not
-    size_t count(const std::string & key) const;
-
-    // appends every value of another array; inserting an array into itself throws
-    void insert(const common_json & vals);
-
-    // a common_json goes through the copy assignment above, everything else becomes a value
-    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
-    common_json & operator=(T && val) {
-        assign(common_json_value(std::forward<T>(val)));
-        return *this;
-    }
-
-    std::string dump(int indent = -1) const;
-
-    // same as dump(), but bad UTF-8 gets replaced instead of throwing
-    std::string dump_safe(int indent = -1) const;
-
-    // walks an array by index, or an object in insertion order
-    // a plain value gives itself once, same as the backing library
-    class iterator {
-      public:
-        using iterator_category = std::forward_iterator_tag;
-        using value_type        = common_json;
-        using difference_type   = std::ptrdiff_t;
-        using pointer           = common_json *;
-        using reference         = common_json &;
-
-        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
-
-        common_json & operator*() const;
-        common_json & value()     const { return **this; }
-        std::string   key()       const;
-
-        iterator & operator++() {
-            idx++;
-            return *this;
-        }
-
-        bool operator!=(const iterator & other) const { return idx != other.idx; }
-        bool operator==(const iterator & other) const { return idx == other.idx; }
-
-      private:
-        common_json * node;
-        size_t        idx;
-    };
-
-    iterator begin() const;
-    iterator end()   const;
-
-    // allows: for (const auto & [key, val] : obj.items())
-    class items_view {
-      public:
-        // the members are public, so an entry also works with structured bindings
-        struct entry {
-            std::string   k;
-            common_json & v;
-
-            const std::string & key()   const { return k; }
-            common_json &       value() const { return v; }
-        };
-
-        items_view(common_json * node, size_t n) : node(node), n(n) {}
-
-        class iterator {
-          public:
-            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
-
-            entry operator*() const;
-
-            iterator & operator++() {
-                idx++;
-                return *this;
-            }
-
-            bool operator!=(const iterator & other) const { return idx != other.idx; }
-
-          private:
-            common_json * node;
-            size_t        idx;
-        };
-
-        iterator begin() const { return iterator(node, 0); }
-        iterator end()   const { return iterator(node, n); }
-
-      private:
-        common_json * node;
-        size_t        n;
-    };
-
-    items_view items() const;
-
-  private:
-    // a negative index must not turn into a huge size_t
-    static size_t to_idx(int idx) {
-        if (idx < 0) {
-            throw common_json_error("negative array index");
-        }
-        return (size_t) idx;
-    }
-
-    // the backing value is built here, json.cpp checks that it fits
-    // it cannot be a pointer: a value inside a tree would then not be a common_json
-    // at() could then only give back a copy instead of a real reference
-    alignas(8) unsigned char storage[32];
-};
-
-using common_json_entry = common_json::items_view::entry;

common/llguidance.cpp

@@ -116,8 +116,6 @@ static llama_sampler_i llama_sampler_llg_i = {
     /* .backend_accept    = */ NULL,
     /* .backend_apply     = */ NULL,
     /* .backend_set_input = */ NULL,
-    /* .backend_reset     = */ NULL,
-    /* .copy_state        = */ NULL,
 };
 
 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,

common/peg-parser.cpp

@@ -10,6 +10,7 @@
 #include <initializer_list>
 #include <map>
 #include <memory>
+#include <nlohmann/json.hpp>
 #include <regex>
 #include <set>
 #include <stdexcept>
@@ -569,34 +570,23 @@ struct parser_executor {
     }
 
     static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
-        auto save = pos;
-
         ++pos; // consume '\'
         if (pos >= ctx.input.size()) {
             if (!ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
             }
-            pos = save; // suppress unmatched '\'
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
         }
 
         char c = ctx.input[pos];
-
         if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
             ++pos;
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
+        } else if (c == 'u') {
+            return handle_unicode_escape(ctx, start, pos);
+        } else {
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
         }
-
-        if (c == 'u') {
-            auto result = handle_unicode_escape(ctx, start, pos);
-            if (result.need_more_input()) {
-                pos = save; // suppress incomplete sequence
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
-            }
-            return result;
-        }
-
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
     }
 
     static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {
@@ -1119,8 +1109,8 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
-common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<nlohmann::ordered_json>(schema), raw}));
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1804,8 +1794,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     }
 }
 
-static common_json serialize_parser_variant(const common_peg_parser_variant & variant) {
-    using json = common_json;
+static nlohmann::json serialize_parser_variant(const common_peg_parser_variant & variant) {
+    using json = nlohmann::json;
 
     return std::visit([](const auto & p) -> json {
         using T = std::decay_t<decltype(p)>;
@@ -1859,7 +1849,7 @@ static common_json serialize_parser_variant(const common_peg_parser_variant & va
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : json(nullptr)},
+                {"schema", p.schema ? *p.schema : nullptr},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1887,19 +1877,19 @@ static common_json serialize_parser_variant(const common_peg_parser_variant & va
     }, variant);
 }
 
-common_json common_peg_arena::to_json() const {
-    auto parsers = common_json::array();
+nlohmann::json common_peg_arena::to_json() const {
+    auto parsers = nlohmann::json::array();
     for (const auto & parser : parsers_) {
         parsers.push_back(serialize_parser_variant(parser));
     }
-    return common_json{
+    return nlohmann::json{
         {"parsers", parsers},
         {"rules", rules_},
         {"root", root_}
     };
 }
 
-static common_peg_parser_variant deserialize_parser_variant(const common_json & j) {
+static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json & j) {
     if (!j.contains("type") || !j["type"].is_string()) {
         throw std::runtime_error("Parser variant JSON missing or invalid 'type' field");
     }
@@ -1968,9 +1958,9 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
         }
         common_peg_chars_parser parser;
         parser.pattern = j["pattern"];
-        parser.negated = j["negated"].get<bool>();
-        parser.min_count = j["min_count"].get<int>();
-        parser.max_count = j["max_count"].get<int>();
+        parser.negated = j["negated"];
+        parser.min_count = j["min_count"];
+        parser.max_count = j["max_count"];
         for (const auto & range_json : j["ranges"]) {
             if (!range_json.contains("start") || !range_json.contains("end")) {
                 throw std::runtime_error("char_range missing 'start' or 'end' field");
@@ -2006,7 +1996,7 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
         if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<common_json>(j["schema"]);
+            parser.schema = std::make_shared<nlohmann::ordered_json>(j["schema"]);
         }
         parser.raw = j["raw"].get<bool>();
         return parser;
@@ -2068,7 +2058,7 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
     throw std::runtime_error("Unknown parser type: " + type);
 }
 
-common_peg_arena common_peg_arena::from_json(const common_json & j) {
+common_peg_arena common_peg_arena::from_json(const nlohmann::json & j) {
     if (!j.contains("parsers") || !j["parsers"].is_array()) {
         throw std::runtime_error("JSON missing or invalid 'parsers' array");
     }
@@ -2108,7 +2098,7 @@ std::string common_peg_arena::save() const {
 }
 
 void common_peg_arena::load(const std::string & data) {
-    *this = from_json(common_json::parse(data));
+    *this = from_json(nlohmann::json::parse(data));
 }
 
 common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {

common/peg-parser.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "json.h"
+#include <nlohmann/json_fwd.hpp>
 
 #include <memory>
 #include <set>
@@ -245,7 +245,7 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<common_json> schema;
+    std::shared_ptr<nlohmann::ordered_json> schema;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -332,8 +332,8 @@ class common_peg_arena {
 
     std::string dump(common_peg_parser_id id) const;
 
-    common_json to_json() const;
-    static common_peg_arena from_json(const common_json & j);
+    nlohmann::json to_json() const;
+    static common_peg_arena from_json(const nlohmann::json & j);
 
     std::string save() const;
     void load(const std::string & data);
@@ -490,7 +490,7 @@ class common_peg_parser_builder {
 
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
-    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.
     // If trigger=true, marks this rule as an entry point for lazy grammar generation.

common/preset.cpp

@@ -322,8 +322,6 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
                     preset.options[opt] = value;
                 }
                 LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
-            } else if (ignore_unknown_keys) {
-                LOG_WRN("ignoring option '%s' from %s: not supported by this program\n", key.c_str(), path.c_str());
             } else {
                 throw std::runtime_error(string_format(
                     "option '%s' not recognized in preset '%s'",
@@ -365,25 +363,8 @@ struct local_model {
     std::string name;
     std::string path;
     std::string path_mmproj;
-    std::string path_draft;
 };
 
-// TODO @ngxson: handle "eagle3-" when it's supported by common_speculative_types_from_gguf()
-static const char * draft_prefixes[] = { "mtp-", "dspark-", "dflash-" };
-
-static bool is_mmproj_file(const std::string & fname) {
-    return fname.find("mmproj") != std::string::npos;
-}
-
-static bool is_draft_file(const std::string & fname) {
-    for (const auto & prefix : draft_prefixes) {
-        if (fname.rfind(prefix, 0) == 0) {
-            return true;
-        }
-    }
-    return false;
-}
-
 common_presets common_preset_context::load_from_models_dir(const std::string & models_dir) const {
     if (!std::filesystem::exists(models_dir) || !std::filesystem::is_directory(models_dir)) {
         throw std::runtime_error(string_format("error: '%s' does not exist or is not a directory\n", models_dir.c_str()));
@@ -395,15 +376,10 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         common_file_info model_file;
         common_file_info first_shard_file;
         common_file_info mmproj_file;
-        common_file_info draft_file;
         for (const auto & file : files) {
             if (string_ends_with(file.name, ".gguf")) {
-                if (is_mmproj_file(file.name)) {
+                if (file.name.find("mmproj") != std::string::npos) {
                     mmproj_file = file;
-                } else if (is_draft_file(file.name)) {
-                    if (draft_file.path.empty()) {
-                        draft_file = file; // first sidecar found wins
-                    }
                 } else if (file.name.find("-00001-of-") != std::string::npos) {
                     first_shard_file = file;
                 } else {
@@ -415,8 +391,7 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         local_model model{
             /* name        */ name,
             /* path        */ first_shard_file.path.empty() ? model_file.path : first_shard_file.path,
-            /* path_mmproj */ mmproj_file.path, // can be empty
-            /* path_draft  */ draft_file.path   // can be empty
+            /* path_mmproj */ mmproj_file.path // can be empty
         };
         if (!model.path.empty()) {
             models.push_back(model);
@@ -428,17 +403,13 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         if (file.is_dir) {
             scan_subdir(file.path, file.name);
         } else if (string_ends_with(file.name, ".gguf")) {
-            if (is_mmproj_file(file.name) || is_draft_file(file.name)) {
-                continue; // companion file, cannot be loaded as a model on its own
-            }
             // single file model
             std::string name = file.name;
             string_replace_all(name, ".gguf", "");
             local_model model{
                 /* name        */ name,
                 /* path        */ file.path,
-                /* path_mmproj */ "",
-                /* path_draft  */ ""
+                /* path_mmproj */ ""
             };
             models.push_back(model);
         }
@@ -453,9 +424,6 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         if (!model.path_mmproj.empty()) {
             preset.set_option(*this, "LLAMA_ARG_MMPROJ", model.path_mmproj);
         }
-        if (!model.path_draft.empty()) {
-            preset.set_option(*this, "LLAMA_ARG_SPEC_DRAFT_MODEL", model.path_draft);
-        }
         out[preset.name] = preset;
     }
 

common/preset.h

@@ -59,10 +59,6 @@ struct common_preset_context {
     bool filter_allowed_keys = false;
     std::set<std::string> allowed_keys;
 
-    // if true, options unknown to the current example are skipped instead of being an error
-    // used for config files shared by all binaries, where each binary only knows a subset of options
-    bool ignore_unknown_keys = false;
-
     // if only_remote_allowed is true, only accept whitelisted keys
     common_preset_context(llama_example ex);
 

common/reasoning-budget.cpp

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

common/sampling.cpp

@@ -518,26 +518,6 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     };
 }
 
-void common_sampler_copy(const common_sampler * src, common_sampler * dst) {
-    if (!src || !dst || src == dst) {
-        return;
-    }
-
-    GGML_ASSERT((src->grmr == nullptr) == (dst->grmr == nullptr));
-    GGML_ASSERT((src->rbudget == nullptr) == (dst->rbudget == nullptr));
-
-    llama_sampler_copy(src->grmr,    dst->grmr);
-    llama_sampler_copy(src->rbudget, dst->rbudget);
-    llama_sampler_copy(src->chain,   dst->chain);
-
-    dst->params     = src->params;
-    dst->prev       = src->prev;
-    dst->cur        = src->cur;
-    dst->cur_p      = src->cur_p;
-    dst->cur_p.data = src->cur_p.data ? dst->cur.data() : nullptr; // re-point to dst's buffer
-    dst->t_total_us = src->t_total_us;
-}
-
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 

common/sampling.h

@@ -47,7 +47,6 @@ void common_sampler_free(struct common_sampler * gsmpl);
 void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated);
 void                    common_sampler_reset (struct common_sampler * gsmpl);
 struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
-void                    common_sampler_copy  (const struct common_sampler * src, struct common_sampler * dst);
 
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);

common/speculative.cpp

@@ -2,7 +2,6 @@
 
 #include "common.h"
 #include "ggml.h"
-#include "ggml-cpp.h"
 #include "llama.h"
 #include "log.h"
 #include "ngram-cache.h"
@@ -172,6 +171,12 @@ struct common_speculative_impl {
     // (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
     virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
     virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}
+
+    // true if this implementation requires the target context to extract post-norm embeddings
+    virtual bool need_embd() const = 0;
+
+    // true if this implementation requires the target context to extract pre-norm embeddings
+    virtual bool need_embd_nextn() const { return false; }
 };
 
 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -188,10 +193,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         auto * ctx_tgt = this->params.ctx_tgt;
 
-        if (!ctx_dft) {
-            throw std::runtime_error("draft-simple requires a draft context");
-        }
-
         SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n");
         SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min);
         SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n",
@@ -384,6 +385,10 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 
@@ -902,6 +907,10 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         pending_g_last[seq_id].resize(n_embd_dec);
         std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 // DFlash: block-diffusion drafting with a draft-side KV cache injection
@@ -913,9 +922,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
     std::vector<common_sampler_ptr> smpls;
 
-    // backend sampler chain per seq, attached to ctx_dft
-    std::vector<llama_sampler *> backend_chains;
-
     int32_t n_embd_dec = 0;  // draft hidden size
     int32_t n_embd_enc = 0;  // target_layer_ids_n * target_hidden_size
     int32_t n_embd_tgt = 0;  // target model hidden size
@@ -926,9 +932,6 @@ 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;
 
@@ -963,20 +966,16 @@ 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, sample_from_anchor=%s\n", __func__,
-                block_size, mask_token_id, target_layer_ids_n, sample_from_anchor ? "true" : "false");
+        LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n);
 
         // DFlash input is [id_last, <mask> * (block_size-1)]: in-place denoising yields at most
-        // 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;
+        // 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;
         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);
@@ -996,22 +995,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             s.reset(common_sampler_init(model_dft, sparams));
         }
 
-        // offload draft sampling to the backend
-        backend_chains.assign(n_seq, nullptr);
-        if (this->params.backend_sampling) {
-            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-                llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
-                llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
-
-                if (!llama_set_sampler(ctx_dft, seq_id, chain)) {
-                    SPC_WRN("backend offload failed for seq_id=%d; using CPU sampler\n", (int) seq_id);
-                    llama_sampler_free(chain);
-                    chain = nullptr;
-                }
-                backend_chains[seq_id] = chain;
-            }
-        }
-
         // turn on extraction of the target layers' input embeddings
         for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
             llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
@@ -1022,18 +1005,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     }
 
     ~common_speculative_impl_draft_dflash() override {
-        auto * ctx_dft = this->params.ctx_dft;
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) backend_chains.size(); ++seq_id) {
-            if (backend_chains[seq_id] == nullptr) {
-                continue;
-            }
-            if (ctx_dft) {
-                llama_set_sampler(ctx_dft, seq_id, nullptr);
-            }
-            llama_sampler_free(backend_chains[seq_id]);
-        }
-        backend_chains.clear();
-
         llama_batch_free(batch);
         llama_batch_free(batch_inject);
     }
@@ -1061,14 +1032,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             return true;
         }
 
-        // Target prefill may contain token IDs or multimodal embeddings. Both
-        // produce the target-layer features used to seed the draft KV cache, so
-        // skipping the embedding batches leaves a hole in the draft's cache and
-        // the next injection fails to initialize.
-        // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
-        const bool has_tokens     = batch_in.token != nullptr;
-        const bool has_embeddings = batch_in.embd  != nullptr;
-        if (has_tokens == has_embeddings) {
+        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
             return true;
         }
 
@@ -1182,7 +1146,7 @@ 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 && sample_from_anchor ? 0 : 1);
+            const int32_t n_block_tokens = n_draft + (is_dspark ? 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) {
@@ -1215,11 +1179,11 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             auto & result = *dp.result;
 
             if (is_dspark) {
-                // DSpark: read from the first draft slot, truncate below the confidence threshold
+                // DSpark predicts the next token from position 0 and optionally truncates
+                // at the first position below the confidence threshold.
                 const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
-                // 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) {
+
+                for (int32_t i = 0; i < n_block_tokens; ++i) {
                     const int32_t idx = beg + i;
 
                     if (conf && conf[(size_t) idx * n_embd_dec] < params.p_min) {
@@ -1276,6 +1240,10 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 struct common_speculative_impl_draft_mtp : public common_speculative_impl {
@@ -1714,6 +1682,14 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
         std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
     }
+
+    bool need_embd() const override {
+        return false;
+    }
+
+    bool need_embd_nextn() const override {
+        return true;
+    }
 };
 
 // state of self-speculation (simple implementation, not ngram-map)
@@ -1760,6 +1736,10 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
@@ -1814,6 +1794,10 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
 
         common_ngram_map_accept(config[seq_id], n_accepted);
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 struct common_speculative_impl_ngram_mod : public common_speculative_impl {
@@ -1989,6 +1973,10 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
             }
         }
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 struct common_speculative_impl_ngram_cache : public common_speculative_impl {
@@ -2128,6 +2116,10 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
+
+    bool need_embd() const override {
+        return false;
+    }
 };
 
 struct common_speculative {
@@ -2235,43 +2227,6 @@ common_speculative_type common_speculative_type_from_name(const std::string & na
     return it->second;
 }
 
-std::vector<common_speculative_type> common_speculative_types_from_gguf(const std::string & path) {
-    struct gguf_init_params gguf_params = {
-        /* .no_alloc = */ true,
-        /* .ctx      = */ nullptr,
-    };
-
-    gguf_context_ptr gguf_ctx(gguf_init_from_file(path.c_str(), gguf_params));
-    if (!gguf_ctx) {
-        return {};
-    }
-
-    const int64_t arch_id = gguf_find_key(gguf_ctx.get(), "general.architecture");
-    if (arch_id < 0 || gguf_get_kv_type(gguf_ctx.get(), arch_id) != GGUF_TYPE_STRING) {
-        return {};
-    }
-
-    const std::string arch = gguf_get_val_str(gguf_ctx.get(), arch_id);
-    if (arch != "dflash") {
-        const uint32_t block_count = gguf_get_val_u32(gguf_ctx.get(), gguf_find_key(gguf_ctx.get(), (arch + ".block_count").c_str()));
-
-        if (gguf_find_tensor(gguf_ctx.get(), ("blk." + std::to_string(block_count - 1) + ".nextn.eh_proj.weight").c_str()) >= 0) {
-            return { COMMON_SPECULATIVE_TYPE_DRAFT_MTP };
-        }
-
-        return {};
-    }
-
-    // the Markov head distinguishes draft-dspark from draft-dflash
-    const auto type = gguf_find_tensor(gguf_ctx.get(), "markov_w1.weight") >= 0
-                    ? COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK
-                    : COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH;
-
-    SPC_INF("auto-detected speculative type '%s' from the draft model metadata\n", common_speculative_type_to_str(type).c_str());
-
-    return { type };
-}
-
 static uint32_t common_get_enabled_speculative_configs(const std::vector<common_speculative_type> & configs) {
     uint32_t result = 0;
     for (size_t i = 0; i < configs.size(); i++) {
@@ -2322,9 +2277,6 @@ common_params common_base_params_to_speculative(const common_params & params) {
     const auto & params_spec = params.speculative.draft;
     common_params result = params;
 
-    result.embedding    = false;
-    result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED;
-
     if (has_draft) {
         result.devices               = params_spec.devices;
         result.model                 = params_spec.mparams;
@@ -2340,24 +2292,6 @@ common_params common_base_params_to_speculative(const common_params & params) {
     result.cache_type_k  = params_spec.cache_type_k;
     result.cache_type_v  = params_spec.cache_type_v;
     result.n_outputs_max = params.n_parallel;
-    result.n_outputs_max_per_seq = 1;
-
-    // dflash/dspark decode the whole noise block in a single pass and sample every block position on the backend
-    // TODO: refactor such properties to be announced by the speculative types
-    //       something like `struct common_speculative_type_props common_speculative_type_get_props(...);`
-    const bool has_block_draft = std::any_of(
-        params.speculative.types.begin(), params.speculative.types.end(),
-        [](common_speculative_type t) {
-            return t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
-        });
-    if (has_block_draft) {
-        // per-seq output positions: DFlash decodes anchor + n_max masks (n_max + 1); DSpark n_max -> +1 covers both
-        const int32_t per_seq = std::max(1, params_spec.n_max + 1);
-        result.n_outputs_max = params.n_parallel * per_seq;
-        if (params_spec.backend_sampling) {
-            result.n_outputs_max_per_seq = per_seq;
-        }
-    }
 
     return result;
 }
@@ -2380,6 +2314,7 @@ common_speculative_init_result::common_speculative_init_result(
     const bool spec_mtp = std::find(params.speculative.types.begin(),
                                     params.speculative.types.end(),
                                     COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+    GGML_ASSERT(has_draft || spec_mtp);
 
     auto mparams = common_model_params_to_llama(params);
     auto cparams = common_context_params_to_llama(params);
@@ -2388,9 +2323,6 @@ common_speculative_init_result::common_speculative_init_result(
         cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
     }
 
-    // the draft context holds as many tokens per sequence as the target context
-    cparams.n_ctx = llama_n_ctx(ctx_tgt);
-
     // note: for small models maybe we can set this to the maximum possible draft from all speculative types
     //       the extra memory for small models is likely negligible?
     cparams.n_rs_seq  = 0;
@@ -2445,17 +2377,6 @@ common_speculative_init_result_ptr common_speculative_init_from_params(common_pa
     return std::make_unique<common_speculative_init_result>(params, model_tgt, ctx_tgt);
 }
 
-common_speculative_output_limits common_speculative_get_output_limits(
-        int32_t n_batch, int32_t n_parallel, int32_t n_draft) {
-    const int64_t per_seq = 1 + (int64_t) std::max(0, n_draft);
-    const int64_t total   = (int64_t) n_parallel * per_seq;
-
-    return {
-        /* .total   = */ (int32_t) std::min<int64_t>(n_batch, total),
-        /* .per_seq = */ (int32_t) std::min<int64_t>(n_batch, per_seq),
-    };
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -2620,6 +2541,34 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
     return result;
 }
 
+bool common_speculative_need_embd(common_speculative * spec) {
+    if (spec == nullptr) {
+        return false;
+    }
+
+    for (auto & impl : spec->impls) {
+        if (impl->need_embd()) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool common_speculative_need_embd_nextn(common_speculative * spec) {
+    if (spec == nullptr) {
+        return false;
+    }
+
+    for (auto & impl : spec->impls) {
+        if (impl->need_embd_nextn()) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 void common_speculative_draft(common_speculative * spec) {
     if (spec == nullptr) {
         return;
@@ -2655,10 +2604,6 @@ 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
@@ -2708,10 +2653,7 @@ void common_speculative_draft(common_speculative * spec) {
 void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
     common_speculative_impl * impl = spec->impl_last[seq_id];
 
-    if (impl == nullptr) {
-        GGML_ASSERT(n_accepted == 0);
-        return;
-    }
+    GGML_ASSERT(impl);
 
     {
         common_time_meas tm(impl->t_accept_us, !impl->gen_perf);

common/speculative.h

@@ -14,9 +14,6 @@ const char * common_speculative_all_types_str();
 // parse user provided types
 std::vector<enum common_speculative_type> common_speculative_types_from_names(const std::vector<std::string> & names);
 
-// infer the spec types from the GGUF metadata of a draft model; empty if unknown
-std::vector<enum common_speculative_type> common_speculative_types_from_gguf(const std::string & path);
-
 // convert string to type
 enum common_speculative_type common_speculative_type_from_name(const std::string & name);
 
@@ -28,15 +25,6 @@ int32_t common_speculative_n_max(const common_params_speculative * spec);
 
 common_params common_base_params_to_speculative(const common_params & params);
 
-struct common_speculative_output_limits {
-    int32_t total;
-    int32_t per_seq;
-};
-
-// return the output limits needed for speculative decoding
-common_speculative_output_limits common_speculative_get_output_limits(
-        int32_t n_batch, int32_t n_parallel, int32_t n_draft);
-
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);
 
 void common_speculative_free(common_speculative * spec);
@@ -70,6 +58,12 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);
 
+// true if any implementation requires target post-norm embeddings to be extracted
+bool common_speculative_need_embd(common_speculative * spec);
+
+// true if any implementation requires target nextn embeddings to be extracted
+bool common_speculative_need_embd_nextn(common_speculative * spec);
+
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);
 

conversion/__init__.py

@@ -27,7 +27,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "BaichuanForCausalLM": "baichuan",
     "BailingMoeForCausalLM": "bailingmoe",
     "BailingMoeV2ForCausalLM": "bailingmoe",
-    "BailingMoeV3ForCausalLM": "bailingmoe3",
     "BambaForCausalLM": "granite",
     "BertForMaskedLM": "bert",
     "BertForSequenceClassification": "bert",
@@ -55,19 +54,12 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekV32ForCausalLM": "deepseek",
     "DFlashDraftModel": "qwen",
     "Qwen3DSparkModel": "qwen",
-    "DSparkDraftModel": "qwen",
-    "DSparkSpeculator": "qwen",
-    "Lfm2DSparkDraftModel": "qwen",
-    "LingDSparkModel": "qwen",
     "DeepseekV4ForCausalLM": "deepseek",
     "DeepseekV4DSparkModel": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
     "DistilBertModel": "bert",
     "Dots1ForCausalLM": "dots1",
-    "Dots3NoteForCausalLM": "dots3",
-    "Dots3NoteForConditionalGeneration": "dots3",
-    "Dots3NoteTextForCausalLM": "dots3",
     "DotsOCRForCausalLM": "qwen",
     "DreamModel": "dream",
     "Ernie4_5ForCausalLM": "ernie",
@@ -111,11 +103,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "GraniteMoeForCausalLM": "granite",
     "GraniteMoeHybridForCausalLM": "granite",
     "GraniteMoeSharedForCausalLM": "granite",
-    "GraniteSwitchForCausalLM": "granite",
     "GraniteSpeechForConditionalGeneration": "granite",
     "GraniteSpeechPlusForConditionalGeneration": "granite",
-    "GraniteSWAForCausalLM": "granite",
-    "GraniteMoeSWAForCausalLM": "granite",
     "Grok1ForCausalLM": "grok",
     "GrokForCausalLM": "grok",
     "GroveMoeForCausalLM": "grovemoe",
@@ -135,7 +124,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "JinaEmbeddingsV5Model": "bert",
     "KORMoForCausalLM": "qwen",
     "KimiK25ForConditionalGeneration": "deepseek",
-    "KimiK3ForConditionalGeneration": "kimi_k3",
     "KimiLinearForCausalLM": "kimi_linear",
     "KimiLinearModel": "kimi_linear",
     "KimiVLForConditionalGeneration": "deepseek",
@@ -172,8 +160,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "MiniCPM3ForCausalLM": "minicpm",
     "MiniCPMForCausalLM": "minicpm",
     "MiniCPMV4_6ForConditionalGeneration": "minicpm",
-    "MiniMaxText01ForCausalLM": "minimax",
-    "MiniMaxM1ForCausalLM": "minimax",
     "MiniMaxM2ForCausalLM": "minimax",
     "MiniMaxM3SparseForCausalLM": "minimax",
     "MiniMaxM3SparseForConditionalGeneration": "minimax",
@@ -196,8 +182,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Olmo3ForCausalLM": "olmo",
     "OlmoForCausalLM": "olmo",
     "OlmoeForCausalLM": "olmo",
-    "MuseGlimmerAssistantModel": "muse_glimmer",
-    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
     "OpenELMForCausalLM": "openelm",
     "OrionForCausalLM": "orion",
     "PLMForCausalLM": "plm",
@@ -227,7 +211,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Qwen3MoeForCausalLM": "qwen",
     "Qwen3NextForCausalLM": "qwen",
     "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
-    "PocketTTSModel": "pockettts",
     "Qwen3TTSForConditionalGeneration": "qwen3tts",
     "Qwen3VLForConditionalGeneration": "qwen3vl",
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -283,8 +266,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "CogVLMForCausalLM": "cogvlm",
     "DeepseekOCR2ForCausalLM": "deepseek",
     "DeepseekOCRForCausalLM": "deepseek",
-    "Dots3NoteForCausalLM": "dots3",
-    "Dots3NoteForConditionalGeneration": "dots3",
     "DotsOCRForCausalLM": "dotsocr",
     "Exaone4_5_ForConditionalGeneration": "exaone",
     "Gemma3ForConditionalGeneration": "gemma",
@@ -316,7 +297,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "MiniCPMV4_6ForConditionalGeneration": "minicpm",
     "Mistral3ForConditionalGeneration": "llava",
     "NemotronH_Nano_VL_V2": "nemotron",
-    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
     "PaddleOCRVisionModel": "ernie",
     "Phi4ForCausalLMV": "phi",
     "Qwen2AudioForConditionalGeneration": "ultravox",
@@ -326,7 +306,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Qwen2_5_VLForConditionalGeneration": "qwenvl",
     "Qwen3ASRForConditionalGeneration": "qwen3vl",
     "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
-    "PocketTTSModel": "pockettts",
     "Qwen3TTSForConditionalGeneration": "qwen3tts",
     "Qwen3VLForConditionalGeneration": "qwen3vl",
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",

conversion/afmoe.py

@@ -13,7 +13,6 @@ from .llama import LlamaModel
 
 
 @ModelBase.register("AfmoeForCausalLM")
-@ModelBase.example("arcee-ai/Trinity-Large-Thinking")
 class AfmoeModel(LlamaModel):
     model_arch = gguf.MODEL_ARCH.AFMOE
 

conversion/arctic.py

@@ -16,7 +16,6 @@ from .llama import LlamaModel
 
 
 @ModelBase.register("ArcticForCausalLM")
-@ModelBase.example("Snowflake/snowflake-arctic-instruct")
 class ArcticModel(TextModel):
     model_arch = gguf.MODEL_ARCH.ARCTIC
 

conversion/baichuan.py

@@ -9,7 +9,6 @@ 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
 

conversion/bailingmoe.py

@@ -11,7 +11,6 @@ from .base import ModelBase, TextModel, gguf
 
 
 @ModelBase.register("BailingMoeForCausalLM")
-@ModelBase.example("inclusionAI/Ling-lite")
 class BailingMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE
 
@@ -109,7 +108,6 @@ class BailingMoeModel(TextModel):
 
 
 @ModelBase.register("BailingMoeV2ForCausalLM")
-@ModelBase.example("inclusionAI/Ling-mini-2.0")
 class BailingMoeV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE2
 
@@ -191,7 +189,6 @@ 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:

conversion/bailingmoe3.py

@@ -1,193 +0,0 @@
-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}")

conversion/base.py

@@ -58,11 +58,6 @@ logger = logging.getLogger("hf-to-gguf")
 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
 
 
-# for checkpoints that ship no config.json, we will try to provide a synthetic one
-HparamsMatcher = Callable[[Path], bool]
-HparamsLoader = Callable[[Path], dict[str, Any]]
-
-
 class SentencePieceTokenTypes(IntEnum):
     NORMAL = 1
     UNKNOWN = 2
@@ -82,7 +77,6 @@ class ModelBase:
         ModelType.TEXT: {},
         ModelType.MMPROJ: {},
     }
-    _hparams_loaders: list[tuple[HparamsMatcher, HparamsLoader]] = []
 
     dir_model: Path
     ftype: gguf.LlamaFileType
@@ -658,43 +652,6 @@ class ModelBase:
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         return ()
 
-    @staticmethod
-    def repack_mxfp4_blocks(packed: Tensor, scale: Tensor) -> np.ndarray:
-        """
-        Repack 4-bit MX weights into ggml `block_mxfp4`. Lossless - only moves bits.
-
-        Source (compressed-tensors "mxfp4-pack-quantized", also used by DeepSeek-V4):
-          packed  uint8 [rows, cols/2]  element 2i in the low nibble, 2i+1 in the high one
-          scale   uint8 [rows, cols/32] one E8M0 biased exponent per 32-element group
-
-        Destination, per group: one scale byte then 16 code bytes, where byte j holds
-        element j in the low nibble and element j+16 in the high one.
-
-        The 4-bit codes need no remapping: both sides index into ggml's kvalues_mxfp4
-        order. ggml doubles the kvalues and halves the scale, so the value is the same.
-        """
-        p = packed.contiguous().view(torch.uint8)
-        s = scale.contiguous().view(torch.uint8)
-
-        rows, packed_cols = p.shape
-        cols = packed_cols * 2
-        if cols % 32 != 0:
-            raise ValueError(f"MXFP4 source row has {cols} values, expected a multiple of 32")
-
-        n_blocks = cols // 32
-        if tuple(s.shape) != (rows, n_blocks):
-            raise ValueError(f"MXFP4 scale shape {tuple(s.shape)} does not match {(rows, n_blocks)}")
-
-        src = p.reshape(rows, n_blocks, 16)
-        lo = src & 0x0F           # elements 0, 2, 4, ...
-        hi = (src >> 4) & 0x0F    # elements 1, 3, 5, ...
-
-        vals = torch.stack((lo, hi), dim=-1).reshape(rows, n_blocks, 32)
-        qs = vals[:, :, :16] | (vals[:, :, 16:] << 4)
-
-        raw = torch.cat((s.unsqueeze(-1), qs.to(torch.uint8)), dim=-1)
-        return raw.reshape(rows, n_blocks * 17).cpu().numpy()
-
     @staticmethod
     def _nvfp4_pack(weight: Tensor, scale: Tensor) -> tuple[np.ndarray, list[int]]:
         """Repack NVFP4 ModelOpt tensors into ggml super-block layout.
@@ -866,7 +823,7 @@ class ModelBase:
             elif any(str(v.get("quant_algo")).endswith("NVFP4") for v in quant_layers.values() if isinstance(v, dict)):
                 quant_algo = "NVFP4"
 
-        self._is_nvfp4 = quant_algo in ("NVFP4", "W4A16_NVFP4")
+        self._is_nvfp4 = quant_algo == "NVFP4"
         self._is_mxfp4 = quant_method == "mxfp4"
 
         # NVFP4 weights are repacked and written directly to gguf_writer.
@@ -1083,24 +1040,6 @@ class ModelBase:
 
         return part_names
 
-    @staticmethod
-    def load_hparams_guess(dir_model: Path) -> dict[str, Any] | None:
-        # some models ship no config.json, will try to guess them
-        from conversion import load_all_models
-        load_all_models()
-
-        for matcher, loader in ModelBase._hparams_loaders:
-            if matcher(dir_model):
-                return loader(dir_model)
-        return None
-
-    @classmethod
-    def register_hparams_loader(cls, matcher: HparamsMatcher) -> Callable[[HparamsLoader], HparamsLoader]:
-        def inner(loader: HparamsLoader) -> HparamsLoader:
-            cls._hparams_loaders.append((matcher, loader))
-            return loader
-        return inner
-
     @staticmethod
     def load_hparams(dir_model: Path, is_mistral_format: bool):
         if is_mistral_format:
@@ -1114,10 +1053,6 @@ class ModelBase:
             config = AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
         except Exception as e:
             logger.warning(f"Failed to load model config from {dir_model}: {e}")
-            if not (dir_model / "config.json").is_file():
-                config = ModelBase.load_hparams_guess(dir_model)
-                if config is not None:
-                    return config
             logger.warning("Trying to load config.json instead")
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 config = json.load(f)
@@ -1149,14 +1084,6 @@ 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():
@@ -2706,10 +2633,7 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # Step3-VL keeps text config under text_config but uses a custom top-level architecture.
     # For text conversion we route to a dedicated text-only class.
     # TODO: refactor this later to avoid adding exception here
-    # Kimi-K3's text_config reports "KimiLinearForCausalLM", which is the older
-    # Kimi-Linear-48B architecture and cannot load K3 (no attention residuals,
-    # latent MoE, situ, ...). Route on the top-level architecture instead.
-    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
+    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration"):
         return arch
 
     # if "architectures" is found in the sub-config, use that instead

conversion/bert.py

@@ -15,7 +15,6 @@ 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
 
@@ -241,7 +240,6 @@ class BertModel(TextModel):
 
 
 @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
-@ModelBase.example("distilbert/distilbert-base-uncased")
 class DistilBertModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -265,7 +263,6 @@ class DistilBertModel(BertModel):
 
 
 @ModelBase.register("RobertaModel", "RobertaForSequenceClassification")
-@ModelBase.example("sentence-transformers/stsb-roberta-base")
 class RobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -315,7 +312,6 @@ class RobertaModel(BertModel):
 
 
 @ModelBase.register("NomicBertModel")
-@ModelBase.example("nomic-ai/nomic-embed-text-v1.5")
 class NomicBertModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -404,7 +400,6 @@ class NomicBertModel(BertModel):
 
 
 @ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
-@ModelBase.example("chandar-lab/NeoBERT")
 class NeoBert(BertModel):
     model_arch = gguf.MODEL_ARCH.NEO_BERT
 
@@ -436,7 +431,6 @@ 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
 
@@ -465,7 +459,6 @@ class EuroBertModel(TextModel):
 
 
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
-@ModelBase.example("BAAI/bge-m3")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
     _lora_files = {}
@@ -568,7 +561,6 @@ 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
 
@@ -596,7 +588,6 @@ class JinaBertV2Model(BertModel):
 
 
 @ModelBase.register("ModernBertModel", "ModernBertForMaskedLM", "ModernBertForSequenceClassification")
-@ModelBase.example("answerdotai/ModernBERT-base")
 class ModernBertModel(BertModel):
     model_arch = gguf.MODEL_ARCH.MODERN_BERT
 

conversion/bitnet.py

@@ -9,7 +9,6 @@ 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
 

conversion/bloom.py

@@ -13,7 +13,6 @@ 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
 

conversion/chameleon.py

@@ -12,8 +12,6 @@ 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
 

conversion/chatglm.py

@@ -9,7 +9,6 @@ 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
 

conversion/codeshell.py

@@ -4,7 +4,6 @@ from .base import ModelBase, TextModel, gguf
 
 
 @ModelBase.register("CodeShellForCausalLM")
-@ModelBase.example("WisdomShell/CodeShell-7B")
 class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
 

conversion/cogvlm.py

@@ -11,7 +11,6 @@ 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):
@@ -30,6 +29,5 @@ 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

conversion/command_r.py

@@ -12,8 +12,6 @@ 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
 
@@ -32,8 +30,6 @@ 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
 
@@ -63,7 +59,6 @@ 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

conversion/dbrx.py

@@ -9,7 +9,6 @@ from .base import ModelBase, TextModel, gguf, logger
 
 
 @ModelBase.register("DbrxForCausalLM")
-@ModelBase.example("alpindale/dbrx-instruct")
 class DbrxModel(TextModel):
     model_arch = gguf.MODEL_ARCH.DBRX
 

conversion/deci.py

@@ -13,7 +13,6 @@ 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
 

conversion/deepseek.py

@@ -18,7 +18,6 @@ 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
@@ -101,13 +100,11 @@ 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
 
@@ -137,7 +134,6 @@ class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
 
 
 @ModelBase.register("DeepseekForCausalLM")
-@ModelBase.example("deepseek-ai/deepseek-moe-16b-chat")
 class DeepseekModel(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK
 
@@ -232,7 +228,6 @@ 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
 
@@ -462,7 +457,6 @@ 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
@@ -523,7 +517,6 @@ 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
@@ -716,6 +709,31 @@ class DeepseekV4Model(TextModel):
         for name in tensors_to_remove:
             del self.model_tensors[name]
 
+    @staticmethod
+    def _pack_mxfp4_blocks(weight: Tensor, scale: Tensor) -> np.ndarray:
+        packed = weight.contiguous().view(torch.uint8)
+        scale_u8 = scale.contiguous().view(torch.uint8)
+
+        out_features, packed_cols = packed.shape
+        logical_cols = packed_cols * 2
+        if logical_cols % 32 != 0:
+            raise ValueError(f"MXFP4 source row has {logical_cols} values, expected a multiple of 32")
+
+        n_blocks = logical_cols // 32
+        if tuple(scale_u8.shape) != (out_features, n_blocks):
+            raise ValueError(f"MXFP4 scale shape {tuple(scale_u8.shape)} does not match {(out_features, n_blocks)}")
+
+        src = packed.reshape(out_features, n_blocks, 16)
+        low = src & 0x0F
+        high = (src >> 4) & 0x0F
+
+        # The safetensors bytes store adjacent values as low/high nibbles.
+        # ggml MXFP4 blocks store values 0..15 in low nibbles and 16..31 in high nibbles.
+        vals = torch.stack((low, high), dim=-1).reshape(out_features, n_blocks, 32)
+        qs = vals[:, :, :16] | (vals[:, :, 16:] << 4)
+        raw = torch.cat((scale_u8.unsqueeze(-1), qs.to(torch.uint8)), dim=-1)
+        return raw.reshape(out_features, n_blocks * 17).cpu().numpy()
+
     def _write_mxfp4_expert_tensor(self, bid: int, proj: str, tensor_key: gguf.MODEL_TENSOR) -> list[str]:
         n_experts = self.hparams["n_routed_experts"]
         data: np.ndarray | None = None
@@ -729,7 +747,7 @@ class DeepseekV4Model(TextModel):
 
             weight = LazyTorchTensor.to_eager(self.model_tensors[weight_name]())
             scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
-            packed = self.repack_mxfp4_blocks(weight, scale)
+            packed = self._pack_mxfp4_blocks(weight, scale)
             if data is None:
                 data = np.empty((n_experts, *packed.shape), dtype=packed.dtype)
             data[eid] = packed
@@ -918,7 +936,6 @@ class DeepseekV4Model(TextModel):
 
 
 @ModelBase.register("DeepseekV4DSparkModel")
-@ModelBase.example("deepseek-ai/DeepSeek-V4-Flash-DSpark")
 class DeepseekV4DSparkModel(DeepseekV4Model):
     model_arch = gguf.MODEL_ARCH.DFLASH
 

conversion/dots1.py

@@ -11,7 +11,6 @@ from .qwen import Qwen2MoeModel
 
 
 @ModelBase.register("Dots1ForCausalLM")
-@ModelBase.example("rednote-hilab/dots.llm1.inst")
 class Dots1Model(Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.DOTS1
 

conversion/dots3.py

@@ -1,323 +0,0 @@
-from __future__ import annotations
-
-import math
-import re
-
-import torch
-
-from typing import TYPE_CHECKING, Any, Callable, Iterable
-
-if TYPE_CHECKING:
-    from torch import Tensor
-
-from .base import MmprojModel, ModelBase, gguf
-
-from .deepseek import DeepseekV2Model
-
-
-@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration", "Dots3NoteTextForCausalLM")
-class Dots3NoteModel(DeepseekV2Model):
-    model_arch = gguf.MODEL_ARCH.DOTS3NOTE
-    skip_mtp = False
-    supports_mtp_export = True
-
-    # trunk layer count, stashed before indexing for filter_tensors (mirrors DeepseekV32Model)
-    _n_main_layers: int | None = None
-
-    def index_tensors(self, remote_hf_model_id: str | None = None):
-        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        hparams = self.hparams
-
-        # config file doesn't specify MTP block, detect it from model weight
-        self.n_nextn = 1 if "model.mtp.embed_tokens.weight" in self.model_tensors else 0
-        if self.n_nextn:
-            self.block_count += self.n_nextn
-            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-        self.layer_types = hparams["layer_types"]
-        if len(self.layer_types) < hparams["num_hidden_layers"]:
-            raise ValueError("layer_types is shorter than num_hidden_layers")
-
-        if hparams.get("use_dsa", True) is not True:
-            raise ValueError("dots3-note conversion requires use_dsa=true")
-        if hparams.get("normalization", "RMSNorm") != "RMSNorm" or hparams.get("final_norm", "RMSNorm") != "RMSNorm":
-            raise ValueError("dots3-note conversion only supports RMSNorm")
-        if hparams.get("k_rope_only_layernorm", True) is not True:
-            raise ValueError("dots3-note conversion requires k_rope_only_layernorm=true")
-        if hparams.get("topk_method", "noaux_tc") != "noaux_tc" or hparams.get("scoring_func") != "sigmoid":
-            raise ValueError("dots3-note conversion only supports noaux_tc/sigmoid expert gating")
-        if hparams.get("n_group", 1) != 1 or hparams.get("topk_group", 1) != 1:
-            raise ValueError("dots3-note conversion does not support grouped expert routing")
-        if hparams.get("use_dynamic_rsf", False) or hparams.get("moe_gating_fp32", False):
-            raise ValueError("dots3-note conversion does not support use_dynamic_rsf/moe_gating_fp32")
-        for key in ("attention_gate_type", "swa_attention_gate_type"):
-            if hparams.get(key, "headwise") != "headwise":
-                raise ValueError(f"dots3-note conversion only supports headwise attention gate, got {key}={hparams.get(key)!r}")
-        if hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"] != hparams.get("swa_head_dim", 256):
-            raise ValueError("swa_head_dim must equal swa_qk_nope_head_dim + swa_qk_rope_head_dim")
-        if hparams["swa_qk_rope_head_dim"] != hparams["qk_rope_head_dim"]:
-            # both layer kinds share a single rope_dimension_count
-            raise ValueError("swa_qk_rope_head_dim must match qk_rope_head_dim")
-
-        self.apply_lora_rescale = hparams.get("apply_mla_qkv_lora_rescale", False)
-
-    def _is_swa_layer(self, bid: int) -> bool:
-        if bid >= self.hparams["num_hidden_layers"]:
-            # note: the NextN/MTP block uses the sliding-attention MLA
-            return True
-        return self.layer_types[bid] == "sliding_attention"
-
-    def set_vocab(self):
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        tokens, toktypes, tokpre = self.get_vocab_base()
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_tokenizer_pre(tokpre)
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab.add_to_gguf(self.gguf_writer)
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        if (titem := super().filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-        if name.startswith(("vision_encoder.", "audio_encoder.")):
-            return None
-
-        assert cls._n_main_layers is not None
-        is_mtp = name.startswith("model.mtp.") or \
-            ((m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers)
-
-        # --no-mtp: drop the NextN/MTP block; --mtp: keep only that block plus the shared embeddings/norm/lm_head
-        if is_mtp and cls.no_mtp:
-            return None
-        if cls.mtp_only and not is_mtp and name not in (
-            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-        ):
-            return None
-
-        return name, gen
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-
-        # head_count is a per-layer array because the two layer kinds have different head counts
-        n_layer = hparams["num_hidden_layers"]
-        hparams["num_attention_heads"] = [
-            hparams["swa_num_attention_heads"] if self._is_swa_layer(il) else hparams["num_attention_heads"]
-            for il in range(self.block_count)
-        ]
-
-        # prevent the base class from emitting key/value_length from the unused head_dim
-        hparams.pop("head_dim", None)
-
-        super().set_gguf_parameters()
-
-        # MLA geometry of the sliding-window layers (rope.freq_base_swa is emitted by the base class)
-        swa_kv_lora_rank = hparams["swa_kv_lora_rank"]
-        self.gguf_writer.add_sliding_window(hparams["sliding_window_size"])
-        self.gguf_writer.add_sliding_window_pattern([self._is_swa_layer(il) for il in range(n_layer)])
-        self.gguf_writer.add_kv_lora_rank_swa(swa_kv_lora_rank)
-        self.gguf_writer.add_key_length_swa(swa_kv_lora_rank + hparams["swa_qk_rope_head_dim"])
-        self.gguf_writer.add_value_length_swa(swa_kv_lora_rank)
-        self.gguf_writer.add_key_length_mla_swa(hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"])
-        self.gguf_writer.add_value_length_mla_swa(hparams["swa_v_head_dim"])
-        if hparams["swa_q_lora_rank"] != hparams["q_lora_rank"]:
-            raise ValueError("dots3-note conversion assumes a shared q_lora_rank for both layer kinds")
-
-        if self.n_nextn:
-            self.gguf_writer.add_nextn_predict_layers(self.n_nextn)
-
-        # DSA indexer (full-attention layers only)
-        self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
-        self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
-        self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
-        self.gguf_writer.add_indexer_types([not self._is_swa_layer(il) for il in range(n_layer)])
-
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # move the MTP token embedding into the NextN block so the standard nextn mapping picks it up
-        if name == "model.mtp.embed_tokens.weight":
-            name = f"model.layers.{self.hparams['num_hidden_layers']}.embed_tokens.weight"
-            bid = self.hparams["num_hidden_layers"]
-
-        # fold the activation rescale sqrt(n_embd/lora_rank) into the preceding RMSNorm weight
-        # this also covers the indexer wq_b, which reads the same rescaled q_lora activation
-        if self.apply_lora_rescale and bid is not None:
-            if name.endswith("q_a_layernorm.weight"):
-                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / self.hparams["q_lora_rank"])
-            elif name.endswith("kv_a_layernorm.weight"):
-                rank = self.hparams["swa_kv_lora_rank"] if self._is_swa_layer(bid) else self.hparams["kv_lora_rank"]
-                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / rank)
-
-        # MLA absorption: split kv_b_proj into k_b (transposed) and v_b, per-layer-kind geometry
-        if name.endswith("kv_b_proj.weight"):
-            assert bid is not None
-            if self._is_swa_layer(bid):
-                n_head = self.hparams["swa_num_attention_heads"]
-                qk_nope_head_dim = self.hparams["swa_qk_nope_head_dim"]
-                v_head_dim = self.hparams["swa_v_head_dim"]
-            else:
-                n_head = self.hparams["num_attention_heads"]
-                qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
-                v_head_dim = self.hparams["v_head_dim"]
-            if isinstance(n_head, list):  # set_gguf_parameters turns this into a per-layer array
-                n_head = n_head[bid]
-
-            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
-
-            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
-            k_b, v_b = kv_b.split([qk_nope_head_dim, v_head_dim], dim=1)
-            k_b = k_b.transpose(1, 2)
-
-            yield from ModelBase.modify_tensors(self, k_b, name.replace("kv_b_proj", "k_b_proj"), bid)
-            yield from ModelBase.modify_tensors(self, v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
-            return
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-
-@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration")
-class Dots3NoteMmprojModel(MmprojModel):
-    has_vision_encoder = True
-    has_audio_encoder = True
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        assert self.hparams_vision is not None
-        assert self.hparams_audio is not None
-
-        # preprocessor_config.json nests the image params under vision_config
-        self.preprocessor_config = {**self.preprocessor_config, **self.preprocessor_config.get("vision_config", {})}
-
-        vis = self.hparams_vision
-        # in this config, hidden_size is the adapter output width; embed_dim is the tower width
-        vis["hidden_size"] = vis["embed_dim"]
-        vis["image_size"] = 0  # dynamic resolution
-        self.pyramid = [max(0, n) for n in vis["pyramid_num_routed"]]
-
-        if vis.get("adapter_type") != "patch_merger" or not vis.get("pre_pixel_shuffle"):
-            raise ValueError("dots3-note vision conversion requires adapter_type=patch_merger and pre_pixel_shuffle")
-        if vis.get("router_scoring_func", "sigmoid") != "sigmoid" or vis.get("router_scale", 1.0) != 1.0:
-            raise ValueError("dots3-note vision conversion only supports sigmoid routing with router_scale=1.0")
-        if vis.get("temporal_patch_size", 1) != 1 or vis.get("use_bias") or not vis.get("use_qk_norm"):
-            raise ValueError("unsupported dots3-note vision config variant")
-
-        aud = self.hparams_audio
-        if not aud.get("use_conv2d_stem") or not aud.get("use_rope") or not aud.get("use_rms_norm") or aud.get("use_causal"):
-            raise ValueError("unsupported dots3-note audio config variant")
-        if aud["whisper_config"].get("activation_function") != "swiglu":
-            raise ValueError("dots3-note audio conversion requires the swiglu activation")
-        if aud.get("merge_factor", 1) != 1 or aud.get("chunk_seconds") != 60:
-            raise ValueError("unsupported dots3-note audio chunking config")
-        # the graph hard-codes these rope parameters
-        rope = aud.get("rope_parameters", {})
-        if rope.get("partial_rotary_factor") != 0.5 or rope.get("rope_theta") != 10000.0:
-            raise ValueError("unsupported dots3-note audio rope config")
-
-    def get_audio_config(self) -> dict[str, Any] | None:
-        cfg = self.global_config.get("audio_config")
-        if cfg is not None:
-            # aliases so MmprojModel.find_aparam() / n_block_keys can resolve them
-            whisper = cfg["whisper_config"]
-            cfg["hidden_size"] = whisper["d_model"]
-            cfg["intermediate_size"] = whisper["encoder_ffn_dim"]
-            cfg["num_attention_heads"] = whisper["encoder_attention_heads"]
-            cfg["num_hidden_layers"] = whisper["encoder_layers"]
-        return cfg
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        assert self.hparams_vision is not None
-        assert self.hparams_audio is not None
-
-        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.DOTS3NOTE_V)
-        self.gguf_writer.add_vision_use_silu(True)
-        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision["rms_norm_eps"])
-        self.gguf_writer.add_vision_spatial_merge_size(self.hparams_vision["spatial_merge_size"])
-        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
-        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
-        # pyramid MoE: per-block routed expert count, 0 = dense block
-        self.gguf_writer.add_vision_expert_count_per_layer(self.pyramid)
-        self.gguf_writer.add_vision_expert_used_count(int(self.hparams_vision["capacity_factor"]))
-
-        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.DOTS3NOTE_A)
-        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["whisper_config"]["num_mel_bins"])
-        self.gguf_writer.add_audio_attention_layernorm_eps(1e-6)  # Dots3NoteAudioRMSNorm default
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, _ = item
-        if not name.startswith(("vision_encoder.", "audio_encoder.")):
-            return None
-        return super().filter_tensors(item)
-
-    _vis_experts: dict[int, dict[str, Tensor]] | None = None
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # router params have no .weight suffix in the checkpoint, but gguf tools expect one
-        if name.endswith((".gate_weight", ".router_bias")):
-            name += ".weight"
-
-        # audio fc1 fuses gate and up for swiglu; split it
-        if ".speech_encoder.layers." in name and ".fc1." in name:
-            gate, up = data_torch.chunk(2, dim=0)
-            yield from super().modify_tensors(gate, name.replace(".fc1.", ".fc1_gate."), bid)
-            yield from super().modify_tensors(up, name.replace(".fc1.", ".fc1_up."), bid)
-            return
-
-        # vision MoE: stack per-expert weights into a single 3D tensor per block
-        if ".mlp.experts." in name:
-            assert bid is not None
-            n_expert = self.pyramid[bid]
-            if self._vis_experts is None:
-                self._vis_experts = {}
-            buf = self._vis_experts.setdefault(bid, {})
-            buf[name] = data_torch
-
-            if len(buf) >= n_expert * 3:
-                for w_name in ("fc1", "fc2", "fc3"):
-                    datas: list[Tensor] = []
-                    for xid in range(n_expert):
-                        ename = f"vision_encoder.blocks.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                        datas.append(buf.pop(ename))
-                    merged = torch.stack(datas, dim=0)
-                    yield from super().modify_tensors(merged, f"vision_encoder.blocks.{bid}.mlp.experts.{w_name}.weight", bid)
-            return
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-        if self._vis_experts is not None:
-            leftover = [k for d in self._vis_experts.values() for k in d.keys()]
-            if leftover:
-                raise ValueError(f"unprocessed vision experts: {leftover}")
-
-    def tensor_force_quant(self, name, new_name, bid, n_dims):
-        # FP32 routing is load-bearing for the vision MoE (near-tied expert scores)
-        if ".ffn_gate_inp." in new_name or ".exp_probs_b." in new_name:
-            return gguf.GGMLQuantizationType.F32
-        if ".conv2d" in new_name or "a.conv_out" in new_name:
-            return gguf.GGMLQuantizationType.F32
-        return super().tensor_force_quant(name, new_name, bid, n_dims)