CI/tests: refactor argument handling, add --errors-only to emit only WARN/ERR messages

2026-10-26 06:27:38 -05:00 · 2026-09-17 18:47:59 +02:00
669 changed files with 58506 additions and 94532 deletions
@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2026.1.1-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
@@ -19,7 +19,7 @@ RUN npm ci
 COPY tools/ui/ ./
 RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build

-FROM docker.io/intel/oneapi-toolkit:$ONEAPI_VERSION AS build
+FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS build

 ARG GGML_SYCL_F16=ON
 ARG LEVEL_ZERO_VERSION=1.28.2
@@ -59,7 +59,7 @@ RUN mkdir -p /app/full \
    && cp requirements.txt /app/full \
    && cp .devops/tools.sh /app/full/tools.sh

-FROM docker.io/intel/oneapi-toolkit:$ONEAPI_VERSION AS base
+FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS base

 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
@@ -1,9 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc4.3.0
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=registry.mthreads.com/mcconline/musa_sdk:5.2.0-devel-ubuntu${UBUNTU_VERSION}-s5000
+ARG BASE_MUSA_DEV_CONTAINER=docker.io/mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64

-ARG BASE_MUSA_RUN_CONTAINER=registry.mthreads.com/mcconline/musa_sdk:5.2.0-runtime-ubuntu${UBUNTU_VERSION}-s5000
+ARG BASE_MUSA_RUN_CONTAINER=docker.io/mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64

 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
@@ -36,10 +37,7 @@ RUN apt-get update && \
    python3-pip \
    git \
    libssl-dev \
-    libgomp1 \
-    musa-mualg-5-2 \
-    musa-muthrust-5-2 \
-    libmthreads-compute
+    libgomp1

 WORKDIR /app

@@ -82,16 +80,13 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
      org.opencontainers.image.source=$IMAGE_SOURCE

 RUN apt-get update \
-    && apt-get install -y libgomp1 curl ffmpeg libmthreads-compute \
+    && apt-get install -y libgomp1 curl ffmpeg \
    && apt autoremove -y \
    && apt clean -y \
    && rm -rf /tmp/* /var/tmp/* \
    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
    && find /var/cache -type f -delete

-# The MUSA runtime image does not register its library directory
-RUN echo "/usr/local/musa/lib" > /etc/ld.so.conf.d/musa-runtime.conf && ldconfig
-
 COPY --from=build /app/lib/ /app

 ### Full
@@ -1,5 +1,5 @@
-ARG OPENVINO_VERSION_MAJOR=2026.4
-ARG OPENVINO_VERSION_FULL=2026.4.0.22959.99c81491cc3
+ARG OPENVINO_VERSION_MAJOR=2026.3.1
+ARG OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
@@ -10,9 +10,9 @@ ARG COMPUTE_RUNTIME_VERSION_FULL=26.31.39395.13-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.38.0
-ARG NPU_DRIVER_FULL=v1.38.0.20260910-34487311128
-ARG LIBZE1_VERSION=1.32.0-1~24.04~ppa1
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -173,7 +173,7 @@ 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 "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260830T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -29,7 +29,7 @@ concurrency:

 jobs:
  android-ndk-snapdragon:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.7'
    defaults:
@@ -59,7 +59,7 @@ jobs:
          path: pkg-snapdragon/llama.cpp

  linux-iot-snapdragon:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.7'
    defaults:
@@ -33,7 +33,7 @@ env:

 jobs:
  default:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest

    steps:
      - name: Clone
@@ -49,7 +49,7 @@ jobs:
          distribution: zulu

      - name: Setup Android SDK
-        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
        with:
          log-accepted-android-sdk-licenses: false

@@ -59,7 +59,7 @@ jobs:
          ./gradlew build --no-daemon

  ndk:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
    defaults:
@@ -93,7 +93,7 @@ jobs:
          path: pkg-adb/llama.cpp

  arm64:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest

    env:
      NDK_VERSION: "29.0.14206865"
@@ -123,7 +123,7 @@ jobs:
          distribution: temurin

      - name: Setup Android SDK
-        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
        with:
          log-accepted-android-sdk-licenses: false

@@ -33,7 +33,6 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  GGML_SCHED_DEBUG_REALLOC: 1
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
  LLAMA_ARG_LOG_TIMESTAMPS: 1
@@ -99,8 +98,7 @@ jobs:
        id: cmake_test
        run: |
          cd build
-          # ref: https://github.com/ggml-org/llama.cpp/pull/19802#issuecomment-4013704023
-          ctest -L main -E "test-llama-archs|test-save-load-state" --verbose --timeout 900
+          ctest -L main -E "test-llama-archs" --verbose --timeout 900

  macos-latest-x64:
    runs-on: macos-15-intel
@@ -41,8 +41,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

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

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

    steps:
      - name: Clone
@@ -5,7 +5,7 @@ on:

 jobs:
  linux:
-    runs-on: [self-hosted, Linux, CPU]
+    runs-on: [self-hosted, Linux]
    steps:
      - uses: actions/checkout@v6
        with:
@@ -37,7 +37,6 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  GGML_SCHED_DEBUG_REALLOC: 1
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
  LLAMA_ARG_LOG_TIMESTAMPS: 1
@@ -89,7 +88,7 @@ jobs:
        run: |
          export PIP_BREAK_SYSTEM_PACKAGES="1"
          python3 -m pip install --upgrade pip setuptools
-          pip3 install ./gguf-py jinja2==3.1.6
+          pip3 install ./gguf-py

      - name: ccache-buckets-restore
        uses: ./.github/actions/ccache-buckets
@@ -125,7 +124,7 @@ jobs:
        id: cmake_test
        run: |
          cd build
-          ctest -L 'main|python' --verbose --timeout 900
+          ctest -L main --verbose --timeout 900

      - name: Test llama2c conversion
        id: llama2c_test
@@ -68,7 +68,7 @@ jobs:
          hf_bucket: ggml-org/cache

      - name: Build with CMake
-        # TODO: Drop GGML_CUDA_CCCL_VERSION when this job uses CTK >= 13.5, which bundles CCCL >= 3.5.
+        # 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: |
          cmake -S . -B build -G Ninja \
            -DLLAMA_FATAL_WARNINGS=ON \
@@ -77,7 +77,7 @@ jobs:
            -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
            -DGGML_NATIVE=OFF \
            -DGGML_CUDA=ON \
-            -DGGML_CUDA_CCCL_VERSION=v3.4.3
+            -DGGML_CUDA_CUB_3DOT2=ON
          cmake --build build

      - name: ccache-buckets-save
@@ -145,7 +145,7 @@ jobs:

  musa:
    runs-on: ubuntu-22.04
-    container: registry.mthreads.com/mcconline/musa_sdk:5.2.0-devel-ubuntu22.04-s5000
+    container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64

    steps:
      - name: Clone
@@ -156,7 +156,7 @@ jobs:
        id: depends
        run: |
          apt-get update
-          apt-get install -y build-essential git cmake libssl-dev jq python3-venv musa-mualg-5-2 musa-muthrust-5-2 libmthreads-compute
+          apt-get install -y build-essential git cmake libssl-dev jq

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
@@ -178,8 +178,8 @@ jobs:
        run: |
          cmake -B build -S . \
            -DGGML_MUSA=ON \
-            -DMUSA_ARCHITECTURES=31
-          cmake --build build --config Release -j $(nproc)
+            -DMUSA_ARCHITECTURES=21
+          time cmake --build build --config Release -j $(nproc)

      - name: ccache-buckets-save
        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -31,16 +31,15 @@ jobs:
    strategy:
      matrix:
        include:
-          # CTK >= 13.5 bundles CCCL >= 3.5; omit GGML_CUDA_CCCL_VERSION for those versions.
          - cuda: '12.4'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
          - cuda: '13.4'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            defines: ''
          - cuda: '13.4'
            arch: arm64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3 -DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -100,46 +100,6 @@ jobs:
          wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
          ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

-  ubuntu-26-zdnn-s390x:
-    name: ubuntu-26-zdnn-s390x
-    runs-on: ubuntu-24.04-s390x
-    container: ubuntu:26.04  # required to get GCC 15.1 and binutils 2.44
-    defaults:
-      run:
-        shell: bash
-
-    steps:
-      - name: Build Dependencies
-        id: build_depends
-        run: |
-          apt-get update
-          apt-get install -y --no-install-recommends \
-            build-essential cmake git ca-certificates \
-            libssl-dev libzdnn-dev
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Toolchain workaround (GCC 15)
-        run: |
-          apt-get install -y gcc-15 g++-15
-          echo "CC=gcc-15" >> "$GITHUB_ENV"
-          echo "CXX=g++-15" >> "$GITHUB_ENV"
-
-      - name: Build with zDNN Backend
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_NATIVE=OFF \
-            -DGGML_VXE=ON \
-            -DGGML_ZDNN=ON \
-            -DGGML_RPC=ON \
-            -DCMAKE_C_FLAGS="-march=arch15" \
-            -DCMAKE_CXX_FLAGS="-march=arch15"
-          time cmake --build build --config Release -j $(nproc)
-
  ubuntu-24-ppc64le:
    runs-on: ubuntu-24.04-ppc64le

@@ -41,8 +41,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

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

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

    steps:
      - name: Clone
@@ -0,0 +1,539 @@
+name: CI (self-hosted)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-self-hosted.yml',
+      'ci/run.sh',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp',
+      '**/*.glsl',
+      '**/*.wgsl'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-self-hosted.yml',
+      'ci/run.sh',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp',
+      '**/*.glsl',
+      '**/*.wgsl'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
+  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  gpu-cuda:
+    runs-on: "hf-jobs-t4-small:cuda13"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y cmake libssl-dev time unzip wget python3 python3-venv python3-pip
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        with:
+          key: self-hosted-gpu-cuda
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          nvidia-smi
+          GG_BUILD_CUDA=1 CUDACXX=/usr/local/cuda/bin/nvcc bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: self-hosted-gpu-cuda
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
+  gpu-rocm:
+    runs-on: [self-hosted, Linux, AMD]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        # HIP_LAUNCH_BLOCKING=1: workaround for an async-execution correctness
+        # issue on integrated RDNA3.5 (gfx1151) where batched inference returns
+        # incorrect output (perplexity ~88 vs ~9.4). Serializing kernel launches
+        # restores correctness. Remove once the underlying ROCm/HIP issue is fixed.
+        env:
+          HIP_LAUNCH_BLOCKING: "1"
+        run: |
+          rocminfo
+          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-vulkan-nvidia-cm:
+    # runs-on: "hf-jobs-t4-small:ubuntu26_04"
+    runs-on: [self-hosted, Linux, NVIDIA]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      # - name: Install dependencies
+      #   run: |
+      #     sudo apt update
+      #     sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
+
+      # - name: ccache
+      #   uses: ggml-org/ccache-action@v1.2.24
+      #   with:
+      #     restore: false
+      #     save: false
+
+      # - name: ccache-buckets-restore
+      #   uses: ./.github/actions/ccache-buckets
+      #   with:
+      #     key: self-hosted-vulkan-nvidia-cm
+      #     folder: llama.cpp
+      #     hf_bucket: ggml-org/cache
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+      # - name: ccache-buckets-save
+      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      #   uses: ./.github/actions/ccache-buckets
+      #   env:
+      #     HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+      #   with:
+      #     key: self-hosted-vulkan-nvidia-cm
+      #     folder: llama.cpp
+      #     evict-old-files: 1d
+      #     hf_bucket: ggml-org/cache
+      #     save: true
+
+  gpu-vulkan-nvidia-cm2:
+    # runs-on: "hf-jobs-t4-small:ubuntu26_04"
+    runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      # - name: Install dependencies
+      #   run: |
+      #     sudo apt update
+      #     sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
+
+      # - name: ccache
+      #   uses: ggml-org/ccache-action@v1.2.24
+      #   with:
+      #     restore: false
+      #     save: false
+
+      # - name: ccache-buckets-restore
+      #   uses: ./.github/actions/ccache-buckets
+      #   with:
+      #     key: self-hosted-vulkan-nvidia-cm2
+      #     folder: llama.cpp
+      #     hf_bucket: ggml-org/cache
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+      # - name: ccache-buckets-save
+      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      #   uses: ./.github/actions/ccache-buckets
+      #   env:
+      #     HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+      #   with:
+      #     key: self-hosted-vulkan-nvidia-cm2
+      #     folder: llama.cpp
+      #     evict-old-files: 1d
+      #     hf_bucket: ggml-org/cache
+      #     save: true
+
+  gpu-webgpu-nvidia:
+    runs-on: "hf-jobs-t4-small:ubuntu26_04"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        with:
+          key: self-hosted-webgpu-nvidia
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v20260908.214631"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_WEBGPU=1 \
+          GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+          GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
+            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: self-hosted-webgpu-nvidia
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
+  # TODO: provision AMX-compatible machine
+  #cpu-amx:
+  #  runs-on: [self-hosted, Linux, CPU, AMX]
+
+  #  steps:
+  #    - name: Clone
+  #      id: checkout
+  #      uses: actions/checkout@v6
+
+  #    - name: Test
+  #      id: ggml-ci
+  #      run: |
+  #        bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  # TODO: provision AMD GPU machine
+  # amd-vulkan:
+  #   runs-on: [self-hosted, Linux, AMD]
+
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v6
+
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         vulkaninfo --summary
+  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  # TODO: provision AMD GPU machine
+  # amd-rocm:
+  #   runs-on: [self-hosted, Linux, AMD]
+
+  #   steps:
+  #     - name: Clone
+  #       id: checkout
+  #       uses: actions/checkout@v6
+
+  #     - name: Test
+  #       id: ggml-ci
+  #       run: |
+  #         amd-smi static
+  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-webgpu-apple:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v20260908.214631"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-vulkan-apple:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-vulkan-intel-linux:
+    runs-on: [self-hosted, Linux, Intel]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  gpu-vulkan-intel-windows:
+    runs-on: [self-hosted, Windows, X64, Intel]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        shell: C:\msys64\usr\bin\bash.exe --noprofile --norc -eo pipefail "{0}"
+        env:
+          MSYSTEM: UCRT64
+          CHERE_INVOKING: 1
+          PATH: C:\msys64\ucrt64\bin;C:\msys64\usr\bin;C:\Windows\System32;${{ env.PATH }}
+        run: |
+          vulkaninfo --summary
+          # Skip python related tests with GG_BUILD_LOW_PERF=1 since Windows MSYS2 UCRT64 currently fails to create
+          # a valid python environment for testing
+          LLAMA_FATAL_WARNINGS=OFF GG_BUILD_NINJA=1 GG_BUILD_VULKAN=1 GG_BUILD_LOW_PERF=1 ./ci/run.sh ./results/llama.cpp ./mnt/llama.cpp
+
+  gpu-openvino-low-perf:
+    runs-on: [self-hosted, Linux, Intel, OpenVINO]
+
+    env:
+      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Setup OpenVINO Toolkit
+        uses: ./.github/actions/linux-setup-openvino
+        with:
+          path: ./openvino_toolkit
+          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
+          version_full: ${{ env.OPENVINO_VERSION_FULL }}
+
+      - name: Install OpenVINO dependencies
+        run: |
+          cd ./openvino_toolkit
+          chmod +x ./install_dependencies/install_openvino_dependencies.sh
+          echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          source ./openvino_toolkit/setupvars.sh
+          GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  cpu-x64-high-perf:
+    runs-on: [self-hosted, Linux, X64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  cpu-arm64-high-perf-graviton4:
+    runs-on: ah-ubuntu_24_04-c8g_8x
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          set -euxo pipefail
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+          apt-get install -y \
+          build-essential \
+          python3-venv \
+          gpg \
+          wget \
+          time \
+          git-lfs
+
+          git lfs install
+
+          # install the latest cmake
+          sudo install -d /usr/share/keyrings
+          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+            | gpg --dearmor \
+            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+            | sudo tee /etc/apt/sources.list.d/kitware.list
+          sudo apt-get update
+          sudo apt-get install -y cmake
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          LLAMA_ARG_THREADS=$(nproc) \
+          GG_BUILD_HIGH_PERF=1 \
+          GG_BUILD_NO_BF16=1 \
+          GG_BUILD_EXTRA_TESTS_0=1 \
+          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  cpu-arm64-graviton4-kleidiai:
+    runs-on: ah-ubuntu_24_04-c8g_8x
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          set -euxo pipefail
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+          apt-get install -y \
+          build-essential \
+          python3-venv \
+          gpg \
+          wget \
+          time \
+          git-lfs
+
+          git lfs install
+
+          # install the latest cmake
+          sudo install -d /usr/share/keyrings
+          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+            | gpg --dearmor \
+            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+            | sudo tee /etc/apt/sources.list.d/kitware.list
+          sudo apt-get update
+          sudo apt-get install -y cmake
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          LLAMA_ARG_THREADS=$(nproc) \
+          GG_BUILD_KLEIDIAI=1 \
+          GG_BUILD_EXTRA_TESTS_0=1 \
+          GG_BUILD_HIGH_PERF=1 \
+          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -48,8 +48,8 @@ jobs:

    env:
      ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2026.1"
-      LEVEL_ZERO_VERSION: "1.33.1"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      LEVEL_ZERO_VERSION: "1.28.2"
      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"

    continue-on-error: true
@@ -63,17 +63,16 @@ jobs:
        shell: bash
        run: |
          cd /tmp
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/5996e26b-f48a-42b1-8db0-b002ad0bd8d7/intel-oneapi-toolkit-2026.1.1.33_offline.sh -O intel-oneapi-toolkit_offline.sh
-          sudo bash intel-oneapi-toolkit_offline.sh -s -a --silent --eula accept
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept

      - name: Install Level Zero SDK
        shell: bash
        run: |
          cd /tmp
-          # v1.33.x renamed the Debian packages to libze1 / libze-dev
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze1_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze1.deb
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze-dev_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze-dev.deb
-          sudo apt-get install -y ./libze1.deb ./libze-dev.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
@@ -102,11 +101,7 @@ jobs:
            -DCMAKE_CXX_COMPILER=icpx \
            -DLLAMA_OPENSSL=OFF \
            -DGGML_NATIVE=OFF \
-            -DGGML_SYCL_F16=${{ matrix.fp16 }} \
-            -DGGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON \
-            -DGGML_SYCL_DNN=ON \
-            -DCMAKE_CXX_FLAGS="-fsycl-unnamed-lambda" \
-            -DCMAKE_EXE_LINKER_FLAGS="-fsycl-unnamed-lambda"
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
          time cmake --build build --config Release -j $(nproc)

      - name: ccache-buckets-save
@@ -129,11 +124,11 @@ jobs:
        shell: bash

    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/0cb67a0d-67f6-410b-868b-f4a0a17ff0cf/intel-oneapi-toolkit-2026.1.1.32_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2026.1"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
    steps:
      - name: Clone
        id: checkout
@@ -31,7 +31,6 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  GGML_SCHED_DEBUG_REALLOC: 1
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
  LLAMA_ARG_LOG_TIMESTAMPS: 1
@@ -19,7 +19,7 @@ on:

 jobs:
  check-vendor:
-    runs-on: ubuntu-slim
+    runs-on: [self-hosted, fast]

    steps:
      - name: Checkout
@@ -1,112 +0,0 @@
-name: CI (self-hosted CPU backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-cpu.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-cpu.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  cpu-x64-high-perf:
-    runs-on: [self-hosted, Linux, X64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  cpu-arm64-high-perf-graviton4:
-    runs-on: ah-ubuntu_24_04-c8g_8x
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          set -euxo pipefail
-          sudo apt-get update
-          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
-          apt-get install -y \
-          build-essential \
-          python3-venv \
-          gpg \
-          wget \
-          time \
-          git-lfs
-
-          git lfs install
-
-          # install the latest cmake
-          sudo install -d /usr/share/keyrings
-          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
-            | gpg --dearmor \
-            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
-          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
-            | sudo tee /etc/apt/sources.list.d/kitware.list
-          sudo apt-get update
-          sudo apt-get install -y cmake
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          LLAMA_ARG_THREADS=$(nproc) \
-          GG_BUILD_HIGH_PERF=1 \
-          GG_BUILD_NO_BF16=1 \
-          GG_BUILD_EXTRA_TESTS_0=1 \
-          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  # TODO: provision AMX-compatible machine
-  #cpu-amx:
-  #  runs-on: [self-hosted, Linux, CPU, AMX]
-
-  #  steps:
-  #    - name: Clone
-  #      id: checkout
-  #      uses: actions/checkout@v6
-
-  #    - name: Test
-  #      id: ggml-ci
-  #      run: |
-  #        bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,124 +0,0 @@
-name: CI (self-hosted CUDA backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-cuda.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-cuda.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**',
-      'ggml/src/ggml-cuda/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  gpu-cuda:
-    runs-on: "hf-jobs-t4-small:cuda13"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y cmake libssl-dev time unzip wget python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: self-hosted-gpu-cuda
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          nvidia-smi
-          GG_BUILD_CUDA=1 CUDACXX=/usr/local/cuda/bin/nvcc bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: self-hosted-gpu-cuda
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-  gpu-rocm:
-    runs-on: [self-hosted, Linux, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        # HIP_LAUNCH_BLOCKING=1: workaround for an async-execution correctness
-        # issue on integrated RDNA3.5 (gfx1151) where batched inference returns
-        # incorrect output (perplexity ~88 vs ~9.4). Serializing kernel launches
-        # restores correctness. Remove once the underlying ROCm/HIP issue is fixed.
-        env:
-          HIP_LAUNCH_BLOCKING: "1"
-        run: |
-          rocminfo
-          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  # TODO: provision AMD GPU machine
-  # amd-rocm:
-  #   runs-on: [self-hosted, Linux, AMD]
-
-  #   steps:
-  #     - name: Clone
-  #       id: checkout
-  #       uses: actions/checkout@v6
-
-  #     - name: Test
-  #       id: ggml-ci
-  #       run: |
-  #         amd-smi static
-  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,85 +0,0 @@
-name: CI (self-hosted KleidiAI backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-kleidiai.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-kleidiai.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  cpu-arm64-graviton4-kleidiai:
-    runs-on: ah-ubuntu_24_04-c8g_8x
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          set -euxo pipefail
-          sudo apt-get update
-          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
-          apt-get install -y \
-          build-essential \
-          python3-venv \
-          gpg \
-          wget \
-          time \
-          git-lfs
-
-          git lfs install
-
-          # install the latest cmake
-          sudo install -d /usr/share/keyrings
-          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
-            | gpg --dearmor \
-            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
-          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
-            | sudo tee /etc/apt/sources.list.d/kitware.list
-          sudo apt-get update
-          sudo apt-get install -y cmake
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          LLAMA_ARG_THREADS=$(nproc) \
-          GG_BUILD_KLEIDIAI=1 \
-          GG_BUILD_EXTRA_TESTS_0=1 \
-          GG_BUILD_HIGH_PERF=1 \
-          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,59 +0,0 @@
-name: CI (self-hosted Metal backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-metal.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.swift',
-      '**/*.m',
-      '**/*.metal'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-metal.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**',
-      'ggml/src/ggml-metal/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  gpu-metal:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,75 +0,0 @@
-name: CI (self-hosted OpenVINO backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-openvino.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-openvino.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**',
-      'ggml/src/ggml-openvino/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  gpu-openvino-low-perf:
-    runs-on: [self-hosted, Linux, Intel, OpenVINO]
-
-    env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Setup OpenVINO Toolkit
-        uses: ./.github/actions/linux-setup-openvino
-        with:
-          path: ./openvino_toolkit
-          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
-          version_full: ${{ env.OPENVINO_VERSION_FULL }}
-
-      - name: Install OpenVINO dependencies
-        run: |
-          cd ./openvino_toolkit
-          chmod +x ./install_dependencies/install_openvino_dependencies.sh
-          echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          source ./openvino_toolkit/setupvars.sh
-          GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,201 +0,0 @@
-name: CI (self-hosted Vulkan backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-vulkan.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.comp',
-      '**/*.glsl'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-vulkan.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**',
-      'ggml/src/ggml-vulkan/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  gpu-vulkan-nvidia-cm:
-    # runs-on: "hf-jobs-t4-small:ubuntu26_04"
-    runs-on: [self-hosted, Linux, NVIDIA]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      # - name: Install dependencies
-      #   run: |
-      #     sudo apt update
-      #     sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
-
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.24
-      #   with:
-      #     restore: false
-      #     save: false
-
-      # - name: ccache-buckets-restore
-      #   uses: ./.github/actions/ccache-buckets
-      #   with:
-      #     key: self-hosted-vulkan-nvidia-cm
-      #     folder: llama.cpp
-      #     hf_bucket: ggml-org/cache
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-      # - name: ccache-buckets-save
-      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-      #   uses: ./.github/actions/ccache-buckets
-      #   env:
-      #     HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-      #   with:
-      #     key: self-hosted-vulkan-nvidia-cm
-      #     folder: llama.cpp
-      #     evict-old-files: 1d
-      #     hf_bucket: ggml-org/cache
-      #     save: true
-
-  gpu-vulkan-nvidia-cm2:
-    # runs-on: "hf-jobs-t4-small:ubuntu26_04"
-    runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      # - name: Install dependencies
-      #   run: |
-      #     sudo apt update
-      #     sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
-
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.24
-      #   with:
-      #     restore: false
-      #     save: false
-
-      # - name: ccache-buckets-restore
-      #   uses: ./.github/actions/ccache-buckets
-      #   with:
-      #     key: self-hosted-vulkan-nvidia-cm2
-      #     folder: llama.cpp
-      #     hf_bucket: ggml-org/cache
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-      # - name: ccache-buckets-save
-      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-      #   uses: ./.github/actions/ccache-buckets
-      #   env:
-      #     HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-      #   with:
-      #     key: self-hosted-vulkan-nvidia-cm2
-      #     folder: llama.cpp
-      #     evict-old-files: 1d
-      #     hf_bucket: ggml-org/cache
-      #     save: true
-
-  gpu-vulkan-apple:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  gpu-vulkan-intel-linux:
-    runs-on: [self-hosted, Linux, Intel]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          persist-credentials: false
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  gpu-vulkan-intel-windows:
-    runs-on: [self-hosted, Windows, X64, Intel]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        shell: C:\msys64\usr\bin\bash.exe --noprofile --norc -eo pipefail "{0}"
-        env:
-          MSYSTEM: UCRT64
-          CHERE_INVOKING: 1
-          PATH: C:\msys64\ucrt64\bin;C:\msys64\usr\bin;C:\Windows\System32;${{ env.PATH }}
-        run: |
-          vulkaninfo --summary
-          # Skip python related tests with GG_BUILD_LOW_PERF=1 since Windows MSYS2 UCRT64 currently fails to create
-          # a valid python environment for testing
-          LLAMA_FATAL_WARNINGS=OFF GG_BUILD_NINJA=1 GG_BUILD_VULKAN=1 GG_BUILD_LOW_PERF=1 ./ci/run.sh ./results/llama.cpp ./mnt/llama.cpp
-
-  # TODO: provision AMD GPU machine
-  # amd-vulkan:
-  #   runs-on: [self-hosted, Linux, AMD]
-
-  #   steps:
-  #     - name: Clone
-  #       id: checkout
-  #       uses: actions/checkout@v6
-
-  #     - name: Test
-  #       id: ggml-ci
-  #       run: |
-  #         vulkaninfo --summary
-  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -1,130 +0,0 @@
-name: CI (self-hosted WebGPU backend)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ci-self-hosted-webgpu.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.wgsl'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ci-self-hosted-webgpu.yml',
-      'ci/run.sh',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      'ggml/src/*',
-      'ggml/src/ggml-cpu/**',
-      'ggml/src/ggml-webgpu/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  gpu-webgpu-nvidia:
-    runs-on: "hf-jobs-t4-small:ubuntu26_04"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: self-hosted-webgpu-nvidia
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v20260908.214631"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_WEBGPU=1 \
-          GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
-          GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
-            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: self-hosted-webgpu-nvidia
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-  gpu-webgpu-apple:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v20260908.214631"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-macos-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
-            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -11,12 +11,10 @@ on:
    paths:
      - .github/workflows/copilot-setup-steps.yml

-cache-mode: none
-
 jobs:
  # The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
  copilot-setup-steps:
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest

    # Set the permissions to the lowest permissions possible needed for your steps.
    # Copilot will be given its own token for its operations.
@@ -33,8 +31,8 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
        with:
-          restore: false
-          save: false
+          key: copilot-setup-steps
+          evict-old-files: 1d

      - name: Dependencies
        id: depends
@@ -90,8 +90,8 @@ jobs:
            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x", "prebuilt_ui": true },
            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.4.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.4.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.3.0", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.3.0", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
            { "tag": "musa", "dockerfile": ".devops/musa.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "intel", "dockerfile": ".devops/intel.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
@@ -0,0 +1,71 @@
+name: Fusion
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/fusion.yml',
+      'ggml/**',
+      'tests/fusion/**',
+      'tests/test-fusion.cpp',
+      'tests/test-llama-archs.cpp',
+      'src/models/**'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/fusion.yml',
+      'ggml/**',
+      'tests/fusion/**',
+      'tests/test-fusion.cpp',
+      'tests/test-llama-archs.cpp',
+      'src/models/**'
+    ]
+
+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:
+  # TODO: add jobs for other backends as they adopt the fusion debug API
+  metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_BLAS=OFF \
+            -DGGML_METAL=ON
+          time cmake --build build --config Release --target test-llama-archs -j $(sysctl -n hw.logicalcpu)
+          time cmake --build build --config Release --target test-fusion -j $(sysctl -n hw.logicalcpu)
+
+      - name: Generate models
+        id: generate_models
+        run: |
+          rm -rf build-ci-models && mkdir -p build-ci-models
+          ./build/bin/test-llama-archs -o build-ci-models --errors-only
+
+      - name: Test fusion
+        id: test_fusion
+        run: |
+          ./build/bin/test-fusion --models build-ci-models --device MTL0 --check tests/fusion/MTL.csv --errors-only
@@ -21,7 +21,7 @@ on:
 jobs:
  deploy:

-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest

    steps:
    - uses: actions/checkout@v6
@@ -13,16 +13,6 @@ on:
        required: true
        type: boolean
        default: true
-      skip_apiabi_check:
-        description: 'Skip API/ABI compatibility check'
-        required: false
-        type: boolean
-        default: false
-      apiabi_compare_tag:
-        description: 'Tag to compare against for API/ABI check (default: latest release)'
-        required: false
-        type: string
-        default: ''

 env:
  GH_TOKEN: ${{ github.token }}
@@ -33,7 +23,7 @@ permissions:

 jobs:
  make-release:
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest

    steps:
      - name: Checkout
@@ -43,18 +33,12 @@ jobs:
          ref: ${{ inputs.commit != '' && inputs.commit || github.ref_name }}
          fetch-depth: 0

-      - name: Install API/ABI check tools
-        if: ${{ github.event.inputs.skip_apiabi_check != 'true' }}
-        run: sudo apt-get install -y abi-compliance-checker abigail-tools
-
      - 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 }}
-          SKIP_APIABI_CHECK: ${{ github.event.inputs.skip_apiabi_check }}
-          APIABI_COMPARE_TAG: ${{ github.event.inputs.apiabi_compare_tag }}

      - name: Create release tag
        if: ${{ github.event.inputs.dry_run == 'false' }}
@@ -1,445 +0,0 @@
-name: Models Backend Check
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/models-check.yml',
-      'ggml/**',
-      'tests/fusion/**',
-      'tests/test-fusion.cpp',
-      'tests/test-llama-archs.cpp',
-      'src/llama-graph.cpp',
-      'src/llama-model*',
-      'src/models/**'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/models-check.yml',
-      'ggml/**',
-      'tests/fusion/**',
-      'tests/test-fusion.cpp',
-      'tests/test-llama-archs.cpp',
-      'src/llama-graph.cpp',
-      'src/llama-model*',
-      'src/models/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  cuda:
-    runs-on: "hf-jobs-t4-medium:cuda13"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y cmake time python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: models-check-cuda
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
-            -DGGML_CUDA=ON
-          time cmake --build build --config Release --target test-llama-archs -j$(nproc)
-          time cmake --build build --config Release --target test-fusion -j$(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: models-check-cuda
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      # - name: Generate models
-      #   id: generate_models
-      #   run: |
-      #     rm -rf build-ci-models && mkdir -p build-ci-models
-      #     ./build/bin/test-llama-archs -o build-ci-models
-
-      # TODO: add for backends as they adopt the fusion debug API
-      # - name: Test fusion
-      #   id: test_fusion
-      #   run: |
-      #     ./build/bin/test-fusion --models build-ci-models --device CUDA0 --check tests/fusion/CUDA.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          GGML_CUDA_DEVICES=1 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=2 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=3 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=4 ./build/bin/test-llama-archs -s 1
-
-  metal:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DGGML_BLAS=OFF \
-            -DGGML_METAL=ON
-          time cmake --build build --config Release --target test-llama-archs -j $(sysctl -n hw.logicalcpu)
-          time cmake --build build --config Release --target test-fusion -j $(sysctl -n hw.logicalcpu)
-
-      - name: Generate models
-        id: generate_models
-        run: |
-          rm -rf build-ci-models && mkdir -p build-ci-models
-          ./build/bin/test-llama-archs -o build-ci-models
-
-      - name: Test fusion
-        id: test_fusion
-        run: |
-          ./build/bin/test-fusion --models build-ci-models --device MTL0 --check tests/fusion/MTL.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          GGML_METAL_DEVICES=1 ./build/bin/test-llama-archs -s 1
-          GGML_METAL_DEVICES=2 ./build/bin/test-llama-archs -s 1
-          GGML_METAL_DEVICES=3 ./build/bin/test-llama-archs -s 1
-          GGML_METAL_DEVICES=4 ./build/bin/test-llama-archs -s 1
-
-  rocm:
-    runs-on: [self-hosted, Linux, gfx1201]
-    container: "rocm/dev-ubuntu-24.04:7.2.4-complete"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          apt update
-          apt install -y build-essential jq cmake time python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: models-check-rocm
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang \
-            -DGPU_TARGETS=gfx1201 \
-            -DGGML_HIP=ON
-          time cmake --build build --config Release --target test-llama-archs -j$(nproc)
-          time cmake --build build --config Release --target test-fusion -j$(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: models-check-rocm
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      # - name: Generate models
-      #   id: generate_models
-      #   run: |
-      #     rm -rf build-ci-models && mkdir -p build-ci-models
-      #     ./build/bin/test-llama-archs -o build-ci-models
-
-      # TODO: add for backends as they adopt the fusion debug API
-      # - name: Test fusion
-      #   id: test_fusion
-      #   run: |
-      #     ./build/bin/test-fusion --models build-ci-models --device CUDA0 --check tests/fusion/CUDA.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          GGML_CUDA_DEVICES=1 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=2 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=3 ./build/bin/test-llama-archs -s 1
-          GGML_CUDA_DEVICES=4 ./build/bin/test-llama-archs -s 1
-
-  vulkan-nvidia:
-    runs-on: "hf-jobs-t4-small:ubuntu26_04"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 time python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: models-check-vulkan-nvidia
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DGGML_VULKAN=ON
-          time cmake --build build --config Release --target test-llama-archs -j$(nproc)
-          time cmake --build build --config Release --target test-fusion -j$(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: models-check-vulkan-nvidia
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      # - name: Generate models
-      #   id: generate_models
-      #   run: |
-      #     rm -rf build-ci-models && mkdir -p build-ci-models
-      #     ./build/bin/test-llama-archs -o build-ci-models
-
-      # TODO: add for backends as they adopt the fusion debug API
-      # - name: Test fusion
-      #   id: test_fusion
-      #   run: |
-      #     ./build/bin/test-fusion --models build-ci-models --device Vulkan0 --check tests/fusion/Vulkan.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          ./build/bin/test-llama-archs -s 1
-
-  vulkan-amd:
-    runs-on: [self-hosted, Linux, gfx1201]
-    container: "ubuntu:26.04"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          apt update
-          apt install -y build-essential jq cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc spirv-headers vulkan-tools mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 time python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: models-check-vulkan-amd
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DGGML_VULKAN=ON
-          time cmake --build build --config Release --target test-llama-archs -j$(nproc)
-          time cmake --build build --config Release --target test-fusion -j$(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: models-check-vulkan-amd
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      # - name: Generate models
-      #   id: generate_models
-      #   run: |
-      #     rm -rf build-ci-models && mkdir -p build-ci-models
-      #     ./build/bin/test-llama-archs -o build-ci-models
-
-      # TODO: add for backends as they adopt the fusion debug API
-      # - name: Test fusion
-      #   id: test_fusion
-      #   run: |
-      #     ./build/bin/test-fusion --models build-ci-models --device Vulkan0 --check tests/fusion/Vulkan.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          ./build/bin/test-llama-archs -s 1
-
-  webgpu-nvidia:
-    runs-on: "hf-jobs-t4-small:ubuntu26_04"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 time python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: models-check-webgpu-nvidia
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v20260908.214631"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DCMAKE_PREFIX_PATH="$GITHUB_WORKSPACE/dawn" \
-            -DDawn_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
-            -DGGML_WEBGPU=ON
-          time cmake --build build --config Release --target test-llama-archs -j$(nproc)
-          time cmake --build build --config Release --target test-fusion -j$(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: models-check-webgpu-nvidia
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      # - name: Generate models
-      #   id: generate_models
-      #   run: |
-      #     rm -rf build-ci-models && mkdir -p build-ci-models
-      #     ./build/bin/test-llama-archs -o build-ci-models
-
-      # TODO: add for backends as they adopt the fusion debug API
-      # - name: Test fusion
-      #   id: test_fusion
-      #   run: |
-      #     ./build/bin/test-fusion --models build-ci-models --device WebGPU --check tests/fusion/WebGPU.csv
-
-      - name: Test archs
-        id: test_archs
-        run: |
-          ./build/bin/test-llama-archs -s 1
@@ -12,7 +12,7 @@ on:

 jobs:
    pre-tokenizer-hashes:
-        runs-on: ubuntu-slim
+        runs-on: [self-hosted, fast]

        steps:
        - name: Checkout repository
@@ -20,7 +20,7 @@ concurrency:

 jobs:
  python-check-requirements:
-    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    runs-on: [self-hosted, CPU, fast]
    name: check-requirements
    steps:
      - name: Check out source repository
@@ -21,7 +21,7 @@ concurrency:

 jobs:
  flake8-lint:
-    runs-on: ubuntu-slim
+    runs-on: [self-hosted, fast]
    name: Lint
    steps:
      - name: Check out source repository
@@ -22,7 +22,7 @@ concurrency:

 jobs:
  python-type-check:
-    runs-on: ubuntu-slim
+    runs-on: [self-hosted, fast]
    name: python type-check
    steps:
      - name: Check out source repository
@@ -31,7 +31,7 @@ jobs:
        uses: actions/setup-python@v6
        with:
          python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.84
+          pip-install: -r requirements/requirements-all.txt ty==0.0.78
      # - name: Type-check with Pyright
      #   uses: jakebailey/pyright-action@v2
      #   with:
@@ -106,7 +106,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-${{ matrix.arch }}
-          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -191,7 +190,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-cpu
-          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -277,7 +275,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-${{ matrix.os }}-vulkan
-          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -323,22 +320,21 @@ jobs:
        include:
          # label = short version used in artifact names / release body
          # cuda  = full container image tag
-          # CTK >= 13.5 bundles CCCL >= 3.5; omit GGML_CUDA_CCCL_VERSION for those versions.
          - build: 'x64'
            os: ubuntu-24.04
            cuda: '12.8.2'
            label: '12.8'
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
          - build: 'x64'
            os: ubuntu-24.04
-            cuda: '13.4.1'
-            label: '13.4'
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            cuda: '13.3.1'
+            label: '13.3'
+            defines: ''
          - build: 'arm64'
            os: ubuntu-24.04-arm
-            cuda: '13.4.1'
-            label: '13.4'
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            cuda: '13.3.1'
+            label: '13.3'
+            defines: ''

    runs-on: ${{ matrix.os }}
    container: nvidia/cuda:${{ matrix.cuda }}-devel-ubuntu24.04
@@ -457,7 +453,7 @@ jobs:
    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest

    #permissions:
    #  actions: write
@@ -485,9 +481,10 @@ jobs:
          distribution: temurin

      - name: Setup Android SDK
-        uses: android-actions/setup-android@be39fa834029ff78f1a44aa3bb0819b8fc2bd8fd # v4.0.4
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
        with:
          log-accepted-android-sdk-licenses: false
+          packages: 'platform-tools'

      - name: Install NDK
        run: |
@@ -504,7 +501,6 @@ jobs:
      #  uses: ggml-org/ccache-action@v1.2.24
      #  with:
      #    key: release-android-arm64
-      #    evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -545,120 +541,6 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz
          name: llama-bin-android-arm64.tar.gz

-  android-arm64-snapdragon:
-    needs: [check-release, ui-build]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: ubuntu-latest
-    container: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.7'
-
-    defaults:
-      run:
-        shell: bash
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      # checkout runs as the host user; in-container steps run as root, so git
-      # refuses to touch a repo it does not own. Mark the workspace as safe.
-      - name: Git safe directory
-        run: git config --global --add safe.directory "$GITHUB_WORKSPACE"
-
-      - name: Download UI build
-        uses: actions/download-artifact@v7
-        with:
-          name: llama-ui.zip
-          path: tools/ui/dist
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cp docs/backend/snapdragon/CMakeUserPresets.json .
-          cmake --preset arm64-android-snapdragon-release -B build \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN;$ORIGIN/../lib' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j $(nproc)
-          cmake --install build --prefix pkg-snapdragon/llama.cpp
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE pkg-snapdragon/llama.cpp/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-android-arm64-snapdragon.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C pkg-snapdragon/llama.cpp .
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-android-arm64-snapdragon.tar.gz
-          name: llama-bin-android-arm64-snapdragon.tar.gz
-
-  linux-arm64-snapdragon:
-    needs: [check-release, ui-build]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: ubuntu-latest
-    container: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.7'
-
-    defaults:
-      run:
-        shell: bash
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      # checkout runs as the host user; in-container steps run as root, so git
-      # refuses to touch a repo it does not own. Mark the workspace as safe.
-      - name: Git safe directory
-        run: git config --global --add safe.directory "$GITHUB_WORKSPACE"
-
-      - name: Download UI build
-        uses: actions/download-artifact@v7
-        with:
-          name: llama-ui.zip
-          path: tools/ui/dist
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cp docs/backend/snapdragon/CMakeUserPresets.json .
-          cmake --preset arm64-linux-snapdragon-release -B build -DGGML_OPENCL=ON \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN;$ORIGIN/../lib' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j $(nproc)
-          cmake --install build --prefix pkg-snapdragon/llama.cpp
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE pkg-snapdragon/llama.cpp/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-linux-arm64-snapdragon.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C pkg-snapdragon/llama.cpp .
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-linux-arm64-snapdragon.tar.gz
-          name: llama-bin-linux-arm64-snapdragon.tar.gz
-
  ubuntu-24-openvino:
    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -673,8 +555,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

    steps:
      - name: Set OpenVINO version output
@@ -697,7 +579,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-          evict-old-files: 1d

      - name: Dependencies
        run: |
@@ -788,8 +669,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.4"
-      OPENVINO_VERSION_FULL: "2026.4.0.22959.99c81491cc3"
+      OPENVINO_VERSION_MAJOR: "2026.3.1"
+      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"

    steps:
      - name: Set OpenVINO version output
@@ -941,7 +822,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-          evict-old-files: 1d

      - name: Build
        shell: cmd
@@ -1186,7 +1066,6 @@ jobs:
      #  uses: ggml-org/ccache-action@v1.2.24
      #  with:
      #    key: release-windows-2025-${{ matrix.arch }}-${{ matrix.backend }}
-      #    evict-old-files: 1d

      - name: Install OpenCL Headers and Libs
        id: install_opencl
@@ -1245,16 +1124,15 @@ jobs:
    strategy:
      matrix:
        include:
-          # CTK >= 13.5 bundles CCCL >= 3.5; omit GGML_CUDA_CCCL_VERSION for those versions.
          - cuda: '12.4'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
          - cuda: '13.4'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            defines: ''
          - cuda: '13.4'
            arch: arm64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3 -DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -1276,11 +1154,11 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
-          evict-old-files: 1d

      - 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' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
@@ -1345,11 +1223,11 @@ jobs:
        shell: bash

    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/0cb67a0d-67f6-410b-868b-f4a0a17ff0cf/intel-oneapi-toolkit-2026.1.1.32_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2026.1"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"

    steps:
      - name: Clone
@@ -1372,7 +1250,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-windows-2022-x64-sycl
-          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -1392,11 +1269,9 @@ jobs:
        run: |
          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"

-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.6.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.3.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_def.3.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_avx2.3.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin

          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
@@ -1411,11 +1286,13 @@ jobs:
            echo "Level Zero loader DLL not found in oneAPI or SDK; relying on system driver/runtime"
          fi

-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl9.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin

          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
@@ -1455,8 +1332,8 @@ jobs:

    env:
      ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2026.1"
-      LEVEL_ZERO_VERSION: "1.33.1"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      LEVEL_ZERO_VERSION: "1.28.2"
      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"

    steps:
@@ -1470,16 +1347,16 @@ jobs:
        shell: bash
        run: |
          cd /tmp
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/5996e26b-f48a-42b1-8db0-b002ad0bd8d7/intel-oneapi-toolkit-2026.1.1.33_offline.sh -O intel-oneapi-toolkit_offline.sh
-          sudo bash intel-oneapi-toolkit_offline.sh -s -a --silent --eula accept
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept

      - name: Install Level Zero SDK
        shell: bash
        run: |
          cd /tmp
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze1_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze1.deb
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze-dev_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze-dev.deb
-          sudo apt-get install -y ./libze1.deb ./libze-dev.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb

      - name: Download UI build
        uses: actions/download-artifact@v7
@@ -1491,7 +1368,6 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-          evict-old-files: 1d

      - name: Build
        id: cmake_build
@@ -1840,8 +1716,6 @@ jobs:
      - ubuntu-24-openvino
      - ubuntu-24-sycl
      - android-arm64
-      - android-arm64-snapdragon
-      - linux-arm64-snapdragon
      - macos-cpu
      - ios-xcode
      #- openEuler-cann
@@ -1971,17 +1845,15 @@ jobs:
            - [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 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-12.8-x64.tar.gz) - [CUDA 12.8 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-12.8-x64.tar.gz)
-            - [Ubuntu x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-x64.tar.gz) - [CUDA 13.4 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-x64.tar.gz)
-            - [Ubuntu arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-arm64.tar.gz) - [CUDA 13.4 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-arm64.tar.gz)
+            - [Ubuntu x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.3-x64.tar.gz) - [CUDA 13.3 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.3-x64.tar.gz)
+            - [Ubuntu arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.3-arm64.tar.gz) - [CUDA 13.3 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.3-arm64.tar.gz)
            - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-10.0-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-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
-            - [Linux arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-linux-arm64-snapdragon.tar.gz) - [setup guide](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/linux.md)

            **Android:**
            - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)
-            - [Android arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64-snapdragon.tar.gz) - [setup guide](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/README.md)

            **Windows:**
            - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
@@ -45,7 +45,7 @@ concurrency:

 jobs:
  server:
-    runs-on: hf-jobs-cpu-performance
+    runs-on: hf-jobs-cpu-upgrade

    strategy:
      matrix:
@@ -116,7 +116,7 @@ jobs:
        run: |
          source .venv/bin/activate
          cd tools/server/tests
-          PYTEST_WORKERS=4 ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
@@ -124,4 +124,4 @@ jobs:
        run: |
          source .venv/bin/activate
          cd tools/server/tests
-          PYTEST_WORKERS=4 SLOW_TESTS=1 ./tests.sh
+          PYTEST_WORKERS=1 SLOW_TESTS=1 ./tests.sh
@@ -16,7 +16,7 @@ on:

 jobs:
    update-ops-docs:
-        runs-on: ubuntu-slim
+        runs-on: [self-hosted, fast, ARM64]

        steps:
        - name: Checkout repository
@@ -8,7 +8,7 @@ on:
 jobs:
  update:
    name: Update Winget Package
-    runs-on: ubuntu-24.04  # previously ubuntu-latest
+    runs-on: ubuntu-latest
    if: github.repository_owner == 'ggml-org'

    steps:
@@ -7,7 +7,6 @@ General:
 - 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
 - When [MODEL] is needed, first try to get it from the `PI_MODEL_NAME` env var before asking the user
- Never read the `AGENTS.md` file

 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
@@ -4,8 +4,8 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 5)
-set(LLAMA_VERSION_PATCH 0)
+set(LLAMA_VERSION_MINOR 4)
+set(LLAMA_VERSION_PATCH 1)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -57,7 +57,7 @@
 /ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
-/ggml/src/ggml-cpu/tiled/               @jbooth @bartowski1182
+/ggml/src/ggml-cpu/iqp.*                @bartowski1182
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
@@ -77,7 +77,7 @@
 /ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-webgpu/                  @ggml-org/ggml-webgpu
 /ggml/src/ggml-zdnn/                    @ggml-org/ggml-zdnn @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml-zendnn/                  @avinashcpandey @Jiten1parmar
+/ggml/src/ggml-zendnn/                  @avinashcpandey @Jiten1parmar @z-vishal
 /ggml/src/ggml.c                        @ggerganov
 /ggml/src/ggml.cpp                      @ggerganov
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
@@ -97,6 +97,7 @@
 /src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-chat.*                      @pwilkin
+/tests/test-llama-archs.cpp             @JohannesGaessler
 /tools/batched-bench/                   @ggerganov
 /tools/cli/                             @ngxson
 /tools/completion/                      @ggerganov
@@ -21,13 +21,13 @@ docker run --privileged -it \
    -v $HOME/llama.cpp/ci-cache:/ci-cache \
    -v $HOME/llama.cpp/ci-results:/ci-results \
    -v $PWD:/ws -w /ws \
-    registry.mthreads.com/mcconline/musa_sdk:5.2.0-devel-ubuntu22.04-s5000
+    mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
 ```

 Inside the container, execute the following commands:

 ```bash
-apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget musa-mualg-5-2 musa-muthrust-5-2 libmthreads-compute
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
 git config --global --add safe.directory /ws
 GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
 ```
@@ -49,6 +49,14 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")

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

 sd=`dirname $0`
@@ -72,8 +80,8 @@ else
 fi

 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    # TODO: Drop GGML_CUDA_CCCL_VERSION when CUDA CI uses CTK >= 13.5, which bundles CCCL >= 3.5.
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DGGML_CUDA_CCCL_VERSION=v3.4.3"
+    # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DGGML_CUDA_CUB_3DOT2=ON"

    if command -v nvidia-smi >/dev/null 2>&1; then
        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
@@ -150,8 +158,8 @@ if [ ! -z ${GG_BUILD_WEBGPU} ]; then
 fi

 if [ ! -z ${GG_BUILD_MUSA} ]; then
-    # Use ph1 by default (MTT S5000)
-    MUSA_ARCH=${MUSA_ARCH:-31}
+    # Use qy1 by default (MTT S80)
+    MUSA_ARCH=${MUSA_ARCH:-21}
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
 fi

@@ -270,17 +278,17 @@ function gg_run_test_llama_archs_tensor_split {
    set -e

    if [ ! -z ${GG_BUILD_CUDA} ]; then
-        GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
    fi

    if [ ! -z ${GG_BUILD_METAL} ]; then
-        GGML_METAL_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_METAL_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_METAL_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_METAL_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
+        GGML_METAL_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 --errors-only 2>&1
    fi

    set +e
@@ -299,7 +307,7 @@ function gg_run_test_llama_archs_models {
        rm -rf build-ci-models && mkdir -p build-ci-models

        # generate the dummy models used by the model-dependent tests
-        ./build-ci-release/bin/test-llama-archs -o build-ci-models 2>&1
+        ./build-ci-release/bin/test-llama-archs -o build-ci-models --errors-only 2>&1
    fi

    set +e
@@ -447,10 +455,10 @@ function gg_run_qwen3_0_6b {

    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload --errors-only) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload --errors-only) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off --errors-only) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on  --errors-only) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

    function check_ppl {
        qnt="$1"
@@ -668,9 +676,9 @@ function gg_run_test_backend_ops {

    # TODO: reduce the test-backend-ops timeout to 1800s
    if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
-        (time timeout 3600 ./bin/test-backend-ops ${args_extra} -b CPU) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra} -b CPU --errors-only) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
    else
-        (time timeout 3600 ./bin/test-backend-ops ${args_extra}       ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra} --errors-only) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
    fi

    set +e
@@ -17,16 +17,14 @@ find_library(llama_LIBRARY llama
    NO_CMAKE_FIND_ROOT_PATH
 )

-if(NOT TARGET llama)
-    add_library(llama UNKNOWN IMPORTED)
-    set_target_properties(llama
-        PROPERTIES
-            INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
-            INTERFACE_LINK_LIBRARIES "ggml::ggml;ggml::ggml-base;"
-            IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
-            IMPORTED_LOCATION "${llama_LIBRARY}"
-            INTERFACE_COMPILE_FEATURES c_std_90
-            POSITION_INDEPENDENT_CODE ON)
-endif()
+add_library(llama UNKNOWN IMPORTED)
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "ggml::ggml;ggml::ggml-base;"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES c_std_90
+        POSITION_INDEPENDENT_CODE ON)

 check_required_components(Llama)
@@ -351,7 +351,7 @@ static bool parse_bool_value(const std::string & value) {
 static std::string get_default_local_path(const std::string & url) {
    auto f = string_split<std::string>(url, '#').front();
    f = string_split<std::string>(f, '?').front();
-    return fs_path_to_utf8(fs_get_cache_file(string_split<std::string>(f, '/').back()));
+    return fs_get_cache_file(string_split<std::string>(f, '/').back());
 }

 static bool spec_types_is_default(const common_params & params) {
@@ -387,9 +387,6 @@ common_models_handler common_models_handler_init(const common_params & params, l
            break;
        }
    }
-    if (curr_ex == LLAMA_EXAMPLE_DOWNLOAD) {
-        use_mmproj = true;
-    }

    opts.bearer_token    = params.hf_token;
    opts.offline         = params.offline;
@@ -720,24 +717,24 @@ void common_models_handler_apply(common_models_handler & handler, common_params
 // 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::filesystem::path> paths;
+    std::vector<std::string> paths;

 #if defined(_WIN32)
-    const std::filesystem::path program_data = common_get_path_from_env("PROGRAMDATA");
+    const std::string program_data = common_get_env("PROGRAMDATA");
    if (!program_data.empty()) {
-        paths.push_back(program_data / "llama.cpp" / "config.ini");
+        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");
+        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::filesystem::path> found;
+    std::vector<std::string> found;
    for (const auto & path : paths) {
        std::error_code ec;
        if (std::filesystem::exists(path, ec)) {
@@ -751,7 +748,7 @@ static void common_params_apply_system_config(common_params & params, llama_exam
    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", fs_path_to_utf8(path).c_str());
+        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);
@@ -2019,7 +2016,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.sampling.temp = std::max(params.sampling.temp, 0.0f);
            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP;
        }
-    ).set_sampling().set_env("LLAMA_ARG_TEMPERATURE"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--top-k"}, "N",
        string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
@@ -2035,7 +2032,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.sampling.top_p = std::stof(value);
            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
        }
-    ).set_sampling().set_env("LLAMA_ARG_TOP_P"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--min-p"}, "N",
        string_format("min-p sampling (default: %.2f, 0.0 = disabled)", (double)params.sampling.min_p),
@@ -2043,7 +2040,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.sampling.min_p = std::stof(value);
            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
        }
-    ).set_sampling().set_env("LLAMA_ARG_MIN_P"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--top-nsigma", "--top-n-sigma"}, "N",
        string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
@@ -2099,7 +2096,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.sampling.penalty_repeat = penalty_repeat;
            params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
        }
-    ).set_sampling().set_env("LLAMA_ARG_REPEAT_PENALTY"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--presence-penalty"}, "N",
        string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
@@ -2110,7 +2107,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            }
            params.sampling.penalty_present = penalty_present;
        }
-    ).set_sampling().set_env("LLAMA_ARG_PRESENCE_PENALTY"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--frequency-penalty"}, "N",
        string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
@@ -2121,7 +2118,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            }
            params.sampling.penalty_freq = penalty_freq;
        }
-    ).set_sampling().set_env("LLAMA_ARG_FREQUENCY_PENALTY"));
+    ).set_sampling());
    add_opt(common_arg(
        {"--dry-multiplier"}, "N",
        string_format("set DRY sampling multiplier (default: %.2f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
@@ -2676,17 +2673,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.video_ffmpeg_bin_dir = value;
        }
    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FFMPEG_DIR"));
-    add_opt(common_arg(
-        {"--rpc"}, "SERVERS",
-        "comma-separated list of RPC servers (host:port)",
-        [](common_params & params, const std::string & value) {
-            if (!llama_supports_rpc()) {
-                throw std::invalid_argument("RPC not supported in this build");
+    if (params.is_gen_docs || llama_supports_rpc()) {
+        add_opt(common_arg(
+            {"--rpc"}, "SERVERS",
+            "comma-separated list of RPC servers (host:port)",
+            [](common_params & params, const std::string & value) {
+                add_rpc_devices(value);
+                GGML_UNUSED(params);
            }
-            add_rpc_devices(value);
-            GGML_UNUSED(params);
-        }
-    ).set_env("LLAMA_ARG_RPC"));
+        ).set_env("LLAMA_ARG_RPC"));
+    }
    add_opt(common_arg(
        {"-lm", "--load-mode"}, "MODE",
        "model loading mode (default: auto)\n"
@@ -3312,18 +3308,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
    add_opt(common_arg(
        {"--host"}, "HOST",
-        string_format("IP addresses to listen on, comma-separated, or UNIX socket paths ending in .sock; with multiple TCP addresses, :: binds IPv6 only; overlapping addresses result in undefined behavior (default: %s)", params.hostnames[0].c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
        [](common_params & params, const std::string & value) {
-            params.hostnames.clear();
-            for (auto & host : parse_csv_row(value)) {
-                host = string_strip(host);
-                if (!host.empty()) {
-                    params.hostnames.push_back(host);
-                }
-            }
-            if (params.hostnames.empty()) {
-                throw std::invalid_argument("--host requires at least one address");
-            }
+            params.hostname = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_HOST"));
    add_opt(common_arg(
@@ -3950,6 +3937,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            common_log_set_verbosity_thold(value);
        }
    ).set_env("LLAMA_ARG_LOG_VERBOSITY"));
+    add_opt(common_arg(
+        {"--errors-only"},
+        "Set verbosity threshold to WARN (only print warnings and errors)",
+        [](common_params & params) {
+            params.verbosity = LOG_LEVEL_WARN;
+            common_log_set_verbosity_thold(LOG_LEVEL_WARN);
+        }
+    ));
    add_opt(common_arg(
        {"--log-prefix"},
        {"--no-log-prefix"},
@@ -122,8 +122,6 @@ struct common_params_context {

 // parse input arguments from CLI
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
-// TODO: this function can load ggml backend (by calling llama_support_rpc)
-//       this is a side-effect that should be avoided
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

 // load all backends and print the list of available (non-CPU) devices to stdout
@@ -318,13 +318,13 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
    bool is_content   = node.tag == common_chat_peg_builder::CONTENT;

    if (is_reasoning) { // GPT OSS can have more than 1 reasoning block, so concatenate here
-        result.reasoning_content += node.sanitized_text();
+        result.reasoning_content += std::string(node.text);
    }

    if (is_content) {
        // Concatenate content from multiple content nodes (e.g., when reasoning markers
        // are preserved before content markers in reasoning_format=NONE mode)
-        result.content += node.sanitized_text();
+        result.content += std::string(node.text);
    }

    // Handle tool-related tags (supporting both JSON and tagged formats)
@@ -1058,12 +1058,12 @@ void common_chat_peg_gemma4_mapper::visit(const common_peg_ast_arena & arena, co
    const auto & node = arena.get(id);

    if (node.tag == "reasoning") {
-        result.reasoning_content += node.sanitized_text();
+        result.reasoning_content += std::string(node.text);
        return;
    }

    if (node.tag == "content") {
-        result.content += node.sanitized_text();
+        result.content += std::string(node.text);
        return;
    }

@@ -1206,12 +1206,12 @@ void common_chat_peg_minimax_m3_mapper::visit(const common_peg_ast_arena & arena
    const auto & node = arena.get(id);

    if (node.tag == common_chat_peg_builder::REASONING) {
-        result.reasoning_content += node.sanitized_text();
+        result.reasoning_content += std::string(node.text);
        return;
    }

    if (node.tag == common_chat_peg_builder::CONTENT) {
-        result.content += node.sanitized_text();
+        result.content += std::string(node.text);
        return;
    }

@@ -1099,12 +1099,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
        return common_chat_params_init_ministral_3(tmpl, params);
    }

-    // LLM-jp-4.1 - GPT-OSS dialect (spaces after special tokens, <|end|>-separated parallel calls)
-    if (src.find("chat_format=llm-jp-harmony-v1") != std::string::npos) {
-        LOG_DBG("Using specialized template: LLM-jp Harmony v1\n");
-        return common_chat_params_init_llm_jp_harmony(tmpl, params);
-    }
-
    // GPT-OSS - has unique channel-based structure that needs dedicated handler
    if (src.find("<|channel|>") != std::string::npos) {
        LOG_DBG("Using specialized template: GPT-OSS\n");
@@ -1139,14 +1133,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
        return common_chat_params_init_kimi_k3(tmpl, params);
    }

-    // Ling 3.0 / Bailing V3 - <role>X</role> sections with <arg_key>/<arg_value> tagged
-    // tool calls. <role> sections are unique to this family among the tagged-arg templates.
-    if (src.find("<role>ASSISTANT</role>") != std::string::npos &&
-        src.find("<arg_key>") != std::string::npos) {
-        LOG_DBG("Using specialized template: Ling 3.0 (Bailing V3)\n");
-        return common_chat_params_init_ling3(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|>).
@@ -1218,9 +1204,7 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
    // Qwen3-Coder XML tool calls, also used by Nemotron Nano 3, Qwen3.5 and StepFun-3.5-Flash
    if (src.find("<tool_call>") != std::string::npos &&
        src.find("<function=") != std::string::npos &&
-        src.find("<parameter=") != std::string::npos &&
-        // Exclude models that don't use \n between tags
-        src.find("'<tool_call><function=' ~ tool_call.name ~ '>'") == std::string::npos) {
+        src.find("<parameter=") != std::string::npos) {
        LOG_DBG("Using specialized template: Qwen3-Coder\n");
        return common_chat_params_init_qwen3_coder(tmpl, params);
    }
@@ -46,10 +46,11 @@
 #include <io.h>
 #else
 #include <sys/ioctl.h>
+#include <sys/stat.h>
 #include <unistd.h>
 #endif

-#if !defined(_WIN32)
+#if defined(__linux__)
 #include <sys/types.h>
 #include <pwd.h>
 #endif
@@ -613,6 +614,34 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
    return buf.str();
 }

+std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch) {
+    std::stringstream buf;
+
+    buf << "[ ";
+
+    bool first = true;
+    for (int i = 0; i < batch.n_tokens; ++i) {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = common_token_to_piece(ctx, batch.token[i]);
+
+        buf << "\n"          << std::to_string(i)
+            << ", token '"   << detokenized << "'"
+            << ", pos "      << std::to_string(batch.pos[i])
+            << ", n_seq_id " << std::to_string(batch.n_seq_id[i])
+            << ", seq_id "   << std::to_string(batch.seq_id[i][0])
+            << ", logits "   << std::to_string(batch.logits[i]);
+    }
+
+    buf << " ]";
+
+    return buf.str();
+}
+
 void string_process_escapes(std::string & input) {
    std::size_t input_len = input.length();
    std::size_t output_idx = 0;
@@ -871,7 +900,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {


 #ifdef _WIN32
-std::wstring utf8_to_wstring(const std::string & str) {
+static std::wstring utf8_to_wstring(const std::string & str) {
    if (str.empty()) {
        return std::wstring();
    }
@@ -887,52 +916,82 @@ std::wstring utf8_to_wstring(const std::string & str) {

    return wstr;
 }
-
-std::string wstring_to_utf8(const std::wstring & str) {
-    if (str.empty()) {
-        return std::string();
-    }
-
-    int size = WideCharToMultiByte(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0, NULL, NULL);
-
-    if (size <= 0) {
-        return std::string();
-    }
-
-    std::string utf8(size, 0);
-    WideCharToMultiByte(CP_UTF8, 0, str.c_str(), (int)str.size(), &utf8[0], size, NULL, NULL);
-
-    return utf8;
-}
 #endif

-// returns the path as a UTF-8 string, preserving its separators
-std::string fs_path_to_utf8(const std::filesystem::path & path) {
-    const auto value = path.u8string();
-    return std::string(value.begin(), value.end());
-}
+// returns true if successful, false otherwise
+bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring wpath = utf8_to_wstring(path);

-void fs_write_atomic(const std::filesystem::path & path, const std::string & data) {
-    std::error_code ec;
-    std::filesystem::path path_tmp = path;
-    path_tmp += ".tmp";
-
-    if (path.has_parent_path()) {
-        std::filesystem::create_directories(path.parent_path(), ec);
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
    }

-    std::ofstream file(path_tmp, std::ios::binary);
-    file << data;
-    file.close();
+    size_t pos_slash = 0;

-    if (!file.fail()) {
-        std::filesystem::rename(path_tmp, path, ec);
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
    }

-    if (file.fail() || ec) {
-        std::filesystem::remove(path_tmp, ec);
-        throw std::runtime_error("failed to write file: " + fs_path_to_utf8(path));
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
 }

 bool fs_is_directory(const std::string & path) {
@@ -957,77 +1016,113 @@ void common_set_env(const std::string & name, const std::string & value) {
 #endif
 }

-std::filesystem::path common_get_path_from_env(const std::string & name) {
-#if defined(_WIN32)
-    const std::wstring wname = utf8_to_wstring(name);
-    const wchar_t * wvalue = _wgetenv(wname.c_str());
-    return wvalue ? std::filesystem::path(wvalue) : std::filesystem::path();
-#else
-    const char * value = std::getenv(name.c_str());
-    return value ? std::filesystem::path(value) : std::filesystem::path();
-#endif
-}
-
-#if !defined(_WIN32)
-static std::filesystem::path get_home_directory() {
-    std::filesystem::path home = common_get_path_from_env("HOME");
-    if (!home.empty()) {
-        return home;
-    }
-    const struct passwd * pw = getpwuid(getuid());
-    if (!pw || !pw->pw_dir || !*pw->pw_dir) {
-        throw std::runtime_error("Failed to find $HOME directory");
-    }
-    return pw->pw_dir;
-}
-#endif
-
-std::filesystem::path fs_get_cache_directory() {
-    std::filesystem::path cache_directory = common_get_path_from_env("LLAMA_CACHE");
-    if (!cache_directory.empty()) {
-        return cache_directory;
-    }
-#if defined(_WIN32)
-    cache_directory = common_get_path_from_env("LOCALAPPDATA");
+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) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    cache_directory = common_get_env("LLAMA_CACHE");
    if (cache_directory.empty()) {
-        throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
-    }
+#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/";
+        } 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");
+            }
+
+            cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
+#else /* defined(__linux__) */
+            throw std::runtime_error("Failed to find $HOME directory");
+#endif /* defined(__linux__) */
+        }
 #elif defined(__APPLE__)
-    cache_directory = get_home_directory() / "Library/Caches";
+        cache_directory = common_get_env("HOME");
+        if (cache_directory.empty()) {
+            throw std::runtime_error("Failed to find $HOME directory");
+        }
+        cache_directory += "/Library/Caches/";
+#elif defined(_WIN32)
+        cache_directory = common_get_env("LOCALAPPDATA");
+        if (cache_directory.empty()) {
+            throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
+        }
+#elif defined(__EMSCRIPTEN__)
+        GGML_ABORT("not implemented on this platform");
 #else
-    cache_directory = common_get_path_from_env("XDG_CACHE_HOME");
-    if (cache_directory.empty()) {
-        cache_directory = get_home_directory() / ".cache";
-    }
+#  error Unknown architecture
 #endif
-    return cache_directory / "llama.cpp";
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
 }

-std::filesystem::path fs_get_config_directory() {
-    std::filesystem::path config_directory;
-#if defined(_WIN32)
-    config_directory = common_get_path_from_env("APPDATA");
+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
-    config_directory = common_get_path_from_env("XDG_CONFIG_HOME");
-    if (config_directory.empty()) {
-        config_directory = get_home_directory() / ".config";
-    }
+#  error Unknown architecture
 #endif
-    return config_directory / "llama.cpp";
+    config_directory = ensure_trailing_slash(config_directory);
+    config_directory += "llama.cpp";
+    return ensure_trailing_slash(config_directory);
 }

-std::filesystem::path fs_get_cache_file(const std::string & filename) {
+std::string fs_get_cache_file(const std::string & filename) {
    GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
-    const std::filesystem::path cache_directory = fs_get_cache_directory();
-    std::error_code ec;
-    std::filesystem::create_directories(cache_directory, ec);
-    if (ec) {
-        throw std::runtime_error("failed to create cache directory: " + fs_path_to_utf8(cache_directory));
+    std::string cache_directory = fs_get_cache_directory();
+    const bool success = fs_create_directory_with_parents(cache_directory);
+    if (!success) {
+        throw std::runtime_error("failed to create cache directory: " + cache_directory);
    }
-    return cache_directory / std::filesystem::u8path(filename);
+    return cache_directory + filename;
 }

 std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
@@ -1432,8 +1527,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
        }

        if (llama_model_has_encoder(model)) {
-            common_batch batch = common_batch_get_one(lctx, tmp);
-            llama_process(lctx, LLAMA_PROCESS_TYPE_ENCODE, batch.get());
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
            llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
            if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
                decoder_start_token_id = bos;
@@ -1442,9 +1536,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
            tmp.push_back(decoder_start_token_id);
        }
        if (llama_model_has_decoder(model)) {
-            tmp.resize(std::min(tmp.size(), (size_t) params.n_batch));
-            common_batch batch = common_batch_get_one(lctx, tmp);
-            llama_process(lctx, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
        }
        llama_memory_clear(llama_get_memory(lctx), true);
        llama_synchronize(lctx);
@@ -1481,7 +1573,9 @@ char * common_get_model_or_exit(int argc, char * argv[]) {

    char * path = getenv("LLAMACPP_TEST_MODELFILE");
    if (!path || strlen(path) == 0) {
-        fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n\033[0m");
+        const bool use_color = common_log_get_verbosity_thold() > LOG_LEVEL_WARN;
+        fprintf(stderr, "%sWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n%s",
+                use_color ? LOG_COL_YELLOW : "", use_color ? LOG_COL_DEFAULT : "");
        exit(EXIT_SUCCESS);
    }

@@ -1508,13 +1602,9 @@ common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
    tmp.push_back(0);
    tmp.push_back(0);

-    int ret;
-    {
-        common_batch batch = common_batch_get_one(ctx, tmp);
-        ret = llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch.get());
-    }
+    int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
    if (ret != 0) {
-        COM_ERR("llama_process() failed: %d\n", ret);
+        COM_ERR("llama_decode() failed: %d\n", ret);
        res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
        goto done;
    }
@@ -1738,6 +1828,33 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
 }

+//
+// Batch utils
+//
+
+void common_batch_clear(struct llama_batch & batch) {
+    batch.n_tokens = 0;
+}
+
+void common_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits) {
+    GGML_ASSERT(batch.seq_id[batch.n_tokens] && "llama_batch size exceeded");
+
+    batch.token   [batch.n_tokens] = id;
+    batch.pos     [batch.n_tokens] = pos;
+    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
+    for (size_t i = 0; i < seq_ids.size(); ++i) {
+        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
+    }
+    batch.logits  [batch.n_tokens] = logits;
+
+    batch.n_tokens++;
+}
+
 //
 // Vocab utils
 //
@@ -2074,137 +2191,18 @@ float lr_opt::get_lr(float epoch) const {
 }

 bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos) {
-    common_batch batch(ctx);
-    batch.add(last_token, pos, 0, true);
-
-    if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch.get())) {
+    llama_batch batch = llama_batch_get_one(&last_token, 1);
+    batch.pos = &pos;
+    if (llama_decode(ctx, batch)) {
        LOG_ERR("%s: failed to replay last token\n", __func__);
        return false;
    }
    return true;
 }

-common_batch::common_batch(llama_context * ctx) : batch(llama_batch_ext_init(ctx)) {
-    const auto rope_type = llama_model_rope_type(llama_get_model(ctx));
-    n_pos = rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? GGML_MROPE_SECTIONS : 1;
-}
-
-void common_batch::clear() {
-    tokens.clear();
-}
-
-int32_t common_batch::add(llama_token id, llama_pos pos, llama_seq_id seq_id, bool output) {
-    tokens.push_back({ id, { pos, 0, 0, 0 }, seq_id, output, { nullptr, 0, 0 }, {} });
-    return size() - 1;
-}
-
-int32_t common_batch::add(llama_token id, llama_pos pos, const std::vector<llama_seq_id> & seq_ids, bool output) {
-    GGML_ASSERT(!seq_ids.empty());
-
-    const int32_t idx = add(id, pos, seq_ids[0], output);
-    for (size_t s = 1; s < seq_ids.size(); ++s) {
-        add_seq(idx, seq_ids[s]);
-    }
-    return idx;
-}
-
-bool common_batch::add_seq(int32_t idx, llama_seq_id seq_id) {
-    if (idx < 0 || idx >= size()) {
-        return false;
-    }
-    tokens[idx].seq_ids_extra.push_back(seq_id);
-    return true;
-}
-
-bool common_batch::set_output(int32_t idx, bool value) {
-    if (idx < 0 || idx >= size()) {
-        return false;
-    }
-    tokens[idx].output = value;
-    return true;
-}
-
-bool common_batch::set_embd(int32_t idx, llama_embd embd) {
-    if (idx < 0 || idx >= size() || tokens[idx].embd.data != nullptr) {
-        return false;
-    }
-    tokens[idx].embd = embd;
-    return true;
-}
-
-int32_t common_batch::add_embd(llama_embd embd, const llama_pos * pos, llama_seq_id seq_id, bool output) {
-    token t = { LLAMA_TOKEN_NULL, { 0, 0, 0, 0 }, seq_id, output, embd, {} };
-    for (int32_t j = 0; j < n_pos; ++j) {
-        t.pos[j] = pos[j];
-    }
-    tokens.push_back(t);
-    return size() - 1;
-}
-
-llama_batch_ext * common_batch::get_sub_batch(int32_t off, int32_t n) {
-    GGML_ASSERT(batch && "common_batch was not initialized with a context");
-    GGML_ASSERT(off >= 0 && n >= 0 && off + n <= size());
-
-    llama_batch_ext * res = batch.get();
-    llama_batch_ext_clear(res);
-
-    for (int32_t i = off; i < off + n; ++i) {
-        const token & t = tokens[i];
-
-        int32_t idx;
-        if (t.id != LLAMA_TOKEN_NULL) {
-            idx = llama_batch_ext_add_token(res, t.seq_id, t.id);
-            if (idx < 0) {
-                GGML_ABORT("%s: failed to add token %d at index %d (error %d, n = %d)\n", __func__, t.id, i, idx, n);
-            }
-            llama_batch_ext_set_pos(res, idx, t.pos.data());
-            if (t.embd.data && !llama_batch_ext_set_embd_token(res, idx, t.embd)) {
-                GGML_ABORT("%s: failed to set the embedding of token %d at index %d\n", __func__, t.id, i);
-            }
-        } else {
-            idx = llama_batch_ext_add_embd(res, t.seq_id, t.embd);
-            if (idx < 0) {
-                GGML_ABORT("%s: failed to add embedding at index %d (error %d, n = %d)\n", __func__, i, idx, n);
-            }
-            llama_batch_ext_set_pos(res, idx, t.pos.data());
-        }
-        GGML_ASSERT(idx == i - off);
-
-        for (const llama_seq_id seq_id : t.seq_ids_extra) {
-            if (!llama_batch_ext_add_seq(res, idx, seq_id)) {
-                GGML_ABORT("%s: failed to add seq %d to the entry at index %d\n", __func__, seq_id, i);
-            }
-        }
-        if (t.output) {
-            llama_batch_ext_set_output_logits(res, idx, true);
-        }
-    }
-
-    return res;
-}
-
-common_batch common_batch_get_one(llama_context * ctx, const llama_token * tokens, int32_t n_tokens) {
-    common_batch batch(ctx);
-
-    auto mem = llama_get_memory(ctx);
-    llama_pos pos = llama_memory_seq_pos_max(mem, 0) + 1; // -1 + 1 == 0 when the memory is empty
-
-    for (int32_t i = 0; i < n_tokens; ++i) {
-        const bool output = i == n_tokens - 1;
-        batch.add(tokens[i], pos, 0, output);
-        pos++;
-    }
-
-    return batch;
-}
-
-common_batch common_batch_get_one(llama_context * ctx, const llama_tokens & tokens) {
-    return common_batch_get_one(ctx, tokens.data(), (int32_t) tokens.size());
-}
-
 bool common_prompt_batch_decode(
              struct llama_context * ctx,
-                const llama_tokens & all_tokens,
+    const std::vector<llama_token> & all_tokens,
                               int   n_new,
                               int & n_past,
                               int   n_batch,
@@ -2225,9 +2223,7 @@ bool common_prompt_batch_decode(
        // Memory implementations in recurrent/hybrid models don't support removing tokens from their
        // memory, so we can't just remove the last token from the memory and replay the last token which
        // is the reason for this logic.
-        llama_tokens prefix_tokens(all_tokens.begin() + offset, all_tokens.begin() + offset + n_tokens_before_last);
-        common_batch batch_prefix = common_batch_get_one(ctx, prefix_tokens);
-        if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch_prefix.get())) {
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_tokens_before_last))) {
            COM_ERR("%s", "failed to eval\n");
            return false;
        }
@@ -2236,18 +2232,18 @@ bool common_prompt_batch_decode(
        llama_state_save_file(ctx, state_path.data(), all_tokens.data(), all_tokens.size());
        COM_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size());

-        common_batch batch_last(ctx);
-        batch_last.add(all_tokens.back(), n_past, 0, true);
+        llama_token last_token = all_tokens.back();
+        llama_batch batch = llama_batch_get_one(&last_token, 1);
+        int32_t pos = n_past;
+        batch.pos = &pos;

-        if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch_last.get())) {
+        if (llama_decode(ctx, batch)) {
            COM_ERR("%s", "failed to eval last token\n");
            return false;
        }
        n_past++;
    } else {
-        llama_tokens new_tokens(all_tokens.begin() + offset, all_tokens.begin() + offset + n_new);
-        common_batch batch = common_batch_get_one(ctx, new_tokens);
-        if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch.get())) {
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_new))) {
            COM_ERR("%s", "failed to eval\n");
            return false;
        }
@@ -8,7 +8,6 @@
 #include "ggml.h"
 #include "llama.h"

-#include <array>
 #include <list>
 #include <set>
 #include <sstream>
@@ -17,7 +16,6 @@
 #include <vector>
 #include <map>
 #include <algorithm>
-#include <filesystem>
 #include <fstream>

 #if defined(_WIN32) && !defined(_WIN32_WINNT)
@@ -633,10 +631,10 @@ struct common_params {
    int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
    int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.

+    std::string hostname      = "127.0.0.1";
    std::string public_path   = "";                                                                         // NOLINT
    std::string api_prefix    = "";                                                                         // NOLINT
    std::string chat_template = "";                                                                         // NOLINT
-    std::vector<std::string> hostnames = {"127.0.0.1"};
    bool use_jinja = true;                                                                                  // NOLINT

    // server CORS params
@@ -810,9 +808,7 @@ static std::vector<T> string_split(const std::string & str, char delim) {
    while (std::getline(str_stream, token, delim)) {
        T value;
        std::istringstream token_stream(token);
-        if (!(token_stream >> value)) {
-            throw std::invalid_argument("invalid value: \"" + token + "\"");
-        }
+        token_stream >> value;
        values.push_back(value);
    }
    return values;
@@ -880,21 +876,10 @@ void string_process_escapes(std::string & input);
 std::string string_from(bool value);
 std::string string_from(const std::vector<int> & values);
 std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
+std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);

 bool glob_match(const std::string & pattern, const std::string & str);

-//
-// Unicode utils
-//
-
-#ifdef _WIN32
-std::wstring utf8_to_wstring(const std::string & str);
-std::string  wstring_to_utf8(const std::wstring & str);
-#endif
-
-// returns the path as a UTF-8 string, preserving its separators
-std::string fs_path_to_utf8(const std::filesystem::path & path);
-
 //
 // Environment utils
 //
@@ -904,19 +889,17 @@ std::string fs_path_to_utf8(const std::filesystem::path & path);
 std::string common_get_env(const std::string & name);
 void        common_set_env(const std::string & name, const std::string & value);

-// reads a path from the environment, an unset variable gives an empty path
-std::filesystem::path common_get_path_from_env(const std::string & name);
-
 //
 // Filesystem utils
 //

 bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
+bool fs_create_directory_with_parents(const std::string & path);
 bool fs_is_directory(const std::string & path);

-std::filesystem::path fs_get_cache_directory();
-std::filesystem::path fs_get_cache_file(const std::string & filename);
-std::filesystem::path fs_get_config_directory();
+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;
@@ -929,8 +912,6 @@ std::vector<common_file_info> fs_list(const std::string & path, bool include_dir
 // fs open, also handle UTF8 on Windows
 std::ifstream fs_open_ifstream(const std::string & fname, std::ios_base::openmode mode);

-void fs_write_atomic(const std::filesystem::path & path, const std::string & data);
-
 //
 // TTY utils
 //
@@ -1031,62 +1012,14 @@ struct common_memory {
 // Batch utils
 //

-// wrapper around llama_batch_ext that provide getter functions for downstream code
-// entries can exceed n_batch, use get_sub_batch() to decode them in chunks
-struct common_batch {
-    struct token {
-        llama_token  id;
-        std::array<llama_pos, GGML_MROPE_SECTIONS> pos; // only pos[0] is used for text tokens
-        llama_seq_id seq_id; // the first sequence id, see add_seq()
-        bool         output;
-        llama_embd   embd; // non-owning view of the data passed to add_embd()/set_embd(), data == NULL if none
-        std::vector<llama_seq_id> seq_ids_extra; // see add_seq()
-    };
+void common_batch_clear(struct llama_batch & batch);

-    std::vector<token> tokens; // mirror of the entries, tokens[i] describes batch index i
-    llama_batch_ext_ptr batch;
-
-    int32_t n_pos = 1; // positions per embedding entry, GGML_MROPE_SECTIONS for MROPE/IMROPE
-
-    common_batch() = default;
-    common_batch(struct llama_context * ctx);
-
-    llama_batch_ext * get() { return get_sub_batch(0, size()); }
-
-    // render entries [off, off + n) into batch, the result is overwritten by the next call
-    llama_batch_ext * get_sub_batch(int32_t off, int32_t n);
-
-    // content type of the batch, all entries carry the same combination
-    bool has_token() const { return !tokens.empty() && tokens[0].id != LLAMA_TOKEN_NULL; }
-    bool has_embd () const { return !tokens.empty() && tokens[0].embd.data != nullptr; }
-
-    void clear();
-
-    // returns the batch index
-    int32_t add(llama_token id, llama_pos pos, llama_seq_id seq_id, bool output);
-
-    // same, with the entry shared by all seq_ids (must not be empty)
-    int32_t add(llama_token id, llama_pos pos, const std::vector<llama_seq_id> & seq_ids, bool output);
-
-    // add the entry at idx to another sequence, tokens[idx].seq_id keeps the first one
-    bool add_seq(int32_t idx, llama_seq_id seq_id);
-
-    bool set_output(int32_t idx, bool value);
-
-    // attach a token embedding to the entry at idx, can only be set once per entry
-    bool set_embd(int32_t idx, llama_embd embd);
-
-    // add an embedding-only entry (no token id)
-    // pos points to n_pos positions
-    int32_t add_embd(llama_embd embd, const llama_pos * pos, llama_seq_id seq_id, bool output);
-
-    int32_t size() const { return (int32_t) tokens.size(); }
-};
-
-// create a single-sequence batch from a list of tokens
-// positions continue from the memory, last token always have output_logits set to true
-common_batch common_batch_get_one(struct llama_context * ctx, const llama_token * tokens, int32_t n_tokens);
-common_batch common_batch_get_one(struct llama_context * ctx, const llama_tokens & tokens);
+void common_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits);

 // decodes a single batch of tokens for a prompt and manages session tokens
 //
@@ -1095,7 +1028,7 @@ common_batch common_batch_get_one(struct llama_context * ctx, const llama_tokens
 // tokens from memory, so this approach works across all model architectures.
 bool common_prompt_batch_decode(
              struct llama_context * ctx,
-                const llama_tokens & all_tokens,
+    const std::vector<llama_token> & all_tokens,
                               int   n_new,
                               int & n_past,
                               int   n_batch,
@@ -1,5 +1,4 @@
 #include "console.h"
-#include "common.h"
 #include "log.h"
 #include <vector>
 #include <iostream>
@@ -1019,7 +1018,6 @@ namespace console {
                    line.clear();
                    pop_cursor();
                }
-                line += '\n';
                has_more = false;
            }
        } else {
@@ -1051,10 +1049,13 @@ namespace console {
        if (!std::getline(std::wcin, wline)) {
            // Input stream is bad or EOF received
            line.clear();
+            GenerateConsoleCtrlEvent(CTRL_C_EVENT, 0);
            return false;
        }

-        line = wstring_to_utf8(wline);
+        int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), NULL, 0, NULL, NULL);
+        line.resize(size_needed);
+        WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), &line[0], size_needed, NULL, NULL);
 #else
        if (!std::getline(std::cin, line)) {
            // Input stream is bad or EOF received
@@ -1065,7 +1066,7 @@ namespace console {
        if (!line.empty()) {
            char last = line.back();
            if (last == '/') { // Always return control on '/' symbol
-                line.back() = '\n';
+                line.pop_back();
                return false;
            }
            if (last == '\\') { // '\\' changes the default action
@@ -46,9 +46,39 @@
 // downloader
 //

+// validate repo name format: owner/repo
+static void write_file(const std::string & fname, const std::string & content) {
+    const std::string fname_tmp = fname + ".tmp";
+    std::ofstream     file(fname_tmp);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+
+    try {
+        file << content;
+        file.close();
+
+        // Makes write atomic
+        if (rename(fname_tmp.c_str(), fname.c_str()) != 0) {
+            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, fname_tmp.c_str(), fname.c_str());
+            // If rename fails, try to delete the temporary file
+            if (remove(fname_tmp.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete temporary file: %s\n", __func__, fname_tmp.c_str());
+            }
+        }
+    } catch (...) {
+        // If anything fails, try to delete the temporary file
+        if (remove(fname_tmp.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete temporary file: %s\n", __func__, fname_tmp.c_str());
+        }
+
+        throw std::runtime_error(string_format("error: failed to write file '%s'\n", fname.c_str()));
+    }
+}
+
 static void write_etag(const std::string & path, const std::string & etag) {
    const std::string etag_path = path + ".etag";
-    fs_write_atomic(std::filesystem::u8path(etag_path), etag);
+    write_file(etag_path, etag);
    LOG_DBG("%s: file etag saved: %s\n", __func__, etag_path.c_str());
 }

@@ -244,12 +274,6 @@ static bool common_pull_file(httplib::Client & cli,
        return false;
    }

-    ofs.close();
-    if (!ofs) {
-        LOG_ERR("%s: error closing file: %s\n", __func__, path_tmp.c_str());
-        return false;
-    }
-
    return true;
 }

@@ -262,7 +286,7 @@ static int common_download_file_single_online(const std::string & url,
    static const int max_attempts        = 3;
    static const int retry_delay_seconds = 2;

-    const bool file_exists = std::filesystem::exists(std::filesystem::u8path(path));
+    const bool file_exists = std::filesystem::exists(path);

    if (file_exists && skip_etag) {
        LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
@@ -453,7 +477,7 @@ int common_download_file_single(const std::string & url,
        return common_download_file_single_online(url, path, online_opts, skip_etag);
    }

-    if (!std::filesystem::exists(std::filesystem::u8path(path))) {
+    if (!std::filesystem::exists(path)) {
        LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
        return -1;
    }
@@ -919,7 +943,7 @@ std::string common_docker_resolve_model(const std::string & docker) {
        std::string model_filename = repo;
        std::replace(model_filename.begin(), model_filename.end(), '/', '_');
        model_filename += "_" + tag + ".gguf";
-        std::string local_path = fs_path_to_utf8(fs_get_cache_file(model_filename));
+        std::string local_path = fs_get_cache_file(model_filename);

        const std::string blob_url = url_prefix + "/blobs/" + gguf_digest;
        common_download_opts opts;
@@ -192,9 +192,9 @@ 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);
+    // size the context for all sequences, but keep minimums and alignment per KV stream
+    const uint32_t n_seq_max  = std::max<uint32_t>(1, cparams->n_seq_max);
+    const uint32_t n_streams  = cparams->kv_unified ? 1 : 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
@@ -264,15 +264,15 @@ static void common_params_fit_impl(
    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_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_seq_max, 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);
+        if (n_seq_max > 1) {
+            LOG_TRC("%s: context size unset -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
+                __func__, n_ctx_max, n_seq_max);
            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
        }
    }
@@ -44,7 +44,8 @@ static fs::path get_cache_directory() {
            {HOME_DIR,                fs::path(".cache") / "huggingface" / "hub"}
        };
        for (const auto & entry : entries) {
-            if (fs::path base = common_get_path_from_env(entry.var); !base.empty()) {
+            if (auto * p = std::getenv(entry.var); p && *p) {
+                fs::path base(p);
                return entry.path.empty() ? base : base / entry.path;
            }
        }
@@ -61,10 +62,6 @@ static fs::path get_cache_directory() {
    return cache;
 }

-std::string get_cache_path() {
-    return fs_path_to_utf8(get_cache_directory());
-}
-
 static std::string folder_name_to_repo(const std::string & folder) {
    constexpr std::string_view prefix = "models--";
    if (folder.rfind(prefix, 0)) {
@@ -172,6 +169,28 @@ static bool is_valid_subpath(const fs::path & path, const fs::path & subpath) {
    return b_end == b.end();
 }

+static void safe_write_file(const fs::path & path, const std::string & data) {
+    fs::path path_tmp = path.string() + ".tmp";
+
+    if (path.has_parent_path()) {
+        fs::create_directories(path.parent_path());
+    }
+
+    std::ofstream file(path_tmp);
+    file << data;
+    file.close();
+
+    std::error_code ec;
+
+    if (!file.fail()) {
+        fs::rename(path_tmp, path, ec);
+    }
+    if (file.fail() || ec) {
+        fs::remove(path_tmp, ec);
+        throw std::runtime_error("failed to write file: " + path.string());
+    }
+}
+
 static common_json api_get(const std::string & url,
                           const std::string & token) {
    auto [cli, parts] = common_http_client(url);
@@ -218,7 +237,6 @@ static std::string get_repo_commit(const std::string & repo_id,
        fs::path refs_path = get_repo_path(repo_id) / "refs";
        std::string name;
        std::string commit;
-        fs::path name_path;

        for (const auto & branch : json["branches"]) {
            if (!branch.is_object() ||
@@ -229,28 +247,24 @@ static std::string get_repo_commit(const std::string & repo_id,
            std::string _name = branch["name"].get<std::string>();
            std::string _commit = branch["targetCommit"].get<std::string>();

-            if (!is_valid_commit(_commit)) {
-                LOG_WRN("%s: skip invalid commit: %s\n", __func__, _commit.c_str());
+            if (!is_valid_subpath(refs_path, _name)) {
+                LOG_WRN("%s: skip invalid branch: %s\n", __func__, _name.c_str());
                continue;
            }
-            const fs::path candidate = fs::u8path(_name);
-
-            if (!is_valid_subpath(refs_path, candidate)) {
-                LOG_WRN("%s: skip invalid branch: %s\n", __func__, _name.c_str());
+            if (!is_valid_commit(_commit)) {
+                LOG_WRN("%s: skip invalid commit: %s\n", __func__, _commit.c_str());
                continue;
            }

            if (_name == "main") {
                name = _name;
                commit = _commit;
-                name_path = candidate;
                break;
            }

            if (name.empty() || commit.empty()) {
                name = _name;
                commit = _commit;
-                name_path = candidate;
            }
        }

@@ -259,7 +273,7 @@ static std::string get_repo_commit(const std::string & repo_id,
            return {};
        }

-        fs_write_atomic(refs_path / name_path, commit);
+        safe_write_file(refs_path / name, commit);
        return commit;

    } catch (const common_json_error & e) {
@@ -308,9 +322,7 @@ hf_files get_repo_files(const std::string & repo_id,
            file.repo_id = repo_id;
            file.path = item["path"].get<std::string>();

-            const fs::path subpath = fs::u8path(file.path);
-
-            if (!is_valid_subpath(commit_path, subpath)) {
+            if (!is_valid_subpath(commit_path, file.path)) {
                LOG_WRN("%s: skip invalid path: %s\n", __func__, file.path.c_str());
                continue;
            }
@@ -330,12 +342,12 @@ hf_files get_repo_files(const std::string & repo_id,

            file.url = endpoint + repo_id + "/resolve/" + commit + "/" + file.path;

-            fs::path final_path = commit_path / subpath;
-            file.final_path = fs_path_to_utf8(final_path);
+            fs::path final_path = commit_path / file.path;
+            file.final_path = final_path.string();

            if (!file.oid.empty() && !fs::exists(final_path)) {
                fs::path local_path = blobs_path / file.oid;
-                file.local_path = fs_path_to_utf8(local_path);
+                file.local_path = local_path.string();
            } else {
                file.local_path = file.final_path;
            }
@@ -381,8 +393,8 @@ static std::string get_cached_ref(const fs::path & repo_path) {
 }

 hf_files get_cached_files(const std::string & repo_id) {
-    const fs::path cache_path = get_cache_directory();
-    if (!fs::exists(cache_path)) {
+    fs::path cache_dir = get_cache_directory();
+    if (!fs::exists(cache_dir)) {
        return {};
    }

@@ -393,7 +405,7 @@ hf_files get_cached_files(const std::string & repo_id) {

    hf_files files;

-    for (const auto & repo : fs::directory_iterator(cache_path)) {
+    for (const auto & repo : fs::directory_iterator(cache_dir)) {
        if (!repo.is_directory()) {
            continue;
        }
@@ -402,7 +414,7 @@ hf_files get_cached_files(const std::string & repo_id) {
        if (!fs::exists(snapshots_path)) {
            continue;
        }
-        std::string _repo_id = folder_name_to_repo(fs_path_to_utf8(repo.path().filename()));
+        std::string _repo_id = folder_name_to_repo(repo.path().filename().string());

        if (!is_valid_repo_id(_repo_id)) {
            continue;
@@ -425,9 +437,8 @@ hf_files get_cached_files(const std::string & repo_id) {
            if (!path.empty()) {
                hf_file file;
                file.repo_id = _repo_id;
-                const auto generic_path = path.generic_u8string();
-                file.path = std::string(generic_path.begin(), generic_path.end());
-                file.local_path = fs_path_to_utf8(entry.path());
+                file.path = path.generic_string();
+                file.local_path = entry.path().string();
                file.final_path = file.local_path;
                files.push_back(std::move(file));
            }
@@ -441,8 +452,8 @@ std::string finalize_file(const hf_file & file) {
    static std::atomic<bool> symlinks_disabled{false};

    std::error_code ec;
-    fs::path local_path = fs::u8path(file.local_path);
-    fs::path final_path = fs::u8path(file.final_path);
+    fs::path local_path(file.local_path);
+    fs::path final_path(file.final_path);

    if (local_path == final_path || fs::exists(final_path, ec)) {
        return file.final_path;
@@ -489,7 +500,7 @@ bool remove_cached_repo(const std::string & repo_id) {
    std::error_code ec;
    auto removed = fs::remove_all(repo_path, ec);
    if (ec) {
-        LOG_ERR("%s: failed to remove repo cache %s: %s\n", __func__, fs_path_to_utf8(repo_path).c_str(), ec.message().c_str());
+        LOG_ERR("%s: failed to remove repo cache %s: %s\n", __func__, repo_path.string().c_str(), ec.message().c_str());
        return false;
    }
    return removed > 0;
@@ -32,7 +32,4 @@ std::string finalize_file(const hf_file & file);
 // Remove the entire cached directory for a repo, returns true if removed
 bool remove_cached_repo(const std::string & repo_id);

-// Returns the HuggingFace hub cache path
-std::string get_cache_path();
-
 } // namespace hf_cache
@@ -429,23 +429,15 @@ private:
            bool negate = false;
            if (is_identifier("not")) { ++current; negate = true; }
            auto test_id = parse_primary_expression();
-            if (is(token::open_paren)) {
-                test_id = parse_call_expression(std::move(test_id));
-            } else if (is(token::numeric_literal) || is(token::string_literal) || is(token::open_curly_bracket) || is(token::open_square_bracket) ||
-                    (is(token::identifier) && !is_identifier("and") && !is_identifier("or") && !is_identifier("else"))) {
-                size_t call_pos = current;
-                statements args;
-                args.push_back(parse_unary_expression());
-                test_id = mk_stmt<call_expression>(call_pos, std::move(test_id), std::move(args));
-            }
+            // FIXME: tests can also be expressed like this: if x is eq 3
+            if (is(token::open_paren)) test_id = parse_call_expression(std::move(test_id));
            operand = mk_stmt<test_expression>(start_pos, std::move(operand), negate, std::move(test_id));
        }
        return operand;
    }

    statement_ptr parse_filter_expression() {
-        // Filters/tests bind outside unary so -n|abs is (-n)|abs, not -(n|abs).
-        auto operand = parse_unary_expression();
+        auto operand = parse_call_member_expression();
        while (is(token::pipe)) {
            size_t start_pos = current;
            ++current; // consume pipe
@@ -456,15 +448,6 @@ private:
        return operand;
    }

-    statement_ptr parse_unary_expression() {
-        if (is(token::unary_operator)) {
-            size_t start_pos = current;
-            auto op = next();
-            return mk_stmt<unary_expression>(start_pos, op, parse_unary_expression());
-        }
-        return parse_call_member_expression();
-    }
-
    statement_ptr parse_call_member_expression() {
        // Handle member expressions recursively
        auto member = parse_member_expression(parse_primary_expression());
@@ -51,7 +51,7 @@ static void ensure_key_type_allowed(const value & val) {
 }

 // execute with error handling
-value statement::execute(context & ctx) const {
+value statement::execute(context & ctx) {
    try {
        return execute_impl(ctx);
    } catch (const continue_statement::signal & /* ex */) {
@@ -80,7 +80,7 @@ value statement::execute(context & ctx) const {
    }
 }

-value identifier::execute_impl(context & ctx) const {
+value identifier::execute_impl(context & ctx) {
    auto it = ctx.get_val(val);
    auto builtins = global_builtins();
    if (!it->is_undefined()) {
@@ -98,7 +98,7 @@ value identifier::execute_impl(context & ctx) const {
    }
 }

-value object_literal::execute_impl(context & ctx) const {
+value object_literal::execute_impl(context & ctx) {
    auto obj = mk_val<value_object>();
    for (const auto & pair : val) {
        value key = pair.first->execute(ctx);
@@ -109,7 +109,7 @@ value object_literal::execute_impl(context & ctx) const {
    return obj;
 }

-value binary_expression::execute_impl(context & ctx) const {
+value binary_expression::execute_impl(context & ctx) {
    value left_val = left->execute(ctx);

    // Logical operators
@@ -317,7 +317,9 @@ static value try_builtin_func(context & ctx, const std::string & name, value & i
    throw std::runtime_error("Unknown (built-in) filter '" + name + "' for type " + input->type());
 }

-static value apply_filter(context & ctx, const statement_ptr & filter, value input) {
+value filter_expression::execute_impl(context & ctx) {
+    value input = operand ? operand->execute(ctx) : val;
+
    JJ_DEBUG("Applying filter to %s", input->type().c_str());

    auto set_filter_alias = [](auto & filter_id) {
@@ -373,21 +375,22 @@ static value apply_filter(context & ctx, const statement_ptr & filter, value inp
    }
 }

-value filter_expression::execute_impl(context & ctx) const {
-    return apply_filter(ctx, filter, operand->execute(ctx));
-}
-
-value filter_statement::execute_impl(context & ctx) const {
+value filter_statement::execute_impl(context & ctx) {
    // eval body as string, then apply filter
    auto body_val = exec_statements(body, ctx);
    value_string parts = mk_val<value_string>();
    gather_string_parts_recursive(body_val, parts);

    JJ_DEBUG("FilterStatement: applying filter to body string of length %zu", parts->val_str.length());
-    return apply_filter(ctx, filter, parts);
+    filter_expression filter_expr(std::move(parts), std::move(filter));
+    value out = filter_expr.execute(ctx);
+
+    // this node can be reused later, make sure filter is preserved
+    this->filter = std::move(filter_expr.filter);
+    return out;
 }

-value test_expression::execute_impl(context & ctx) const {
+value test_expression::execute_impl(context & ctx) {
    // NOTE: "value is something" translates to function call "test_is_something(value)"
    const auto & builtins = global_builtins();

@@ -436,7 +439,7 @@ value test_expression::execute_impl(context & ctx) const {
    }
 }

-value unary_expression::execute_impl(context & ctx) const {
+value unary_expression::execute_impl(context & ctx) {
    value operand_val = argument->execute(ctx);
    JJ_DEBUG("Executing unary expression with operator '%s'", op.value.c_str());

@@ -450,17 +453,12 @@ value unary_expression::execute_impl(context & ctx) const {
        } else {
            throw std::runtime_error("Unary - operator requires numeric operand");
        }
-    } else if (op.value == "+") {
-        if (is_val<value_int>(operand_val) || is_val<value_float>(operand_val)) {
-            return operand_val;
-        }
-        throw std::runtime_error("Unary + operator requires numeric operand");
    }

    throw std::runtime_error("Unknown unary operator '" + op.value + "'");
 }

-value if_statement::execute_impl(context & ctx) const {
+value if_statement::execute_impl(context & ctx) {
    value test_val = test->execute(ctx);

    auto out = mk_val<value_array>();
@@ -481,14 +479,20 @@ value if_statement::execute_impl(context & ctx) const {
    return str;
 }

-value for_statement::execute_impl(context & ctx) const {
+value for_statement::execute_impl(context & ctx) {
    context scope(ctx); // new scope for loop variables

-    const jinja::select_expression * select_expr = cast_stmt<select_expression>(iterable);
+    jinja::select_expression * select_expr = cast_stmt<select_expression>(iterable);
    statement_ptr test_expr_nullptr;

-    const statement_ptr & iter_expr = select_expr ? select_expr->lhs : iterable;
-    const statement_ptr & test_expr = select_expr ? select_expr->test : test_expr_nullptr;
+    statement_ptr & iter_expr = [&]() -> statement_ptr & {
+        auto tmp = cast_stmt<select_expression>(iterable);
+        return tmp ? tmp->lhs : iterable;
+    }();
+    statement_ptr & test_expr = [&]() -> statement_ptr & {
+        auto tmp = cast_stmt<select_expression>(iterable);
+        return tmp ? tmp->test : test_expr_nullptr;
+    }();

    JJ_DEBUG("Executing for statement, iterable type: %s", iter_expr->type().c_str());

@@ -537,6 +541,8 @@ value for_statement::execute_impl(context & ctx) const {

    std::vector<value> filtered_items;
    for (size_t i = 0; i < items.size(); ++i) {
+        context loop_scope(scope);
+
        value current = items[i];

        std::function<void(context&)> scope_update_fn = [](context &) { /* no-op */};
@@ -582,7 +588,6 @@ value for_statement::execute_impl(context & ctx) const {
        }

        if (select_expr && test_expr) {
-            context loop_scope(scope);
            scope_update_fn(loop_scope);
            value test_val = test_expr->execute(loop_scope);
            if (!test_val->as_bool()) {
@@ -640,7 +645,7 @@ value for_statement::execute_impl(context & ctx) const {
    return str;
 }

-value set_statement::execute_impl(context & ctx) const {
+value set_statement::execute_impl(context & ctx) {
    auto rhs = val ? val->execute(ctx) : exec_statements(body, ctx);

    if (is_stmt<identifier>(assignee)) {
@@ -739,7 +744,7 @@ static inline void bind_parameters(const std::string & name, const statements &
    }
 }

-value macro_statement::execute_impl(context & ctx) const {
+value macro_statement::execute_impl(context & ctx) {
    if (!is_stmt<identifier>(this->name)) {
        throw std::runtime_error("Macro name must be an identifier");
    }
@@ -762,7 +767,7 @@ value macro_statement::execute_impl(context & ctx) const {
    return mk_val<value_undefined>();
 }

-value call_statement::execute_impl(context & ctx) const {
+value call_statement::execute_impl(context & ctx) {
    auto call_expr = cast_stmt<call_expression>(this->call);
    if (!call_expr) {
        throw std::runtime_error("Call statement requires a valid call expression");
@@ -802,7 +807,7 @@ value call_statement::execute_impl(context & ctx) const {
    return callee_func->invoke(args);
 }

-value member_expression::execute_impl(context & ctx) const {
+value member_expression::execute_impl(context & ctx) {
    value object = this->object->execute(ctx);

    value property;
@@ -887,7 +892,7 @@ value member_expression::execute_impl(context & ctx) const {
        JJ_DEBUG("Accessed property '%s' value, got type: %s", key.c_str(), val->type().c_str());

    } else if (is_val<value_array>(object) || is_val<value_string>(object)) {
-        if (is_val<value_int>(property) || is_val<value_bool>(property)) {
+        if (is_val<value_int>(property)) {
            int64_t index = property->as_int();
            JJ_DEBUG("Accessing %s index %d", object->type().c_str(), (int)index);
            if (is_val<value_array>(object)) {
@@ -910,6 +915,8 @@ value member_expression::execute_impl(context & ctx) const {
            JJ_DEBUG("Accessing %s built-in '%s'", is_val<value_array>(object) ? "array" : "string", key.c_str());
            val = try_builtin_func(ctx, key, object, true);

+        } else {
+            throw std::runtime_error("Cannot access property with non-string/non-number: got " + property->type());
        }
    } else {
        if (!is_val<value_string>(property)) {
@@ -923,17 +930,17 @@ value member_expression::execute_impl(context & ctx) const {
        value_t::stats_t::mark_used(val);
        value_t::stats_t::mark_used(object);
        value_t::stats_t::mark_used(property);
-        if (is_val<value_object>(object) || is_val<value_string>(property) || is_val<value_float>(property) || is_val<value_array>(property) || is_val<value_none>(property)) {
-            object->stats.ops.insert("object_access");
-        } else if (is_val<value_int>(property) || is_val<value_bool>(property)) {
+        if (is_val<value_int>(property)) {
            object->stats.ops.insert("array_access");
+        } else if (is_val<value_string>(property)) {
+            object->stats.ops.insert("object_access");
        }
    }

    return val;
 }

-value call_expression::execute_impl(context & ctx) const {
+value call_expression::execute_impl(context & ctx) {
    // gather arguments
    func_args args(ctx);
    for (auto & arg_stmt : this->args) {
@@ -951,7 +958,7 @@ value call_expression::execute_impl(context & ctx) const {
    return callee_func->invoke(args);
 }

-value keyword_argument_expression::execute_impl(context & ctx) const {
+value keyword_argument_expression::execute_impl(context & ctx) {
    if (!is_stmt<identifier>(key)) {
        throw std::runtime_error("Keyword argument key must be identifiers");
    }
@@ -975,7 +982,7 @@ std::string runtime::debug_dump_program(const program & prog, const std::string
        return std::string(lvl * 2, ' ');
    };

-    ctx.visitor = [&](bool is_leaf, const statement * node, std::vector<visitor_pair> children) {
+    ctx.visitor = [&](bool is_leaf, statement * node, std::vector<visitor_pair> children) {
        oss << indent(lvl) << node->type() << ":\n";
        lvl++;
        if (is_leaf) {
@@ -48,9 +48,9 @@ const T * cast_stmt(const statement_ptr & ptr) {
 void enable_debug(bool enable);

 // for visiting AST nodes
-// function signature: void(bool is_leaf, const statement * node, pair of <label, children>)
-using visitor_pair = std::pair<std::string, std::vector<const statement *>>;
-using visitor_fn = std::function<void(bool, const statement *, std::vector<visitor_pair>)>;
+// function signature: void(bool is_leaf, statement * node, pair of <label, children>)
+using visitor_pair = std::pair<std::string, std::vector<statement *>>;
+using visitor_fn = std::function<void(bool, statement *, std::vector<visitor_pair>)>;

 struct context {
    std::shared_ptr<std::string> src; // for debugging; use shared_ptr to avoid copying on scope creation
@@ -107,8 +107,8 @@ private:
 };

 // utils for visiting AST nodes
-static std::vector<const statement *> stmts_to_ptr(const statements & stmts) {
-    std::vector<const statement *> children;
+static std::vector<statement *> stmts_to_ptr(const statements & stmts) {
+    std::vector<statement *> children;
    for (const auto & stmt : stmts) {
        children.push_back(stmt.get());
    }
@@ -117,18 +117,17 @@ static std::vector<const statement *> stmts_to_ptr(const statements & stmts) {

 /**
 * Base class for all nodes in the AST.
- * The AST is shared between threads, so visit and execute must be const.
 */
 struct statement {
    size_t pos; // position in source, for debugging
    virtual ~statement() = default;
    virtual std::string type() const { return "Statement"; }
-    virtual void visit(context & ctx) const { ctx.visitor(true, this, {}); }
+    virtual void visit(context & ctx) { ctx.visitor(true, this, {}); }

    // execute_impl must be overridden by derived classes
-    virtual value execute_impl(context &) const { throw_exec_error(); }
+    virtual value execute_impl(context &) { throw_exec_error(); }
    // execute is the public method to execute a statement with error handling
-    value execute(context &) const;
+    value execute(context &);

 private:
    [[noreturn]] void throw_exec_error() const {
@@ -167,7 +166,7 @@ struct program : public statement {
    program() = default;
    explicit program(statements && body) : body(std::move(body)) {}
    std::string type() const override { return "Program"; }
-    [[noreturn]] value execute_impl(context &) const override {
+    [[noreturn]] value execute_impl(context &) override {
        throw std::runtime_error("Cannot execute program directly, use jinja::runtime instead");
    }
 };
@@ -183,8 +182,8 @@ struct if_statement : public statement {
    }

    std::string type() const override { return "If"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"test", {test.get()}},
            {"body", stmts_to_ptr(body)},
@@ -214,8 +213,8 @@ struct for_statement : public statement {
    }

    std::string type() const override { return "For"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"loopvar", {loopvar.get()}},
            {"iterable", {iterable.get()}},
@@ -234,7 +233,7 @@ struct break_statement : public statement {
        }
    };

-    [[noreturn]] value execute_impl(context &) const override {
+    [[noreturn]] value execute_impl(context &) override {
        throw break_statement::signal();
    }
 };
@@ -248,7 +247,7 @@ struct continue_statement : public statement {
        }
    };

-    [[noreturn]] value execute_impl(context &) const override {
+    [[noreturn]] value execute_impl(context &) override {
        throw continue_statement::signal();
    }
 };
@@ -256,7 +255,7 @@ struct continue_statement : public statement {
 // do nothing
 struct noop_statement : public statement {
    std::string type() const override { return "Noop"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_undefined>();
    }
 };
@@ -273,8 +272,8 @@ struct set_statement : public statement {
    }

    std::string type() const override { return "Set"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"assignee", {assignee.get()}},
            {"value", {val.get()}},
@@ -295,8 +294,8 @@ struct macro_statement : public statement {
    }

    std::string type() const override { return "Macro"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"name", {name.get()}},
            {"args", stmts_to_ptr(args)},
@@ -309,7 +308,7 @@ struct comment_statement : public statement {
    std::string val;
    explicit comment_statement(const std::string & v) : val(v) {}
    std::string type() const override { return "Comment"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_undefined>();
    }
 };
@@ -319,7 +318,7 @@ struct comment_statement : public statement {
 // Represents an omitted expression in a computed member, e.g. `a[]`.
 struct blank_expression : public expression {
    std::string type() const override { return "BlankExpression"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_undefined>();
    }
 };
@@ -335,8 +334,8 @@ struct member_expression : public expression {
        chk_type<expression>(this->property);
    }
    std::string type() const override { return "MemberExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"object", {object.get()}},
            {"property", {property.get()}}
@@ -354,8 +353,8 @@ struct call_expression : public expression {
        for (const auto& arg : this->args) chk_type<expression>(arg);
    }
    std::string type() const override { return "CallExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"callee", {callee.get()}},
            {"args", stmts_to_ptr(args)}
@@ -370,7 +369,7 @@ struct identifier : public expression {
    std::string val;
    explicit identifier(const std::string & val) : val(val) {}
    std::string type() const override { return "Identifier"; }
-    value execute_impl(context & ctx) const override;
+    value execute_impl(context & ctx) override;
 };

 // Literals
@@ -379,7 +378,7 @@ struct integer_literal : public expression {
    int64_t val;
    explicit integer_literal(int64_t val) : val(val) {}
    std::string type() const override { return "IntegerLiteral"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_int>(val);
    }
 };
@@ -388,7 +387,7 @@ struct float_literal : public expression {
    double val;
    explicit float_literal(double val) : val(val) {}
    std::string type() const override { return "FloatLiteral"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_float>(val);
    }
 };
@@ -397,7 +396,7 @@ struct string_literal : public expression {
    std::string val;
    explicit string_literal(const std::string & val) : val(val) {}
    std::string type() const override { return "StringLiteral"; }
-    value execute_impl(context &) const override {
+    value execute_impl(context &) override {
        return mk_val<value_string>(val);
    }
 };
@@ -408,7 +407,7 @@ struct array_literal : public expression {
        for (const auto& item : this->val) chk_type<expression>(item);
    }
    std::string type() const override { return "ArrayLiteral"; }
-    value execute_impl(context & ctx) const override {
+    value execute_impl(context & ctx) override {
        auto arr = mk_val<value_array>();
        for (const auto & item_stmt : val) {
            arr->push_back(item_stmt->execute(ctx));
@@ -423,7 +422,7 @@ struct tuple_literal : public expression {
        for (const auto& item : this->val) chk_type<expression>(item);
    }
    std::string type() const override { return "TupleLiteral"; }
-    value execute_impl(context & ctx) const override {
+    value execute_impl(context & ctx) override {
        auto arr = mk_val<value_array>();
        for (const auto & item_stmt : val) {
            arr->push_back(item_stmt->execute(ctx));
@@ -442,7 +441,7 @@ struct object_literal : public expression {
        }
    }
    std::string type() const override { return "ObjectLiteral"; }
-    value execute_impl(context & ctx) const override;
+    value execute_impl(context & ctx) override;
 };

 // Complex Expressions
@@ -463,8 +462,8 @@ struct binary_expression : public expression {
        chk_type<expression>(this->right);
    }
    std::string type() const override { return "BinaryExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"left", {left.get()}},
            {"right", {right.get()}}
@@ -477,7 +476,10 @@ struct binary_expression : public expression {
 * Operator precedence: https://github.com/pallets/jinja/issues/379#issuecomment-168076202
 */
 struct filter_expression : public expression {
+    // either an expression or a value is allowed
    statement_ptr operand;
+    value_string val; // will be set by filter_statement
+
    statement_ptr filter;

    filter_expression(statement_ptr && operand, statement_ptr && filter)
@@ -486,9 +488,14 @@ struct filter_expression : public expression {
        chk_type<identifier, call_expression>(this->filter);
    }

+    filter_expression(value_string && val, statement_ptr && filter)
+        : val(std::move(val)), filter(std::move(filter)) {
+        chk_type<identifier, call_expression>(this->filter);
+    }
+
    std::string type() const override { return "FilterExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"operand", {operand.get()}},
            {"filter", {filter.get()}}
@@ -505,8 +512,8 @@ struct filter_statement : public statement {
        chk_type<identifier, call_expression>(this->filter);
    }
    std::string type() const override { return "FilterStatement"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"filter", {filter.get()}},
            {"body", stmts_to_ptr(body)}
@@ -530,14 +537,14 @@ struct select_expression : public expression {
        chk_type<expression>(this->test);
    }
    std::string type() const override { return "SelectExpression"; }
-    value execute_impl(context & ctx) const override {
+    value execute_impl(context & ctx) override {
        auto predicate = test->execute_impl(ctx);
        if (!predicate->as_bool()) {
            return mk_val<value_undefined>();
        }
        return lhs->execute_impl(ctx);
    }
-    void visit(context & ctx) const override {
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"lhs", {lhs.get()}},
            {"test", {test.get()}}
@@ -560,8 +567,8 @@ struct test_expression : public expression {
        chk_type<identifier, call_expression>(this->test);
    }
    std::string type() const override { return "TestExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"operand", {operand.get()}},
            {"test", {test.get()}}
@@ -581,8 +588,8 @@ struct unary_expression : public expression {
        chk_type<expression>(this->argument);
    }
    std::string type() const override { return "UnaryExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"argument", {argument.get()}}
        });
@@ -601,10 +608,10 @@ struct slice_expression : public expression {
        chk_type<expression>(this->step_expr);
    }
    std::string type() const override { return "SliceExpression"; }
-    [[noreturn]] value execute_impl(context &) const override {
+    [[noreturn]] value execute_impl(context &) override {
        throw std::runtime_error("must be handled by MemberExpression");
    }
-    void visit(context & ctx) const override {
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"start_expr", {start_expr.get()}},
            {"stop_expr", {stop_expr.get()}},
@@ -623,8 +630,8 @@ struct keyword_argument_expression : public expression {
        chk_type<expression>(this->val);
    }
    std::string type() const override { return "KeywordArgumentExpression"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"key", {key.get()}},
            {"val", {val.get()}}
@@ -638,7 +645,7 @@ struct spread_expression : public expression {
        chk_type<expression>(this->argument);
    }
    std::string type() const override { return "SpreadExpression"; }
-    void visit(context & ctx) const override {
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"argument", {argument.get()}}
        });
@@ -656,8 +663,8 @@ struct call_statement : public statement {
        for (const auto & arg : this->caller_args) chk_type<expression>(arg);
    }
    std::string type() const override { return "CallStatement"; }
-    value execute_impl(context & ctx) const override;
-    void visit(context & ctx) const override {
+    value execute_impl(context & ctx) override;
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"call", {call.get()}},
            {"caller_args", stmts_to_ptr(caller_args)},
@@ -678,7 +685,7 @@ struct ternary_expression : public expression {
        chk_type<expression>(this->false_expr);
    }
    std::string type() const override { return "Ternary"; }
-    value execute_impl(context & ctx) const override {
+    value execute_impl(context & ctx) override {
        value cond_val = condition->execute(ctx);
        if (cond_val->as_bool()) {
            return true_expr->execute(ctx);
@@ -686,7 +693,7 @@ struct ternary_expression : public expression {
            return false_expr->execute(ctx);
        }
    }
-    void visit(context & ctx) const override {
+    void visit(context & ctx) override {
        ctx.visitor(false, this, {
            {"condition", {condition.get()}},
            {"true_expr", {true_expr.get()}},
@@ -149,13 +149,6 @@ static value test_type_fn(const func_args & args) {
    JJ_DEBUG("test_type_fn: type=%s, %s or %s result=%d", typeid(T).name(), typeid(U).name(), typeid(V).name(), is_type ? 1 : 0);
    return mk_val<value_bool>(is_type);
 }
-template<typename T, typename U, typename V, typename W>
-static value test_type_fn(const func_args & args) {
-    args.ensure_count(1);
-    bool is_type = is_val<T>(args.get_pos(0)) || is_val<U>(args.get_pos(0)) || is_val<V>(args.get_pos(0)) || is_val<W>(args.get_pos(0));
-    JJ_DEBUG("test_type_fn: type=%s, %s, %s or %s result=%d", typeid(T).name(), typeid(U).name(), typeid(V).name(), typeid(W).name(), is_type ? 1 : 0);
-    return mk_val<value_bool>(is_type);
-}
 template<value_compare_op op>
 static value test_compare_fn(const func_args & args) {
    args.ensure_count(2, 2);
@@ -268,30 +261,6 @@ static value tojson(const func_args & args) {
    return mk_val<value_string>(json_str);
 }

-static value & get_attribute(const value & val, const value & attr, value & default_val) {
-    if (!attr->is_undefined()) {
-        if (is_val<value_array>(val)) {
-            value idx = attr;
-
-            if (is_val<value_string>(attr)) {
-                const std::string s = attr->as_string().str();
-                if (!s.empty() && std::all_of(s.begin(), s.end(), [](unsigned char c) { return std::isdigit(c); })) {
-                    try {
-                        idx = mk_val<value_int>(std::stoll(s));
-                    } catch (...) {
-                        idx = mk_val<value_undefined>();
-                    }
-                }
-            }
-
-            return val->at(idx, default_val);
-        } else if (is_val<value_object>(val)) {
-            return val->at(attr, default_val);
-        }
-    }
-    return default_val;
-}
-
 template<bool is_reject>
 static value selectattr(const func_args & args) {
    args.ensure_count(2, 4);
@@ -305,7 +274,10 @@ static value selectattr(const func_args & args) {
    if (args.count() == 2) {
        // example: array | selectattr("active")
        for (const auto & item : arr) {
-            value attr_val = get_attribute(item, attribute, val_default);
+            if (!is_val<value_object>(item)) {
+                throw raised_exception("selectattr: item is not an object");
+            }
+            value attr_val = item->at(attribute, val_default);
            bool is_selected = attr_val->as_bool();
            if constexpr (is_reject) is_selected = !is_selected;
            if (is_selected) out->push_back(item);
@@ -346,7 +318,10 @@ static value selectattr(const func_args & args) {
        }
        auto test_fn = it->second;
        for (const auto & item : arr) {
-            value attr_val = get_attribute(item, attribute, val_default);
+            if (!is_val<value_object>(item)) {
+                throw raised_exception("selectattr: item is not an object");
+            }
+            value attr_val = item->at(attribute, val_default);
            func_args test_args(args.ctx);
            test_args.push_back(attr_val); // attribute value
            test_args.push_back(extra_arg); // extra argument
@@ -373,43 +348,6 @@ static value default_value(const func_args & args) {
    return no_value ? args.get_pos(1) : args.get_pos(0);
 }

-static value toobject(const func_args & args) {
-    auto out = mk_val<value_object>();
-    value iter = args.get_pos(0, mk_val<value_undefined>());
-    bool iter_first = false;
-    if (is_val<value_array>(iter)) {
-        iter_first = true;
-        for (const auto & it : iter->as_array()) {
-            if (is_val<value_array>(it) && it->as_array().size() == 2) {
-                auto tuple = it->as_array();
-                auto key = tuple[0];
-                auto val = tuple[1];
-                JJ_DEBUG("namespace/dict: adding key '%s'", key->as_string().str().c_str());
-                out->insert(key, val);
-            } else {
-                throw raised_exception("namespace/dict() iterable argument must consist of tuples, not " + it->type());
-            }
-        }
-    } else if (is_val<value_object>(iter)) {
-        iter_first = true;
-        for (const auto & pair : iter->as_ordered_object()) {
-            JJ_DEBUG("namespace/dict: adding key '%s'", pair.first->as_string().str().c_str());
-            out->insert(pair.first, pair.second);
-        }
-    }
-    for (const auto & arg : args.get_args()) {
-        if (is_val<value_kwarg>(arg)) {
-            auto kwarg = cast_val<value_kwarg>(arg);
-            JJ_DEBUG("namespace/dict: adding key '%s'", kwarg->key.c_str());
-            out->insert(kwarg->key, kwarg->val);
-        } else if (!iter_first) {
-            throw raised_exception("namespace/dict() arguments must be kwargs, dict and/or iterable of tuples, not " + arg->type());
-        }
-        iter_first = false;
-    }
-    return out;
-}
-
 const func_builtins & global_builtins() {
    static const func_builtins builtins = {
        {"raise_exception", [](const func_args & args) -> value {
@@ -417,8 +355,18 @@ const func_builtins & global_builtins() {
            std::string msg = args.get_pos(0)->as_string().str();
            throw raised_exception("Jinja Exception: " + msg);
        }},
-        {"dict", toobject},
-        {"namespace", toobject},
+        {"namespace", [](const func_args & args) -> value {
+            auto out = mk_val<value_object>();
+            for (const auto & arg : args.get_args()) {
+                if (!is_val<value_kwarg>(arg)) {
+                    throw raised_exception("namespace() arguments must be kwargs");
+                }
+                auto kwarg = cast_val<value_kwarg>(arg);
+                JJ_DEBUG("namespace: adding key '%s'", kwarg->key.c_str());
+                out->insert(kwarg->key, kwarg->val);
+            }
+            return out;
+        }},
        {"strftime_now", [](const func_args & args) -> value {
            args.ensure_vals<value_string>();
            std::string format = args.get_pos(0)->as_string().str();
@@ -503,8 +451,8 @@ const func_builtins & global_builtins() {
        {"test_is_integer", test_type_fn<value_int>},
        {"test_is_float", test_type_fn<value_float>},
        {"test_is_number", test_type_fn<value_int, value_float>},
-        {"test_is_iterable", test_type_fn<value_object, value_array, value_string, value_undefined>},
-        {"test_is_sequence", test_type_fn<value_object, value_array, value_string, value_undefined>},
+        {"test_is_iterable", test_type_fn<value_array, value_string, value_undefined>},
+        {"test_is_sequence", test_type_fn<value_array, value_string, value_undefined>},
        {"test_is_mapping", test_type_fn<value_object>},
        {"test_is_lower", [](const func_args & args) -> value {
            args.ensure_vals<value_string>();
@@ -567,28 +515,8 @@ const func_builtins & global_builtins() {
        }},
        {"test_is_sameas", [](const func_args & args) -> value {
            // Check if an object points to the same memory address as another object
-            args.ensure_count(2);
-            auto a = args.get_pos(0);
-            auto b = args.get_pos(1);
-            bool res = false;
-            if (!is_val<value_undefined>(a) && !is_val<value_undefined>(b)) {
-                if (is_val<value_none>(a) && is_val<value_none>(b)) {
-                    res = true;
-                } else if (is_val<value_bool>(a) && is_val<value_bool>(b)) {
-                    if (a->as_bool() == b->as_bool()) {
-                        res = true;
-                    }
-                } else if (is_val<value_int>(a) && is_val<value_int>(b)) {
-                    const int64_t x = a->as_int();
-                    // Allow comparison within small-int cache range
-                    if (x >= -5 && x <= 256 && x == b->as_int()) {
-                        res = true;
-                    }
-                } else if (a == b) {
-                    res = true;
-                }
-            }
-            return mk_val<value_bool>(res);
+            (void)args;
+            throw not_implemented_exception("sameas test not implemented");
        }},
        {"test_is_escaped", [](const func_args & args) -> value {
            (void)args;
@@ -1093,14 +1021,22 @@ const func_builtins & value_array_t::get_builtins() const {
            }
            value val_delim = args.get_kwarg_or_pos("d",         1);
            value attribute = args.get_kwarg_or_pos("attribute", 2);
-            value undef = mk_val<value_undefined>();
            const auto & arr = args.get_pos(0)->as_array();
+            const bool attr_is_int = is_val<value_int>(attribute);
+            if (!attribute->is_undefined() && !is_val<value_string>(attribute) && !attr_is_int) {
+                throw raised_exception("join() attribute must be string or integer");
+            }
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
            const std::string delim = val_delim->is_undefined() ? "" : val_delim->as_string().str();
            std::string result;
            for (size_t i = 0; i < arr.size(); ++i) {
                value val_arr = arr[i];
                if (!attribute->is_undefined()) {
-                    val_arr = get_attribute(val_arr, attribute, undef);
+                    if (attr_is_int && is_val<value_array>(val_arr)) {
+                        val_arr = val_arr->at(attr_int);
+                    } else if (!attr_is_int && is_val<value_object>(val_arr)) {
+                        val_arr = val_arr->at(attribute);
+                    }
                }
                if (!is_val<value_string>(val_arr) && !is_val<value_int>(val_arr) && !is_val<value_float>(val_arr)) {
                    throw raised_exception("join() can only join arrays of strings or numerics");
@@ -1132,11 +1068,21 @@ const func_builtins & value_array_t::get_builtins() const {
            }
            value val       = args.get_pos(0);
            value attribute = args.get_kwarg_or_pos("attribute", 1);
+            const bool attr_is_int = is_val<value_int>(attribute);
+            if (!is_val<value_string>(attribute) && !attr_is_int) {
+                throw raised_exception("map: attribute must be string or integer");
+            }
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
            value default_val = args.get_kwarg("default", mk_val<value_undefined>());
            auto out = mk_val<value_array>();
            auto arr = val->as_array();
            for (const auto & item : arr) {
-                value attr_val = get_attribute(item, attribute, default_val);
+                value attr_val;
+                if (attr_is_int) {
+                    attr_val = is_val<value_array>(item) ? item->at(attr_int, default_val) : default_val;
+                } else {
+                    attr_val = is_val<value_object>(item) ? item->at(attribute, default_val) : default_val;
+                }
                out->push_back(attr_val);
            }
            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(out->as_array())) : out;
@@ -1173,14 +1119,22 @@ const func_builtins & value_array_t::get_builtins() const {
            // FIXME: sorting is currently always case sensitive
            //const bool case_sensitive = val_case->as_bool(); // undefined == false
            const bool reverse = val_reverse->as_bool(); // undefined == false
-            value undef = mk_val<value_undefined>();
+            const bool attr_is_int = is_val<value_int>(attribute);
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
            std::vector<value> arr = val->as_array(); // copy
            std::sort(arr.begin(), arr.end(),[&](const value & a, const value & b) {
                value val_a = a;
                value val_b = b;
                if (!attribute->is_undefined()) {
-                    val_a = get_attribute(a, attribute, undef);
-                    val_b = get_attribute(b, attribute, undef);
+                    if (attr_is_int && is_val<value_array>(a) && is_val<value_array>(b)) {
+                        val_a = a->at(attr_int);
+                        val_b = b->at(attr_int);
+                    } else if (!attr_is_int && is_val<value_object>(a) && is_val<value_object>(b)) {
+                        val_a = a->at(attribute);
+                        val_b = b->at(attribute);
+                    } else {
+                        throw raised_exception("sort: unsupported object attribute comparison between " + a->type() + " and " + b->type());
+                    }
                }
                return value_compare(val_a, val_b, reverse ? value_compare_op::gt : value_compare_op::lt);
            });
@@ -1198,23 +1152,19 @@ const func_builtins & value_array_t::get_builtins() const {
            args.ensure_vals<value_array>();
            value val_case    = args.get_kwarg_or_pos("case_sensitive", 1);
            value attribute   = args.get_kwarg_or_pos("attribute",      2);
+            if (!attribute->is_undefined()) {
+                throw not_implemented_exception("min: attribute not implemented");
+            }
            // FIXME: min is currently always case sensitive
            (void) val_case;
-            value undef = mk_val<value_undefined>();
            const auto & arr = args.get_pos(0)->as_array();
            if (arr.empty()) {
-                return undef;
+                return mk_val<value_undefined>();
            }
            value result = arr[0];
-            for (const auto & item : arr) {
-                value val_arr = item;
-                value val_cmp = result;
-                if (!attribute->is_undefined()) {
-                    val_arr = get_attribute(val_arr, attribute, undef);
-                    val_cmp = get_attribute(val_cmp, attribute, undef);
-                }
-                if (value_compare(val_arr, val_cmp, value_compare_op::lt)) {
-                    result = item;
+            for (size_t i = 1; i < arr.size(); ++i) {
+                if (value_compare(arr[i], result, value_compare_op::lt)) {
+                    result = arr[i];
                }
            }
            return result;
@@ -1224,23 +1174,19 @@ const func_builtins & value_array_t::get_builtins() const {
            args.ensure_vals<value_array>();
            value val_case    = args.get_kwarg_or_pos("case_sensitive", 1);
            value attribute   = args.get_kwarg_or_pos("attribute",      2);
+            if (!attribute->is_undefined()) {
+                throw not_implemented_exception("max: attribute not implemented");
+            }
            // FIXME: max is currently always case sensitive
            (void) val_case;
-            value undef = mk_val<value_undefined>();
            const auto & arr = args.get_pos(0)->as_array();
            if (arr.empty()) {
-                return undef;
+                return mk_val<value_undefined>();
            }
            value result = arr[0];
-            for (const auto & item : arr) {
-                value val_arr = item;
-                value val_cmp = result;
-                if (!attribute->is_undefined()) {
-                    val_arr = get_attribute(val_arr, attribute, undef);
-                    val_cmp = get_attribute(val_cmp, attribute, undef);
-                }
-                if (value_compare(val_arr, val_cmp, value_compare_op::gt)) {
-                    result = item;
+            for (size_t i = 1; i < arr.size(); ++i) {
+                if (value_compare(arr[i], result, value_compare_op::gt)) {
+                    result = arr[i];
                }
            }
            return result;
@@ -433,12 +433,6 @@ struct value_array_t : public value_t {
        }
        return val_arr[index];
    }
-    virtual value & at(const value & index, value & default_val) override {
-        if (!is_val<value_int>(index) && !is_val<value_bool>(index)) {
-            return default_val;
-        }
-        return at(index->as_int(), default_val);
-    }
    virtual const func_builtins & get_builtins() const override;
    virtual bool is_hashable() const override {
        if (std::all_of(val_arr.begin(), val_arr.end(), [&](auto & val) -> bool {
@@ -321,8 +321,7 @@ static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
        case 'x': n_hex = 2; break;
        case 'u': n_hex = 4; break;
        case 'U': n_hex = 8; break;
-        // keep in sync with parse_char() in src/llama-grammar.cpp
-        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']': case '-':
+        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
            return 2;
        default:
            return 0;
@@ -82,9 +82,6 @@ struct common_json_value {
    // 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_enum<T>::value, int>::type = 0>
-    common_json_value(T val) : common_json_value((typename std::underlying_type<T>::type) val) {}
-
    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) {
@@ -114,7 +111,6 @@ struct common_json_item {
 // 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_enum<T>::value ||
    std::is_arithmetic<T>::value ||
    std::is_same<T, std::nullptr_t>::value ||
    std::is_same<T, std::string>::value ||
@@ -104,12 +104,6 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
    const std::string GEN_PROMPT   = "<｜Assistant｜>";
    const std::string TC_SEPARATOR = "\n\n";

-    // lets the server find user turns in the prompt and place context checkpoints there
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PROMPT },
-        { COMMON_CHAT_ROLE_USER, "<｜User｜>" },
-    };
-
    data.prompt = common_chat_template_direct_apply_impl(
        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
    data.generation_prompt = common_chat_template_generation_prompt_impl(
@@ -272,10 +272,6 @@ common_chat_params common_chat_params_init_gemma4(const common_chat_template &
                /* max = */ inputs.parallel_tool_calls ? -1 : 1
            ));

-            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return start + thought + tool_call;
-            }
-
            auto scan_to_toolcall = p.rule("scan-to-toolcall", p.until("<|tool_call>"));
            auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>", "<|tool_call>"})));
            auto message = p.rule("message", thought + content);
@@ -1,194 +0,0 @@
-#include "parsers.h"
-
-// Ling 3.0 / Bailing V3 - <role>X</role> sections with tagged tool calls:
-//   assistant := [<think> ... </think>] [content] {<tool_call>name
-//                  <arg_key>k</arg_key>\n<arg_value>v</arg_value> ...</tool_call>}
-// The generation prompt ends with "<role>ASSISTANT</role>\n<think>", so the model
-// never emits the opening think tag, and a tool call can arrive before any
-// </think>. Reasoning therefore terminates at the think close tag or at a tool
-// call start, like the Qwen3-Coder and Kimi K3 parsers. With thinking off the
-// template pre-closes the think block instead, and the model emits bare content.
-common_chat_params common_chat_params_init_ling3(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 ROLE        = "<role>ASSISTANT</role>";
-    const std::string THINK_START = "<think>";
-    const std::string THINK_END   = "</think>";
-    const std::string CALL_START  = "<tool_call>";
-    const std::string CALL_END    = "</tool_call>";
-    const std::string ARG_KEY     = "<arg_key>";
-    const std::string ARG_KEY_END = "</arg_key>";
-    const std::string ARG_VAL     = "<arg_value>";
-    const std::string ROLE_END    = "<|role_end|>";
-    const std::string ARG_VAL_END = "</arg_value>";
-
-    data.preserved_tokens = {
-        THINK_START, THINK_END, CALL_START, CALL_END,
-        ARG_KEY, ARG_KEY_END, ARG_VAL, ARG_VAL_END, ROLE_END,
-    };
-
-    data.thinking_start_tag = THINK_START;
-    // Support both </think> and <tool_call> as reasoning end sequences: a call
-    // can be emitted before the think block is closed.
-    data.thinking_end_tags  = { THINK_END, CALL_START };
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<role>ASSISTANT</role>" },
-        { COMMON_CHAT_ROLE_USER,      "<role>HUMAN</role>"     },
-        { COMMON_CHAT_ROLE_TOOL,      "<role>OBSERVATION</role>" },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<role>SYSTEM</role>"    },
-    };
-
-    // the model may spell the end-of-turn control token out as text tokens,
-    // which does not stop generation; a literal stop string catches it either
-    // way (as the Laguna patch does for its </assistant> token)
-    data.additional_stops = { ROLE_END };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = ROLE + "\n" + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    // The generation prompt pre-opens the think block when thinking is on, so
-    // the opening tag is optional here and reasoning runs until </think> or a
-    // tool call start; with thinking off the template pre-closes the block and
-    // everything the model emits is content.
-    bool think_open = false;
-    if (inputs.has_continuation()) {
-        think_open = inputs.continue_final_message != COMMON_CHAT_CONTINUATION_CONTENT;
-    } else {
-        auto last_open  = data.generation_prompt.rfind(THINK_START);
-        auto last_close = data.generation_prompt.rfind(THINK_END);
-        think_open      = last_open != std::string::npos &&
-                     (last_close == std::string::npos || last_open > last_close);
-    }
-
-    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;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto end = p.end();
-
-        // the effective parse input is generation_prompt + model output, so the
-        // assistant opener is optionally consumed here
-        auto opener = p.optional(p.literal(ROLE) + p.optional(p.space()));
-
-        // the generation prompt pre-opens the think block, so the opening tag
-        // is optional; a missing close tag does not swallow a tool call
-        auto body_end   = think_open ? p.until_one_of({ THINK_END, CALL_START }) : p.until_one_of({ THINK_END });
-        auto think_body = extract_reasoning ? p.reasoning(body_end) : p.content(body_end);
-
-        auto reasoning = p.optional(p.optional(p.literal(THINK_START)) + think_body +
-                                    p.optional(p.literal(THINK_END)));
-
-        // content between the think block and the first tool call, plus any
-        // trailing text after the last tool call, are plain content
-        auto content = p.optional(p.content(p.until_one_of({ CALL_START })));
-
-        // a trailing end-of-turn token is consumed instead of leaking into content
-        auto tail = p.optional(p.content(p.until(ROLE_END))) + p.optional(p.literal(ROLE_END));
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return opener + reasoning + tail + end;
-        }
-
-        auto tool_choices = p.choice();
-        auto arg_close    = p.tool_arg_close(p.literal(ARG_VAL_END));
-        auto arg_string   = p.rule("ling3-arg-string",
-                                   p.tool_arg_string_value(p.until(ARG_VAL_END)) + arg_close);
-
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-
-            std::vector<common_peg_parser> required_args;
-            std::vector<common_peg_parser> optional_args;
-
-            // each argument may be preceded by whitespace: the model emits
-            // newlines between arguments, the template history does not
-            foreach_parameter(function, [&](const common_chat_schema_property & param, const common_chat_schema_document_ptr & doc) {
-                auto rule_name = "ling3-arg-" + name + "-" + param.name;
-
-                auto types = param.schema->value_types();
-
-                // string arguments are raw text up to the closing tag, other
-                // types parse as JSON per their schema; each alternative
-                // consumes the closing tag itself so a JSON prefix can not
-                // commit the choice before the tag matches
-                auto arg_value = p.eps();
-                if (!types.has(common_chat_schema::TYPE_STRING)) {
-                    arg_value = p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", doc, *param.schema)) + arg_close;
-                } else if (types.is_only(common_chat_schema::TYPE_STRING)) {
-                    arg_value = arg_string;
-                } else {
-                    // the parser tries the JSON alternative first to type the value
-                    arg_value = p.gbnf(p.atomic(p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", doc, *param.schema)) + arg_close) | arg_string,
-                                       "ling3-arg-string");
-                }
-
-                auto arg = p.rule(rule_name,
-                    p.optional(p.space()) +
-                    p.tool_arg(p.tool_arg_open(p.literal(ARG_KEY) + p.tool_arg_name(p.literal(param.name)) +
-                                               p.literal(ARG_KEY_END)) +
-                               p.optional(p.space()) + p.literal(ARG_VAL) +
-                               arg_value));
-
-                (param.required ? required_args : optional_args).push_back(arg);
-            });
-
-            // required arguments in any order (as Qwen3-Coder does), then
-            // optional ones in any order and number
-            auto args = p.permute("ling3-" + name + "-args", required_args);
-            if (!optional_args.empty()) {
-                args = args + p.zero_or_more(p.choice(optional_args));
-            }
-
-            auto call = p.tool(p.tool_open(p.literal(CALL_START) + p.tool_name(p.literal(name)) +
-                                           p.optional(p.space())) +
-                               p.tool_args(args) +
-                               p.tool_close(p.optional(p.space()) + p.literal(CALL_END)));
-
-            tool_choices |= p.rule("ling3-tool-" + name, call);
-        });
-
-        auto calls = inputs.parallel_tool_calls ?
-                     tool_choices + p.zero_or_more(p.space() + tool_choices) :
-                     tool_choices;
-
-        auto tools_section = p.trigger_rule("ling3-tool-call", calls + p.space() +
-                                        p.optional(p.content(p.until(ROLE_END))) + p.optional(p.literal(ROLE_END)));
-
-        auto tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? tools_section :
-                                                                             p.optional(tools_section);
-
-        return opener + reasoning + content + tools + tail + 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) {
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, CALL_START },
-        };
-    }
-
-    return data;
-}
@@ -1,164 +0,0 @@
-#include "parsers.h"
-
-// LLM-jp-4.1: the GPT-OSS (Harmony) format with two differences
-//  - the tokenizer emits a space after every special token: "<|channel|> analysis<|message|> ..."
-//  - parallel tool calls are consecutive assistant messages, all but the last closed by <|end|>
-common_chat_params common_chat_params_init_llm_jp_harmony(const common_chat_template &          tmpl,
-                                                          const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    // Copy reasoning to the "thinking" field as expected by the template
-    auto adjusted_messages = json::array();
-    for (auto msg : inputs.messages) {
-        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
-            msg["thinking"] = msg.at("reasoning_content");
-            if (msg.contains("tool_calls") && msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty()) {
-                msg.erase("content");
-            }
-        }
-        adjusted_messages.push_back(msg);
-    }
-
-    auto prompt = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-
-    // Check if we need to replace the return token with end token during
-    // inference and without generation prompt. For more details see:
-    // https://github.com/ggml-org/llama.cpp/issues/15417
-    if (inputs.is_inference && !inputs.add_generation_prompt) {
-        static constexpr std::string_view return_token = "<|return|>";
-        static constexpr std::string_view end_token    = "<|end|>";
-        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
-            prompt.replace(pos, return_token.length(), end_token);
-        }
-    }
-
-    data.prompt            = prompt;
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
-        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>developer" },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
-        { COMMON_CHAT_ROLE_TOOL,      "<|start|>functions" },
-    };
-
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    data.thinking_start_tag = "<|channel|>analysis<|message|>";
-    data.thinking_end_tags  = {"<|end|>"};
-
-    // These special tokens are required to parse properly, so we include them
-    // even if parse_tool_calls is false.
-    data.preserved_tokens = {
-        "<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
-    };
-
-    // Adjust prompt for continuation
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|start|>assistant<|channel|>analysis<|message|>" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<|end|><|start|>assistant<|channel|>final<|message|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        // tokenizer space after special tokens; not p.space() since GBNF `space` allows one space only
-        auto sp          = p.chars("[ ]", 0, -1);
-        auto channel_tag = p.literal("<|channel|>") + sp;
-        // one space only: keep an intentional leading space in the body
-        auto message     = p.literal("<|message|>") + p.optional(p.literal(" "));
-
-        auto start          = p.rule("start", p.literal("<|start|>") + sp + p.literal("assistant"));
-        auto end            = p.rule("end", p.literal("<|end|>"));
-        auto content        = p.rule("message-content", p.until("<|end|>"));
-        auto channel        = channel_tag + (p.literal("commentary") | p.literal("analysis"));
-        auto constrain_type = p.chars("[A-Za-z0-9_-]", 1, -1);
-        auto constraint     = p.optional(p.space() + p.optional(p.literal("<|constrain|>") + sp) + constrain_type);
-
-        auto start_analysis = channel_tag + p.literal("analysis") + message;
-        if (extract_reasoning) {
-            p.rule("analysis", start_analysis + p.reasoning(content) + end);
-        } else {
-            p.rule("analysis", p.content(start_analysis + content + end));
-        }
-
-        auto analysis  = p.ref("analysis");
-        auto preamble  = p.rule("preamble", channel_tag + p.literal("commentary") + message + p.content(content) + end);
-        auto final_msg = p.rule("final", channel_tag + p.literal("final") + message + p.content(content));
-
-        auto any = p.rule("any", preamble | analysis);
-
-        if (has_response_format) {
-            auto response_format = p.rule("response-format",
-                channel_tag + p.literal("final") + constraint + message +
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
-
-            return p.zero_or_more(start + analysis) + start + response_format;
-        }
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto tool_choice = p.choice();
-
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string  name     = function.at("name");
-                const auto   params   = common_chat_tool_parameters(function);
-
-                auto func_name = p.literal(" to=functions.") + p.tool_name(p.literal(name));
-                auto args      = p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", params));
-
-                // recipient in role header
-                //   <|start|>assistant to=functions.NAME<|channel|>(commentary|analysis)[constraint]<|message|>ARGS
-                auto tool_in_role = p.tool(p.tool_open(func_name + channel + constraint + message) + args);
-
-                // recipient in channel header
-                //   <|channel|>(commentary|analysis) to=functions.NAME[constraint]<|message|>ARGS
-                auto tool_in_channel = p.tool(p.tool_open(channel + func_name + constraint + message) + args);
-
-                tool_choice |= p.rule("tool-" + name, tool_in_role | tool_in_channel);
-            });
-
-            // parallel calls are separated by <|end|>; inside the trigger rule so the lazy grammar covers all of them
-            auto tool_calls = inputs.parallel_tool_calls
-                ? tool_choice + p.zero_or_more(end + start + tool_choice)
-                : tool_choice;
-            auto tool_call  = p.trigger_rule("tool-call", tool_calls);
-
-            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return p.zero_or_more(start + any) + start + tool_call;
-            }
-
-            return p.zero_or_more(start + any) + start + (tool_call | final_msg);
-        }
-
-        return p.zero_or_more(start + any) + start + final_msg;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^\\s+to$" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^<\\|channel\\|>\\s*(?:commentary|analysis)\\s+to=functions$" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>\\s*assistant(\\s+to)" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>\\s*assistant(<\\|channel\\|>\\s*(?:commentary|analysis)\\s+to)" }
-        };
-    }
-
-    return data;
-}
@@ -43,10 +43,9 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa

    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;

-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    // Constrained grammar whenever tools are offered or a response format is requested.
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_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"));
@@ -66,15 +65,6 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa
        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));

-        if (has_response_format) {
-            auto response_json   = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
-            auto response_format = p.rule("response-format",
-                recipient + p.literal("<|message|>") +
-                ((p.literal("```json") + p.space() + response_json + p.space() + p.literal("```")) | response_json));
-
-            return p.zero_or_more(start + analysis) + start + response_format;
-        }
-
        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>")),
@@ -134,13 +124,13 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa
    data.parser = parser.save();

    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
            parser.build_grammar(builder, data.grammar_lazy);
        });
        data.grammar_triggers = {
            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-              "(?:^|<\\|start\\|>assistant)( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
+              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
        };
    }

@@ -63,13 +63,9 @@ common_chat_params common_chat_params_init_kimi_k2(const common_chat_template &

 common_chat_params common_chat_params_init_kimi_k3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);

-common_chat_params common_chat_params_init_ling3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
 // tool_list_tokens preserves the LFM2 system tool-list markers; LFM2.5 renders without them
 common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const autoparser::generation_params & inputs, bool tool_list_tokens);

-common_chat_params common_chat_params_init_llm_jp_harmony(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
 common_chat_params common_chat_params_init_minicpm5(const common_chat_template & tmpl, const autoparser::generation_params & inputs);

 common_chat_params common_chat_params_init_minimax_m3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
@@ -11,9 +11,7 @@ set(LLAMA_CHAT_PARSERS_SOURCES
    ${CMAKE_CURRENT_LIST_DIR}/gpt-oss.cpp
    ${CMAKE_CURRENT_LIST_DIR}/kimi-k2.cpp
    ${CMAKE_CURRENT_LIST_DIR}/kimi-k3.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/ling3.cpp
    ${CMAKE_CURRENT_LIST_DIR}/lfm2.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/llm-jp-harmony.cpp
    ${CMAKE_CURRENT_LIST_DIR}/minicpm5.cpp
    ${CMAKE_CURRENT_LIST_DIR}/minimax-m3.cpp
    ${CMAKE_CURRENT_LIST_DIR}/ministral3.cpp
@@ -166,25 +166,6 @@ common_peg_ast_id common_peg_ast_arena::find_by_rule(const common_peg_ast_node &
    return COMMON_PEG_INVALID_AST_ID;
 }

-std::string common_peg_ast_node::sanitized_text() const {
-    if (invalid_utf8.empty()) {
-        return std::string(text);
-    }
-
-    std::string out;
-    out.reserve(text.size() + 2 * invalid_utf8.size());
-
-    size_t seg_start = start;
-    for (const auto & invalid : invalid_utf8) {
-        out.append(text.data() + (seg_start - start), invalid.pos - seg_start);
-        out.append("\xEF\xBF\xBD");
-        seg_start = invalid.pos + invalid.len;
-    }
-    out.append(text.data() + (seg_start - start), end - seg_start);
-
-    return out;
-}
-
 void common_peg_ast_arena::visit(common_peg_ast_id id, const common_peg_ast_visitor & visitor) const {
    if (id == COMMON_PEG_INVALID_AST_ID) {
        return;
@@ -301,7 +282,6 @@ struct parser_executor {

        auto pos = start_pos;
        std::vector<common_peg_ast_id> nodes;
-        std::vector<common_peg_invalid_utf8> invalid_utf8;

        for (size_t i = 0; i < p.children.size(); i++) {
            const auto & child_id = p.children[i];
@@ -326,14 +306,13 @@ struct parser_executor {
            if (!result.nodes.empty()) {
                nodes.insert(nodes.end(), result.nodes.begin(), result.nodes.end());
            }
-            invalid_utf8.insert(invalid_utf8.end(), result.invalid_utf8.begin(), result.invalid_utf8.end());

            if (result.need_more_input()) {
                ctx.parse_depth--;
                if (ctx.is_debug()) {
                    fprintf(stderr, "%sSEQ -> NEED_MORE\n", debug_indent().c_str());
                }
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes), std::move(invalid_utf8));
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes));
            }

            pos = result.end;
@@ -343,7 +322,7 @@ struct parser_executor {
        if (ctx.is_debug()) {
            fprintf(stderr, "%sSEQ -> SUCCESS at %zu->%zu\n", debug_indent().c_str(), start_pos, pos);
        }
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes), std::move(invalid_utf8));
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes));
    }

    common_peg_parse_result operator()(const common_peg_choice_parser & p) {
@@ -391,7 +370,6 @@ struct parser_executor {
        auto pos = start_pos;
        int match_count = 0;
        std::vector<common_peg_ast_id> nodes;
-        std::vector<common_peg_invalid_utf8> invalid_utf8;

        // Try to match up to max_count times (or unlimited if max_count is -1)
        while (p.max_count == -1 || match_count < p.max_count) {
@@ -422,7 +400,6 @@ struct parser_executor {
                if (!result.nodes.empty()) {
                    nodes.insert(nodes.end(), result.nodes.begin(), result.nodes.end());
                }
-                invalid_utf8.insert(invalid_utf8.end(), result.invalid_utf8.begin(), result.invalid_utf8.end());

                pos = result.end;
                match_count++;
@@ -433,14 +410,13 @@ struct parser_executor {
                if (!result.nodes.empty()) {
                    nodes.insert(nodes.end(), result.nodes.begin(), result.nodes.end());
                }
-                invalid_utf8.insert(invalid_utf8.end(), result.invalid_utf8.begin(), result.invalid_utf8.end());

                ctx.parse_depth--;
                if (ctx.is_debug()) {
                    fprintf(stderr, "%sREPEAT -> NEED_MORE (count=%d, nodes=%zu)\n", debug_indent().c_str(),
                            match_count, nodes.size());
                }
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes), std::move(invalid_utf8));
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes));
            }

            // Child failed - stop trying
@@ -458,7 +434,7 @@ struct parser_executor {
                    fprintf(stderr, "%sREPEAT -> NEED_MORE (not enough matches: %d < %d)\n", debug_indent().c_str(),
                            match_count, p.min_count);
                }
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos, std::move(nodes), std::move(invalid_utf8));
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos, std::move(nodes));
            }
            if (ctx.is_debug()) {
                fprintf(stderr, "%sREPEAT -> FAIL (not enough matches: %d < %d)\n", debug_indent().c_str(), match_count,
@@ -472,7 +448,7 @@ struct parser_executor {
            fprintf(stderr, "%sREPEAT -> SUCCESS (count=%d, nodes=%zu)\n", debug_indent().c_str(), match_count,
                    nodes.size());
        }
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes), std::move(invalid_utf8));
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes));
    }

    common_peg_parse_result operator()(const common_peg_and_parser & p) {
@@ -688,23 +664,23 @@ struct parser_executor {
        // Scan input and check for delimiters
        size_t pos = start_pos;
        size_t last_valid_pos = start_pos;
-        std::vector<common_peg_invalid_utf8> invalid_utf8;

        while (pos < ctx.input.size()) {
            auto utf8_result = common_parse_utf8_codepoint(ctx.input, pos);

-            if (utf8_result.status == utf8_parse_result::INCOMPLETE && ctx.is_lenient()) {
-                // The rest of the sequence may still arrive, return what we have so far
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, last_valid_pos, {}, std::move(invalid_utf8));
+            if (utf8_result.status == utf8_parse_result::INCOMPLETE) {
+                // Incomplete UTF-8 sequence
+                if (!ctx.is_lenient()) {
+                    // Input is complete but UTF-8 is incomplete = malformed
+                    return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
+                }
+                // Return what we have so far (before incomplete sequence)
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, last_valid_pos);
            }

-            if (utf8_result.status != utf8_parse_result::SUCCESS) {
-                // Malformed UTF-8, or a sequence truncated by the end of a complete input.
-                // A delimiter cannot start inside bytes that fail to decode, so consume them and move on
-                invalid_utf8.push_back({pos, utf8_result.bytes_consumed});
-                pos += utf8_result.bytes_consumed;
-                last_valid_pos = pos;
-                continue;
+            if (utf8_result.status == utf8_parse_result::INVALID) {
+                // Malformed UTF-8
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
            }

            // Check if a delimiter starts at this position
@@ -712,12 +688,12 @@ struct parser_executor {

            if (match == common_trie::COMPLETE_MATCH) {
                // Found a complete delimiter, return everything before it
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, {}, std::move(invalid_utf8));
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
            }

            if (match == common_trie::PARTIAL_MATCH) {
                // Found a partial match extending to end of input, return everything before it
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, {}, std::move(invalid_utf8));
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
            }

            pos += utf8_result.bytes_consumed;
@@ -726,9 +702,9 @@ struct parser_executor {

        if (last_valid_pos == ctx.input.size() && ctx.is_lenient()) {
            // Reached the end of a partial stream, there might still be more input that we need to consume.
-            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, last_valid_pos, {}, std::move(invalid_utf8));
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, last_valid_pos);
        }
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, last_valid_pos, {}, std::move(invalid_utf8));
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, last_valid_pos);
    }

    common_peg_parse_result operator()(const common_peg_schema_parser & p) {
@@ -752,11 +728,10 @@ struct parser_executor {
                result.end,
                text,
                std::move(result.nodes),
-                result.need_more_input(),
-                result.invalid_utf8
+                result.need_more_input()
            );

-            return common_peg_parse_result(result.type, result.start, result.end, { node_id }, std::move(result.invalid_utf8));
+            return common_peg_parse_result(result.type, result.start, result.end, { node_id });
        }

        return result;
@@ -782,11 +757,10 @@ struct parser_executor {
                result.end,
                text,
                std::move(result.nodes),
-                result.need_more_input(),
-                result.invalid_utf8
+                result.need_more_input()
            );

-            return common_peg_parse_result(result.type, result.start, result.end, { node_id }, std::move(result.invalid_utf8));
+            return common_peg_parse_result(result.type, result.start, result.end, { node_id });
        }

        return result;
@@ -72,12 +72,6 @@ enum common_peg_parse_result_type {

 const char * common_peg_parse_result_type_name(common_peg_parse_result_type type);

-// A run of input bytes that does not decode as UTF-8
-struct common_peg_invalid_utf8 {
-    size_t pos;
-    size_t len;
-};
-
 struct common_peg_ast_node {
    common_peg_ast_id id;
    std::string rule;
@@ -88,12 +82,6 @@ struct common_peg_ast_node {
    std::vector<common_peg_ast_id> children;

    bool is_partial = false;
-
-    // Invalid UTF-8 inside the node, in ascending order
-    std::vector<common_peg_invalid_utf8> invalid_utf8;
-
-    // Returns the text with every invalid run replaced by U+FFFD
-    std::string sanitized_text() const;
 };

 struct common_peg_parse_result;
@@ -110,11 +98,10 @@ class common_peg_ast_arena {
        size_t end,
        std::string_view text,
        std::vector<common_peg_ast_id> children,
-        bool is_partial = false,
-        std::vector<common_peg_invalid_utf8> invalid_utf8 = {}
+        bool is_partial = false
    ) {
        common_peg_ast_id id = nodes_.size();
-        nodes_.push_back({id, rule, tag, start, end, text, std::move(children), is_partial, std::move(invalid_utf8)});
+        nodes_.push_back({id, rule, tag, start, end, text, std::move(children), is_partial});
        return id;
    }

@@ -140,9 +127,6 @@ struct common_peg_parse_result {

    std::vector<common_peg_ast_id> nodes;

-    // Invalid UTF-8 consumed by this result, carried up to the enclosing AST nodes
-    std::vector<common_peg_invalid_utf8> invalid_utf8;
-
    common_peg_parse_result() = default;

    common_peg_parse_result(common_peg_parse_result_type type, size_t start)
@@ -151,8 +135,8 @@ struct common_peg_parse_result {
    common_peg_parse_result(common_peg_parse_result_type type, size_t start, size_t end)
        : type(type), start(start), end(end) {}

-    common_peg_parse_result(common_peg_parse_result_type type, size_t start, size_t end, std::vector<common_peg_ast_id> nodes, std::vector<common_peg_invalid_utf8> invalid_utf8 = {})
-        : type(type), start(start), end(end), nodes(std::move(nodes)), invalid_utf8(std::move(invalid_utf8)) {}
+    common_peg_parse_result(common_peg_parse_result_type type, size_t start, size_t end, std::vector<common_peg_ast_id> nodes)
+        : type(type), start(start), end(end), nodes(std::move(nodes)) {}

    bool fail() const { return type == COMMON_PEG_PARSE_RESULT_FAIL; }
    bool need_more_input() const { return type == COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT; }
@@ -446,7 +430,6 @@ class common_peg_parser_builder {
    common_peg_parser space() { return add(common_peg_space_parser{}); }

    // Matches all characters until a delimiter is found (delimiter not consumed).
-    // Invalid UTF-8 is consumed and recorded on the AST nodes.
    //   S -> (!delim .)*
    common_peg_parser until(const std::string & delimiter) { return add(common_peg_until_parser{{delimiter}}); }

@@ -167,16 +167,16 @@ void common_preset::apply_to_params(common_params & params, const std::set<std::
    }
 }

-static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_file(const std::filesystem::path & path) {
+static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_file(const std::string & path) {
    std::map<std::string, std::map<std::string, std::string>> parsed;

    if (!std::filesystem::exists(path)) {
-        throw std::runtime_error("preset file does not exist: " + fs_path_to_utf8(path));
+        throw std::runtime_error("preset file does not exist: " + path);
    }

    std::ifstream file(path);
    if (!file.good()) {
-        throw std::runtime_error("failed to open server preset file: " + fs_path_to_utf8(path));
+        throw std::runtime_error("failed to open server preset file: " + path);
    }

    std::string contents((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
@@ -225,7 +225,7 @@ static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_
    common_peg_parse_context ctx(contents);
    const auto result = parser.parse(ctx);
    if (!result.success()) {
-        throw std::runtime_error("failed to parse server config file: " + fs_path_to_utf8(path));
+        throw std::runtime_error("failed to parse server config file: " + path);
    }

    std::string current_section = COMMON_PRESET_DEFAULT_NAME;
@@ -282,7 +282,7 @@ common_preset_context::common_preset_context(llama_example ex)
    key_to_opt = get_map_key_opt(ctx_params);
 }

-common_presets common_preset_context::load_from_ini(const std::filesystem::path & path, common_preset & global) const {
+common_presets common_preset_context::load_from_ini(const std::string & path, common_preset & global) const {
    common_presets out;
    auto ini_data = parse_ini_from_file(path);

@@ -323,7 +323,7 @@ common_presets common_preset_context::load_from_ini(const std::filesystem::path
                }
                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(), fs_path_to_utf8(path).c_str());
+                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'",
@@ -67,7 +67,7 @@ struct common_preset_context {
    common_preset_context(llama_example ex);

    // load presets from INI file
-    common_presets load_from_ini(const std::filesystem::path & path, common_preset & global) const;
+    common_presets load_from_ini(const std::string & path, common_preset & global) const;

    // generate presets from cached models
    common_presets load_from_cache() const;
@@ -214,7 +214,7 @@ struct common_sampler * common_sampler_init(
 #ifdef LLAMA_USE_LLGUIDANCE
        grmr = llama_sampler_init_llg(vocab, "lark", grammar_str.c_str());
 #else
-        throw std::runtime_error("failed to parse grammar: llguidance is not enabled");
+        GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
    } else {
        std::vector<std::string> trigger_patterns;
@@ -681,8 +681,6 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
    std::vector<llama_token> result;
    result.reserve(idxs.size());

-    const llama_vocab * vocab = llama_model_get_vocab(llama_get_model(ctx));
-
    size_t i = 0;
    for (; i < draft.size(); i++) {
        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
@@ -691,9 +689,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample

        result.push_back(id);

-        // do not accept draft tokens after an EOG - they are not output but would stay in the context
-        // on replay the last token is from the target and can be EOG, so a trailing EOG is still accepted
-        if (draft[i] != id || (llama_vocab_is_eog(vocab, id) && i + 1 < draft.size())) {
+        if (draft[i] != id) {
            break;
        }
    }
@@ -165,7 +165,7 @@ struct common_speculative_impl {

    virtual void begin(llama_seq_id seq_id, const llama_tokens & prompt) = 0;

-    virtual bool process(const common_batch & batch) = 0;
+    virtual bool process(const llama_batch & batch) = 0;

    virtual void draft(common_speculative_draft_params_vec & dparams) = 0;

@@ -179,11 +179,7 @@ struct common_speculative_impl {
 struct common_speculative_impl_draft_simple : public common_speculative_impl {
    common_params_speculative_draft params;

-    common_batch batch;
-
-    // zero row at the draft input width, stands in for target embeddings the draft cannot read
-    std::vector<float> zeros;
-    bool zeros_warned = false; // the substitution is reported once
+    llama_batch batch;

    std::vector<common_sampler_ptr> smpls;

@@ -198,8 +194,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
            throw std::runtime_error("draft-simple requires a draft context");
        }

-        zeros.assign(llama_model_n_embd_inp(llama_get_model(ctx_dft)), 0.0f);
-
        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",
@@ -210,7 +204,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
                ctx_dft ? "yes" : "no",
                common_speculative_get_devices_str(this->params.devices).c_str());

-        batch = common_batch(ctx_dft);
+        batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);

        // TODO: optimize or pass from outside?
        // {
@@ -234,7 +228,9 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
            common_params_sampling params;
            params.no_perf = false;
            params.top_k = 10;
-            params.samplers.assign(1, COMMON_SAMPLER_TYPE_TOP_K);
+            params.samplers = {
+                COMMON_SAMPLER_TYPE_TOP_K,
+            };

            smpl.reset(common_sampler_init(llama_get_model(ctx_dft), params));
        }
@@ -255,46 +251,21 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        }
    }

+    ~common_speculative_impl_draft_simple() override {
+        llama_batch_free(batch);
+    }
+
    void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
        // noop
    }

-    bool process(const common_batch & batch_in) override {
+    bool process(const llama_batch & batch) override {
        auto * ctx_dft = params.ctx_dft;

-        // copy the entries to a batch owned by the draft context, only the last token is output
-        batch.clear();
-        const int32_t n_tokens = batch_in.size();
-        for (int32_t k = 0; k < n_tokens; ++k) {
-            const auto & t = batch_in.tokens[k];
-            const bool output = k == n_tokens - 1;
-            if (t.id != LLAMA_TOKEN_NULL) {
-                const int32_t idx = batch.add(t.id, t.pos[0], t.seq_id, output);
-                if (t.embd.data) {
-                    batch.set_embd(idx, t.embd);
-                }
-            } else {
-                // mtmd input is projected by the target encoder, a draft with a different width cannot read it
-                // it gets zeros instead, keeping its positions contiguous
-                // ref: https://github.com/ggml-org/llama.cpp/pull/29385#discussion_r4124743243
-                const size_t n_embd = t.embd.n_rows * t.embd.n_embd;
-                const bool same_width = n_embd == zeros.size();
-                if (!same_width && !zeros_warned) {
-                    SPC_WRN("target embeddings of size %zu do not fit the draft input width %zu, "
-                            "the draft receives zero rows for them and drafts after multimodal input will be poor\n",
-                            n_embd, zeros.size());
-                    zeros_warned = true;
-                }
-                const llama_embd embd = same_width ? t.embd : llama_embd{ zeros.data(), 1, zeros.size() };
-                batch.add_embd(embd, t.pos.data(), t.seq_id, output);
-            }
-        }
+        llama_batch batch_dft = batch;
+        batch_dft.logits = nullptr;

-        if (batch.size() == 0) {
-            return true;
-        }
-
-        const int ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+        const int ret = llama_decode(ctx_dft, batch_dft);

        if (ret != 0) {
            SPC_ERR("failed to decode draft batch, ret = %d\n", ret);
@@ -308,7 +279,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
    void draft(common_speculative_draft_params_vec & dparams) override {
        auto & ctx_dft = params.ctx_dft;

-        batch.clear();
+        common_batch_clear(batch);

        // keep track of which sequences are still drafting
        int n_drafting = 0;
@@ -325,12 +296,12 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
            drafting[seq_id] = true;
            common_sampler_reset(smpls[seq_id].get());

-            batch.add(dp.id_last, dp.pos0, seq_id, true);
+            common_batch_add(batch, dp.id_last, dp.pos0, { seq_id }, true);
        }

-        int ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+        int ret = llama_decode(ctx_dft, batch);
        if (ret != 0) {
-            SPC_ERR("llama_process returned %d\n", ret);
+            SPC_ERR("llama_decode returned %d\n", ret);
            return;
        }

@@ -339,7 +310,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        while (n_drafting > 0) {
            int i_batch = 0;

-            batch.clear();
+            common_batch_clear(batch);

            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                if (!drafting[seq_id]) {
@@ -384,17 +355,17 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
                    continue;
                }

-                batch.add(id, dp.pos0 + i + 1, seq_id, true);
+                common_batch_add(batch, id, dp.pos0 + i + 1, { seq_id }, true);
            }

-            if (batch.size() == 0) {
+            if (batch.n_tokens == 0) {
                break;
            }

            // evaluate the drafted tokens on the draft model
-            ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+            ret = llama_decode(ctx_dft, batch);
            if (ret != 0) {
-                SPC_ERR("llama_process[%d] returned %d\n", i, ret);
+                SPC_ERR("llama_decode[%d] returned %d\n", i, ret);
                break;
            }

@@ -454,8 +425,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
 //      encoder+decoder on n_accepted+1 rows).
 struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
    common_params_speculative_draft params;
-    common_batch batch;     // decoder input, (token, g_embd) pairs
-    common_batch batch_enc; // encoder input, built from the extracted target features
+    llama_batch batch;

    std::vector<common_sampler_ptr> smpls;

@@ -509,8 +479,11 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt;
        n_layer_tgt = llama_model_n_layer(model_tgt);

-        batch     = common_batch(ctx_dft);
-        batch_enc = common_batch(ctx_dft);
+        const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
+        batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd_dec, /*n_seq_max=*/ 1);
+        // llama_batch_init allocates only one of token/embd; eagle3 decoder needs both.
+        // TODO: fix, how to call without malloc
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_b);

        smpls.resize(n_seq);
        for (auto & s : smpls) {
@@ -572,6 +545,12 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
            llama_sampler_free(backend_chains[seq_id]);
        }
        backend_chains.clear();
+
+        if (batch.token != nullptr) {
+            free(batch.token);
+            batch.token = nullptr;
+        }
+        llama_batch_free(batch);
    }

    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
@@ -590,16 +569,16 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        }
    }

-    bool process(const common_batch & batch_in) override {
-        if (batch_in.size() <= 0) {
+    bool process(const llama_batch & batch_in) override {
+        if (batch_in.n_tokens <= 0) {
            return true;
        }

-        if (!batch_in.has_token() || batch_in.has_embd()) {
+        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
            return true;
        }

-        const int32_t n_tokens = batch_in.size();
+        const int32_t n_tokens = batch_in.n_tokens;

        // i_batch_beg[seq] / i_batch_end[seq]: inclusive batch indices of this seq's
        // first/last token in batch_in. Assumes per-seq tokens are contiguous within
@@ -607,7 +586,8 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        std::vector<int32_t> i_batch_beg(n_seq, -1);
        std::vector<int32_t> i_batch_end(n_seq, -1);
        for (int k = 0; k < n_tokens; ++k) {
-            const llama_seq_id seq_id = batch_in.tokens[k].seq_id;
+            GGML_ASSERT(batch_in.n_seq_id[k] == 1);
+            const llama_seq_id seq_id = batch_in.seq_id[k][0];
            if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
                continue;
            }
@@ -641,23 +621,24 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {

        g_embd_buf.resize((size_t) n_tokens * n_embd_dec);

-        // llama_process() requires the full encoder batch to fit in n_ubatch.
+        // llama_encode() requires the full encoder batch to fit in n_ubatch.
        // Allow batch > ubatch: eagle3's per-token encoder can be chunked safely.
        const int32_t n_ubatch_dft = (int32_t) llama_n_ubatch(ctx_dft);
        for (int32_t i = 0; i < n_tokens; i += n_ubatch_dft) {
            const int32_t n_chunk = std::min(n_ubatch_dft, n_tokens - i);

-            // the per-token encoder does not use positions, generate placeholder ones from the memory state
-            batch_enc.clear();
-            llama_pos pos = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), 0) + 1;
-            for (int32_t j = 0; j < n_chunk; ++j) {
-                batch_enc.add_embd({ features_buf.data() + (size_t) (i + j) * n_embd_enc, 1, (size_t) n_embd_enc }, &pos, 0, true);
-                pos++;
-            }
-
-            const int32_t rc = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_ENCODE, batch_enc.get());
+            llama_batch enc_batch = {
+                /*.n_tokens =*/ n_chunk,
+                /*.token    =*/ nullptr,
+                /*.embd     =*/ features_buf.data() + (size_t) i * n_embd_enc,
+                /*.pos      =*/ nullptr,
+                /*.n_seq_id =*/ nullptr,
+                /*.seq_id   =*/ nullptr,
+                /*.logits   =*/ nullptr,
+            };
+            const int32_t rc = llama_encode(ctx_dft, enc_batch);
            if (rc != 0) {
-                SPC_ERR("llama_process(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                SPC_ERR("llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
                        rc, (int) n_chunk, (int) i);
                return false;
            }
@@ -685,7 +666,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        //       deferred boundary, completed by the next process() or draft() call.
        //   (c) refresh deferred state — stash this ubatch's full g_embd into verify_g,
        //       update pending_g_last / pending_pos_last to the last row.
-        batch.clear();
+        common_batch_clear(batch);

        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
            const int32_t beg = i_batch_beg[seq_id];
@@ -700,34 +681,36 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
            //   2) pending_pos_last + 1 == pos[beg]
            //   3) pending_pos_last > dft_pos_max // TODO: is this check needed?
            const llama_pos pending_pos = pending_pos_last[seq_id];
-            if (pending_pos >= 0 && pending_pos + 1 == batch_in.tokens[beg].pos[0]) {
+            if (pending_pos >= 0 && pending_pos + 1 == batch_in.pos[beg]) {
                const llama_pos dft_pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
                if (pending_pos > dft_pos_max) {
-                    const int32_t idx = batch.add(batch_in.tokens[beg].id, pending_pos, seq_id, /*output=*/ false);
-                    batch.set_embd(idx, { pending_g_last[seq_id].data(), 1, (size_t) n_embd_dec });
+                    common_batch_add(batch, batch_in.token[beg], pending_pos, { seq_id }, /*logits=*/ false);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                                pending_g_last[seq_id].data(), row_bytes);
                }
            }

            for (int32_t k = beg; k < end; ++k) {
-                const int32_t idx = batch.add(batch_in.tokens[k + 1].id, batch_in.tokens[k].pos[0], seq_id, /*output=*/ false);
-                batch.set_embd(idx, { g_embd + (size_t) k * n_embd_dec, 1, (size_t) n_embd_dec });
+                common_batch_add(batch, batch_in.token[k + 1], batch_in.pos[k], { seq_id }, /*logits=*/ false);
+                std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                            g_embd + (size_t) k * n_embd_dec, row_bytes);
            }

            // refresh deferred state
            const int32_t n_rows = end - beg + 1;
-            verify_pos_first[seq_id] = batch_in.tokens[beg].pos[0];
-            pending_pos_last[seq_id] = batch_in.tokens[end].pos[0];
+            verify_pos_first[seq_id] = batch_in.pos[beg];
+            pending_pos_last[seq_id] = batch_in.pos[end];
            verify_g_rows[seq_id]    = n_rows;
            verify_g[seq_id].resize((size_t) n_rows * n_embd_dec, 0.0f);
            std::memcpy(verify_g[seq_id].data(),       g_embd + (size_t) beg * n_embd_dec, row_bytes * n_rows);
            std::memcpy(pending_g_last[seq_id].data(), g_embd + (size_t) end * n_embd_dec, row_bytes);
        }

-        if (batch.size() > 0) {
-            const int32_t rc = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+        if (batch.n_tokens > 0) {
+            const int32_t rc = llama_decode(ctx_dft, batch);
            if (rc != 0) {
-                SPC_ERR("llama_process(ctx_dft) failed rc=%d (n_tokens=%d, ubatch_pos[0]=%d)\n",
-                        rc, (int) batch.size(), (int) batch_in.tokens[0].pos[0]);
+                SPC_ERR("llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, ubatch_pos[0]=%d)\n",
+                        rc, (int) batch.n_tokens, (int) batch_in.pos[0]);
                return false;
            }
        }
@@ -738,12 +721,14 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
    void draft(common_speculative_draft_params_vec & dparams) override {
        auto & ctx_dft = params.ctx_dft;

-        batch.clear();
+        common_batch_clear(batch);

        // keep track of which sequences are still drafting
        int n_drafting = 0;
        std::vector<bool> drafting(n_seq);

+        const size_t row_bytes = (size_t) n_embd_dec * sizeof(float);
+
        // Complete the deferred boundary pair (dp.id_last, pending_g_last) at memory
        // pos pending_pos_last. dp.id_last is target's freshest sample (= corrected
        // token after verify, or first generated token after prefill), matching the
@@ -764,17 +749,19 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {

            llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, pending_pos_last[seq_id], -1);

-            const int32_t idx = batch.add(dp.id_last, pending_pos_last[seq_id], seq_id, true);
-            batch.set_embd(idx, { pending_g_last[seq_id].data(), 1, (size_t) n_embd_dec });
+            common_batch_add(batch, dp.id_last, pending_pos_last[seq_id], { seq_id }, true);
+            std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec,
+                        pending_g_last[seq_id].data(),
+                        row_bytes);
        }

-        if (batch.size() == 0) {
+        if (batch.n_tokens == 0) {
            return;
        }

-        int ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+        int ret = llama_decode(ctx_dft, batch);
        if (ret != 0) {
-            SPC_ERR("llama_process returned %d\n", ret);
+            SPC_ERR("llama_decode returned %d\n", ret);
            return;
        }

@@ -783,7 +770,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        while (n_drafting > 0) {
            int i_batch = 0;

-            batch.clear();
+            common_batch_clear(batch);

            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                if (!drafting[seq_id]) {
@@ -829,17 +816,17 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
                    continue;
                }

-                const int32_t idx = batch.add(id, pending_pos_last[seq_id] + (i + 1), seq_id, true);
-                batch.set_embd(idx, { prenorm, 1, (size_t) n_embd_dec });
+                common_batch_add(batch, id, pending_pos_last[seq_id] + (i + 1), { seq_id }, true);
+                std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd_dec, prenorm, row_bytes);
            }

-            if (batch.size() == 0) {
+            if (batch.n_tokens == 0) {
                break;
            }

-            ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+            ret = llama_decode(ctx_dft, batch);
            if (ret != 0) {
-                SPC_ERR("llama_process[%d] returned %d\n", i, ret);
+                SPC_ERR("llama_decode[%d] returned %d\n", i, ret);
                break;
            }

@@ -923,10 +910,8 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
 struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    common_params_speculative_draft params;

-    common_batch batch;        // noise tokens
-    common_batch batch_inject; // target features for KV cache injection
-
-    std::vector<float> features_buf; // [n_chunk, n_embd_enc] gathered target features
+    llama_batch batch;        // noise tokens
+    llama_batch batch_inject; // target features for KV cache injection

    std::vector<common_sampler_ptr> smpls;

@@ -1022,11 +1007,15 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
        }
        this->n_max = this->params.n_max;

-        batch        = common_batch(ctx_dft);
-        batch_inject = common_batch(ctx_dft);
+        batch        = llama_batch_init(llama_n_batch(ctx_dft), 0,          n_seq);
+        batch_inject = llama_batch_init(llama_n_ubatch(ctx_dft), n_embd_enc, n_seq);

-        // embd batches on an M-RoPE draft carry 4 position rows per token
+        // embd batches on an M-RoPE draft need 4 position rows per token
        is_mrope = llama_model_rope_type(model_dft) == LLAMA_ROPE_TYPE_MROPE;
+        if (is_mrope) {
+            free(batch_inject.pos);
+            batch_inject.pos = (llama_pos *) malloc(sizeof(llama_pos) * 4 * llama_n_batch(ctx_dft));
+        }

        smpls.resize(n_seq);
        for (auto & s : smpls) {
@@ -1075,6 +1064,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            llama_sampler_free(backend_chains[seq_id]);
        }
        backend_chains.clear();
+
+        llama_batch_free(batch);
+        llama_batch_free(batch_inject);
    }

    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
@@ -1095,8 +1087,8 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
        }
    }

-    bool process(const common_batch & batch_in) override {
-        if (batch_in.size() <= 0) {
+    bool process(const llama_batch & batch_in) override {
+        if (batch_in.n_tokens <= 0) {
            return true;
        }

@@ -1104,19 +1096,20 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
        // produce the target-layer features used to seed the draft KV cache, so
        // embeddings are injected too, except the pinned ones skipped below.
        // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
-        const bool has_tokens     = batch_in.has_token();
-        const bool has_embeddings = batch_in.has_embd();
+        const bool has_tokens     = batch_in.token != nullptr;
+        const bool has_embeddings = batch_in.embd  != nullptr;
        if (has_tokens == has_embeddings) {
            return true;
        }

-        const int32_t n_tokens = batch_in.size();
+        const int32_t n_tokens = batch_in.n_tokens;

        // per-seq inclusive batch range (assumes each seq's tokens are contiguous in the batch)
        std::vector<int32_t> i_batch_beg(n_seq, -1);
        std::vector<int32_t> i_batch_end(n_seq, -1);
        for (int32_t k = 0; k < n_tokens; ++k) {
-            const llama_seq_id seq_id = batch_in.tokens[k].seq_id;
+            GGML_ASSERT(batch_in.n_seq_id[k] == 1);
+            const llama_seq_id seq_id = batch_in.seq_id[k][0];
            if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
                continue;
            }
@@ -1139,7 +1132,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

            // an M-RoPE image pins all its rows to one position, so a windowed draft
            // cache cannot free cells for it - skip it, the draft can jump over the gap
-            const bool pos_pinned = batch_in.tokens[i_batch_beg[seq_id]].pos[0] == batch_in.tokens[i_batch_end[seq_id]].pos[0];
+            const bool pos_pinned = batch_in.pos[i_batch_beg[seq_id]] == batch_in.pos[i_batch_end[seq_id]];
            if (has_embeddings && n_rows > 1 && pos_pinned) {
                continue;
            }
@@ -1149,28 +1142,34 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

                // gather target features per extract layer; the fused decode encodes and
                // injects them into the K/V cache at the target positions
-                features_buf.resize((size_t) n_chunk * n_embd_enc);
+                batch_inject.n_tokens = n_chunk;
                for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
                    const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
                    if (!layer) {
                        GGML_ABORT("DFlash: target layer %d input not extracted.", target_layer_ids[k]);
                    }
                    for (int32_t i = 0; i < n_chunk; ++i) {
-                        float       * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
+                        float       * dst = batch_inject.embd + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
                        const float * src = layer + (size_t) (i_batch_beg[seq_id] + offset + i) * n_embd_tgt;
                        std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float));
                    }
                }

-                batch_inject.clear();
                for (int32_t i = 0; i < n_chunk; ++i) {
-                    const llama_pos p = batch_in.tokens[i_batch_beg[seq_id] + offset + i].pos[0];
-                    const llama_pos pos_arr[4] = { p, p, p, 0 };
-                    batch_inject.add_embd({ features_buf.data() + (size_t) i * n_embd_enc, 1, (size_t) n_embd_enc }, pos_arr, seq_id, false);
+                    const llama_pos p = batch_in.pos[i_batch_beg[seq_id] + offset + i];
+                    batch_inject.pos[i] = p;
+                    if (is_mrope) {
+                        batch_inject.pos[1 * n_chunk + i] = p;
+                        batch_inject.pos[2 * n_chunk + i] = p;
+                        batch_inject.pos[3 * n_chunk + i] = 0;
+                    }
+                    batch_inject.n_seq_id[i]  = 1;
+                    batch_inject.seq_id[i][0] = seq_id;
+                    batch_inject.logits[i]    = false;
                }
-                const int32_t rc = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch_inject.get());
+                const int32_t rc = llama_decode(ctx_dft, batch_inject);
                if (rc != 0) {
-                    LOG_ERR("%s: llama_process(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                    LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
                            __func__, rc, (int) n_chunk, (int) offset);
                    return false;
                }
@@ -1183,7 +1182,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    void draft(common_speculative_draft_params_vec & dparams) override {
        auto & ctx_dft = params.ctx_dft;

-        batch.clear();
+        common_batch_clear(batch);

        // build one batch holding every drafting sequence's noise block into a single decode)
        // record where each block starts and its size
@@ -1203,21 +1202,21 @@ 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);
-            i_block_beg[seq_id] = batch.size();
+            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) {
-                batch.add(i == 0 ? dp.id_last : mask_token_id, n + i, seq_id, !is_dflash2);
+                common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, !is_dflash2);
            }
        }

-        if (batch.size() == 0) {
+        if (batch.n_tokens == 0) {
            return;
        }

        // decode all sequence's noise block in a single batch
-        int ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+        int ret = llama_decode(ctx_dft, batch);
        if (ret != 0) {
-            LOG_WRN("%s: llama_process returned %d\n", __func__, ret);
+            LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
            return;
        }

@@ -1331,7 +1330,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 struct common_speculative_impl_draft_mtp : public common_speculative_impl {
    common_params_speculative_draft params; // reuses the draft-model params slot (ctx_tgt/ctx_dft)

-    common_batch batch;
+    llama_batch batch;

    std::vector<common_sampler_ptr> smpls;

@@ -1387,7 +1386,11 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                ctx_dft ? "yes" : "no",
                common_speculative_get_devices_str(this->params.devices).c_str());

-        batch = common_batch(ctx_dft);
+        const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
+        batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd, /*n_seq_max=*/ 1);
+        // llama_batch_init allocates only one of token/embd; MTP needs both.
+        // TODO: fix, how to call without malloc
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_b);

        smpls.resize(n_seq);
        for (auto & s : smpls) {
@@ -1452,6 +1455,12 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
            llama_sampler_free(backend_chains[seq_id]);
        }
        backend_chains.clear();
+
+        if (batch.token != nullptr) {
+            free(batch.token);
+            batch.token = nullptr;
+        }
+        llama_batch_free(batch);
    }

    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
@@ -1466,23 +1475,23 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
        if (pos_max < N - 1 && !is_mem_shared) {
            SPC_WRN("ctx_dft pos_max=%d < N-1=%d - "
                    "process() hook may not have run on every prefill ubatch "
-                    "(need_embd / output flag on every prompt position?). "
+                    "(need_embd / logits=1 on every prompt position?). "
                    "Drafts may degrade.\n",
                    (int) pos_max, N - 1);
        }
    }

-    bool process(const common_batch & batch_in) override {
-        if (batch_in.size() <= 0) {
+    bool process(const llama_batch & batch_in) override {
+        if (batch_in.n_tokens <= 0) {
            return true;
        }

        // TODO: how to make it work with vision tokens?
-        if (!batch_in.has_token() || batch_in.has_embd()) {
+        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
            return true;
        }

-        const int32_t n_tokens = batch_in.size();
+        const int32_t n_tokens = batch_in.n_tokens;

        // remember the first and last batch index for each sequence
        std::fill(i_batch_beg.begin(), i_batch_beg.end(), -1);
@@ -1490,7 +1499,9 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {

        for (int k = 0; k < n_tokens; ++k) {
            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-                if (batch_in.tokens[k].seq_id == seq_id) {
+                GGML_ASSERT(batch_in.n_seq_id[k] == 1);
+
+                if (batch_in.seq_id[k][0] == seq_id) {
                    i_batch_end[seq_id] = k;
                    if (i_batch_beg[seq_id] < 0) {
                        i_batch_beg[seq_id] = k;
@@ -1506,26 +1517,33 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {

        // if kv is shared with target (e.g Gemma4), then we can skip this catch-up decode
        if (!is_mem_shared) {
-            batch.clear();
+            common_batch_clear(batch);

-            // pair each token with the tgt embedding shifted right by one position, and
-            // the first token of each sequence with the pending embedding from a previous run
+            for (int k = 0; k < n_tokens; ++k) {
+                common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
+            }
+
+            // shift the tgt embeddings to the right by one position
            // assumes that the tokens in the batch are sequential for each sequence
            // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
            //                                                       ^--- this is a problem
            // TODO:this is generally true, but would be nice to assert it
-            const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
+            {
+                const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
+                std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
+            }

-            for (int k = 0; k < n_tokens; ++k) {
-                const llama_seq_id seq_id = batch_in.tokens[k].seq_id;
+            // fill the pending embeddings from a previous run
+            auto set_h = [&](int idx, const float * h_row) {
+                std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
+            };

-                const int32_t idx = batch.add(batch_in.tokens[k].id, batch_in.tokens[k].pos[0], seq_id, false);
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                if (i_batch_beg[seq_id] < 0) {
+                    continue;
+                }

-                const float * h_row = k == i_batch_beg[seq_id]
-                    ? pending_h[seq_id].data()
-                    : h_tgt + (size_t) (k - 1) * n_embd;
-
-                batch.set_embd(idx, { h_row, 1, (size_t) n_embd });
+                set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
            }

            auto * mem_dft = llama_get_memory(ctx_dft);
@@ -1538,15 +1556,15 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                        if (i_batch_beg[seq_id] < 0) {
                            continue;
                        }
-                        llama_memory_seq_rm(mem_dft, seq_id, batch_in.tokens[i_batch_beg[seq_id]].pos[0], -1);
+                        llama_memory_seq_rm(mem_dft, seq_id, batch_in.pos[i_batch_beg[seq_id]], -1);
                    }
                    llama_set_nextn_layer_offset(ctx_dft, head);
                }

-                const int32_t rc = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+                const int32_t rc = llama_decode(ctx_dft, batch);
                if (rc != 0) {
-                    SPC_ERR("llama_process(ctx_dft) head=%d failed rc=%d (pos=%d)\n",
-                            head, (int) rc, (int) batch_in.tokens[0].pos[0]);
+                    SPC_ERR("llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n",
+                            head, (int) rc, (int) batch_in.pos[0]);
                    ok = false;
                    break;
                }
@@ -1584,12 +1602,14 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
    void draft(common_speculative_draft_params_vec & dparams) override {
        auto & ctx_dft = params.ctx_dft;

-        batch.clear();
+        common_batch_clear(batch);

        // keep track of which sequences are still drafting
        int n_drafting = 0;
        std::vector<bool> drafting(n_seq);

+        const size_t row_bytes = (size_t) n_embd * sizeof(float);
+
        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
            auto & dp = dparams[seq_id];

@@ -1601,10 +1621,10 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
            drafting[seq_id] = true;
            common_sampler_reset(smpls[seq_id].get());

-            const int32_t idx = batch.add(dp.id_last, dp.pos0, seq_id, true);
-            batch.set_embd(idx, { pending_h[seq_id].data(), 1, (size_t) n_embd });
+            common_batch_add(batch, dp.id_last, dp.pos0, { seq_id }, true);
+            std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, pending_h[seq_id].data(), row_bytes);

-            i_last[seq_id] = idx;
+            i_last[seq_id] = batch.n_tokens - 1;

            if (chain_heads) {
                chain_h[seq_id].assign(pending_h[seq_id].begin(), pending_h[seq_id].end());
@@ -1630,16 +1650,16 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                llama_set_nextn_layer_offset(ctx_dft, i);
            }

-            int ret = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch.get());
+            int ret = llama_decode(ctx_dft, batch);
            if (ret != 0) {
-                SPC_ERR("llama_process[%d] returned %d\n", i, ret);
+                SPC_ERR("llama_decode[%d] returned %d\n", i, ret);
                break;
            }

            // rebuild the batch for the next step: the growing-KV paths re-add only the
            // new token (the KV already holds the prefix), while chained heads re-add the
            // whole prefix at the next head. dropped sequences are simply not re-added.
-            batch.clear();
+            common_batch_clear(batch);

            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                if (!drafting[seq_id]) {
@@ -1690,24 +1710,24 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                    const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far
                    for (int t = 0; t < n_rows; ++t) {
                        const llama_token tok = (t == 0) ? dp.id_last : result[t - 1];
-                        const int32_t idx = batch.add(tok, dp.pos0 + t, seq_id, t == n_rows - 1);
-                        batch.set_embd(idx, { chain_h[seq_id].data() + (size_t) t * n_embd, 1, (size_t) n_embd });
-                        i_last[seq_id] = idx;
+                        common_batch_add(batch, tok, dp.pos0 + t, { seq_id }, t == n_rows - 1);
+                        std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd,
+                                    chain_h[seq_id].data() + (size_t) t * n_embd, row_bytes);
                    }
                } else if (is_mem_shared) {
                    // note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
                    // ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
-                    const int32_t idx = batch.add(id, dp.pos0, seq_id, true);
-                    batch.set_embd(idx, { h_row, 1, (size_t) n_embd });
-                    i_last[seq_id] = idx;
+                    common_batch_add(batch, id, dp.pos0, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                } else {
-                    const int32_t idx = batch.add(id, dp.pos0 + i + 1, seq_id, true);
-                    batch.set_embd(idx, { h_row, 1, (size_t) n_embd });
-                    i_last[seq_id] = idx;
+                    common_batch_add(batch, id, dp.pos0 + i + 1, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                }
+
+                i_last[seq_id] = batch.n_tokens - 1;
            }

-            if (batch.size() == 0) {
+            if (batch.n_tokens == 0) {
                break;
            }

@@ -1769,7 +1789,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
        // noop
    }

-    bool process(const common_batch & /*batch*/) override {
+    bool process(const llama_batch & /*batch*/) override {
        // TODO: implement
        return true;
    }
@@ -1817,7 +1837,7 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
        common_ngram_map_begin(config[seq_id], prompt);
    }

-    bool process(const common_batch & /*batch*/) override {
+    bool process(const llama_batch & /*batch*/) override {
        // TODO: implement
        return true;
    }
@@ -1975,7 +1995,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
        sinfo.n_draft_last = result.size();
    }

-    bool process(const common_batch & /*batch*/) override {
+    bool process(const llama_batch & /*batch*/) override {
        // TODO: implement
        return true;
    }
@@ -2137,7 +2157,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
        }
    }

-    bool process(const common_batch & /*batch*/) override {
+    bool process(const llama_batch & /*batch*/) override {
        // TODO: implement
        return true;
    }
@@ -2539,7 +2559,7 @@ common_speculative_init_result::common_speculative_init_result(
        model_path = params.speculative.draft.mparams.path;
        LOG_INF("%s: loading draft model '%s'\n", __func__, model_path.c_str());

-        llama_model * model_dft = llama_model_load_from_file(model_path.c_str(), mparams);
+        llama_model * model_dft = llama_model_load_from_file(params.model.path.c_str(), mparams);
        if (model_dft == NULL) {
            LOG_ERR("%s: failed to load draft model, '%s'\n", __func__, model_path.c_str());
            return;
@@ -2770,7 +2790,7 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
    }
 }

-bool common_speculative_process(common_speculative * spec, const common_batch & batch) {
+bool common_speculative_process(common_speculative * spec, const llama_batch & batch) {
    bool result = true;

    if (spec == nullptr) {
@@ -77,7 +77,7 @@ common_speculative_draft_params & common_speculative_get_draft_params(common_spe
 void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, const llama_tokens & prompt);

 // process the batch and update the internal state of the speculative context
-bool common_speculative_process(common_speculative * spec, const common_batch & batch);
+bool common_speculative_process(common_speculative * spec, const llama_batch & batch);

 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);
@@ -26,16 +26,16 @@ utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t off

    // Invalid: continuation byte as first byte
    if (!(input[offset] & 0x40)) {
-        return utf8_parse_result(utf8_parse_result::INVALID, 0, 1);
+        return utf8_parse_result(utf8_parse_result::INVALID);
    }

    // 2-byte sequence
    if (!(input[offset] & 0x20)) {
        if (offset + 1 >= input.size()) {
-            return utf8_parse_result(utf8_parse_result::INCOMPLETE, 0, 1);
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
        }
        if ((input[offset + 1] & 0xc0) != 0x80) {
-            return utf8_parse_result(utf8_parse_result::INVALID, 0, 1);
+            return utf8_parse_result(utf8_parse_result::INVALID);
        }
        auto result = ((input[offset] & 0x1f) << 6) | (input[offset + 1] & 0x3f);
        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 2);
@@ -43,14 +43,11 @@ utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t off

    // 3-byte sequence
    if (!(input[offset] & 0x10)) {
-        // Check one byte at a time so a bad byte is reported before a short input
-        for (size_t i = 1; i < 3; i++) {
-            if (offset + i >= input.size()) {
-                return utf8_parse_result(utf8_parse_result::INCOMPLETE, 0, i);
-            }
-            if ((input[offset + i] & 0xc0) != 0x80) {
-                return utf8_parse_result(utf8_parse_result::INVALID, 0, i);
-            }
+        if (offset + 2 >= input.size()) {
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+        }
+        if ((input[offset + 1] & 0xc0) != 0x80 || (input[offset + 2] & 0xc0) != 0x80) {
+            return utf8_parse_result(utf8_parse_result::INVALID);
        }
        auto result = ((input[offset] & 0x0f) << 12) | ((input[offset + 1] & 0x3f) << 6) | (input[offset + 2] & 0x3f);
        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 3);
@@ -58,20 +55,18 @@ utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t off

    // 4-byte sequence
    if (!(input[offset] & 0x08)) {
-        for (size_t i = 1; i < 4; i++) {
-            if (offset + i >= input.size()) {
-                return utf8_parse_result(utf8_parse_result::INCOMPLETE, 0, i);
-            }
-            if ((input[offset + i] & 0xc0) != 0x80) {
-                return utf8_parse_result(utf8_parse_result::INVALID, 0, i);
-            }
+        if (offset + 3 >= input.size()) {
+            return utf8_parse_result(utf8_parse_result::INCOMPLETE);
+        }
+        if ((input[offset + 1] & 0xc0) != 0x80 || (input[offset + 2] & 0xc0) != 0x80 || (input[offset + 3] & 0xc0) != 0x80) {
+            return utf8_parse_result(utf8_parse_result::INVALID);
        }
        auto result = ((input[offset] & 0x07) << 18) | ((input[offset + 1] & 0x3f) << 12) | ((input[offset + 2] & 0x3f) << 6) | (input[offset + 3] & 0x3f);
        return utf8_parse_result(utf8_parse_result::SUCCESS, result, 4);
    }

    // Invalid first byte
-    return utf8_parse_result(utf8_parse_result::INVALID, 0, 1);
+    return utf8_parse_result(utf8_parse_result::INVALID);
 }

 bool common_utf8_is_complete(const std::string & s) {
@@ -9,7 +9,7 @@

 struct utf8_parse_result {
    uint32_t codepoint;      // Decoded codepoint (only valid if status == SUCCESS)
-    size_t bytes_consumed;   // How many bytes this codepoint uses (1-4), or the length of the valid prefix if status != SUCCESS
+    size_t bytes_consumed;   // How many bytes this codepoint uses (1-4)
    enum status { SUCCESS, INCOMPLETE, INVALID } status;

    utf8_parse_result(enum status s, uint32_t cp = 0, size_t bytes = 0)
@@ -28,7 +28,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "BailingMoeForCausalLM": "bailingmoe",
    "BailingMoeV2ForCausalLM": "bailingmoe",
    "BailingMoeV3ForCausalLM": "bailingmoe3",
-    "BailingMoeV3VLForConditionalGeneration": "bailingmoe3",
    "BambaForCausalLM": "granite",
    "BertForMaskedLM": "bert",
    "BertForSequenceClassification": "bert",
@@ -95,7 +94,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Gemma3nForCausalLM": "gemma",
    "Gemma3nForConditionalGeneration": "gemma",
    "Gemma4AssistantForCausalLM": "gemma",
-    "Gemma4DSparkModel": "gemma",
    "Gemma4ForConditionalGeneration": "gemma",
    "Gemma4ForCausalLM": "gemma",
    "Gemma4UnifiedForConditionalGeneration": "gemma",
@@ -106,7 +104,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Glm4MoeLiteForCausalLM": "glm",
    "Glm4vForConditionalGeneration": "glm",
    "Glm4vMoeForConditionalGeneration": "glm",
-    "Glm5NextForConditionalGeneration": "glm",
    "GlmForCausalLM": "chatglm",
    "GlmMoeDsaForCausalLM": "glm",
    "GlmOcrForConditionalGeneration": "glm",
@@ -304,9 +301,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Gemma4ForConditionalGeneration": "gemma",
    "Gemma4UnifiedForConditionalGeneration": "gemma",
    "Glm4vForConditionalGeneration": "qwen3vl",
-    "BailingMoeV3VLForConditionalGeneration": "bailingmoe3",
    "Glm4vMoeForConditionalGeneration": "qwen3vl",
-    "Glm5NextForConditionalGeneration": "qwen3vl",
    "Glm5vForConditionalGeneration": "kimivl",
    "GlmOcrForConditionalGeneration": "qwen3vl",
    "GlmasrModel": "ultravox",
@@ -9,9 +9,7 @@ import torch
 if TYPE_CHECKING:
    from torch import Tensor

-from .base import ModelBase, MmprojModel, TextModel, gguf
-
-from .qwen3vl import Qwen3VLVisionModel
+from .base import ModelBase, TextModel, gguf


@ModelBase.register("BailingMoeV3ForCausalLM")
@@ -76,7 +74,7 @@ class BailingMoeV3Model(TextModel):

        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.get("num_shared_experts", 1))
+        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"])
@@ -193,111 +191,3 @@ class BailingMoeV3Model(TextModel):
            experts = [name for layer in self._experts for name in layer]
            if experts:
                raise ValueError(f"Unprocessed experts: {experts}")
-
-
-@ModelBase.register("BailingMoeV3VLForConditionalGeneration")
-@ModelBase.example("inclusionAI/Ling-3.0-flash-VL")
-class BailingMoeV3VLModel(BailingMoeV3Model):
-    model_arch = gguf.MODEL_ARCH.BAILINGMOE3
-
-    def index_tensors(self, remote_hf_model_id: str | None = None):
-        # hoist text_config before the shared BailingMoeV3 logic runs:
-        # ModelBase.__init__ calls this with the raw VL config, where the text
-        # dims still live under text_config
-        if "text_config" in self.hparams:
-            self.hparams = {**self.hparams, **self.hparams["text_config"]}
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        mrope_section = self.hparams.get("mrope_section")
-        if mrope_section is None:
-            raise ValueError("BailingMoeV3VL requires mrope_section in the config")
-        if sum(mrope_section[:3]) * 2 != self.hparams["qk_rope_head_dim"]:
-            raise ValueError(
-                f"mrope_section {mrope_section[:3]} counts rope pairs and must sum to"
-                f" qk_rope_head_dim / 2 = {self.hparams['qk_rope_head_dim'] // 2}"
-            )
-        # mrope_section is [t, h, w]; pad to the 4-wide sections array
-        self.gguf_writer.add_rope_dimension_sections(list(mrope_section[:3]) + [0])
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        # Skip projector tensors; the vision tower is skipped by TextModel.filter_tensors
-        if name.startswith("linear_proj"):
-            return None
-
-        return super().filter_tensors(item)
-
-
-@ModelBase.register("BailingMoeV3VLForConditionalGeneration")
-@ModelBase.example("inclusionAI/Ling-3.0-flash-VL")
-class BailingMoeV3VLVisionModel(Qwen3VLVisionModel):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        assert self.hparams_vision is not None
-
-        if self.hparams_vision.get("disable_merger_proj") is not True:
-            raise ValueError("BailingMoeV3VL requires disable_merger_proj=true")
-
-        # out_hidden_size is the vision encoder output (post spatial merge, pre linear_proj)
-        self.image_emb_dim = self.hparams_vision.get("out_hidden_size")
-        if self.image_emb_dim is None:
-            raise ValueError("BailingMoeV3VL vision config requires out_hidden_size")
-
-    def set_gguf_parameters(self):
-        assert self.hparams_vision is not None
-        MmprojModel.set_gguf_parameters(self)  # skip Qwen3VLVisionModel parameters
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LING3VL)
-        self.gguf_writer.add_vision_use_gelu(True)
-
-        merge_size = self.hparams_vision.get("spatial_merge_size")
-        if merge_size is not None:
-            self.gguf_writer.add_vision_spatial_merge_size(int(merge_size))
-
-        rms_norm_eps = self.global_config.get("text_config", {}).get("rms_norm_eps", 1e-6)
-        self.gguf_writer.add_vision_attention_layernorm_eps(rms_norm_eps)
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        if name.startswith("lm_head."):
-            return None
-
-        if name.startswith("linear_proj"):
-            # top-level projector MLP: linear_proj.0 -> mm.0, linear_proj.2 -> mm.2
-            parts = name.split(".")
-            if len(parts) != 3:
-                raise ValueError(f"Unexpected linear_proj tensor: {name}")
-            idx, suffix = int(parts[1]), parts[2]
-            name = f"mm.{idx}.{suffix}"
-            # the qwen3vl filter keeps only visual.*; skip it for the renamed projector tensors
-            return MmprojModel.filter_tensors((name, gen))
-
-        if name.startswith("model.visual."):
-            name = name.replace("model.visual.", "visual.", 1)
-
-        if not name.startswith("visual."):
-            return None
-
-        return super().filter_tensors((name, gen))
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        assert self.hparams_vision is not None
-
-        if name.startswith("mm.0.") or name.startswith("mm.2."):
-            # top-level projector MLP (linear_proj.0 / linear_proj.2, renamed by filter_tensors)
-            yield (name, data_torch)
-            return
-
-        if name == "visual.merger.norm.weight" or name == "visual.merger.norm.bias":
-            # the merger is norm-only for Ling: per-patch LayerNorm before the spatial merge
-            new_name = f"mm.input_norm.{name.split('.')[-1]}"
-            yield (new_name, data_torch)
-            return
-
-        # Ling has no patch bias; the Conv3D split below matches the stock qwen3vl path
-        yield from Qwen3VLVisionModel.modify_tensors(self, data_torch, name, bid)
@@ -170,9 +170,6 @@ class ModelBase:
        self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
        self._is_nvfp4 = False
        self._is_mxfp4 = False
-        self._nvfp4_global_algo: str | None = None # checkpoint-wide NVFP4 quant_algo
-        self._nvfp4_layer_algo: dict[str, str | None] = {} # per-layer quant_algo, keyed by HF module path
-        self._prec_a4: dict[str, bool] = {} # gguf tensor name -> can use 4-bit (A4) activations
        self._fp8_as_q8 = fp8_as_q8
        self._fp8_dequantized: set[str] = set()

@@ -234,7 +231,7 @@ class ModelBase:

        prefix = "model" if not self.is_mistral_format else "consolidated"
        part_names: list[str] = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
-        is_safetensors: bool = len(part_names) > 0 or (not self.is_mistral_format and (self.dir_model / "model.safetensors.index.json").is_file())
+        is_safetensors: bool = len(part_names) > 0
        if not is_safetensors:
            part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")

@@ -667,18 +664,6 @@ class ModelBase:
                if bias_types:
                    self._fusable_qkv_bias_layers.add(bid)

-    def _tag_prec_a4(self, hf_name: str, gguf_name: str) -> None:
-        # W4A16_NVFP4 should not use 4-bit activations
-        name = hf_name.removesuffix(".weight").removesuffix(".bias")
-        algo = self._nvfp4_global_algo
-        while name:
-            if name in self._nvfp4_layer_algo:
-                algo = self._nvfp4_layer_algo[name]
-                break
-            name = name.rpartition(".")[0]
-        if algo == "W4A16_NVFP4":
-            self._prec_a4[gguf_name] = False
-
    def set_gguf_parameters(self):
        raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")

@@ -791,36 +776,6 @@ class ModelBase:
        raw = torch.cat((s.unsqueeze(-1), qs.to(torch.uint8)), dim=-1)
        return raw.reshape(rows, n_blocks * 17).cpu().numpy()

-    def _mxfp4_expert_tensor(self, loaders: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]):
-        """
-        One stacked [n_expert, rows, cols] MXFP4 tensor, built lazily.
-
-        gguf_writer holds every added tensor until the final write, so building
-        this eagerly (like the DeepSeek-V4 path does) keeps every expert in
-        memory at once. lazy means only the tensor being written is resident.
-        """
-        # meta shapes, so this does not read any weights
-        rows, packed_cols = loaders[0][0]().shape
-        n_blocks = (packed_cols * 2) // 32
-        byte_shape = (len(loaders), rows, n_blocks * 17)
-
-        def load(fns: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]) -> np.ndarray:
-            out = np.empty(byte_shape, dtype=np.uint8)
-            for eid, (packed_fn, scale_fn) in enumerate(fns):
-                out[eid] = self.repack_mxfp4_blocks(
-                    LazyTorchTensor.to_eager(packed_fn()),
-                    LazyTorchTensor.to_eager(scale_fn()),
-                )
-            return out
-
-        # loaders goes through args, not the closure, so that `func` matches
-        # LazyBase's single-argument shape
-        return gguf.LazyNumpyTensor(
-            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(np.uint8, byte_shape),
-            args=(loaders,),
-            func=load,
-        )
-
    @staticmethod
    def _nvfp4_pack(weight: Tensor, scale: Tensor) -> tuple[np.ndarray, list[int]]:
        """Repack NVFP4 ModelOpt tensors into ggml super-block layout.
@@ -852,7 +807,6 @@ class ModelBase:
        raw, shape = self._nvfp4_pack(weight, scale)
        logger.info(f"Repacked {new_name} with shape {shape} and quantization NVFP4")
        self.gguf_writer.add_tensor(new_name, raw, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
-        self._tag_prec_a4(name, new_name)

        self._write_scale_tensor(new_name.replace(".weight", ".scale"), scale2)
        self._write_scale_tensor(new_name.replace(".weight", ".input_scale"), input_scale)
@@ -945,7 +899,6 @@ class ModelBase:
        new_name = self.map_tensor_name(merged_name)
        logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
        self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
-        self._tag_prec_a4(merged_name, new_name)

        scales.sort(key=lambda x: x[0])
        self._write_scales_tensor(new_name.replace(".weight", ".scale"), [s[1] for s in scales])
@@ -988,9 +941,6 @@ class ModelBase:
            and bool(quant_groups)
            and all(g.get("format") == "nvfp4-pack-quantized" for g in quant_groups.values() if isinstance(g, dict))
        )
-
-        self._nvfp4_global_algo = quant_algo
-
        if quant_algo != "NVFP4":
            if nvfp4_compressed_tensors:
                quant_algo = "NVFP4"
@@ -1000,22 +950,6 @@ class ModelBase:
        self._is_nvfp4 = quant_algo in ("NVFP4", "W4A16_NVFP4")
        self._is_mxfp4 = quant_method == "mxfp4"

-        # Per-tensor NVFP4 precision.
-        self._nvfp4_layer_algo = {}
-        if quant_layers:
-            # store all possible module paths and assert if a quantized layer is not in the model
-            modules: set[str] = set()
-            for name in self.model_tensors:
-                while name := name.rpartition(".")[0]:
-                    modules.add(name)
-
-            for layer_name, entry in quant_layers.items():
-                if not isinstance(entry, dict):
-                    continue
-                if titem := self.filter_tensors((layer_name, lambda: torch.empty(0))):
-                    assert titem[0] in modules, f"quantized_layers entry {layer_name!r} is not in the model tensors"
-                    self._nvfp4_layer_algo[titem[0]] = entry.get("quant_algo")
-
        # NVFP4 weights are repacked and written directly to gguf_writer.
        # This must run before dequant_model so NVFP4 tensors are removed
        # from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
@@ -1221,12 +1155,6 @@ class ModelBase:
        logger.info("Set model quantization version")
        self.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)

-        if self._prec_a4:
-            names = sorted(self._prec_a4.keys())
-            values = [self._prec_a4[n] for n in names]
-            logger.info(f"Set prec_a4 metadata for {len(names)} tensor(s)")
-            self.gguf_writer.add_tensor_extra_prec_a4(names, values)
-
    def write_vocab(self):
        raise NotImplementedError("write_vocab() must be implemented in subclasses")

@@ -1933,9 +1861,6 @@ class TextModel(ModelBase):
        if chkhsh == "972da7b59cec44d1f0a490a86c96df53859e486e481563e5dddac155013d87ac":
            # ref: https://huggingface.co/poolside/Laguna-XS.2
            res = "laguna"
-        if chkhsh == "653660222fb704f61cbf2b618a8ae6502b7f8b20c980f9a5de07ed78e13319cd":
-            # ref: https://huggingface.co/ufakai/ufakzeka-1
-            res = "ufakzeka"

        if res is None:
            logger.warning("\n")
@@ -2326,12 +2251,6 @@ class TextModel(ModelBase):
                raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
            self.gguf_writer.add_pooling_type(pooling_type)

-        # pooling before a classification head (e.g. ModernBertForSequenceClassification)
-        if (classifier_pooling := self.hparams.get("classifier_pooling")) is not None:
-            if classifier_pooling not in ("cls", "mean"):
-                raise NotImplementedError(f"Unsupported classifier_pooling: {classifier_pooling}")
-            self.gguf_writer.add_classifier_pooling_type(mode_mapping[classifier_pooling])
-
    def _set_vocab_glmedge(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -2566,11 +2485,6 @@ class TextModel(ModelBase):

        self.gguf_writer.add_add_space_prefix(False)

-        if (add_bos := tokenizer_config.get("add_bos_token")) is not None:
-            self.gguf_writer.add_add_bos_token(add_bos)
-        if (add_eos := tokenizer_config.get("add_eos_token")) is not None:
-            self.gguf_writer.add_add_eos_token(add_eos)
-

 class MmprojModel(ModelBase):
    model_type = ModelType.MMPROJ
@@ -341,7 +341,7 @@ class NomicBertModel(BertModel):
        else:
            raise ValueError(f"unrecognized parameters: n_positions={npos}, max_trained_positions={mtp}")

-        assert self.hparams["activation_function"] == ("gelu" if self.is_moe else "swiglu")
+        assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu"

        # this doesn't do anything in the HF version
        assert self.hparams["causal"] is False
@@ -11,7 +11,6 @@ if TYPE_CHECKING:
    from torch import Tensor

 from .base import MmprojModel, ModelBase, TextModel, gguf, logger
-from .qwen import DFlashModel


@ModelBase.register("GemmaForCausalLM")
@@ -810,105 +809,6 @@ class Gemma4Model(Gemma3Model):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Gemma4DSparkModel")
-class Gemma4DSparkModel(DFlashModel):
-    model_arch = gguf.MODEL_ARCH.DFLASH
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        if not self.hparams.get("attention_k_eq_v", False):
-            raise ValueError("Gemma4 DSpark currently requires attention_k_eq_v")
-        if self.hparams.get("layer_types") != ["full_attention"] * self.block_count:
-            raise ValueError("Gemma4 DSpark currently requires uniform full_attention layer types")
-        if self.hparams.get("hidden_activation", "gelu_pytorch_tanh") != "gelu_pytorch_tanh":
-            raise ValueError("Gemma4 DSpark currently requires hidden_activation=gelu_pytorch_tanh")
-        if self.hparams.get("attention_bias", False) or self.hparams.get("enable_moe_block", False):
-            raise ValueError("Gemma4 DSpark attention bias and MoE are not supported")
-        if (self.hparams.get("draft_vocab_size") or self.hparams["vocab_size"]) != self.hparams["vocab_size"]:
-            raise ValueError("Gemma4 DSpark currently requires a full draft vocabulary")
-        if "model.lm_head.weight" not in self.model_tensors and self.hparams.get("tie_word_embeddings") is not True:
-            raise ValueError("Gemma4 DSpark requires lm_head.weight unless tie_word_embeddings is true")
-
-        self.dflash_config = self.hparams.get("dflash_config", {})
-        markov_type = self.dflash_config.get("markov_head_type", self.hparams.get("markov_head_type", "vanilla"))
-        if markov_type != "vanilla":
-            raise ValueError("Gemma4 DSpark currently requires a vanilla Markov head")
-
-        # Gemma4TextConfig supplies these defaults when rope_parameters is absent.
-        rope = self.hparams.get("rope_parameters") or {
-            "full_attention": {"rope_type": "proportional", "partial_rotary_factor": 0.25, "rope_theta": 1000000.0},
-        }
-        self.rope_parameters = rope.get("full_attention", rope)
-        if self.rope_parameters.get("rope_type") not in ("default", "proportional"):
-            raise ValueError("Gemma4 DSpark requires default or proportional RoPE")
-
-    def set_vocab(self):
-        super().set_vocab()
-        mask_id = self.dflash_config.get("mask_token_id", self.hparams.get("mask_token_id"))
-        if mask_id is None:
-            raise ValueError("Gemma4 DSpark requires mask_token_id")
-        if "mask_token_id" not in self.dflash_config:
-            self.gguf_writer.add_mask_token_id(mask_id)
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        head_dim = int(self.hparams["global_head_dim"])
-        self.gguf_writer.add_head_count_kv(self.hparams["num_global_key_value_heads"])
-        self.gguf_writer.add_key_length(head_dim)
-        self.gguf_writer.add_value_length(head_dim)
-        self.gguf_writer.add_rope_dimension_count(head_dim)
-        self.gguf_writer.add_embedding_scale(self.hparams["hidden_size"] ** 0.5)
-        self.gguf_writer.add_attention_scale(1.0)
-        self.gguf_writer.add_hidden_act("gelu_pytorch_tanh")
-
-        self.gguf_writer.add_sample_from_anchor(self.hparams.get("sample_from_anchor", True))
-        target_layers = self.dflash_config.get("target_layer_ids", self.hparams.get("target_layer_ids"))
-        if not target_layers:
-            raise ValueError("Gemma4 DSpark requires target_layer_ids")
-        self.gguf_writer.add_has_confidence_head(any("confidence_head.proj" in name for name in self.model_tensors))
-
-        if self.hparams.get("final_logit_softcapping"):
-            raise ValueError("Gemma4 DSpark logit softcapping is not supported")
-        # The top-level sliding_window is inert unless the draft enables SWA.
-        if self.dflash_config.get("use_swa", False):
-            window = self.dflash_config["swa_window_size"]
-            if window <= 0:
-                raise ValueError("Gemma4 DSpark swa_window_size must be positive")
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-        if not name.startswith("model."):
-            name = "model." + name
-        if name.endswith(".layer_scalar"):
-            name += ".weight"
-        name = name.replace("model.confidence_proj.", "model.confidence_head.proj.")
-        return super().filter_tensors((name, gen))
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # The shared DFlash map assigns this name to Qwen's pre-FFN norm.
-        if name.endswith(".post_attention_layernorm.weight"):
-            name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_POST_NORM, bid)
-        elif name.endswith(".pre_feedforward_layernorm.weight"):
-            name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_NORM, bid)
-        yield from super().modify_tensors(data_torch, name, bid)
-
-    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        if self.rope_parameters["rope_type"] == "proportional":
-            # Keep the unrotated dimensions in place, as in the Gemma4 converter.
-            head_dim = int(self.hparams["global_head_dim"])
-            fraction_value = self.rope_parameters.get("partial_rotary_factor", 0.25)
-            if not isinstance(fraction_value, (int, float)):
-                raise ValueError("Gemma4 DSpark partial_rotary_factor must be numeric")
-            fraction = float(fraction_value)
-            n_rot = int(head_dim * fraction / 2)
-            if not 0 < fraction <= 1 or head_dim * fraction != 2 * n_rot:
-                raise ValueError("Gemma4 DSpark rotary dimension count must be positive and even")
-            factors = torch.tensor([1.0] * n_rot + [1e30] * (head_dim // 2 - n_rot), dtype=torch.float32)
-            yield self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), factors
-
-
@ModelBase.register("Gemma4UnifiedForConditionalGeneration")
@ModelBase.example("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")
 class Gemma4UnifiedModel(Gemma4Model):
@@ -402,197 +402,3 @@ class SolarOpenModel(Glm4MoeModel):
        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<unk>"])  # ty: ignore[unresolved-attribute]
        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|startoftext|>"])  # ty: ignore[unresolved-attribute]
        special_vocab.add_to_gguf(self.gguf_writer)
-
-
-@ModelBase.register("Glm5NextForConditionalGeneration")
-@ModelBase.example("zai-org/GLM-5.3-Flash")
-class Glm5NextModel(TextModel):
-
-    model_arch = gguf.MODEL_ARCH.GLM5_NEXT
-    supports_mtp_export = True
-
-    _experts: list[dict[str, Tensor]] | None = None
-    _n_main_layers: int | None = None
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        self.n_nextn_layers = self.hparams.get("num_nextn_predict_layers", 0)
-        self.skip_mtp = self.no_mtp or self.n_nextn_layers == 0
-
-        if not self.skip_mtp:
-            self.block_count += self.n_nextn_layers
-            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-        self.hparams.pop("head_dim", None)
-
-    def set_vocab(self):
-        # requires transformers >= 5, tokpre hash-resolves to glm4
-        return self._set_vocab_glm()
-
-    def index_tensors(self, remote_hf_model_id: str | None = None):
-        hp = self.hparams.get("text_config", self.hparams)
-        type(self)._n_main_layers = hp["num_hidden_layers"]
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
-
-    @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
-
-        assert cls._n_main_layers is not None
-        m = re.match(r"model\.layers\.(\d+)\.", name)
-        is_mtp = m is not None and int(m.group(1)) >= cls._n_main_layers
-
-        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):
-        super().set_gguf_parameters()
-        hp = self.hparams
-
-        layer_types = hp["layer_types"]
-        n_kv_heads = [0 if t == "linear_attention" else 1 for t in layer_types]
-        assert len(n_kv_heads) == hp["num_hidden_layers"]
-        # Pad to block_count
-        n_kv_heads += [1] * (self.block_count - len(n_kv_heads))
-        self.gguf_writer.add_head_count_kv(n_kv_heads)
-        self.gguf_writer.add_vocab_size(hp["vocab_size"])
-        self.gguf_writer.add_layer_norm_eps(1e-6)
-
-        if not self.skip_mtp:
-            self.gguf_writer.add_nextn_predict_layers(self.n_nextn_layers)
-
-        # KDA
-        lin = hp["linear_attn_config"]
-        assert lin["num_heads"] == hp["num_attention_heads"]
-        self.gguf_writer.add_ssm_conv_kernel(lin["short_conv_kernel_size"])
-        self.gguf_writer.add_kda_head_dim(lin["head_dim"])
-        if (lb := lin.get("gate_lower_bound")) is not None:
-            self.gguf_writer.add_kda_gate_lower_bound(lb)
-
-        # MLA (nope only)
-        assert hp.get("mla_use_nope") and hp["qk_rope_head_dim"] == 0, "expected nope-only MLA"
-        kv_lora_rank = hp["kv_lora_rank"]
-        qk_rope = hp["qk_rope_head_dim"]
-        self.gguf_writer.add_q_lora_rank(hp["q_lora_rank"])
-        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
-        self.gguf_writer.add_rope_dimension_count(qk_rope)
-        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope)
-        self.gguf_writer.add_value_length(kv_lora_rank)
-        self.gguf_writer.add_key_length_mla(hp["qk_nope_head_dim"] + qk_rope)
-        self.gguf_writer.add_value_length_mla(hp["v_head_dim"])
-
-        # DSA indexer with k-pool compression
-        self.gguf_writer.add_indexer_head_count(hp["index_n_heads"])
-        self.gguf_writer.add_indexer_key_length(hp["index_head_dim"])
-        self.gguf_writer.add_indexer_top_k(hp["index_topk"])
-        self.gguf_writer.add_indexer_kpool(hp["index_kpool"])
-        self.gguf_writer.add_indexer_kpool_select_tail(hp.get("index_kpool_always_select_tail", True))
-        if (indexer_types := hp.get("indexer_types")) is not None:
-            self.gguf_writer.add_indexer_types([t == "full" for t in indexer_types])
-
-        # mHC
-        assert hp.get("mhc", True)
-        self.gguf_writer.add_hyper_connection_count(hp["hc_mult"])
-        self.gguf_writer.add_hyper_connection_sinkhorn_iterations(hp["hc_sinkhorn_iters"])
-        self.gguf_writer.add_hyper_connection_epsilon(hp["hc_eps"])
-
-        # MoE
-        self.gguf_writer.add_leading_dense_block_count(hp["first_k_dense_replace"])
-        self.gguf_writer.add_expert_feed_forward_length(hp["moe_intermediate_size"])
-        self.gguf_writer.add_expert_shared_count(hp["n_shared_experts"])
-        self.gguf_writer.add_expert_weights_scale(hp["routed_scaling_factor"])
-        self.gguf_writer.add_expert_weights_norm(hp["norm_topk_prob"])
-        if (limit := hp.get("swiglu_limit")) is not None:
-            self.gguf_writer.add_swiglu_clamp_exp([limit] * self.block_count)
-            self.gguf_writer.add_swiglu_clamp_shexp([limit] * self.block_count)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if name == "lm_head.weight" and self.hparams.get("tie_word_embeddings", False):
-            return
-
-        # routed experts
-        if ".mlp.experts." in name:
-            n_experts = self.hparams["n_routed_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:
-                return
-            for w_name in ("down_proj", "gate_proj", "up_proj"):
-                datas: list[Tensor] = []
-                for xid in range(n_experts):
-                    ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                    datas.append(self._experts[bid].pop(ename))
-                merged = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                yield from super().modify_tensors(torch.stack(datas, dim=0), merged, bid)
-            return
-
-        # MLA absorption
-        if name.endswith("kv_b_proj.weight"):
-            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)
-            yield from super().modify_tensors(k_b.transpose(1, 2), name.replace("kv_b_proj", "k_b_proj"), bid)
-            yield from super().modify_tensors(v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
-            return
-
-        # KDA conv1d
-        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
-            if data_torch.ndim == 3:
-                d_inner, _, d_conv = data_torch.shape
-            elif data_torch.ndim == 2:
-                d_inner, d_conv = data_torch.shape
-            else:
-                raise ValueError(f"unexpected conv1d rank {data_torch.ndim} for {name}")
-            data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
-
-        if name.endswith(".A_log"):
-            n_head = self.hparams["num_attention_heads"]
-            data_torch = -torch.exp(data_torch.float().flatten()[:n_head])
-
-        if name.endswith(".dt_bias"):
-            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
-
-        if re.search(r"\.(hc_(?:attn|ffn)_(?:fn|base|scale)|index_kpool_compress_(?:ape|gate))$", name):
-            yield self.map_tensor_name(name) + ".weight", data_torch
-            return
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        # keep the small mHC / gating parameters exact
-        exact_keys = ("hc_attn_", "hc_ffn_", "indexer_compressor_", "ssm_a", "ssm_dt", "exp_probs_b")
-        if new_name.startswith(("blk.", "output_hc")) and any(k in new_name for k in exact_keys):
-            return gguf.GGMLQuantizationType.F32
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
-    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 prepare_tensors(self):
-        super().prepare_tensors()
-        if self._experts is not None:
-            leftover = [k for d in self._experts for k in d.keys()]
-            if leftover:
-                raise ValueError(f"Unprocessed experts: {leftover}")
@@ -159,14 +159,32 @@ class HunYuanMoEModel(TextModel):
 class HunYuanModel(TextModel):
    model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE

-    def set_vocab(self):
-        # Also called by draft models (e.g. DFlash), with dir_model pointing at
-        # the target model.
-        config = ModelBase.load_hparams(self.dir_model, self.is_mistral_format)
-        config = {**config, **config.get("text_config", {})}
-        self.hparams["pad_token_id"] = config.get("pad_token_id")
-        self.hparams["eod_token_id"] = config.get("eod_token_id")
+    def _get_eod_token_id(self) -> int | None:
+        """Get the actual end-of-generation token from config (eod_token_id)."""
+        return self.hparams.get("eod_token_id")

+    def _get_eot_token_id(self) -> int | None:
+        """Get the end-of-turn token from generation_config.json.
+        This is the first entry in eos_token_id when it's a list."""
+        gen_cfg_path = self.dir_model / "generation_config.json"
+        if gen_cfg_path.is_file():
+            with open(gen_cfg_path, encoding="utf-8") as f:
+                gen_cfg = json.load(f)
+            eos = gen_cfg.get("eos_token_id")
+            if isinstance(eos, list) and len(eos) >= 2:
+                return eos[0]
+        return None
+
+    def _fix_special_tokens(self):
+        """Fix EOS/EOT tokens that are incorrect in upstream configs."""
+        eod_id = self._get_eod_token_id()
+        if eod_id is not None:
+            self.gguf_writer.add_eos_token_id(eod_id)
+        eot_id = self._get_eot_token_id()
+        if eot_id is not None:
+            self.gguf_writer.add_eot_token_id(eot_id)
+
+    def set_vocab(self):
        if (self.dir_model / "tokenizer.json").is_file():
            tokens, toktypes, tokpre = self.get_vocab_base()
            self.gguf_writer.add_tokenizer_model("gpt2")
@@ -181,6 +199,7 @@ class HunYuanModel(TextModel):
                token_types = ('bos', 'eos', 'unk', 'sep', 'cls', 'mask')
            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True, special_token_types=token_types)
            special_vocab.add_to_gguf(self.gguf_writer)
+            self._fix_special_tokens()
        else:
            from transformers import AutoTokenizer
            tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
@@ -232,18 +251,7 @@ class HunYuanModel(TextModel):
            # FIX for BOS token: Overwrite incorrect id read from config.json
            if self.hparams['hidden_size'] == 4096:
                self.gguf_writer.add_bos_token_id(127958) # only for 7b dense, fix <|bos|> token
-
-        # Fix EOS/EOT tokens that are incorrect in upstream configs.
-        eod_id = self.hparams.get("eod_token_id")
-        if eod_id is not None:
-            self.gguf_writer.add_eos_token_id(eod_id)
-
-        gen_cfg = self.dir_model / "generation_config.json"
-        if gen_cfg.is_file():
-            with open(gen_cfg, encoding="utf-8") as f:
-                eos = json.load(f).get("eos_token_id")
-            if isinstance(eos, list) and len(eos) >= 2:
-                self.gguf_writer.add_eot_token_id(eos[0])
+            self._fix_special_tokens()

    def set_gguf_parameters(self):
        # Some HunYuanVL variants set num_experts=1 (not real MoE);
@@ -2,14 +2,15 @@ from __future__ import annotations

 import re
 from pathlib import Path
-from typing import Iterable, Iterator, TYPE_CHECKING
+from typing import Callable, Iterable, Iterator, TYPE_CHECKING

+import numpy as np
 import torch

 if TYPE_CHECKING:
    from torch import Tensor

-from .base import ModelBase, TextModel, gguf, logger
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf, logger

 from .kimi_linear import KimiLinearModel

@@ -103,6 +104,36 @@ class KimiK3Model(TextModel):
                "only the routed experts have a repack path"
            )

+    def _mxfp4_expert_tensor(self, loaders: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]):
+        """
+        One stacked [n_expert, rows, cols] MXFP4 tensor, built lazily.
+
+        gguf_writer holds every added tensor until the final write, so building
+        this eagerly (like the DeepSeek-V4 path does) keeps all ~1.38 TB of
+        experts in memory. lazy means only the tensor being written is resident.
+        """
+        # meta shapes, so this does not read any weights
+        rows, packed_cols = loaders[0][0]().shape
+        n_blocks = (packed_cols * 2) // 32
+        byte_shape = (len(loaders), rows, n_blocks * 17)
+
+        def load(fns: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]) -> np.ndarray:
+            out = np.empty(byte_shape, dtype=np.uint8)
+            for eid, (packed_fn, scale_fn) in enumerate(fns):
+                out[eid] = self.repack_mxfp4_blocks(
+                    LazyTorchTensor.to_eager(packed_fn()),
+                    LazyTorchTensor.to_eager(scale_fn()),
+                )
+            return out
+
+        # loaders goes through args, not the closure, so that `func` matches
+        # LazyBase's single-argument shape
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(np.uint8, byte_shape),
+            args=(loaders,),
+            func=load,
+        )
+
    def _write_mxfp4_experts(self) -> None:
        n_experts = self.hparams["num_experts"]

@@ -10,14 +10,13 @@ import torch
 if TYPE_CHECKING:
    from torch import Tensor

-from .base import MmprojModel, ModelBase, TextModel, gguf, logger
+from .base import MmprojModel, ModelBase, TextModel, gguf


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

    # MiMo V2-Flash, V2.5 and V2.5-Pro all ship 3 trained MTP layers under model.mtp.layers.{0,1,2}.
    # The HF config does not expose the count, so it's hardcoded to match the count found in the safetensors.
@@ -26,8 +25,6 @@ class MimoV2Model(TextModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

-        if self.no_mtp:
-            self._n_nextn = 0
        self.block_count = self.hparams["num_hidden_layers"] + self._n_nextn
        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)

@@ -104,7 +101,7 @@ class MimoV2Model(TextModel):
        qkv_overrides: dict[str, tuple[Callable, Callable, int]] = {}
        qc = self.hparams.get("quantization_config")
        if isinstance(qc, dict) and qc.get("quant_method") == "fp8":
-            pat = re.compile(r"^model\.(mtp\.)?layers\.(\d+)\.self_attn\.qkv_proj\.weight_scale_inv$")
+            pat = re.compile(r"^model\.layers\.(\d+)\.self_attn\.qkv_proj\.weight_scale_inv$")
            for name in list(self.model_tensors.keys()):
                m = pat.match(name)
                if not m:
@@ -112,13 +109,10 @@ class MimoV2Model(TextModel):
                weight_name = name.removesuffix("_scale_inv")
                if weight_name not in self.model_tensors:
                    continue
-                bid = int(m.group(2))
-                if m.group(1) is not None:
-                    bid += self.hparams["num_hidden_layers"]
                qkv_overrides[weight_name] = (
                    self.model_tensors[weight_name],
                    self.model_tensors[name],
-                    bid,
+                    int(m.group(1)),
                )

        super().dequant_model()
@@ -171,86 +165,7 @@ class MimoV2Model(TextModel):
        if v_scale is not None:
            self.gguf_writer.add_attn_value_scale(float(v_scale))

-        if self._n_nextn > 0:
-            self.gguf_writer.add_nextn_predict_layers(self._n_nextn)
-
-    _MXFP4_EXPERT_RE = re.compile(
-        r"^model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.(gate|up|down)_proj\.weight$"
-    )
-    _MXFP4_PROJ = {
-        "gate": gguf.MODEL_TENSOR.FFN_GATE_EXP,
-        "up":   gguf.MODEL_TENSOR.FFN_UP_EXP,
-        "down": gguf.MODEL_TENSOR.FFN_DOWN_EXP,
-    }
-
-    def _is_mxfp4_packed(self) -> bool:
-        quant_config = self.hparams.get("quantization_config") or {}
-        if quant_config.get("store_dtype") != "mxfp4":
-            return False
-        # repack_mxfp4_blocks assumes ggml's 32-element group
-        block_size = quant_config.get("mxfp4_block_size", 32)
-        if block_size != 32:
-            raise NotImplementedError(
-                f"MXFP4 block size {block_size} is not ggml's QK_MXFP4 (32)")
-        return True
-
-    def _write_mxfp4_experts(self) -> None:
-        n_experts = self.hparams["n_routed_experts"]
-
-        # the FP8 half uses `weight_scale_inv` and is left to dequant_model
-        stray = [n for n in self.model_tensors
-                 if n.endswith(".weight_scale") and not self._MXFP4_EXPERT_RE.match(n.removesuffix("_scale"))]
-        if stray:
-            raise NotImplementedError(
-                f"{len(stray)} MXFP4 tensor(s) outside the routed experts, e.g. {stray[0]!r}; "
-                "only the routed experts have a repack path"
-            )
-
-        # (bid, proj) -> {expert id: (weight name, scale name)}
-        groups: dict[tuple[int, str], dict[int, tuple[str, str]]] = {}
-        for name in self.model_tensors:
-            m = self._MXFP4_EXPERT_RE.match(name)
-            if m is None:
-                continue
-            bid, eid, proj = int(m.group(1)), int(m.group(2)), m.group(3)
-            scale_name = name + "_scale"
-            if scale_name not in self.model_tensors:
-                raise KeyError(f"missing {scale_name} for {name}")
-            groups.setdefault((bid, proj), {})[eid] = (name, scale_name)
-
-        consumed: list[str] = []
-        for (bid, proj), experts in sorted(groups.items()):
-            missing = [e for e in range(n_experts) if e not in experts]
-            if missing or len(experts) != n_experts:
-                raise KeyError(
-                    f"layer {bid} {proj}_proj: {len(experts)} of {n_experts} experts present"
-                    + (f", first missing is {missing[0]}" if missing else "")
-                )
-
-            loaders = []
-            for eid in range(n_experts):
-                weight_name, scale_name = experts[eid]
-                loaders.append((self.model_tensors[weight_name], self.model_tensors[scale_name]))
-                consumed += [weight_name, scale_name]
-
-            data = self._mxfp4_expert_tensor(loaders)
-            new_name = self.format_tensor_name(self._MXFP4_PROJ[proj], bid)
-            shape = gguf.quant_shape_from_byte_shape(data.shape, gguf.GGMLQuantizationType.MXFP4)
-            logger.info(
-                f"{new_name}: repacked {n_experts} experts to MXFP4, "
-                f"shape = {{{', '.join(str(n) for n in reversed(shape))}}}"
-            )
-            self.gguf_writer.add_tensor(new_name, data, raw_dtype=gguf.GGMLQuantizationType.MXFP4)
-
-        for name in consumed:
-            del self.model_tensors[name]
-
-    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        # not a generator on purpose: base.py chains this with get_tensors(), so the
-        # tensors used here must be removed from model_tensors before that starts
-        if self._is_mxfp4_packed():
-            self._write_mxfp4_experts()
-        return ()
+        self.gguf_writer.add_nextn_predict_layers(self._n_nextn)

    _experts: list[dict[str, Tensor]] | None = None

@@ -258,32 +173,11 @@ class MimoV2Model(TextModel):
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item

-        is_mtp = name.startswith("model.mtp.layers.")
-        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
-
        if "attention_sink" in name and not name.endswith(".weight"):
            name += ".weight"

        return super().filter_tensors((name, gen))

-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
    def modify_tensors(self, data_torch, name, bid):
        # Remap MTP/NextN tensors to additional layer slots so the standard tensor map handles them.
        # HF: model.mtp.layers.{i}.foo  ->  model.layers.{n_layer_text + i}.foo
@@ -298,7 +192,7 @@ class MimoV2Model(TextModel):
            bid = new_bid

        # process the experts separately
-        if ".mlp.experts." in name and name.endswith(".weight"):
+        if name.find("mlp.experts") != -1:
            n_experts = self.hparams["n_routed_experts"]
            assert bid is not None

@@ -335,10 +229,6 @@ class MimoV2Model(TextModel):
            if len(experts) > 0:
                raise ValueError(f"Unprocessed experts: {experts}")

-        if self._is_mxfp4_packed():
-            self._is_mxfp4 = True
-            self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
-

@ModelBase.register("MiMoV2ForCausalLM")
@ModelBase.example("XiaomiMiMo/MiMo-V2.5")
@@ -492,8 +382,6 @@ class MiMoV2VisionAudioModel(MmprojModel):
            "_codebook.inited",
        )
        for name, tensor in state_dict.items():
-            if name.startswith("decoder."):
-                continue
            if name.endswith(skip_suffixes):
                continue
            if m := codebook_re.match(name):
@@ -424,6 +424,12 @@ class NemotronHModel(GraniteHybridModel):
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        special_vocab.add_to_gguf(self.gguf_writer)

+        # The tokenizer _does_ add a BOS token (via post_processor type
+        # TemplateProcessing) but does not set add_bos_token to true in the
+        # config, so we need to explicitly override it here.
+        if not self.is_moe:
+            self.gguf_writer.add_add_bos_token(True)
+
    _MTP_SPECIAL_RENAMES = {
        "mtp.layers.0.enorm.weight":           "model.layers.{bid}.enorm.weight",
        "mtp.layers.0.hnorm.weight":           "model.layers.{bid}.hnorm.weight",
@@ -154,21 +154,6 @@ class Plamo2Model(TextModel):
 class Plamo3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.PLAMO3

-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # PLaMo-3 builds rope_parameters from flat config keys at runtime; mirror the YaRN settings for GGUF.
-        rope_scaling_factor = self.hparams.get("rope_scaling_factor", 1)
-        if rope_scaling_factor != 1 and "rope_type" not in self.rope_parameters:
-            self.rope_parameters.update({
-                "rope_type": "yarn",
-                "factor": float(rope_scaling_factor),
-                "original_max_position_embeddings": int(self.hparams["initial_context_length"]),
-                "beta_fast": 32.0,
-                "beta_slow": 1.0,
-                "truncate": False,
-            })
-
    def set_vocab(self):
        self._set_vocab_plamo()

@@ -10,7 +10,7 @@ import torch
 if TYPE_CHECKING:
    from torch import Tensor

-from .base import LazyTorchTensor, ModelBase, ModelType, TextModel, get_model_architecture, gguf, logger
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf, logger


@ModelBase.register("QWenLMHeadModel")
@@ -469,21 +469,6 @@ class _LinearAttentionVReorderBase(Qwen3NextModel):
        shape = list(tensor.shape)
        if dim < 0:
            dim += len(shape)
-
-        # LoRA tensors (W ≈ B @ A) cannot reshape their row dimension.
-        # Instead, build a permutation index and apply it to A (column reorder) or B (row reorder) directly.
-        if hasattr(tensor, 'get_lora_A_B'):
-            n = shape[dim]
-            idx = torch.arange(n).reshape(num_k_heads, num_v_per_k, head_dim)
-            idx = idx.permute(1, 0, 2).contiguous().reshape(n)
-            lora_A, lora_B = tensor.get_lora_A_B()  # ty: ignore[call-non-callable]
-            if dim == len(shape) - 1:
-                return type(tensor)(lora_A[:, idx], lora_B)
-            elif dim == 0:
-                return type(tensor)(lora_A, lora_B[idx])
-            else:
-                raise NotImplementedError(f"_reorder_v_heads on dim={dim} not supported for LoRA tensors")
-
        new_shape = shape[:dim] + [num_k_heads, num_v_per_k, head_dim] + shape[dim + 1:]
        tensor = tensor.reshape(*new_shape)
        perm = list(range(len(new_shape)))
@@ -681,7 +666,7 @@ class DFlashModel(Qwen3Model):
        from . import get_model_class
        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
            target_hparams = json.load(f)
-        target_arch = get_model_architecture(target_hparams, ModelType.TEXT)
+            target_arch = target_hparams["architectures"][0]
        target_cls = get_model_class(target_arch)

        if target_cls is not type(self):
@@ -701,12 +686,6 @@ class DFlashModel(Qwen3Model):
        super().set_gguf_parameters()

        dflash_config = self.hparams.get("dflash_config", {})
-        if (partial_rotary_factor := self.rope_parameters.get("partial_rotary_factor")) is not None:
-            head_dim = self.hparams.get("head_dim") or self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-            self.gguf_writer.add_rope_dimension_count(int(head_dim * partial_rotary_factor))
-        if (value_scale := dflash_config.get("attention_value_scale")) is not None:
-            self.gguf_writer.add_attn_value_scale(float(value_scale))
-
        block_size = dflash_config.get("block_size", self.hparams.get("block_size", 16))
        self.gguf_writer.add_block_size(block_size)

@@ -732,7 +711,7 @@ class DFlashModel(Qwen3Model):
        if embedding_scale is not None:
            self.gguf_writer.add_embedding_scale(float(embedding_scale))

-        target_layer_ids = dflash_config.get("target_layer_ids", self.hparams.get("target_layer_ids", []))
+        target_layer_ids = dflash_config.get("target_layer_ids", [])
        if target_layer_ids:
            extract_layer_ids = [i + 1 for i in target_layer_ids]
            self.gguf_writer.add_target_layers(extract_layer_ids)
@@ -740,9 +719,8 @@ class DFlashModel(Qwen3Model):
        use_sliding_window = self.hparams.get("use_sliding_window", False) or dflash_config.get("use_swa", False)
        sliding_window = dflash_config.get("swa_window_size") or self.hparams.get("sliding_window")
        layer_types = self.hparams.get("layer_types")
-        if use_sliding_window and sliding_window:
-            is_swa = ([True] * self.block_count if dflash_config.get("use_swa", False)
-                      else [lt == "sliding_attention" for lt in layer_types or []])
+        if use_sliding_window and sliding_window and layer_types:
+            is_swa = [lt == "sliding_attention" for lt in layer_types]
            self.gguf_writer.add_sliding_window(sliding_window)
            self.gguf_writer.add_sliding_window_pattern(is_swa)

@@ -758,62 +736,6 @@ class DFlashModel(Qwen3Model):
            head_dim = self.hparams.get("head_dim") or self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
            self.gguf_writer.add_rope_dimension_sections([head_dim // 2, 0, 0, 0])

-    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        yield from super().generate_extra_tensors()
-
-        mask_path = self.dir_model / "mask_embedding.pt"
-        if not mask_path.is_file():
-            return
-
-        mask = torch.load(mask_path, map_location="cpu", weights_only=True)
-        mask_id = self.hparams.get("dflash_config", {}).get("mask_token_id")
-        if mask_id is None or mask["mask_token_id"] != mask_id:
-            raise ValueError("mask_embedding.pt mask_token_id does not match dflash_config")
-        if tuple(mask["embedding"].shape) != (self.hparams["hidden_size"],):
-            raise ValueError("mask_embedding.pt has an unexpected embedding shape")
-        if not 0 <= mask_id < self.hparams["vocab_size"]:
-            raise ValueError("mask_embedding.pt mask_token_id is outside the vocabulary")
-
-        def target_tensor(name: str) -> Tensor:
-            if self.target_model_dir is None:
-                raise ValueError("mask_embedding.pt requires --target-model-dir with the target embeddings and output head")
-            index_path = self.target_model_dir / "model.safetensors.index.json"
-            if index_path.is_file():
-                with open(index_path, encoding="utf-8") as f:
-                    weight_map = json.load(f)["weight_map"]
-                part_names = [weight_map[name]]
-            else:
-                part_names = self.get_model_part_names(self.target_model_dir, "model", ".safetensors")
-
-            for part_name in part_names:
-                with gguf.utility.SafetensorsLocal(self.target_model_dir / part_name) as part:
-                    if name in part:
-                        return LazyTorchTensor.from_local_tensor(part[name])
-            raise ValueError(f"Target tensor {name!r} was not found in safetensors")
-
-        embedding_name = "model.embed_tokens.weight"
-        if embedding_name in self.model_tensors:
-            embeddings = self.model_tensors.pop(embedding_name)()
-        else:
-            embeddings = target_tensor(embedding_name)
-
-        if "model.lm_head.weight" not in self.model_tensors:
-            if self.target_model_dir is None:
-                raise ValueError("mask_embedding.pt requires --target-model-dir to obtain the output head")
-            target_config = ModelBase.load_hparams(self.target_model_dir, False)
-            target_config = {**target_config, **target_config.get("text_config", {})}
-            head_name = embedding_name if target_config.get("tie_word_embeddings", False) else "lm_head.weight"
-            # Keep the output head separate from the patched input embedding table.
-            yield "model.lm_head.weight", target_tensor(head_name)
-
-        embeddings = LazyTorchTensor.to_eager(embeddings).clone()
-        if tuple(embeddings.shape) != (self.hparams["vocab_size"], self.hparams["hidden_size"]):
-            raise ValueError("Target token embedding shape does not match the DFlash draft")
-        # MiMo's target mask row is untrained; the draft provides its own vector.
-        embeddings[mask_id] = mask["embedding"].to(embeddings.dtype)
-        self.hparams["has_embed_tokens"] = True
-        yield embedding_name, embeddings
-
    def _target_uses_mrope(self) -> bool:
        if self.target_model_dir is None:
            return False
@@ -918,6 +840,13 @@ class DSparkModel(DFlashModel):
            return None
        return super().filter_tensors(item)

+    _ROPE_PERMUTE_SUFFIXES = (
+        "self_attn.q_proj.weight",
+        "self_attn.k_proj.weight",
+        "self_attn.q_norm.weight",
+        "self_attn.k_norm.weight",
+    )
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name == "model.d2t":
            self._d2t = data_torch
@@ -926,6 +855,12 @@ class DSparkModel(DFlashModel):
        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith("lm_head.weight"):
            return

+        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
+        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
+            head_dim = self.hparams["head_dim"]
+            shape = data_torch.shape
+            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
+
        yield from super().modify_tensors(data_torch, name, bid)

    def prepare_tensors(self):
@@ -228,12 +228,10 @@ class Qwen3ASRMmprojModel(Qwen3OmniMmprojModel):
@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")
@ModelBase.example("zai-org/GLM-4.1V-9B-Thinking", "zai-org/GLM-4.5V")
 class Glm4VVisionModel(Qwen3VLVisionModel):
-    projector_type = gguf.VisionProjectorType.GLM4V
-
    def set_gguf_parameters(self):
        MmprojModel.set_gguf_parameters(self) # skip Qwen3VLVisionModel parameters
        assert self.hparams_vision is not None
-        self.gguf_writer.add_clip_projector_type(self.projector_type)
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GLM4V)

        hidden_act = str(self.hparams_vision.get("hidden_act", "")).lower()
        if hidden_act == "gelu":
@@ -251,32 +249,6 @@ class Glm4VVisionModel(Qwen3VLVisionModel):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Glm5NextForConditionalGeneration")
-@ModelBase.example("zai-org/GLM-5.3-Flash")
-class Glm5NextVisionModel(Glm4VVisionModel):
-    # GLM-5.3-Flash vision tower. glm4v layout with per-head qk-norm, no post-conv norm and no learned position embeddings.
-    # Images are placed on a ceil aligned canvas with padding.
-
-    projector_type = gguf.VisionProjectorType.GLM5V
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        assert self.hparams_vision is not None
-        self.gguf_writer.add_vision_spatial_merge_size(int(self.hparams_vision.get("spatial_merge_size", 2)))
-        if (limit := self.hparams_vision.get("swiglu_limit")) is not None:
-            self.gguf_writer.add_vision_swiglu_clamp(float(limit))
-
-        # image token budget from the processor, stored as single-frame pixel counts
-        pc = self.preprocessor_config
-        patch = int(pc.get("patch_size", 14))
-        merge = int(pc.get("merge_size", 2))
-        pixels_per_token = (patch * merge) ** 2
-        if (min_tok := pc.get("min_image_tokens")) is not None:
-            self.gguf_writer.add_vision_min_pixels(int(min_tok) * pixels_per_token)
-        if (max_tok := pc.get("max_image_tokens")) is not None:
-            self.gguf_writer.add_vision_max_pixels(int(max_tok) * pixels_per_token)
-
-
@ModelBase.register("Qwen3VLForConditionalGeneration")
@ModelBase.example("Qwen/Qwen3-VL-4B-Instruct")
 class Qwen3VLTextModel(Qwen3Model):
@@ -25,34 +25,15 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):

    model_arch = gguf.MODEL_ARCH.QWEN4EXP

-    # the MTP head: one full-attention QSA block after the trunk, fed by the trunk's hc-wide residual
-    supports_mtp_export = True
-
-    # MTP tensors the shared Qwen remapper does not know
-    _MTP_EXTRA = {
-        "fc_embedding":           "nextn_fc_embedding",
-        "fc_hidden":              "nextn_fc_hidden",
-        "hyper_connection_mixer": "nextn_hc_head",
-    }
+    # the MTP block is a separate draft head; vLLM drops it too
+    supports_mtp_export = False
+    no_mtp = True

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        # only the shard names, so the table itself is never held
        self._ple_shards: dict[int, str] = {}
        self._ple_row_dim: int | None = None
-        self._mtp_fc: dict[str, Tensor] = {}
-
-    @classmethod
-    def filter_tensors(cls, item):
-        name, gen = item
-        part = name.split(".")[1] if name.startswith("mtp.") else None
-        if part in cls._MTP_EXTRA:
-            if cls.no_mtp:
-                return None
-            assert cls._original_block_count is not None
-            rest = name.split(".", 2)[2]
-            return f"model.layers.{cls._original_block_count}.{cls._MTP_EXTRA[part]}.{rest}", gen
-        return super().filter_tensors(item)

    def _read_hash_constants(self, suffix: str) -> list[int]:
        """Read an int64 PLE constant straight from the checkpoint.
@@ -82,17 +63,14 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
        self.gguf_writer.add_indexer_top_k(hp["indexer_budget"])
        ratio = hp["indexer_compress_ratio"]
        layer_types = hp["layer_types"]
-        # the MTP block is a full-attention QSA layer too
        self.gguf_writer.add_attention_compress_ratios(
            [ratio if layer_types[i] == "full_attention" else 0 for i in range(n_layer)]
-            + [ratio] * (self.block_count - n_layer)
        )

        # ple_layer_ids is 1-based in the HF config; empty means no n-gram table,
        # so emit no PLE keys rather than optional ones
-        # the MTP head never reads PLE, so an MTP-only file carries none of it
        ple_layers = [i - 1 for i in hp["ple_layer_ids"]]
-        if not ple_layers or self.mtp_only:
+        if not ple_layers:
            return
        self.gguf_writer.add_ple_layers(ple_layers)
        self.gguf_writer.add_ple_ngram_size(hp["ngram_size"])
@@ -142,14 +120,6 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
        if ".ngram_embedding.shard_" in name:
            return self._place_ple_shard(data_torch, name)

-        # eh_proj([e ; h_s]) = fc_embedding(e) + fc_hidden(h_s) for every hc stream s
-        if name.endswith((".nextn_fc_embedding.weight", ".nextn_fc_hidden.weight")):
-            self._mtp_fc[name.rsplit(".", 2)[1]] = data_torch
-            if len(self._mtp_fc) < 2:
-                return []
-            eh = torch.cat([self._mtp_fc.pop("nextn_fc_embedding"), self._mtp_fc.pop("nextn_fc_hidden")], dim=1)
-            return [(self.format_tensor_name(gguf.MODEL_TENSOR.NEXTN_EH_PROJ, bid, ".weight"), eh)]
-
        # one projection feeds indexer q and k; split it, as minimax-m3 does
        if ".indexer.index_qk_proj.weight" in name:
            n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
@@ -212,8 +182,6 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):

    def prepare_tensors(self):
        super().prepare_tensors()
-        if self._mtp_fc:
-            raise ValueError(f"MTP projection missing its other half: {sorted(self._mtp_fc)}")
        n_parts = self.hparams.get("split_ngram_parts", 0)
        if self._ple_shards and len(self._ple_shards) != n_parts:
            raise ValueError(
@@ -163,7 +163,6 @@ models = [
    {"name": "granite-embed-multi-311m", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-embedding-311m-multilingual-r2", },
    {"name": "mellum2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Base"},
    {"name": "laguna",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/poolside/Laguna-XS.2", },
-    {"name": "ufakzeka",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ufakai/ufakzeka-1", },
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -509,7 +509,6 @@ The following templates have active tests in `tests/test-chat.cpp`:
 | Kimi-K2 / Kimi-K2-Instruct | JSON_NATIVE | JSON tools with special markers |
 | Llama 3.1/3.2/3.3 | JSON_NATIVE | Standard Llama tool format |
 | OpenAI GPT-OSS | Specialized | Channel-based (dedicated handler) |
-| LLM-jp-4.1 | Specialized | GPT-OSS dialect (dedicated handler) |
 | Apriel 1.5 | JSON_NATIVE | `<tool_calls>` wrapper with JSON array |
 | Apriel 1.6 Thinker | Reasoning | Implicit reasoning start |
 | Mistral Small 3.2 | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` with call ID |
@@ -1,140 +0,0 @@
-# AMD AOCL-BLAS
-
-> [!NOTE]
-> The [ZenDNN backend](ZenDNN.md) is the recommended path for inference on AMD CPUs. Refer to its documentation for the currently supported operations and data types. This page covers AOCL-BLAS as a vendor option for the generic `GGML_BLAS` backend.
-
-AOCL-BLAS is AMD's BLAS library, optimized for AMD EPYC and Ryzen CPUs.
-llama.cpp can link against it through the existing BLAS backend (`GGML_BLAS`).
-
-The BLAS backend can use AOCL-BLAS for eligible large prompt GEMMs and generally does not participate in token generation.
-F32 weights are passed to `cblas_sgemm` directly. Other types are converted to F32 first, so a quantized model can be slower than the native CPU kernels.
-See [BLAS Build](../build.md#blas-build).
-
-AOCL download / install: https://www.amd.com/en/developer/aocl.html
-
-Use the Quick start for a short command list. The later sections cover install layout, the single-threaded tree, threading, and checks.
-
-### Quick start (MT, 64 threads)
-
-If AOCL is not installed yet, follow [Prepare](#prepare) first. Adjust the `amd-libs.cfg` path to your install. CMake flags are set once. Source `amd-libs.cfg` again in every new shell before you launch, or the loader will not find AOCL-BLAS.
-
-```bash
-source /opt/aocl/<version>/aocc/MT/amd-libs.cfg
-
-cmake -B build \
-  -DGGML_BLAS=ON \
-  -DGGML_BLAS_VENDOR=AOCL_mt \
-  -DBLAS_INCLUDE_DIRS="${AOCL_ROOT}/include" \
-  -DGGML_NATIVE=ON
-cmake --build build --config Release
-
-source /opt/aocl/<version>/aocc/MT/amd-libs.cfg
-./build/bin/llama-cli -m model.gguf -t 64
-```
-
-`-t 64` is the thread count to pass on each launch (`--threads 64` is the same flag). Details, the single-threaded tree, and NUMA binding are below.
-
-CMake vendor names `AOCL` and `AOCL_mt` are recognized by [FindBLAS](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) (CMake 3.27+).
-When either vendor is selected, llama.cpp enables the BLIS code path (`GGML_BLAS_USE_BLIS`): it includes `blis.h` and calls `bli_thread_set_num_threads()` before GEMM. The same vendors also set `GGML_BLAS_USE_AOCL`, which only changes the device description to `AOCL-BLAS`. Upstream BLIS (`FLAME`) sets `GGML_BLAS_USE_BLIS` alone, so its description stays `BLIS`.
-
-### Prepare
-
-1. Install AOCL from AMD (package or tarball). Current releases ship **ST** (single-threaded) and **MT** (multi-threaded) libraries in separate folders. The default `AOCL_ROOT` is the **MT** tree. A typical layout (replace `<version>` and `aocc` with your install):
-
-```
-<aocl-prefix>/<version>/aocc/MT/
-<aocl-prefix>/<version>/aocc/ST/
-```
-
-2. Source `amd-libs.cfg` from the tree you want. Current AOCL versions use this file (not a separate `aocl-env.sh`). Adjust the prefix, version, and compiler (`aocc` vs `gcc`) to match your install:
-
-```bash
-# Multi-threaded (default AOCL_ROOT):
-source /opt/aocl/<version>/aocc/MT/amd-libs.cfg
-
-# Single-threaded:
-# source /opt/aocl/<version>/aocc/ST/amd-libs.cfg
-```
-
-This sets library and include paths so the linker can find AOCL-BLAS. Skipping it is a common cause of BLAS not found / unresolved symbol errors.
-
-Optional, if your install provides an environment module:
-
-```bash
-cd /opt/aocl/<version>/aocc/MT
-module load ./aocl-linux-aocc-<version>_module
-# module unload ./aocl-linux-aocc-<version>_module
-```
-
-3. Prefer the **MT** libraries for llama.cpp. Use `-DGGML_BLAS_VENDOR=AOCL_mt` after sourcing the MT `amd-libs.cfg`. Use `-DGGML_BLAS_VENDOR=AOCL` if you sourced the ST tree instead.
-
-### llama.cpp compilation
-
-Requires **CMake 3.27 or newer** for `-DGGML_BLAS_VENDOR=AOCL` / `AOCL_mt`.
-
-FindBLAS does not detect AOCL headers via pkg-config. After sourcing `amd-libs.cfg`, `AOCL_ROOT` is set and `$AOCL_ROOT/include` is a symlink to the active integer ABI (`include_LP64` by default). Pass that path to CMake:
-
-```bash
-source /opt/aocl/<version>/aocc/MT/amd-libs.cfg   # adjust path
-
-cmake -B build \
-  -DGGML_BLAS=ON \
-  -DGGML_BLAS_VENDOR=AOCL_mt \
-  -DBLAS_INCLUDE_DIRS="${AOCL_ROOT}/include" \
-  -DGGML_NATIVE=ON
-
-cmake --build build --config Release
-```
-
-#### CMake older than 3.27
-
-`AOCL` / `AOCL_mt` may be unknown to FindBLAS. After sourcing the AOCL env, you can try:
-
-```bash
-cmake -B build \
-  -DGGML_BLAS=ON \
-  -DGGML_BLAS_VENDOR=Generic \
-  -DBLAS_LIBRARIES="-lblis -lm" \
-  -DBLAS_INCLUDE_DIRS="${AOCL_ROOT}/include" \
-  -DGGML_NATIVE=ON
-```
-
-Library names differ between AOCL packages (`blis`, `blis-mt`, etc.). Pass whatever your install provides.
-`GGML_BLAS_VENDOR=Generic` does not enable the BLIS header and thread path (`GGML_BLAS_USE_BLIS`). Upgrade CMake so `AOCL` or `AOCL_mt` is recognized.
-
-### llama.cpp execution
-
-`--threads` / `--threads-batch` are the thread budget for **every** backend that implements `set_n_threads` (CPU and BLAS).
-
-On each large BLAS `MUL_MAT`, the backend calls `bli_thread_set_num_threads()` with that **same** value, so BLIS GEMM may use up to `--threads-batch` threads during prompt processing. Other ops stay on the CPU backend with the same limit. Token generation usually does not use BLAS.
-
-`BLIS_NUM_THREADS` is **not** a reliable way to cap BLIS here: the per-GEMM `bli_thread_set_num_threads()` call overrides it. To use fewer cores, lower `--threads` and/or `--threads-batch`.
-
-Thread scaling depends on the CPU, NUMA layout, model, and batch size. Benchmark the thread counts used for deployment. Nested OpenMP (ggml type conversion, then BLIS GEMM, both using OpenMP) can still oversubscribe even though those two steps are sequential.
-
-On a multi-socket machine, bind the process to one NUMA node. To skip SMT, bind to that node's physical cores only (check `lscpu -e`; on many AMD layouts the first range is the physical cores and a higher range is the sibling threads):
-
-```bash
-numactl --physcpubind=0-127 --membind=0 ./build/bin/llama-cli -m model.gguf
-```
-
-Keep the sourced AOCL env (or `LD_LIBRARY_PATH`) set when running binaries, or dynamic linking to AOCL libs will fail. Source the same `amd-libs.cfg` you used at build time.
-
-### Verify
-
- Configure output should show BLAS found, with libraries under the AOCL MT tree and includes at `$AOCL_ROOT/include`.
- `ldd` on `llama-bench` should list that same AOCL-BLAS library; `blis-mt` or `blis`.
- `--list-devices` prints the description `AOCL-BLAS` (`BLAS: AOCL-BLAS`). Upstream BLIS (`FLAME`) still prints `BLIS`.
- `llama-bench` reports the backend as `BLAS` for this build and `CPU` for a build with `-DGGML_BLAS=OFF`.
- `test-backend-ops -b BLAS` checks that BLAS GEMMs match the CPU reference.
-
-### Notes
-
- Optional `-march=znver3` / `znver4` / `znver5` (or similar) can be passed via `CMAKE_C_FLAGS` / `CMAKE_CXX_FLAGS` for a known CPU, but `-DGGML_NATIVE=ON` is usually enough and is safer across Ryzen / EPYC generations.
- For building AOCL-BLAS (AMD's BLIS fork) from source instead of AOCL packages, see https://github.com/amd/blis
-
-### Reference
-
-1. https://www.amd.com/en/developer/aocl.html
-2. https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors
-3. https://github.com/amd/blis
@@ -12,8 +12,6 @@ The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a t
 - Compiles and caches the model for the target device.
 - Binds GGML tensor memory to OpenVINO inference tensors and runs inference.

-For guidance on contributing to the OpenVINO backend, see the [OpenVINO Backend Contributing Guide](https://github.com/ravi9/llamacpp-ov-dev-guide/blob/main/contributing-llamacpp-ov.md).
-
 ## Contents

 - [Supported Devices](#supported-devices)
@@ -98,7 +96,7 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
  - **SL** = Stateless (`GGML_OPENVINO_STATEFUL_EXECUTION=0`)
  - **SF** = Stateful (`GGML_OPENVINO_STATEFUL_EXECUTION=1`)
  - Note: The NPU operates in stateless mode only.
- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel Graphics Compiler 2.41.5 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.38.0.
+- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.35.0.
 - See [Known Limitations](#known-limitations) for context on observed failures.

 | Model | CPU (SL / SF) | GPU (SL / SF) | NPU (SL) |
@@ -119,9 +117,9 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [lmstudio-community/Qwen3.5-9B-Q4_K_M](https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
 |  |  |  |  |
 | [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
-| [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
+| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
+| [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
+| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -134,9 +132,9 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [bartowski/DeepSeek-R1-Distill-Llama-8B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
-| [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✗ / ✗ | ✓ |
 | [ibm-granite/granite-4.0-micro-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-micro-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [ibm-granite/granite-4.0-1b-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF) | ✓ / ✓ | ✓ / ✓ | ✗ |
+| [ibm-granite/granite-4.0-1b-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF) | ✓ / ✓ | ✗ / ✗ | ✗ |
 | [ibm-research/granite-3.2-8b-instruct-Q4_K_M](https://huggingface.co/ibm-research/granite-3.2-8b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [HuggingFaceTB/smollm2-1.7b-instruct-q4_k_m](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -244,8 +242,8 @@ chmod +x build-llamacpp-ov.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.4"
-OPENVINO_VERSION_FULL="2026.4.0.22959.99c81491cc3"
+OPENVINO_VERSION_MAJOR="2026.3.1"
+OPENVINO_VERSION_FULL="2026.3.1.22476.56d9685302d"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -342,7 +340,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.4` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -372,8 +370,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.4"
-set "OPENVINO_VERSION_FULL=2026.4.0.22959.99c81491cc3"
+set "OPENVINO_VERSION_MAJOR=2026.3.1"
+set "OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -552,7 +550,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.4` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.

 </details>

@@ -52,10 +52,6 @@ The packages for FP32 and FP16 would have different accuracy and performance on

 ## News

- 2026.09
-  - Update the CI build environment for oneAPI 2026.1 (unified oneAPI Toolkit). oneDNN is removed from the Deep Learning Essentials package in 2026.0, so the CI now uses the oneAPI Toolkit installer which still includes oneDNN.
-  - oneAPI 2026.1 improves the SYCL build performance: measured with the same code on Arc B570, prompt processing 1331 vs 434 t/s (3.1x) vs the 2025.3-based release build.
-
 - 2026.04-05
  - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q6_K, Q8_0.
  - Fused MoE.
@@ -261,7 +257,7 @@ Platform #0: Intel(R) OpenCL HD Graphics
 `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```

-2. **Install Intel® oneAPI Toolkit**
+2. **Install Intel® oneAPI Base toolkit**

 SYCL backend depends on:
  - Intel® oneAPI DPC++/C++ compiler/running-time.
@@ -271,11 +267,11 @@ SYCL backend depends on:

 - **For Intel GPU**

-With the 2026.0 release, the Intel® oneAPI Base toolkit and the HPC toolkit are combined into the **Intel® oneAPI Toolkit**, and **oneDNN is removed from the Intel® Deep Learning Essentials** package (oneDNN is distributed separately since then). The **Intel® oneAPI Toolkit** includes oneDNN until 2027.0.
+All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.

-It's recommended to install the **Intel® oneAPI Toolkit**.
+It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.

-The **Intel® oneAPI Toolkit** can be obtained from the official [Intel® oneAPI Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html) page.
+The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.

 Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.

@@ -285,7 +281,6 @@ Upon a successful installation, SYCL is enabled for the available Intel devices,

 |Verified release|
 |-|
-|2026.1 |
 |2025.3.3 |
 |2025.2.1|
 |2025.1|
@@ -816,9 +811,6 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. Unsupported types and layouts fall back to the standalone op kernels. See `ggml_sycl_can_fuse()`. |
 | GGML_SYCL_ENABLE_ESIMD | 0 or 1 (default)| Enable ESIMD kernels when available. |
-| GGML_SYCL_SPARSE_FA | 0 (default) or 1 | Enable Sparse Flash-attention.|
-| GGML_SYCL_SPARSE_FA_DEBUG | 0 (default) or 1 | Enable to debug for Sparse Flash-attention.|
-| GGML_SYCL_SPARSE_FA_MARGIN | [0,..] default:256 | Set the margin value for Sparse Flash-attention.|
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
 | GGML_SYCL_USM_SYSTEM | 0 (default) or 1 | Enable experimental support for [USM system allocations](https://github.khronos.org/SYCL_Reference/iface/usm_basic_concept.html#system-allocations) for large GPU buffers. This requires enough host memory for model weights and caches, an Intel Xe2+ GPU such as BMG or newer and supported on Linux only, with CONFIG_DRM_XE_GPUSVM enabled. |
@@ -77,8 +77,6 @@

    { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
    { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
-    { "name": "arm64-android-snapdragon-relwithdebinfo", "inherits": [ "arm64-android-snapdragon-release" ],
-      "cacheVariables": { "GGML_HEXAGON_HTP_BUILD_TYPE": "RelWithDebInfo" } },

    { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] },
@@ -327,9 +327,6 @@ on 4 physical NPUs, or `--devices 'HTP0[0-1:0],HTP1[0-1:1]'` on 2 physical NPUs
 - `GGML_HEXAGON_HOSTBUF=1` (default: 0, disabled)
  Enables allocating host buffers for debugging. By default, host buffers are disabled.

- `GGML_HEXAGON_DMA64=0` (default: enabled on v81+)
-  Disables 64-bit DMA for model weights. Set to `1` to enable it explicitly on a supported architecture.
-
 - `GGML_HEXAGON_VERBOSE=1`
  Enables verbose logging of Ops from the backend. Example output:

@@ -146,28 +146,6 @@ Writing high-performance operators for Hexagon requires following specific guide
  python3 scripts/snapdragon/ggml-hexagon-align-macros.py --fix ggml/src/ggml-hexagon/htp/
  ```

-### Binary Inspection and Spill Analysis
-
-Use [`scripts/snapdragon/ggml-hexagon-inspect.py`](../../../scripts/snapdragon/ggml-hexagon-inspect.py) to audit Hexagon binaries for register
-spills, unexpected float promotions, or disassembly:
-
- Always verify that compute kernels have zero in-loop vector spills (`--spills --strict`) and no float promotions (`--promotions`).
- Avoid excessive loop unrolling (`#pragma unroll`), which increases register pressure and causes spills.
-
-```bash
-# Check for vector and scalar register spills
-python3 scripts/snapdragon/ggml-hexagon-inspect.py --spills --strict --func "^compute_"
-
-# Check for float promotions
-python3 scripts/snapdragon/ggml-hexagon-inspect.py --promotions --func "^compute_"
-
-# Disassemble with annotated loops and spill markers
-python3 scripts/snapdragon/ggml-hexagon-inspect.py --disasm compute_same_shape_div_f32
-
-# Resolve crash addresses to function symbols and lines
-python3 scripts/snapdragon/ggml-hexagon-inspect.py --addr2line 0x51a30 0x5ba54
-```
-
 ## Multi-Device Partitioning (mdev)

 Multi-device (mdev) mode enables row-level tensor parallel execution across multiple physical NPU cores or virtual NPU
--- a/Show More
+++ b/Show More