RPC: populate use_count to enable fusion inside backends (#27142 )

* RPC: populate use_count to enable fusion inside backends * bump minor version
ui: Services consolidation refactor (#27239 )
2026-08-18 18:52:40 +02:00 · 2026-08-18 21:08:57 +05:30 · 2026-08-18 16:39:32 +02:00 · 2026-08-18 16:37:26 +02:00 · 2026-08-18 16:23:43 +02:00 · 2026-08-18 16:22:32 +02:00
170 changed files with 2760 additions and 1686 deletions
@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04

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

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

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

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

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

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

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

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

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

    steps:
      - name: Clone
@@ -21,6 +21,7 @@ on:
    paths: [
      '.github/workflows/build-cpu.yml',
      '.github/workflows/build-cmake-pkg.yml',
+      'ggml/src/ggml-rpc/**',
      '**/CMakeLists.txt',
      '**/.cmake',
      '**/*.h',
@@ -123,7 +124,6 @@ jobs:
    env:
      OPENBLAS_VERSION: 0.3.23
      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.357.0

    strategy:
      matrix:
@@ -134,9 +134,6 @@ jobs:
          - build: 'x64-openblas'
            arch: 'x64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'x64-vulkan'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
          - build: 'arm64'
            arch: 'arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
@@ -167,15 +164,6 @@ jobs:
          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll

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

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

    permissions:
@@ -29,7 +30,16 @@ jobs:

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

    steps:
      - name: Clone
@@ -39,12 +49,13 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -54,26 +65,21 @@ jobs:
      - name: Build
        id: cmake_build
        shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

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

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

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

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

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

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

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

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

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

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

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

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

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -133,3 +127,56 @@ jobs:
          # This is using llvmpipe and runs slower than other backends
          # test-backend-ops is too slow on llvmpipe, skip it
          ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.357.0
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-windows-2025-x64-vulkan
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_NATIVE=OFF `
+            -DLLAMA_BUILD_SERVER=ON `
+            -DGGML_RPC=ON `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
+            -DGGML_VULKAN=ON `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
@@ -49,6 +49,33 @@ jobs:
          git push origin "${VERSION}"
          echo "Created and pushed tag ${VERSION}"

+      - name: Generate release description
+        id: desc
+        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
+        env:
+          GITHUB_REPOSITORY: ${{ github.repository }}
+
+      - name: Create release
+        if: ${{ github.event.inputs.dry_run == 'false' }}
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        with:
+          tag_name: ${{ steps.checks.outputs.version }}
+          # TODO: remove the prerelease flag once the semantic versioning workflow is ready
+          # ref: https://github.com/ggml-org/ggml/discussions/1579
+          prerelease: true
+          body: |
+            > [!NOTE]
+            > Semantic versioning is still work in progress.
+            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+
+            ${{ steps.desc.outputs.nightly }}
+
+            ## ${{ steps.desc.outputs.changelog_title }}
+
+            ${{ steps.desc.outputs.changelog }}
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
@@ -446,8 +446,8 @@ jobs:

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

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

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

    steps:
      - name: Set OpenVINO version output
@@ -5,7 +5,7 @@ include(CheckIncludeFileCXX)
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
 set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 1)
+set(LLAMA_VERSION_PATCH 2)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

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

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

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

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

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

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

 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
+- [mackron/miniaudio](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [sheredom/subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
@@ -290,6 +290,7 @@ combine_static_libraries() {
        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
+        "${base_dir}/${build_dir}/vendor/hash/${release_dir}/libvendor-hash.a"
    )

    # Create temporary directory for processing
@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

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

 ## helpers
@@ -126,7 +126,8 @@ set_target_properties(${TARGET} PROPERTIES
    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )

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

 if (LLAMA_SUBPROCESS)
@@ -1750,6 +1750,18 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const commo
    return tpp;
 }

+namespace {
+
+bool can_share_threadpool(const ggml_threadpool_params & tpp1, const ggml_threadpool_params & tpp2) {
+    // n_threads does not matter -> we'll use what's larger
+    ggml_threadpool_params tpp_comparison = tpp1;
+    tpp_comparison.n_threads = tpp2.n_threads;
+
+    return ggml_threadpool_params_match(&tpp_comparison, &tpp2);
+}
+
+}  // namespace
+
 common_threadpools::~common_threadpools() {
    if (!free_fn) {
        return;
@@ -1778,7 +1790,9 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    struct ggml_threadpool_params tpp =
            ggml_threadpool_params_from_cpu_params(params.cpuparams);

-    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+    if (can_share_threadpool(tpp, tpp_batch)) {
+        tpp.n_threads = std::max(tpp.n_threads, tpp_batch.n_threads);
+    } else {
        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
        if (!threadpool_batch) {
            COM_WRN("batch threadpool create failed : n_threads %d\n", tpp_batch.n_threads);
@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.2.1"
-OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` 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` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.2.1"
-set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` 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**.
+> The script pins OpenVINO `2026.3` 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>

@@ -2,5 +2,5 @@ set(TARGET llama-gguf-hash)
 add_executable(${TARGET} gguf-hash.cpp)
 install(TARGETS ${TARGET} RUNTIME)

-target_link_libraries(${TARGET} PRIVATE vendor-hash ggml ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE vendor::hash ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
@@ -17,14 +17,17 @@
 extern "C" {
 #endif

-#include "xxhash/xxhash.h"
-#include "sha1/sha1.h"
-#include "sha256/sha256.h"
+#include "hash/xxhash/xxhash.h"
+#include "hash/sha256/sha256.h"

 #ifdef __cplusplus
 }
 #endif

+// sha1 is compiled as C++ and lives in a namespace, see scripts/sync_vendor.py
+#include "hash/sha1/sha1.h"
+using namespace vendor_hash;
+

 // uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp')
 #define UUID_NAMESPACE_LLAMA_CPP "ef001206-dadc-5f6d-a15f-3359e577d4e5"
@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_PATCH 2)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -7,7 +7,7 @@ extern "C" {
 #endif

 #define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    1
 #define RPC_PROTO_PATCH_VERSION    0

 #ifdef  __cplusplus
@@ -4,6 +4,7 @@
 #include "vecdotq.cuh"

 #include <cstdint>
+#include <type_traits>

 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);

@@ -69,7 +70,8 @@ enum mmvq_parameter_table_id {
    MMVQ_PARAMETERS_GCN,
    MMVQ_PARAMETERS_RDNA2,
    MMVQ_PARAMETERS_RDNA3_0,
-    MMVQ_PARAMETERS_RDNA4
+    MMVQ_PARAMETERS_RDNA4,
+    MMVQ_PARAMETERS_GB10
 };

 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
@@ -83,6 +85,8 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
    return MMVQ_PARAMETERS_GCN;
 #elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING && __CUDA_ARCH__ < GGML_CUDA_CC_AMPERE
    return MMVQ_PARAMETERS_TURING;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+    return MMVQ_PARAMETERS_GB10;
 #else
    return MMVQ_PARAMETERS_GENERIC;
 #endif
@@ -104,6 +108,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING && ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_AMPERE) {
        return MMVQ_PARAMETERS_TURING;
    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_DGX_SPARK) {
+        return MMVQ_PARAMETERS_GB10;
+    }
    return MMVQ_PARAMETERS_GENERIC;
 }

@@ -351,7 +358,7 @@ static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
 #endif
 }

-static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id, bool small_k = false, bool halve_iters = false) {
    if (table_id == MMVQ_PARAMETERS_GENERIC) {
        switch (ncols_dst) {
            case 1:
@@ -454,11 +461,32 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                return 1;
        }
    }
+    if (table_id == MMVQ_PARAMETERS_GB10) {
+        const int generic = calc_nwarps(type, ncols_dst, MMVQ_PARAMETERS_GENERIC);
+        // Only worth the wider block when it actually retires the K loop in half the trips (Observation)
+        if (ncols_dst == 1 && !small_k && halve_iters) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                    return 2 * generic;
+                default:
+                    break;
+            }
+        }
+        return generic;
+    }
    return 1;
 }

 static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
-    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING || table_id == MMVQ_PARAMETERS_GB10) {
        switch (ncols_dst) {
            case 1:
                return small_k ? nwarps : 1;
@@ -477,8 +505,8 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
    return 1;
 }

-template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
-__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false, bool halve_iters = false>
+__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id(), small_k, halve_iters)*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
        const void * vx_ptr, const void * vy_ptr, const int32_t * ids_ptr, const ggml_cuda_mm_fusion_args_device fusion, float * dst_ptr,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -495,7 +523,7 @@ static __global__ void mul_mat_vec_q(
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
    constexpr int vdr = get_vdr_mmvq(type);
    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

@@ -773,8 +801,8 @@ static __global__ void mul_mat_vec_q_moe(
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
-        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false) {
-    const int nwarps = calc_nwarps(type, ncols_dst, table_id);
+        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false, const bool halve_iters = false) {
+    const int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    const int rpb = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    const int64_t nblocks = (nrows_x + rpb - 1) / rpb;
    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
@@ -782,7 +810,7 @@ static std::pair<dim3, dim3> calc_launch_params(
    return {block_nums, block_dims};
 }

-template<ggml_type type, int c_ncols_dst, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false, bool halve_iters = false>
 static void mul_mat_vec_q_switch_fusion(
        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -797,7 +825,7 @@ static void mul_mat_vec_q_switch_fusion(
    if constexpr (c_ncols_dst == 1) {
        if (has_fusion) {
            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k>, launch_params,
+            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k, halve_iters>, launch_params,
                 vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -808,7 +836,7 @@ static void mul_mat_vec_q_switch_fusion(
    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");

    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k>, launch_params,
+    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k, halve_iters>, launch_params,
        vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -860,16 +888,18 @@ static void mul_mat_vec_q_switch_ncols_dst(

    const bool has_ids = ids != nullptr;

+    // How the K loop divides up at the baseline block width, both decisions below use these.
+    constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr                   = get_vdr_mmvq(type);
+    const int     blocks_per_row_x      = ncols_x / qk;
+    const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+
    const auto should_use_small_k = [&](int c_ncols_dst) {
        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
-        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
-        constexpr int vdr                   = get_vdr_mmvq(type);
-        const int     blocks_per_row_x      = ncols_x / qk;
-        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
-        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
-        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+        const int  nwarps = calc_nwarps(type, c_ncols_dst, table_id);
+        bool       use    = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;

        constexpr std::array<ggml_type, 2> iq_slow_turing = {
            GGML_TYPE_IQ3_XXS,
@@ -902,6 +932,28 @@ static void mul_mat_vec_q_switch_ncols_dst(
        return use;
    };

+    // Whether doubling nwarps pays off on the ncols_dst == 1 path, where K sets the K loop trip count.
+    const auto should_halve_iters = [&] {
+        if (table_id != MMVQ_PARAMETERS_GB10) {
+            return false;
+        }
+
+        // Expert rows are gathered per token, so a wider block adds reduction work without reuse.
+        if (has_ids) {
+            return false;
+        }
+
+        const int blocks_per_iter = calc_nwarps(type, 1, table_id) * blocks_per_iter_1warp;
+        const int iters           = (blocks_per_row_x + blocks_per_iter - 1) /  blocks_per_iter;
+        const int iters_wide      = (blocks_per_row_x + blocks_per_iter * 2 - 1) / (blocks_per_iter * 2);
+
+        // An odd trip count leaves half the wider block idle for its last iteration, that tail is
+        // only affordable once the loop is long enough to dilute it to an eighth of the work (observation).
+        const int idle = iters_wide * 2 - iters;
+
+        return idle * 8 <= iters_wide * 2;
+    };
+
    if (has_ids && ncols_dst > 1) {
        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
        mul_mat_vec_q_moe_launch<type>(
@@ -914,26 +966,34 @@ static void mul_mat_vec_q_switch_ncols_dst(

    switch (ncols_dst) {
        case 1: {
-            constexpr int c_ncols_dst = 1;
+            // static, else MSVC lambda capture breaks the constexpr uses below
+            static constexpr int c_ncols_dst = 1;

-            bool use_small_k = should_use_small_k(c_ncols_dst);
+            // Tag types keep the flags compile-time, so __launch_bounds__ matches what is launched.
+            const auto launch = [&](auto small_k_tag, auto halve_iters_tag) {
+                constexpr bool c_small_k = decltype(small_k_tag)::value;
+                // Types the table does not promote would compile a second, identical kernel.
+                constexpr bool c_promoted =
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, true) !=
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, false);

-            if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
+                constexpr bool c_halve_iters = decltype(halve_iters_tag)::value && c_promoted;
+
+                const std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                              nsamples_dst, warp_size, table_id, c_small_k, c_halve_iters);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, c_small_k, c_halve_iters>(
                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
                    stream);
+            };
+
+            if (should_use_small_k(c_ncols_dst)) {
+                launch(std::true_type{},  std::false_type{});
+            } else if (should_halve_iters()) {
+                launch(std::false_type{}, std::true_type{});
            } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
-                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
-                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
-                    stream);
+                launch(std::false_type{}, std::false_type{});
            }
        } break;
        case 2: {
@@ -47,7 +47,7 @@ struct rpc_tensor {
    uint64_t data;
    char name[GGML_MAX_NAME];

-    char padding[4];
+    int32_t use_count;
 };

 static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
@@ -447,7 +447,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {

    // Avoid sending uninitialized data over the wire
    memset(result.name, 0, sizeof(result.name));
-    memset(result.padding, 0, sizeof(result.padding));
+    result.use_count = 0;

    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
    return result;
@@ -675,7 +675,7 @@ static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
    // this is no-op because we don't have any async operations
 }

-static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
+static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
    if (tensor == nullptr) {
        return;
    }
@@ -684,10 +684,15 @@ static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors,
    }
    visited.insert(tensor);
    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        add_tensor(tensor->src[i], tensors, visited);
+        add_tensor(tensor->src[i], cgraph, tensors, visited);
    }
-    add_tensor(tensor->view_src, tensors, visited);
-    tensors.push_back(serialize_tensor(tensor));
+    add_tensor(tensor->view_src, cgraph, tensors, visited);
+    rpc_tensor result = serialize_tensor(tensor);
+    const size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, tensor);
+    if (hash_pos != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        result.use_count = cgraph->use_counts[hash_pos];
+    }
+    tensors.push_back(result);
 }

 static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
@@ -695,7 +700,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
    std::vector<rpc_tensor> tensors;
    std::unordered_set<ggml_tensor*> visited;
    for (uint32_t i = 0; i < n_nodes; i++) {
-        add_tensor(cgraph->nodes[i], tensors, visited);
+        add_tensor(cgraph->nodes[i], cgraph, tensors, visited);
    }
    // serialization format:
    // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
@@ -1451,6 +1456,10 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {
            GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
            return false;
        }
+        if (graph->nodes[i] != nullptr) {
+            const size_t hash_pos = ggml_hash_insert(&graph->visited_hash_set, graph->nodes[i]);
+            graph->use_counts[hash_pos] = tensor_ptrs.at(id)->use_count;
+        }
    }
    ggml_status status = ggml_backend_graph_compute(backends[device], graph);
    GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");
@@ -0,0 +1,119 @@
+#include "fwht.hpp"
+
+#include <cmath>
+
+template <int N>
+static void fwht_kernel(const float * __restrict__ src, float * __restrict__ dst, const int64_t n_rows,
+                        const float scale, const sycl::nd_item<2> & item) {
+    const sycl::sub_group sg = item.get_sub_group();
+
+    const int64_t r = item.get_global_id(0);
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    constexpr int el_w = N / WARP_SIZE;
+    static_assert(el_w >= 1 && N % WARP_SIZE == 0, "row must be a whole number of sub-group widths");
+
+    float     reg[el_w];
+    const int lane = sg.get_local_linear_id();
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        reg[i] = src[i * WARP_SIZE + lane] * scale;
+    }
+
+    // Butterflies inside the sub-group. The partner of a lane with bit h clear is the
+    // lower index of the pair, so it takes the sum and the upper takes lower - upper.
+#pragma unroll
+    for (int h = 1; h < WARP_SIZE; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; ++j) {
+            const float val  = reg[j];
+            const float val2 = dpct::permute_sub_group_by_xor(sg, val, h, WARP_SIZE);
+
+            reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+        }
+    }
+
+    // Butterflies across registers: h is a multiple of WARP_SIZE, so the partner of
+    // element i*WARP_SIZE + lane lives in reg[i + h/WARP_SIZE] on the same lane.
+#pragma unroll
+    for (int h = WARP_SIZE; h < N; h *= 2) {
+        const int step = h / WARP_SIZE;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int k = 0; k < step; ++k) {
+                const float x = reg[j + k];
+                const float y = reg[j + k + step];
+
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        dst[i * WARP_SIZE + lane] = reg[i];
+    }
+}
+
+template <int N>
+static void launch_fwht(const float * src, float * dst, const int64_t n_rows, const float scale,
+                        dpct::queue_ptr stream) {
+    constexpr int rows_per_block = 4;
+
+    const int64_t num_blocks = (n_rows + rows_per_block - 1) / rows_per_block;
+
+    // dim 1 is the fastest-varying, so a sub-group is exactly one row's WARP_SIZE lanes.
+    const sycl::range<2> global(num_blocks * rows_per_block, WARP_SIZE);
+    const sycl::range<2> local(rows_per_block, WARP_SIZE);
+
+    stream->parallel_for(sycl::nd_range<2>(global, local),
+                         [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             fwht_kernel<N>(src, dst, n_rows, scale, item);
+                         });
+}
+
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    if (src->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!ggml_are_same_shape(src, dst)) {
+        return false;
+    }
+    if (!ggml_is_contiguous(src) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    const int     n    = (int) src->ne[0];
+    const int64_t rows = ggml_nrows(src);
+
+    const float *   src_d  = (const float *) src->data;
+    float *         dst_d  = (float *) dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    const float scale = 1.0f / std::sqrt((float) n);
+
+    switch (n) {
+        case 64:
+            launch_fwht<64>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 128:
+            launch_fwht<128>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 256:
+            launch_fwht<256>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 512:
+            launch_fwht<512>(src_d, dst_d, rows, scale, stream);
+            return true;
+        default:
+            return false;
+    }
+}
@@ -0,0 +1,12 @@
+#ifndef GGML_SYCL_FWHT_HPP
+#define GGML_SYCL_FWHT_HPP
+
+#include "common.hpp"
+
+// Fast Walsh-Hadamard transform, the fast path for a MUL_MAT whose src0 ggml has
+// tagged GGML_HINT_SRC0_IS_HADAMARD. src0 is not read at all. Returns false if the
+// shape is not one this can serve, in which case the caller must fall through to the
+// ordinary mat-mul dispatch.
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst);
+
+#endif  // GGML_SYCL_FWHT_HPP
@@ -58,6 +58,7 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
+#include "ggml-sycl/fwht.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml-sycl/norm.hpp"
@@ -4473,6 +4474,18 @@ static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor *

 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    // Handle HADAMARAD hint given from further up the pipeline and pass it to the correct
+    // kernel.
+    //
+    // The op check is not redundant: this backend also routes MUL_MAT_ID through here with a
+    // stack copy of dst, which carries MUL_MAT_ID's own op_params. ggml_mul_mat_set_hint()
+    // asserts GGML_OP_MUL_MAT for the same reason.
+    if (dst->op == GGML_OP_MUL_MAT && ggml_get_op_params_i32(dst, 1) == GGML_HINT_SRC0_IS_HADAMARD &&
+        ggml_sycl_op_fwht(ctx, src1, dst)) {
+        return;
+    }
+
    const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
    int64_t min_compute_capability = INT_MAX;

@@ -0,0 +1,87 @@
+#!/bin/bash
+# Generate the description of a release: the previous release version, the
+# change log and the link to the nightly release corresponding to the commit being released.
+#
+# Usage: make-release-desc.sh <version>
+#   <version>: current release version (v<maj>.<min>.<pat>, the leading v is optional)
+#
+# The previous version is the highest plain semver tag (v<maj>.<min>.<pat>)
+# strictly below <version>. The change log lists all commits between the
+# previous version tag and the release commit, one line per commit.
+#
+# The release commit is the commit <version> points at when the tag exists,
+# HEAD otherwise. The nightly release is the b* tag pointing at that commit
+# (release.yml tags the same commit); the link is only generated when that
+# tag exists.
+#
+# Env (when running in GitHub Actions):
+#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog and nightly are written here
+#   GITHUB_REPOSITORY: owner/repo, used to build the nightly release URL (skipped when unset)
+set -euo pipefail
+
+if [[ $# -ne 1 ]]; then
+    echo "Usage: $(basename "$0") <version>"
+    exit 1
+fi
+VERSION="$1"
+
+# Accept the version with or without the leading v, reject anything else
+if [[ "${VERSION}" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+    VERSION="v${VERSION}"
+elif [[ ! "${VERSION}" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+    echo "Error: invalid version '${VERSION}' (expected v<maj>.<min>.<pat>)"
+    exit 1
+fi
+
+# Make sure all remote tags are available locally (skipped on local runs without origin)
+if ! git fetch --tags origin 2>/dev/null; then
+    echo "Warning: could not fetch tags from origin (local run?)"
+fi
+
+# Release commit: the commit <version> points at when the tag exists, HEAD otherwise.
+if ! RELEASE_COMMIT="$(git rev-parse -q --verify "refs/tags/${VERSION}^{commit}" 2>/dev/null)"; then
+    RELEASE_COMMIT="$(git rev-parse HEAD)"
+fi
+
+echo "Release commit: $(git rev-parse --short "${RELEASE_COMMIT}")"
+
+PREV="$( { git tag --list; echo "${VERSION}"; } \
+    | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' \
+    | sort -V \
+    | awk -v cur="${VERSION}" '$0 == cur { exit } { prev = $0 } END { print prev }')"
+
+if [[ -n "${PREV}" ]]; then
+    CHANGELOG="$(git log --oneline "${PREV}..${RELEASE_COMMIT}")"
+    CHANGELOG_TITLE="Change log since ${PREV}"
+else
+    CHANGELOG="(no previous release tag found)"
+    CHANGELOG_TITLE="Change log"
+fi
+
+# Nightly release: the b* tag pointing at the release commit (|| true: no match is not an error)
+NIGHTLY_TAG="$(git tag --points-at "${RELEASE_COMMIT}" | grep -E '(^|-)b[0-9]+(-[0-9a-f]{7})?$' | head -n 1 || true)"
+
+NIGHTLY=""
+if [[ -n "${NIGHTLY_TAG}" ]]; then
+    if [[ -n "${GITHUB_REPOSITORY:-}" ]]; then
+        NIGHTLY_URL="https://github.com/${GITHUB_REPOSITORY}/releases/tag/${NIGHTLY_TAG}"
+        NIGHTLY="**Nightly build:** [${NIGHTLY_TAG}](${NIGHTLY_URL})"
+        echo "Nightly release: ${NIGHTLY_URL}"
+    fi
+else
+    echo "No nightly release found for commit $(git rev-parse --short "${RELEASE_COMMIT}")"
+fi
+
+echo "Previous version: ${PREV:-none}"
+echo "${CHANGELOG}"
+
+if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
+    {
+        echo "previous_tag=${PREV}"
+        echo "changelog_title=${CHANGELOG_TITLE}"
+        echo "nightly=${NIGHTLY}"
+        echo "changelog<<CHANGELOG_EOF"
+        echo "${CHANGELOG}"
+        echo "CHANGELOG_EOF"
+    } >> "${GITHUB_OUTPUT}"
+fi
@@ -1 +1 @@
-3834fd814e74e8af277939dabd69ecc780affd21
+8c63e70982c95ceb862e3a1073a2c1beef75d60a
@@ -56,6 +56,44 @@ patches = {
        '    && (defined(_MSC_VER) && (_MSC_VER >= 1000) || !defined(_MSC_VER)) /* >= C11 */\n'
    )],

+    # sha1 exports a bare "SHA1" symbol, which clashes with the boringssl one at link time.
+    # we compile it as C++ (see vendor/hash/CMakeLists.txt) and put it in a namespace.
+    "vendor/hash/sha1/sha1.h": [
+        (
+            '#if defined(__cplusplus)\n'
+            'extern "C" {\n'
+            '#endif\n',
+
+            'namespace vendor_hash {\n'
+        ),
+        (
+            '#if defined(__cplusplus)\n'
+            '}\n'
+            '#endif\n',
+
+            '} // namespace vendor_hash\n'
+        ),
+    ],
+
+    "vendor/hash/sha1/sha1.c": [
+        (
+            '#include "sha1.h"\n',
+
+            '#include "sha1.h"\n'
+            '\n'
+            'namespace vendor_hash {\n'
+        ),
+        (
+            '    SHA1Final((unsigned char *)hash_out, &ctx);\n'
+            '}\n',
+
+            '    SHA1Final((unsigned char *)hash_out, &ctx);\n'
+            '}\n'
+            '\n'
+            '} // namespace vendor_hash\n'
+        ),
+    ],
+
    # silence a maybe-uninitialized warning
    "vendor/hash/sha256/sha256.c": [(
        "  uint32_t W[16];\n",
@@ -1395,6 +1395,11 @@ void llama_model_loader::get_mapping_range(size_t * first, size_t * last, void *
    }
 }

+void llama_model_loader::unmap_weight(const llama_tensor_weight & w) const {
+    if (!use_mmap) { return; }
+    mappings.at(w.idx)->unmap_fragment(w.offs, w.offs + ggml_nbytes(w.tensor));
+}
+
 void llama_model_loader::load_data_for(struct ggml_tensor * cur) const {
    const auto & w = require_weight(ggml_get_name(cur));

@@ -194,6 +194,9 @@ struct llama_model_loader {

    void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;

+    // release a weight's mmap pages
+    void unmap_weight(const llama_tensor_weight & w) const;
+
    // for backwards compatibility, does not support ggml-backend
    void load_data_for(struct ggml_tensor * cur) const;

@@ -1270,7 +1270,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
            total_size_org += tensor_size;
            total_size_new += new_size;

-            // update the gguf meta data as we go
+            // update the gguf metadata as we go
            gguf_set_tensor_type(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_type);
            GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), metadata[i].name.c_str())) == new_size);
            gguf_set_tensor_data(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_data);
@@ -1278,6 +1278,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
            // write tensor data + padding
            fout.write((const char *) new_data, new_size);
            zeros(fout, GGML_PAD(new_size, align) - new_size);
+
+            // unmap the tensor to free memory
+            if (ml.use_mmap) { ml.unmap_weight(weight); }
+
        } // no --dry-run
    } // main loop

@@ -2428,17 +2428,24 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
    const uint32_t n_tokens = gguf_get_arr_n(ctx, token_idx);

    const float * scores = nullptr;
+    const int * iscores = nullptr;
    const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str());
    if (score_idx != -1) {
-        if (gguf_get_kv_type(ctx, score_idx) != GGUF_TYPE_ARRAY ||
-            gguf_get_arr_type(ctx, score_idx) != GGUF_TYPE_FLOAT32) {
+        const gguf_type kv_type = gguf_get_kv_type(ctx, score_idx);
+        const gguf_type arr_type = kv_type == GGUF_TYPE_ARRAY ? gguf_get_arr_type(ctx, score_idx) : GGUF_TYPE_COUNT;
+        if (arr_type != GGUF_TYPE_INT32 &&
+            arr_type != GGUF_TYPE_FLOAT32) {
            throw std::runtime_error(format("invalid gguf type for %s", kv(LLM_KV_TOKENIZER_SCORES).c_str()));
        }
        const uint32_t n_scores = gguf_get_arr_n(ctx, score_idx);
        if (n_scores < n_tokens) {
            throw std::runtime_error("Index out of array bounds for scores (" + std::to_string(n_scores) + " < " + std::to_string(n_tokens) + ")\n");
        }
-        scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+        if (arr_type == GGUF_TYPE_INT32) {
+            iscores = (const int *) gguf_get_arr_data(ctx, score_idx);
+        } else {
+            scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+        }
    }

    const int * toktypes = nullptr;
@@ -2469,7 +2476,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

        auto & token_data = id_to_token[i];
        token_data.text  = std::move(word);
-        token_data.score = scores ? scores[i] : 0.0f;
+        if (scores) {
+            token_data.score = scores[i];
+        } else if (iscores) {
+            token_data.score = static_cast<float>(iscores[i]);
+        } else {
+            token_data.score = 0.0f;
+        }
        token_data.attr  = LLAMA_TOKEN_ATTR_NORMAL;

        if (toktypes) {  //TODO: remove, required until per token attributes are available from GGUF file
@@ -1241,7 +1241,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
        { unicode_cpt_flags::LETTER,      "\x41-\x5A\x61-\x7A" }, // A-Za-z
        { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
        { unicode_cpt_flags::ACCENT_MARK, "" }, // no sub-128 codepoints
-        { unicode_cpt_flags::SYMBOL,      "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|
+        { unicode_cpt_flags::SYMBOL,      "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C\\\x7E" }, // $+<=>^`|~
    };

    // compute collapsed codepoints only if needed by at least one regex
@@ -116,6 +116,8 @@ function(llama_build_and_test source)
    set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${LLAMA_TEST_LABEL})
 endfunction()

+llama_build_and_test(test-unicode.cpp)
+
 # build test-tokenizer-0 target once and add many tests
 llama_build(test-tokenizer-0.cpp)

@@ -310,6 +312,9 @@ llama_build_and_test(test-mtmd-c-api.c)
 target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd)
 unset(LLAMA_TEST_NAME)

+llama_build_and_test(test-mtmd-impl.cpp)
+target_link_libraries(test-mtmd-impl PRIVATE mtmd)
+
 # GGUF model data fetcher library for tests that need real model metadata
 # Only compile when cpp-httplib has SSL support (CPPHTTPLIB_OPENSSL_SUPPORT)
 if (TARGET cpp-httplib)
@@ -0,0 +1,88 @@
+#include "testing.h"
+
+#include "mtmd-image.h"
+
+#include <iostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+// this test file contains:
+// 1. test cases for mtmd helpers
+// 2. test cases for internal mtmd components
+// internal headers can be included here
+
+struct test_registry {
+    using fn_t = void (*)(testing &);
+
+    struct entry {
+        std::string name;
+        fn_t fn;
+    };
+
+    static std::vector<entry> & all() {
+        static std::vector<entry> entries;
+        return entries;
+    }
+
+    test_registry(const char * name, fn_t fn) {
+        all().push_back({ name, fn });
+    }
+};
+
+#define MAKE_TEST(name)                                               \
+    static void name(testing & t);                                    \
+    static const test_registry test_registry_ ## name(#name, &name);  \
+    static void name(testing & t)
+
+
+//
+// mtmd_image
+//
+
+MAKE_TEST(test_image_preprocessor_lfm2) {
+    clip_hparams hparams;
+    hparams.patch_size = 16;
+    hparams.n_merge = 2;
+    hparams.set_limit_image_tokens(64, 256);
+
+    // { image size, expected tiling }
+    const std::vector<std::pair<clip_image_size, bool>> cases = {
+        { {  704, 704 }, false },
+        // 720 / (patch_size * n_merge) is exactly 22.5, so this only matches HF
+        // if round_by_factor rounds half to even (22) instead of away from zero (23)
+        { {  720, 720 }, false },
+        { {  736, 736 }, true  },
+        { { 1024, 977 }, true  },
+        { { 1056, 384 }, false },
+    };
+
+    for (const auto & [size, expected] : cases) {
+        const bool actual = mtmd_image_preprocessor_lfm2::should_tile(hparams, size);
+
+        t.assert_equal(
+            "tiling for " + std::to_string(size.width) + "x" + std::to_string(size.height),
+            std::string(expected ? "tiled" : "single"),
+            std::string(actual   ? "tiled" : "single"));
+    }
+}
+
+//
+// main
+//
+
+int main(int argc, char ** argv) {
+    testing t(std::cout);
+    t.verbose = true;
+
+    // usage: test-mtmd-impl [filter_regex]
+    for (int i = 1; i < argc; i++) {
+        t.set_filter(argv[i]);
+    }
+
+    for (const auto & e : test_registry::all()) {
+        t.test(e.name, e.fn);
+    }
+
+    return t.summary();
+}
@@ -0,0 +1,24 @@
+#include "../src/unicode.h"
+
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main() {
+    const std::vector<std::string> regex_exprs = {
+        "[~][A-Za-z]+| ?[\\p{S}]+|\\s+",
+    };
+    const std::vector<std::string> expected = { " ~", "foo" };
+    const auto actual = unicode_regex_split(" ~foo", regex_exprs, false);
+
+    if (actual != expected) {
+        fprintf(stderr, "unexpected split:");
+        for (const auto & piece : actual) {
+            fprintf(stderr, " [%s]", piece.c_str());
+        }
+        fprintf(stderr, "\n");
+        return 1;
+    }
+
+    return 0;
+}
@@ -78,10 +78,8 @@ set_target_properties(mtmd PROPERTIES
 )

 target_link_libraries     (mtmd PUBLIC ggml llama)
-target_link_libraries     (mtmd PRIVATE Threads::Threads)
+target_link_libraries     (mtmd PRIVATE Threads::Threads vendor::hash vendor::miniaudio vendor::stb vendor::sheredom)
 target_include_directories(mtmd PUBLIC  .)
-target_include_directories(mtmd PRIVATE ../..)
-target_include_directories(mtmd PRIVATE ../../vendor)
 target_compile_features   (mtmd PRIVATE cxx_std_17)

 if (MTMD_VIDEO)
@@ -92,6 +90,9 @@ if (BUILD_SHARED_LIBS)
    set_target_properties     (mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
    target_compile_definitions(mtmd PRIVATE LLAMA_BUILD)
    target_compile_definitions(mtmd PUBLIC  LLAMA_SHARED)
+
+    # export all symbols so that internal components can be tested by test-mtmd-impl
+    set_target_properties     (mtmd PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 endif()

 set(MTMD_PUBLIC_HEADERS
@@ -858,6 +858,9 @@ static std::ifstream open_ifstream_binary(const std::string & fname) {
 }
 #endif

+// in test-mtmd-impl, we include woth common.h and this file, and these functions are duplicated
+// this is a quick fix to avoid compilation errors
+#ifndef DIRECTORY_SEPARATOR
 static std::string string_format(const char * fmt, ...) {
    va_list ap;
    va_list ap2;
@@ -915,6 +918,7 @@ inline bool string_ends_with(std::string_view str, std::string_view suffix) {
    return str.size() >= suffix.size() &&
           str.compare(str.size() - suffix.size(), suffix.size(), suffix) == 0;
 }
+#endif

 //
 // gguf utils
@@ -12,6 +12,8 @@
 #include "mtmd-helper-common.h"
 #include "llama.h"

+#include "hash/hash.h"
+
 #include <algorithm>
 #include <cinttypes>
 #include <vector>
@@ -356,25 +358,14 @@ static bool decode_audio_from_buf(const unsigned char * buf_in, size_t len, int

 } // namespace audio_helpers

-// Computes FNV-1a hash of the data
-static std::string fnv_hash(const uint8_t * data, size_t len) {
-    const uint64_t fnv_prime = 0x100000001b3ULL;
-    uint64_t hash = 0xcbf29ce484222325ULL;
-
-    for (size_t i = 0; i < len; ++i) {
-        hash ^= data[i];
-        hash *= fnv_prime;
-    }
-    return std::to_string(hash);
-}
-
 mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder) {
    // calculate the hash if needed
    std::string id;
    mtmd_bitmap * result = nullptr;

    if (!placeholder) {
-        id = fnv_hash(buf, len);
+        // use sha256 to prevent cache poisoning
+        id = hash_sha256_hex(buf, len);
    }

    if (audio_helpers::is_audio_file((const char *)buf, len)) {
@@ -49,7 +49,7 @@ MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtm
 // note:
 //   - for now, video input is only supported via C++ helper functions
 //   - audio files will be auto-detected based on magic bytes
-//   - output bitmap will have FNV hash as the ID
+//   - output bitmap will have SHA-256 hash (hex string) as the ID
 // returns nullptr on failure
 // this function is thread-safe
 MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder);
@@ -999,14 +999,45 @@ mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_mi
 // mtmd_image_preprocessor_lfm2
 //

+mtmd_image_preproc_out mtmd_image_preprocessor_lfm2::preprocess(const clip_image_u8 & img) {
+    auto const inst = get_slice_instructions(img.get_size());
+    if (!inst.slices.empty()) {
+        return mtmd_image_preprocessor_llava_uhd::preprocess(img);
+    }
+
+    // single tile: no thumbnail
+    // note: not using output.overview here because it will emit <|img_thumbnail|> token, which we don't want in this case
+    auto sliced = slice_image(img, inst);
+    mtmd_image_preproc_out output;
+    output.append(hparams, sliced.overview, true);
+    return output;
+}
+
+bool mtmd_image_preprocessor_lfm2::should_tile(
+        const clip_hparams & hparams,
+        const clip_image_size & original_size) {
+    const int align_size = hparams.patch_size * hparams.n_merge;
+
+    const auto round_by_factor = [align_size](float x) {
+        // see https://github.com/ggml-org/llama.cpp/pull/27057#discussion_r3796264887
+        return static_cast<int>(std::nearbyint(static_cast<double>(x) / align_size)) * align_size;
+    };
+
+    const int h_bar = std::max(hparams.patch_size, round_by_factor(original_size.height));
+    const int w_bar = std::max(hparams.patch_size, round_by_factor(original_size.width));
+
+    return static_cast<double>(h_bar) * static_cast<double>(w_bar) >
+           static_cast<double>(hparams.image_max_pixels) * max_pixels_tolerance;
+}
+
 mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_lfm2::get_slice_instructions(const clip_image_size & original_size) {
    mtmd_image_preprocessor_llava_uhd::slice_instructions inst;
    const int align_size = hparams.patch_size * hparams.n_merge;
    inst.overview_size = img_tool::calc_size_preserved_ratio(
                            original_size,
                            { align_size, hparams.image_min_pixels, hparams.image_max_pixels, 0 });
-    // tile if either dimension exceeds tile_size with tolerance
-    const bool needs_tiling = original_size.width > tile_size * max_pixels_tolerance || original_size.height > tile_size * max_pixels_tolerance;
+
+    const bool needs_tiling = should_tile(hparams, original_size);

    if (!needs_tiling) {
        inst.refined_size = clip_image_size{0, 0};
@@ -145,8 +145,11 @@ struct mtmd_image_preprocessor_lfm2 : mtmd_image_preprocessor_llava_uhd {
    static constexpr int   tile_size            = 512;

    using mtmd_image_preprocessor_llava_uhd::mtmd_image_preprocessor_llava_uhd;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
    slice_instructions get_slice_instructions(const clip_image_size & original_size) override;

+    static bool should_tile(const clip_hparams & hparams, const clip_image_size & original_size);
+
 private:
    clip_image_size find_closest_aspect_ratio(
            float aspect_ratio,
@@ -2322,23 +2322,12 @@ void mtmd_input_chunk_free(mtmd_input_chunk * chunk) {
    }
 }

-int32_t mtmd_input_chunk_save(const mtmd_input_chunk * chunk, char * out_buf, size_t out_len, size_t * expected_out_len) {
+// returns 0 on success
+static int32_t mtmd_input_chunk_save_impl(const mtmd_input_chunk * chunk, std::vector<char> & out_buf) {
    try {
        mtmd_serialization ser(MTMD_SERIALIZATION_VERSION);
        chunk->serialize(ser);
-
-        if (expected_out_len) {
-            *expected_out_len = ser.data.size();
-        }
-        if (!out_buf) {
-            // caller is only querying the required size
-            return 0;
-        }
-        if (out_len < ser.data.size()) {
-            LOG_ERR("%s: out_buf is too small, need %zu bytes, got %zu\n", __func__, ser.data.size(), out_len);
-            return -1;
-        }
-        std::memcpy(out_buf, ser.data.data(), ser.data.size());
+        out_buf = std::move(ser.data);
        return 0;
    } catch (const std::exception & e) {
        LOG_ERR("%s: %s\n", __func__, e.what());
@@ -2346,6 +2335,35 @@ int32_t mtmd_input_chunk_save(const mtmd_input_chunk * chunk, char * out_buf, si
    }
 }

+mtmd_input_chunk * mtmd_input_chunk_get_placeholder(const mtmd_input_chunk * chunk) {
+    // this is hacky, but still faster than copy the whole batch data
+    std::vector<char> buf;
+    if (mtmd_input_chunk_save_impl(chunk, buf) != 0) {
+        return nullptr;
+    }
+    return mtmd_input_chunk_load(buf.data(), buf.size());
+}
+
+int32_t mtmd_input_chunk_save(const mtmd_input_chunk * chunk, char * out_buf, size_t out_len, size_t * expected_out_len) {
+    std::vector<char> buf;
+    if (mtmd_input_chunk_save_impl(chunk, buf) != 0) {
+        return -1;
+    }
+    if (expected_out_len) {
+        *expected_out_len = buf.size();
+    }
+    if (!out_buf) {
+        // caller is only querying the required size
+        return 0;
+    }
+    if (out_len < buf.size()) {
+        LOG_ERR("%s: out_buf is too small, need %zu bytes, got %zu\n", __func__, buf.size(), out_len);
+        return -1;
+    }
+    std::memcpy(out_buf, buf.data(), buf.size());
+    return 0;
+}
+
 mtmd_input_chunk * mtmd_input_chunk_load(const char * buf, size_t len) {
    try {
        mtmd_serialization ser(MTMD_SERIALIZATION_VERSION, buf, len);
@@ -233,6 +233,9 @@ MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd
 MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk);
 MTMD_API void               mtmd_input_chunk_free(mtmd_input_chunk * chunk);

+// similar to mtmd_input_chunk_copy, but returns a placeholder chunk
+MTMD_API mtmd_input_chunk * mtmd_input_chunk_get_placeholder(const mtmd_input_chunk * chunk);
+
 // save/load an input chunk to/from a buffer (useful for KV save/load)
 // important: only chunk's metadata will be saved, the actual image/audio data will not be saved
 // the loaded chunk will always be a placeholder, cannot be used for mtmd_encode() or mtmd_batch_encode()
@@ -2023,7 +2023,6 @@ int llama_perplexity(int argc, char ** argv) {
    }

    const int32_t n_ctx = params.n_ctx;
-
    if (n_ctx <= 0) {
        LOG_ERR("%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
        return 1;
@@ -189,7 +189,7 @@ This endpoint is intended to be used internally by the Web UI and subject to cha
 Get a list of tools, each tool has these fields:
 - `tool` (string): the ID name of the tool, to be used in POST call. Example: `read_file`
 - `display_name` (string): the name to be displayed on UI. Example: `Read file`
- `type` (string): `"builtin"` for a built-in tool, or `"mcp"` for a tool exposed by an MCP server
+- `type` (string): `"server"` for a server tool, or `"mcp"` for a tool exposed by an MCP server
 - `permissions` (object): a mapping string --> boolean that indicates the permission required by this tool. This is useful for the UI to ask the user before calling the tool. For now, the only permission supported is `"write"`
 - `definition` (object): the OAI-compat definition of this tool

@@ -196,11 +196,11 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--ui-config, --webui-config JSON` | JSON that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG) |
 | `--ui-config-file, --webui-config-file PATH` | JSON file that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG_FILE) |
 | `--ui-mcp-proxy, --webui-mcp-proxy, --no-ui-mcp-proxy, --no-webui-mcp-proxy` | experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)<br/>(env: LLAMA_ARG_UI_MCP_PROXY) |
-| `--tools TOOL1,TOOL2,...` | experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)<br/>specify "all" to enable all tools<br/>available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_info<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_TOOLS) |
+| `--tools TOOL1,TOOL2,...` | experimental: whether to enable server tools for AI agents - do not enable in untrusted environments (default: no tools)<br/>specify "all" to enable all tools<br/>available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_info<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_TOOLS) |
 | `--tools-runtime OPTION` | experimental: run tools in a separate runtime environment (default: none, use host environment)<br/>available options:<br/>  'docker:<image>', 'podman:<image>': spin up a new container and reuse it for all invocations, clean up on server exit<br/>  'docker-container:<id>', 'podman-container:<id>': use an existing container by ID, won't stop on server exit<br/>  'ssh:<target>': run tools on a remote POSIX host over SSH, key-based auth and a trusted host key are required<br/><br/>(env: LLAMA_ARG_TOOLS_RUNTIME) |
 | `--mcp-servers-config PATH` | experimental: path to JSON file with MCP server definitions (Cursor-compatible format) - do not enable in untrusted environments (default: none)<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_MCP_SERVERS_CONFIG) |
 | `--mcp-servers-json JSON` | experimental: inline JSON with MCP server definitions (Cursor-compatible format) - do not enable in untrusted environments (default: none)<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_MCP_SERVERS_JSON) |
-| `-ag, --agent, -no-ag, --no-agent` | whether to enable CORS proxy and all built-in tools - do not enable in untrusted environments (default: disabled)<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_AGENT) |
+| `-ag, --agent, -no-ag, --no-agent` | whether to enable CORS proxy and all server tools - do not enable in untrusted environments (default: disabled)<br/>note: for security reasons, this will limit --cors-origins to localhost by default<br/>(env: LLAMA_ARG_AGENT) |
 | `--ui, --webui, --no-ui, --no-webui` | whether to enable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_UI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--rerank, --reranking` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
@@ -337,12 +337,64 @@ It is currently available in the following endpoints:

 For more details, please refer to [multimodal documentation](../../docs/multimodal.md)

-### Built-in tools support
+### Server tools support

-The server includes a set of built-in tools that enable the LLM to access the local file system directly from the Web UI.
+The server includes a set of server tools that enable the LLM to access the local file system directly from the Web UI.

 To use this feature, start the server with `--tools all`. You can also enable only specific tools by passing a comma-separated list: `--tools name1,name2,...`. Run `--help` for the full list of available tool names.

+### MCP servers
+
+Besides the built-in tools, the server can expose tools coming from MCP servers, added in [#26062](https://github.com/ggml-org/llama.cpp/pull/26062). Only the stdio transport is supported: such a server is a child process reading JSON-RPC messages on its stdin and writing replies on its stdout, so nothing has to be started or maintained outside `llama-server`.
+
+Servers are declared in a Cursor-compatible JSON file:
+
+```json
+{
+  "mcpServers": {
+    "example": { "command": "/path/to/server", "args": [] }
+  }
+}
+```
+
+```sh
+llama-server -m model.gguf --mcp-servers-config mcp.json
+```
+
+The same JSON can be passed inline with `--mcp-servers-json`. Each entry under `mcpServers` accepts:
+
+| Key | Explanation |
+| --- | ----------- |
+| `command` | executable to spawn, required, entries without it are skipped |
+| `args` | array of arguments |
+| `env` | object merged over the parent environment |
+| `cwd` | working directory of the child process |
+| `timeout_ms` | per-tool-call timeout (default: 30000) |
+
+Every server is spawned once at startup to list its tools, then stopped, and respawned on demand when one of its tools is called. Tools are exposed as `<server>_<tool>` alongside the built-in ones: they show up in the Web UI and in `GET /tools`, and the model calls them like any other tool. A name colliding with an already registered tool is skipped. This is independent of `--tools`, MCP servers can be the only tools available.
+
+The child process runs with the same privileges as the server, so only declare commands you trust. As with `--tools`, `--cors-origins` then defaults to `localhost`.
+
+Note: `--ui-mcp-proxy` is unrelated, it only lets the Web UI reach remote MCP servers from the browser.
+
+Any server written against the [MCP specification](https://modelcontextprotocol.io) works as is, whether it uses an official SDK or not: the transport is one JSON-RPC message per line on stdio, so a script wrapping an existing program is a valid server too.
+
+### CORS
+
+By default the server reflects any `Origin` header back with credentials allowed. This matches the old, always-on `*` behavior and is fine as long as the server only exposes stateless, read-only endpoints.
+
+Enabling `--tools` or `--agent` exposes file read/write over the API, so in that case `--cors-origins` defaults to `localhost` instead: only pages served from localhost can reach the server. Pass `--cors-origins` explicitly to override either default.
+
+Recommended `--cors-origins` setting, depending on where the server runs:
+
+| Deployment | Recommendation |
+| ---------- | --------------- |
+| Public | set an API key, put the server behind a reverse proxy, `--cors-origins` optional |
+| Local network | set `--cors-origins` to your frontend's origin |
+| Same machine | `--cors-origins localhost` (default once `--agent` is set) |
+
+Related flags: `--cors-origins`, `--cors-methods`, `--cors-headers`, `--cors-credentials` / `--no-cors-credentials`. Background and rationale: [#25655](https://github.com/ggml-org/llama.cpp/pull/25655).
+
 ## Build

 `llama-server` is built alongside everything else from the root of the project
@@ -1579,9 +1631,9 @@ curl http://localhost:8080/v1/messages/count_tokens \
 {"input_tokens": 10}
 ```

-## Server built-in tools
+## Server tools

-The server exposes a REST API under `/tools` that allows the Web UI to call built-in tools. This endpoint is intended to be used internally by the Web UI and subject to change or to be removed in the future.
+The server exposes a REST API under `/tools` that allows the Web UI to call server tools. This endpoint is intended to be used internally by the Web UI and subject to change or to be removed in the future.

 **Please do NOT use this endpoint in a downstream application**

@@ -507,6 +507,23 @@ void server_tokens::push_back(const mtmd_input_chunk * chunk) {
    }
 }

+void server_tokens::push_back_placeholder(const mtmd_input_chunk * chunk) {
+    auto type = mtmd_input_chunk_get_type(chunk);
+    if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE || type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        GGML_ASSERT(has_mtmd);
+        mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_get_placeholder(chunk));
+        GGML_ASSERT(new_chunk != nullptr && "failed to create placeholder chunk");
+        const size_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
+        size_t start_idx = tokens.size();
+        for (size_t i = 0; i < n_tokens; ++i) {
+            tokens.emplace_back(LLAMA_TOKEN_NULL);
+        }
+        map_idx_to_media[start_idx] = std::move(new_chunk);
+    } else {
+        push_back(chunk);
+    }
+}
+
 void server_tokens::push_back(server_tokens & tokens) {
    size_t start_idx = size();
    for (size_t i = 0; i < tokens.size(); i++) {
@@ -195,6 +195,10 @@ public:
    // will create a copy of the chunk if it contains non-text data
    void push_back(const mtmd_input_chunk * chunk);

+    // same as push_back, but media chunks are stored as placeholders (no image/audio data)
+    // only use this if the chunk will never be encoded again (e.g. it is already in the KV cache)
+    void push_back_placeholder(const mtmd_input_chunk * chunk);
+
    // appends server tokens, updates the media map. copies media chunks.
    void push_back(server_tokens & tokens);

@@ -3416,7 +3416,8 @@ private:
                        // add the mtmd chunk to cache
                        {
                            const auto & chunk = input_tokens.find_chunk(cur_token_idx);
-                            slot.prompt.tokens.push_back(chunk.get()); // copy
+                            // the chunk is already in the KV cache at this point, so we don't need to keep its data around
+                            slot.prompt.tokens.push_back_placeholder(chunk.get());
                        }

                        has_mtmd = true;
@@ -2035,7 +2035,7 @@ void server_tools::setup(const std::vector<std::string> & enabled_tools,
        }
    }

-    // append MCP tools, skipping any that collide with a built-in or another MCP tool of the same "<server>_<tool>" name
+    // append MCP tools, skipping any that collide with a server tool or another MCP tool of the same "<server>_<tool>" name
    if (!mcp_mgr.empty()) {
        std::unordered_set<std::string> seen_names;
        for (auto & t : tools) {
@@ -18,7 +18,7 @@ struct server_tool {

    virtual ~server_tool() = default;
    virtual json get_definition() const = 0;
-    virtual std::string type() const { return "builtin"; }
+    virtual std::string type() const { return "server"; }

    struct stream {
        server_response & qr;
@@ -346,7 +346,7 @@ int llama_server(common_params & params, int argc, char ** argv) {
        ctx_http.get ("/tools",           ex_wrapper(tools.handle_get));
        ctx_http.post("/tools",           ex_wrapper(tools.handle_post));
        if (!params.server_tools.empty()) {
-            warn_names.push_back("built-in tools (experimental)");
+            warn_names.push_back("server tools (experimental)");
        }
        if (!params.server_tools_runtime.empty()) {
            warn_names.push_back("tools runtime (experimental)");
@@ -259,6 +259,8 @@ int main(int argc, char ** argv) {
            }
            cpp += fmt("static const unsigned char asset_%d_data[] = {", i);
            append_bytes_hex(cpp, bytes);
+
+            // note: this is a simple hash for cache busting, not a cryptographic hash; fnv is enough here
            const auto hash = fnv_hash(bytes.data(), bytes.size());

            cpp += fmt("};\nstatic const std::size_t   asset_%d_size = %zu;\n",
@@ -61,6 +61,9 @@ export default ts.config(
 				{ blankLine: 'always', next: ['return', 'throw', 'break', 'continue'], prev: '*' }
 			],

+			// Alphabetical order for enum members
+			'perfectionist/sort-enums': ['error', { type: 'natural' }],
+
 			'perfectionist/sort-objects': ['error', { type: 'natural' }],

 			// Alphabetical order for variable declarations and object keys
@@ -48,6 +48,7 @@
 		containsFileMentionLink,
 		findCommandToken,
 		findMentionToken,
+		getConversationModel,
 		isIMEComposing,
 		isOffsetInCodeBlock,
 		parseClipboardContent,
@@ -190,31 +191,9 @@

 	let isRouter = $derived(serverStore.isRouterMode);
 	let conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);
-	let activeModelId = $derived.by(() => {
-		const options = modelsStore.models;
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-
-			if (model) return model.model;
-		}
-
-		return null;
-	});
+	let activeModelId = $derived(modelsStore.activeModelId);

 	let hasModelSelected = $derived(
 		!isRouter || !!conversationModel || !!modelsStore.selectedModelId
@@ -35,7 +35,7 @@
 						<span>
 							Run llama-server with <code>{CLI_FLAGS.TOOLS}</code> flag to enable

-							<strong>Built-in Tools</strong>.
+							<strong>Server Tools</strong>.
 						</span>
 					</span>

@@ -2,10 +2,10 @@
 	import ChatFormActionAddButton from './ChatFormActionAddButton.svelte';
 	import ChatFormActionAddDropdown from './ChatFormActionAddDropdown.svelte';
 	import ChatFormActionAddSheet from './ChatFormActionAddSheet.svelte';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 </script>

-{#if isMobile.current}
+{#if deviceStore.isMobile}
 	<ChatFormActionAddSheet>
 		{#snippet trigger({ disabled, onclick })}
 			<ChatFormActionAddButton {disabled} {onclick} />
@@ -1,6 +1,7 @@
 <script lang="ts">
 	import { ModelsSelectorDropdown, ModelsSelectorSheet } from '$lib/components/app';
-	import { chatStore, conversationsStore, isMobile, modelsStore, serverStore } from '$lib/stores';
+	import { conversationsStore, deviceStore, modelsStore, serverStore } from '$lib/stores';
+	import { getConversationModel } from '$lib/utils';

 	interface Props {
 		disabled?: boolean;
@@ -30,7 +31,7 @@
 	let isOffline = $derived(!!serverStore.error);

 	let conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);

 	let lastSyncedConversationModel: string | null = null;
@@ -74,29 +75,7 @@
 		}
 	});

-	let activeModelId = $derived.by(() => {
-		const options = modelsStore.models;
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-
-			if (model) return model.model;
-		}
-
-		return null;
-	});
+	let activeModelId = $derived(modelsStore.activeModelId);

 	let modelPropsVersion = $state(0); // Used to trigger reactivity after fetch

@@ -170,7 +149,7 @@
 	}
 </script>

-{#if isMobile.current}
+{#if deviceStore.isMobile}
 	<ModelsSelectorSheet
 		disabled={disabled || isOffline}
 		bind:this={selectorModelRef}
@@ -1,15 +1,14 @@
 <script lang="ts">
 	import ContextGaugeDial from './ContextGaugeDial.svelte';
-	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
 	import {
-		chatStore,
-		conversationsStore,
 		gaugeTriggerClick,
 		gaugeTriggerEnter,
 		gaugeTriggerKeydown,
 		gaugeTriggerLeave,
 		gaugeTriggerPointerDown
-	} from '$lib/stores';
+	} from './gauge-popup.svelte';
+	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
+	import { chatStore, conversationsStore } from '$lib/stores';
 	import { untrack } from 'svelte';

 	const gauge = useContextGauge();
@@ -1,9 +1,9 @@
 <script lang="ts">
 	import ContextGaugeDetailRow from './ContextGaugeDetailRow.svelte';
+	import { gaugePopup } from './gauge-popup.svelte';
 	import { ChevronDown } from '@lucide/svelte';
 	import * as Collapsible from '$lib/components/ui/collapsible';
 	import { STATS_UNITS } from '$lib/constants';
-	import { gaugePopup } from '$lib/stores/context-gauge-popup.svelte';

 	interface Props {
 		currentRead: number;
@@ -2,8 +2,13 @@
 	import { colorLevelBgClass, colorLevelTextClass } from './context-gauge';
 	import ContextGaugeDetails from './ContextGaugeDetails.svelte';
 	import ContextGaugeLoadModel from './ContextGaugeLoadModel.svelte';
+	import {
+		gaugeCardEnter,
+		gaugeCardLeave,
+		gaugePopup,
+		gaugePopupClose
+	} from './gauge-popup.svelte';
 	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
-	import { gaugeCardEnter, gaugeCardLeave, gaugePopup, gaugePopupClose } from '$lib/stores';
 	import { formatParameters } from '$lib/utils/formatters';

 	const gauge = useContextGauge();
@@ -62,7 +62,7 @@
 	// it, the picker still opens for manual entry but explains why search is
 	// unavailable instead of firing searches that would only fail. Browse is
 	// hidden too: it resolves the picked folder name through the same tool.
-	const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.FILE_GLOB_SEARCH));
+	const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.SERVER_FILE_GLOB_SEARCH));
 	const fileSearchEnabled = $derived(
 		fileSearchKey !== null && toolsStore.isToolEnabled(fileSearchKey)
 	);
@@ -212,7 +212,7 @@
 	// so the caller fails visibly instead of committing a bare leaf name.
 	async function resolveNativeName(name: string): Promise<string | null> {
 		try {
-			const res = await ToolsService.executeToolRaw(BuiltInTool.FILE_GLOB_SEARCH, {
+			const res = await ToolsService.executeToolRaw(BuiltInTool.SERVER_FILE_GLOB_SEARCH, {
 				include: buildCaseInsensitiveGlob(name),
 				limit: SEARCH.NATIVE_LIMIT,
 				max_depth: SEARCH.NATIVE_MAX_DEPTH,
@@ -1,5 +1,5 @@
 <script lang="ts">
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import { autoResizeTextarea } from '$lib/utils';
 	import { onMount } from 'svelte';

@@ -37,7 +37,7 @@
 	}

 	export function focus() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;

 		textareaElement?.focus({ preventScroll: true });
 	}
@@ -1,7 +1,7 @@
 <script lang="ts">
 	import { CODE_BLOCK, CODE_TOKEN_ATTR, UI_DATA_ATTRS } from '$lib/constants';
 	import { BooleanString, ChatFormInputRichTokenKind, ColorMode } from '$lib/enums';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import type { ChatFormInputRichToken } from '$lib/types';
 	import type { SourceHistoryEntry } from '$lib/utils';
 	import {
@@ -750,7 +750,7 @@
 		syncEmptyState();
 		document.addEventListener('selectionchange', handleSelectionChange);

-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			rootElement?.focus({ preventScroll: true });
 		}
 	});
@@ -792,7 +792,7 @@
 	}

 	export function focus() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;

 		rootElement?.focus({ preventScroll: true });
 	}
@@ -8,7 +8,7 @@
 	import { BuiltInTool, FileMentionEntryType, GlobSearchType, KeyboardKey } from '$lib/enums';
 	import { useDebouncedSearch } from '$lib/hooks/use-debounced-search.svelte';
 	import { usePickerNavigation } from '$lib/hooks/use-picker-navigation.svelte';
-	import { isMobile, settingsStore, toolsStore } from '$lib/stores';
+	import { deviceStore, settingsStore, toolsStore } from '$lib/stores';
 	import type { FileMentionEntry, GlobEntryResult } from '$lib/types';
 	import { abbreviateHome, runGlobSearchWithChildren } from '$lib/utils';

@@ -51,7 +51,7 @@
 	// When the server does not expose file_glob_search (started without
 	// --tools) or the user disabled it, the picker still opens but explains
 	// why instead of firing searches that would only fail.
-	const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.FILE_GLOB_SEARCH));
+	const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.SERVER_FILE_GLOB_SEARCH));
 	const fileSearchEnabled = $derived(
 		fileSearchKey !== null && toolsStore.isToolEnabled(fileSearchKey)
 	);
@@ -130,7 +130,7 @@
 		return searchError ? `Search failed - ${searchError}` : 'No matching files or folders';
 	});

-	const showTooltip = $derived(!isMobile.current);
+	const showTooltip = $derived(!deviceStore.isMobile);

 	$effect(() => {
 		if (typeof window === 'undefined') return;
@@ -11,7 +11,7 @@
 	import { setChatMessageActionsContext, setChatMessageEditContext } from '$lib/contexts';
 	import { AgenticSectionType, AttachmentType, MessageRole } from '$lib/enums';
 	import { DatabaseService } from '$lib/services/database.service';
-	import { chatStore, conversationsStore, isMobile } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore } from '$lib/stores';
 	import type {
 		ChatMessageActions,
 		ChatMessageDeletionInfo,
@@ -304,7 +304,7 @@

 	// After the system message flow ends, hand focus to the main chat form
 	function focusMainChatForm() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;

 		document.querySelector<HTMLTextAreaElement>('.chat-screen-form-wrapper textarea')?.focus();
 	}
@@ -35,19 +35,19 @@

 {#if isSearchCall}
 	<ChatMessageToolCallBlockSearchResults {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.GET_DATETIME}
+{:else if section.toolName === BuiltInTool.BROWSER_GET_DATETIME}
 	<ChatMessageToolCallBlockGetDatetime {section} {isStreaming} />
-{:else if section.toolName === BuiltInTool.GET_INFO}
+{:else if section.toolName === BuiltInTool.SERVER_GET_INFO}
 	<ChatMessageToolCallBlockGetInfo {section} {isStreaming} />
-{:else if section.toolName === BuiltInTool.READ_FILE}
+{:else if section.toolName === BuiltInTool.SERVER_READ_FILE}
 	<ChatMessageToolCallBlockReadFile {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.READ_MEDIA}
+{:else if section.toolName === BuiltInTool.BROWSER_READ_MEDIA}
 	<ChatMessageToolCallBlockReadMedia {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.EDIT_FILE}
+{:else if section.toolName === BuiltInTool.SERVER_EDIT_FILE}
 	<ChatMessageToolCallBlockEditFile {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.WRITE_FILE}
+{:else if section.toolName === BuiltInTool.SERVER_WRITE_FILE}
 	<ChatMessageToolCallBlockWriteFile {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.EXEC_SHELL_COMMAND}
+{:else if section.toolName === BuiltInTool.SERVER_EXEC_SHELL_COMMAND}
 	<ChatMessageToolCallBlockExecShellCommand
 		{section}
 		{open}
@@ -56,11 +56,11 @@
 		{attachments}
 		{onToggle}
 	/>
-{:else if section.toolName === BuiltInTool.FILE_GLOB_SEARCH}
+{:else if section.toolName === BuiltInTool.SERVER_FILE_GLOB_SEARCH}
 	<ChatMessageToolCallBlockFileGlobSearch {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.GREP_SEARCH}
+{:else if section.toolName === BuiltInTool.SERVER_GREP_SEARCH}
 	<ChatMessageToolCallBlockGrepSearch {section} {open} {isStreaming} {onToggle} />
-{:else if section.toolName === BuiltInTool.RUN_JAVASCRIPT}
+{:else if section.toolName === BuiltInTool.BROWSER_RUN_JAVASCRIPT}
 	<ChatMessageToolCallBlockRunJavascript {section} {open} {isStreaming} {onToggle} />
 {:else}
 	<ChatMessageToolCallBlockDefault {section} {open} {isStreaming} {attachments} {onToggle} />
@@ -12,7 +12,7 @@
 	import {
 		classifyToolResult,
 		formatJsonPretty,
-		getBuiltinToolUi,
+		getToolUi,
 		parseToolResultWithMedia
 	} from '$lib/utils';
 	import { createBase64DataUrl } from '$lib/utils/data-url';
@@ -27,7 +27,7 @@

 	let { attachments, isStreaming, onToggle, open, section }: Props = $props();

-	const title = $derived(getBuiltinToolUi(section.toolName)?.label ?? section.toolName ?? '');
+	const title = $derived(getToolUi(section.toolName)?.label ?? section.toolName ?? '');
 	const outputKind = $derived(classifyToolResult(section.toolResult));
 	const parsedLines: ToolResultLine[] = $derived(
 		section.toolResult ? parseToolResultWithMedia(section.toolResult, attachments) : []
@@ -6,7 +6,7 @@
 	import { MAX_HEIGHT_CODE_BLOCK } from '$lib/constants';
 	import { FileTypeText } from '$lib/enums';
 	import type { AgenticSection } from '$lib/types';
-	import { getBuiltinToolUi } from '$lib/utils';
+	import { getToolUi } from '$lib/utils';

 	interface Props {
 		section: AgenticSection;
@@ -18,7 +18,7 @@
 	let { isStreaming, onToggle, open, section }: Props = $props();

 	const runJsMeta = $derived(parseRunJavascriptMeta(section));
-	const title = $derived(getBuiltinToolUi(section.toolName)?.label ?? section.toolName ?? '');
+	const title = $derived(getToolUi(section.toolName)?.label ?? section.toolName ?? '');
 </script>

 <ToolCallBlock {section} {open} {isStreaming} meta={runJsMeta} {title} {onToggle}>
@@ -14,8 +14,8 @@
 	import { ICON_CLASS_DEFAULT, ICON_CLASS_SPIN } from '$lib/constants';
 	import { AgenticSectionType } from '$lib/enums';
 	import { mcpStore } from '$lib/stores';
-	import type { AgenticSection, BuiltinToolUiEntry } from '$lib/types';
-	import { getBuiltinToolUi } from '$lib/utils';
+	import type { AgenticSection, ToolUiEntry } from '$lib/types';
+	import { getToolUi } from '$lib/utils';
 	import type { Component, Snippet } from 'svelte';

 	type ToolCallBlockMetaWithError = TMeta & { errorMessage?: string };
@@ -82,7 +82,7 @@
 	const showSpinner = $derived(isPending || (isStreamingCall && isStreaming) || extraLiveStreaming);
 	const isCodeStreaming = $derived(isStreaming && (isPending || isStreamingCall));

-	const toolUi: BuiltinToolUiEntry | null = $derived(getBuiltinToolUi(section.toolName));
+	const toolUi: ToolUiEntry | null = $derived(getToolUi(section.toolName));
 	const toolIcon: Component = $derived(
 		spinIconWhenActive && showSpinner ? Loader2 : (toolUi?.icon ?? Wrench)
 	);
@@ -24,7 +24,7 @@ export type EditFileMeta = {
 };

 export function parseEditFileMeta(section: AgenticSection): EditFileMeta | null {
-	const args = parseToolArgs(BuiltInTool.EDIT_FILE, section, { partial: true });
+	const args = parseToolArgs(BuiltInTool.SERVER_EDIT_FILE, section, { partial: true });

 	if (!args) return null;

@@ -14,7 +14,7 @@ export type ExecShellCommandMeta = {
 };

 export function parseExecShellCommandMeta(section: AgenticSection): ExecShellCommandMeta | null {
-	const args = parseToolArgs(BuiltInTool.EXEC_SHELL_COMMAND, section);
+	const args = parseToolArgs(BuiltInTool.SERVER_EXEC_SHELL_COMMAND, section);

 	if (!args) return null;

@@ -19,7 +19,7 @@ export type FileGlobSearchMeta = {
 };

 export function parseFileGlobSearchMeta(section: AgenticSection): FileGlobSearchMeta | null {
-	const args = parseToolArgs(BuiltInTool.FILE_GLOB_SEARCH, section);
+	const args = parseToolArgs(BuiltInTool.SERVER_FILE_GLOB_SEARCH, section);

 	if (!args) return null;

@@ -28,7 +28,7 @@ export type GrepSearchMeta = {
 };

 export function parseGrepSearchMeta(section: AgenticSection): GrepSearchMeta | null {
-	const args = parseToolArgs(BuiltInTool.GREP_SEARCH, section);
+	const args = parseToolArgs(BuiltInTool.SERVER_GREP_SEARCH, section);

 	if (!args) return null;

@@ -16,7 +16,7 @@ export type ReadFileMeta = {
 };

 export function parseReadFileMeta(section: AgenticSection): ReadFileMeta | null {
-	const args = parseToolArgs(BuiltInTool.READ_FILE, section, { partial: true });
+	const args = parseToolArgs(BuiltInTool.SERVER_READ_FILE, section, { partial: true });

 	if (!args) return null;

@@ -16,7 +16,7 @@ export type RunJavascriptMeta = {
 };

 export function parseRunJavascriptMeta(section: AgenticSection): RunJavascriptMeta | null {
-	const args = parseToolArgs(BuiltInTool.RUN_JAVASCRIPT, section);
+	const args = parseToolArgs(BuiltInTool.BROWSER_RUN_JAVASCRIPT, section);

 	if (!args) return null;

@@ -20,7 +20,7 @@ export type WriteFileMeta = {
 };

 export function parseWriteFileMeta(section: AgenticSection): WriteFileMeta | null {
-	const args = parseToolArgs(BuiltInTool.WRITE_FILE, section, { partial: true });
+	const args = parseToolArgs(BuiltInTool.SERVER_WRITE_FILE, section, { partial: true });

 	if (!args) return null;

@@ -61,8 +61,8 @@
 				{:else}
 					{@const source = toolsStore.getToolSource(toolName)}
 					{@const providerName =
-						source === ToolSource.BUILTIN
-							? TOOL_SERVER_LABELS[ToolSource.BUILTIN]
+						source === ToolSource.SERVER
+							? TOOL_SERVER_LABELS[ToolSource.SERVER]
 							: source === ToolSource.CUSTOM
 								? TOOL_SERVER_LABELS[ToolSource.CUSTOM]
 								: 'MCP Tools'}
@@ -21,8 +21,7 @@
 	import {
 		chatStore,
 		conversationsStore,
-		device,
-		isMobile,
+		deviceStore,
 		serverStore,
 		settingsStore
 	} from '$lib/stores';
@@ -32,7 +31,7 @@
 	let { showCenteredEmpty = false } = $props();

 	let disableAutoScroll = $derived(
-		Boolean(settingsStore.config.disableAutoScroll) || isMobile.current
+		Boolean(settingsStore.config.disableAutoScroll) || deviceStore.isMobile
 	);
 	let isMobileUserScrolledUp = $state(false);
 	let mobileScrollDownHint = $state(false);
@@ -52,11 +51,11 @@
 	let hasPropsError = $derived(!!serverStore.error);
 	let isCurrentConversationLoading = $derived(chatStore.isLoading || chatStore.isStreaming());
 	let chatFormBottomPosition = $derived.by(() => {
-		if (!isMobile.current) return '1rem';
+		if (!deviceStore.isMobile) return '1rem';

-		if (device.isStandalone) return '1.5rem';
+		if (deviceStore.isStandalone) return '1.5rem';

-		if (device.isIOSSafari) return '0.25rem';
+		if (deviceStore.isIOSSafari) return '0.25rem';

 		return '0.5rem';
 	});
@@ -84,7 +83,7 @@
 	});

 	function handleMobileScroll() {
-		if (!isMobile.current) return;
+		if (!deviceStore.isMobile) return;

 		const container = scroll.chatScrollContainer;

@@ -184,7 +183,7 @@
 	}

 	function handleSendLikeScroll() {
-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			autoScroll.enable();
 		}

@@ -197,7 +196,7 @@
 				'.chat-message:nth-last-child(2) .chat-message-user .chat-message-user-bubble'
 			) as HTMLElement | null;

-			if (isMobile.current) {
+			if (deviceStore.isMobile) {
 				// Keep the last user message bubble just above the input on mobile
 				const bubbleHeight = lastUserBubble?.scrollHeight ?? 0;
 				const baseHeight = container.scrollHeight - innerHeight;
@@ -220,7 +219,7 @@
 			}
 		}, 100);

-		if (isMobile.current) {
+		if (deviceStore.isMobile) {
 			autoScroll.setDisabled(disableAutoScroll);
 			mobileScrollDownHint = true;
 			mobileScrollDownHintLockedUntil = Date.now() + 500;
@@ -243,7 +242,8 @@

 	$effect(() => {
 		const shouldDisableAutoScroll =
-			settingsStore.config.disableAutoScroll || (isMobile.current && isCurrentConversationLoading);
+			settingsStore.config.disableAutoScroll ||
+			(deviceStore.isMobile && isCurrentConversationLoading);

 		autoScroll.setDisabled(shouldDisableAutoScroll);

@@ -266,7 +266,7 @@
 			autoScroll.enable();
 		}

-		if (isMobile.current && isCurrentConversationLoading) {
+		if (deviceStore.isMobile && isCurrentConversationLoading) {
 			mobileScrollDownHint = true;
 			mobileScrollDownHintLockedUntil = Date.now() + 500;
 		}
@@ -318,9 +318,9 @@
 		<div
 			class={[
 				'pointer-events-none md:sticky fixed  mt-auto transition-all duration-200',
-				device.isStandalone
+				deviceStore.isStandalone
 					? 'bottom-6 right-4 left-4'
-					: device.isIOSSafari
+					: deviceStore.isIOSSafari
 						? 'bottom-1 left-2 right-2'
 						: 'bottom-2 right-2 left-2',
 				isEmpty ? 'md:bottom-[calc(50dvh-7rem)] 2xl:bottom-[calc(50dvh-4rem)]' : 'md:bottom-4'
@@ -336,7 +336,7 @@
 			{/if}

 			<div class="pointer-events-none flex flex-col gap-6 items-center w-full">
-				{#if (isMobile.current ? mobileScrollDownHint || isMobileUserScrolledUp : autoScroll.userScrolledUp) && page.url.hash.includes(ROUTES.CHAT) && page.params.id}
+				{#if (deviceStore.isMobile ? mobileScrollDownHint || isMobileUserScrolledUp : autoScroll.userScrolledUp) && page.url.hash.includes(ROUTES.CHAT) && page.params.id}
 					<ChatScreenActionScrollDown
 						onclick={() => {
 							mobileScrollDownHint = false;
@@ -3,7 +3,7 @@
 	import { page } from '$app/state';
 	import { ChatForm } from '$lib/components/app';
 	import { useDraftMessages } from '$lib/hooks/use-draft-messages.svelte';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import { onMount } from 'svelte';

 	interface Props {
@@ -120,13 +120,13 @@
 	}

 	onMount(() => {
-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			setTimeout(focusFormUnlessCaptured, 100);
 		}
 	});

 	afterNavigate((navigation) => {
-		if (navigation?.from != null && !isMobile.current) {
+		if (navigation?.from != null && !deviceStore.isMobile) {
 			setTimeout(focusFormUnlessCaptured, 100);
 		}
 	});
@@ -278,7 +278,7 @@ export { default as ChatFormInput } from './ChatForm/ChatFormInput/ChatFormInput
 /**
 * Working directory selector for agent mode. Renders a chip below the chat
 * form; clicking it opens a popover with a directory picker backed by the
- * server's `file_glob_search` built-in tool (POST /tools). The picked
+ * server's `file_glob_search` server tool (POST /tools). The picked
 * directory is exposed via `bind:directory`; changing it records a
 * synthetic "Set working directory to ..." user message into chat history
 * and is enforced on tool calls via the `x-tool-cwd` request header.
@@ -380,7 +380,7 @@ export { default as ChatFormPickerListItemSkeleton } from './ChatForm/ChatFormPi

 /**
 * `@`-triggered file/folder mention picker. Resolves `@<query>` in the chat
- * input to a filesystem match via the server's `file_glob_search` built-in
+ * input to a filesystem match via the server's `file_glob_search` server tool
 * tool, scoped to the conversation cwd (or server home when unset).
 * Selection splices a `[name](file:///<abs path>)` link into the input.
 */
@@ -14,7 +14,7 @@
 	import { useKeyboardShortcuts } from '$lib/hooks/use-keyboard-shortcuts.svelte';
 	import { useMarqueeSelection } from '$lib/hooks/use-marquee-selection.svelte';
 	import { RouterService } from '$lib/services/router.service';
-	import { chatStore, conversationsStore, device, isMobile, settingsStore } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore, settingsStore } from '$lib/stores';
 	import { buildConversationTree } from '$lib/utils';
 	import { circIn } from 'svelte/easing';
 	import { SvelteSet } from 'svelte/reactivity';
@@ -36,7 +36,7 @@
 	let logoHovered = $state(false);

 	const isStripExpanded = $derived(isExpandedMode || hoveredTooltip !== null);
-	const isOnMobile = $derived(isMobile.current);
+	const isOnMobile = $derived(deviceStore.isMobile);
 	const alwaysShowOnDesktop = $derived(settingsStore.config.alwaysShowSidebarOnDesktop as boolean);

 	$effect(() => {
@@ -65,7 +65,7 @@
 	});

 	$effect(() => {
-		if (isMobile.current && page.url.hash.includes(ROUTES.SEARCH)) {
+		if (deviceStore.isMobile && page.url.hash.includes(ROUTES.SEARCH)) {
 			isExpandedMode = false;
 		}
 	});
@@ -227,7 +227,7 @@
 	}

 	async function selectConversation(id: string) {
-		if (isMobile.current) {
+		if (deviceStore.isMobile) {
 			scheduleMobileCollapse();
 		}

@@ -315,9 +315,9 @@
 			'fixed md:sticky top-2 left-2 md:left-0 md:ml-2 md:mt-2 pt-2 z-10 w-[calc(100dvw-1rem)]',
 			'md:h-[calc(100dvh-1.125rem)]',
 			isExpandedMode &&
-				(device.isStandalone
+				(deviceStore.isStandalone
 					? 'h-[calc(100dvh-2rem)]'
-					: device.isIOSDevice
+					: deviceStore.isIOSDevice
 						? 'h-[calc(100dvh-0.5rem)]'
 						: 'h-[calc(100dvh-1rem)]'),
 			'rounded-3xl md:rounded-2xl',
@@ -353,7 +353,7 @@

 			{#if isOnMobile || (isExpandedMode && !alwaysShowOnDesktop)}
 				<div
-					class="flex items-center transition-all duration-150 ease-out {isMobile.current &&
+					class="flex items-center transition-all duration-150 ease-out {deviceStore.isMobile &&
 					!isExpandedMode
 						? 'opacity-0 h-0!'
 						: ''}"
@@ -361,7 +361,7 @@
 					out:fade={{ duration: 100 }}
 				>
 					<ActionIcon
-						icon={isMobile.current ? X : PanelLeftClose}
+						icon={deviceStore.isMobile ? X : PanelLeftClose}
 						size="lg"
 						iconSize="h-4.5 w-4.5 md:h-4 md:w-4"
 						class="backdrop-blur-none md:h-9 md:w-9 h-10 w-10 rounded-full mr-1 hover:bg-accent!"
@@ -375,9 +375,9 @@
 		</div>

 		<div
-			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {isMobile.current
+			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {deviceStore.isMobile
 				? 'transition-[opacity,height] duration-200 ease-out'
-				: ''} {isMobile.current && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
+				: ''} {deviceStore.isMobile && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
 			in:fade={{ duration: 200 }}
 			out:fade={{ duration: 200 }}
 		>
@@ -395,7 +395,7 @@
 					isSearchModeActive = true;
 				}}
 				onNewChat={() => {
-					if (isMobile.current) {
+					if (deviceStore.isMobile) {
 						scheduleMobileCollapse();
 					}
 				}}
@@ -12,7 +12,7 @@
 		SIDEBAR_ACTIONS_ITEMS
 	} from '$lib/constants';
 	import { TooltipSide } from '$lib/enums';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import type { Component } from 'svelte';
 	import { onMount } from 'svelte';
 	import { circIn } from 'svelte/easing';
@@ -42,7 +42,7 @@
 	let showIcons = $state(false);
 	let searchInputRef = $state<HTMLInputElement | null>(null);

-	const isOnMobile = $derived(isMobile.current);
+	const isOnMobile = $derived(deviceStore.isMobile);

 	$effect(() => {
 		if (isSearchModeActive && searchInputRef) {
@@ -107,7 +107,7 @@
 	>
 		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
-			{@const isSearchOnMobile = item.icon === Search && isMobile.current}
+			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemHref = isSearchOnMobile ? ROUTES.SEARCH : item.route}
 			{@const itemOnClick = item.route
 				? () => {
@@ -156,7 +156,7 @@
 	<div class="{className} flex-col gap-1 hidden md:flex">
 		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
-			{@const isSearchOnMobile = item.icon === Search && isMobile.current}
+			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemOnClick = item.route
 				? () => {
 						onNewChat?.();
@@ -8,6 +8,7 @@
 	} from '$lib/components/app';
 	import SettingsGroup from '$lib/components/app/settings/SettingsGroup.svelte';
 	import { ConversationSelectionMode, FileExtensionText, HtmlInputType } from '$lib/enums';
+	import { ConversationTransferService } from '$lib/services';
 	import { conversationsStore, settingsStore } from '$lib/stores';
 	import { createMessageCountMap } from '$lib/utils';
 	import { fade } from 'svelte/transition';
@@ -147,9 +148,9 @@
 			);

 			if (allData.length === 1) {
-				conversationsStore.downloadConversationFile(allData[0]);
+				ConversationTransferService.downloadConversationFile(allData[0]);
 			} else {
-				conversationsStore.downloadConversationsArchive(allData);
+				ConversationTransferService.downloadConversationsArchive(allData);
 			}

 			exportedConversations = selectedConversations;
@@ -177,7 +178,7 @@
 				if (!file) return;

 				try {
-					const importedData = await conversationsStore.parseImportFile(file);
+					const importedData = await ConversationTransferService.parseImportFile(file);

 					if (importedData.length === 0) {
 						throw new Error('No conversations found in file');
@@ -6,7 +6,7 @@
 	import { ICON_CLASS_DEFAULT } from '$lib/constants';
 	import { ToolSource } from '$lib/enums/tools.enums';
 	import { mcpStore, permissionsStore, toolsStore } from '$lib/stores';
-	import { getBuiltinToolUi } from '$lib/utils';
+	import { getToolUi } from '$lib/utils';
 	import { SvelteSet } from 'svelte/reactivity';

 	let expandedGroups = new SvelteSet<string>();
@@ -69,12 +69,12 @@

 						{#each group.tools as entry (entry.key)}
 							{@const toolName = entry.definition.function.name}
-							{@const builtinUi =
-								entry.source === ToolSource.BUILTIN || entry.source === ToolSource.FRONTEND
-									? getBuiltinToolUi(toolName)
+							{@const toolUi =
+								entry.source === ToolSource.SERVER || entry.source === ToolSource.BROWSER
+									? getToolUi(toolName)
 									: null}
-							{@const displayLabel = builtinUi?.label ?? toolName}
-							{@const IconComponent = builtinUi?.icon ?? null}
+							{@const displayLabel = toolUi?.label ?? toolName}
+							{@const IconComponent = toolUi?.icon ?? null}
 							{@const isEnabled = toolsStore.isToolEnabled(entry.key)}
 							{@const permissionKey = entry.key}
 							{@const isAlwaysAllowed = permissionsStore.hasTool(permissionKey)}
@@ -69,7 +69,7 @@ export { default as SettingsChatFields } from './SettingsChat/SettingsChatFields
 /**
 * **SettingsChatToolsTab** - Tools configuration tab for chat settings
 *
- * Displays available tools grouped by source (built-in, MCP, custom) with
+ * Displays available tools grouped by source (server, browser, MCP, custom) with
 * toggles to enable/disable individual tools and tool groups. Shows MCP
 * server favicons and permission management controls.
 */
@@ -31,5 +31,11 @@ export const API_STREAM = {
 	LOOKUP: './v1/streams/lookup'
 };

+// query params for the resumable stream routes
+export const STREAM_QUERY_PARAMS = {
+	CONV_ID: 'conv_id',
+	FROM: 'from'
+} as const;
+
 /** CORS proxy endpoint path */
 export const CORS_PROXY_ENDPOINT = '/cors-proxy';
@@ -2,7 +2,9 @@ import { CLI_FLAGS } from './cli-flags.constants';
 import { BuiltInTool, JsonSchemaType, ToolCallType } from '$lib/enums';
 import type { OpenAIToolDefinition } from '$lib/types';

-export const BROWSER_INFO_TOOL_NAME = BuiltInTool.GET_INFO;
+// get_info is served by the server, but the browser falls back to this
+// implementation when the server does not provide it - same wire name.
+export const BROWSER_INFO_TOOL_NAME = BuiltInTool.SERVER_GET_INFO;

 /** UA token to OS name, first match wins - Android and iOS UAs also carry the Linux / Mac OS X tokens */
 export const BROWSER_INFO_OS_UA_PATTERNS: readonly [RegExp, string][] = [
@@ -1,52 +0,0 @@
-// Registry of built-in and frontend (browser) tools whose renderer
-// shows a recognizable icon and friendly label inline in the chat UI.
-//
-// To add a new built-in tool, add an entry to BUILTIN_TOOL_UI. To give a
-// tool a custom title or body renderer, add a dedicated component under
-// ChatMessageToolCall/ and route it in ChatMessageToolCallBlock.svelte
-// (see ChatMessageToolCallBlockGetDatetime and
-// ChatMessageToolCallBlockSearchResults for prior art).
-
-import {
-	Braces,
-	Clock,
-	Eye,
-	FilePen,
-	FilePlus,
-	FileSearch,
-	FileText,
-	Info,
-	SearchCode,
-	Terminal
-} from '@lucide/svelte';
-import { BuiltInTool, ToolSource } from '$lib/enums';
-import type { BuiltinToolUiEntry } from '$lib/types';
-
-export const BUILTIN_TOOL_UI: Readonly<Record<BuiltInTool, BuiltinToolUiEntry>> = {
-	[BuiltInTool.EDIT_FILE]: { icon: FilePen, label: 'Edit file', source: ToolSource.BUILTIN },
-	[BuiltInTool.EXEC_SHELL_COMMAND]: {
-		icon: Terminal,
-		label: 'Run command',
-		source: ToolSource.BUILTIN
-	},
-	[BuiltInTool.FILE_GLOB_SEARCH]: {
-		icon: FileSearch,
-		label: 'Search files',
-		source: ToolSource.BUILTIN
-	},
-	[BuiltInTool.GET_DATETIME]: { icon: Clock, label: 'Current time', source: ToolSource.FRONTEND },
-	[BuiltInTool.GET_INFO]: { icon: Info, label: 'Runtime info', source: ToolSource.BUILTIN },
-	[BuiltInTool.GREP_SEARCH]: {
-		icon: SearchCode,
-		label: 'Search in files',
-		source: ToolSource.BUILTIN
-	},
-	[BuiltInTool.READ_FILE]: { icon: FileText, label: 'Read file', source: ToolSource.BUILTIN },
-	[BuiltInTool.READ_MEDIA]: { icon: Eye, label: 'Read media', source: ToolSource.FRONTEND },
-	[BuiltInTool.RUN_JAVASCRIPT]: {
-		icon: Braces,
-		label: 'Run JavaScript',
-		source: ToolSource.FRONTEND
-	},
-	[BuiltInTool.WRITE_FILE]: { icon: FilePlus, label: 'Write file', source: ToolSource.BUILTIN }
-} as const;
@@ -1,7 +1,7 @@
 import { BuiltInTool, JsonSchemaType, ToolCallType } from '$lib/enums';
 import type { OpenAIToolDefinition } from '$lib/types';

-export const GET_DATETIME_TOOL_NAME = BuiltInTool.GET_DATETIME;
+export const GET_DATETIME_TOOL_NAME = BuiltInTool.BROWSER_GET_DATETIME;

 export function buildGetDatetimeToolDefinition(): OpenAIToolDefinition {
 	return {
@@ -15,7 +15,7 @@ export * from './context-gauge-popup.constants';
 export * from './conversation-import.constants';
 export * from './binary-detection.constants';
 export * from './content-detection.constants';
-export * from './built-in-tools.constants';
+export * from './tool-ui.constants';
 export * from './cache.constants';
 export * from './chat-form.constants';
 export * from './cli-flags.constants';
@@ -7,7 +7,7 @@ import {
 } from '$lib/enums';
 import type { OpenAIToolDefinition } from '$lib/types';

-export const READ_MEDIA_TOOL_NAME = BuiltInTool.READ_MEDIA;
+export const READ_MEDIA_TOOL_NAME = BuiltInTool.BROWSER_READ_MEDIA;

 // header lines of the tool result, parsed back by the read_media renderer
 export const PREFIX_FILE = 'File: ';
@@ -1,6 +1,6 @@
 import { BuiltInTool } from '$lib/enums';

-export const SANDBOX_TOOL_NAME = BuiltInTool.RUN_JAVASCRIPT;
+export const SANDBOX_TOOL_NAME = BuiltInTool.BROWSER_RUN_JAVASCRIPT;

 export const SANDBOX_TIMEOUT_MS_DEFAULT = 10000;

@@ -0,0 +1,60 @@
+// Registry of server and browser tools whose renderer
+// shows a recognizable icon and friendly label inline in the chat UI.
+//
+// To add a new tool, add an entry to TOOL_UI. To give a
+// tool a custom title or body renderer, add a dedicated component under
+// ChatMessageToolCall/ and route it in ChatMessageToolCallBlock.svelte
+// (see ChatMessageToolCallBlockGetDatetime and
+// ChatMessageToolCallBlockSearchResults for prior art).
+
+import {
+	Braces,
+	Clock,
+	Eye,
+	FilePen,
+	FilePlus,
+	FileSearch,
+	FileText,
+	Info,
+	SearchCode,
+	Terminal
+} from '@lucide/svelte';
+import { BuiltInTool, ToolSource } from '$lib/enums';
+import type { ToolUiEntry } from '$lib/types';
+
+export const TOOL_UI: Readonly<Record<BuiltInTool, ToolUiEntry>> = {
+	[BuiltInTool.BROWSER_GET_DATETIME]: {
+		icon: Clock,
+		label: 'Current time',
+		source: ToolSource.BROWSER
+	},
+	[BuiltInTool.BROWSER_READ_MEDIA]: { icon: Eye, label: 'Read media', source: ToolSource.BROWSER },
+	[BuiltInTool.BROWSER_RUN_JAVASCRIPT]: {
+		icon: Braces,
+		label: 'Run JavaScript',
+		source: ToolSource.BROWSER
+	},
+	[BuiltInTool.SERVER_EDIT_FILE]: { icon: FilePen, label: 'Edit file', source: ToolSource.SERVER },
+	[BuiltInTool.SERVER_EXEC_SHELL_COMMAND]: {
+		icon: Terminal,
+		label: 'Run command',
+		source: ToolSource.SERVER
+	},
+	[BuiltInTool.SERVER_FILE_GLOB_SEARCH]: {
+		icon: FileSearch,
+		label: 'Search files',
+		source: ToolSource.SERVER
+	},
+	[BuiltInTool.SERVER_GET_INFO]: { icon: Info, label: 'Runtime info', source: ToolSource.SERVER },
+	[BuiltInTool.SERVER_GREP_SEARCH]: {
+		icon: SearchCode,
+		label: 'Search in files',
+		source: ToolSource.SERVER
+	},
+	[BuiltInTool.SERVER_READ_FILE]: { icon: FileText, label: 'Read file', source: ToolSource.SERVER },
+	[BuiltInTool.SERVER_WRITE_FILE]: {
+		icon: FilePlus,
+		label: 'Write file',
+		source: ToolSource.SERVER
+	}
+} as const;
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Aman Gupta	af5172627d	RPC: populate use_count to enable fusion inside backends (#27142 ) * RPC: populate use_count to enable fusion inside backends * bump minor version	2026-08-18 21:08:57 +05:30
Aleksander Grygier	3dc7285b4f	ui: Services consolidation refactor (#27239 ) * ui: Move stream lookup and replay fetches into ChatService chatStore called fetch() directly for /v1/streams/lookup and the /v1/stream replay. These now live next to the other stream-session methods in ChatService, so services stay the only API I/O layer. * ui: Move /models/sse feed reader into ModelsService ModelsService.watchModelEvents owns the byte stream, reconnect loop and SSE record parsing; modelsStore keeps only event routing and state. * ui: Extract conversation import/export into ConversationTransferService The JSONL session format, ZIP archiving and browser downloads are pure I/O with no store state, so they move out of conversationsStore. The store keeps the DB orchestration (bulkExportConversations, downloadConversation, importConversationsData) and delegates the format work. * ui: Consolidate active model resolution into modelsStore.activeModelId The same resolution chain was duplicated in useChatScreenActiveModel, ChatForm, ChatFormActionModels and contextStatsStore, with slight drift in the single-model fallback. The canonical getter now lives in modelsStore, and the shared last-assistant-model lookup moved to utils as getConversationModel. * ui: Initialize stores explicitly via initStores() Store constructors and module-level side effects ran migrations and localStorage reads in import order. Migrations rename and rewrite localStorage keys, so a settings load racing ahead of them could clobber migrated values. initStores() is called once from the root layout and runs migrations first, then the stores that read localStorage, then the conversations DB load. * refactor: Constants for stream query params	2026-08-18 16:39:32 +02:00
Aleksander Grygier	fdf4c64604	ui: Stores consolidation refactor (#27238 ) * ui: Remove dead code from stores - persisted() helper was exported but never used - messageUpdateCallback / registerMessageUpdateCallback were never wired up - conversationsStore.initialize() alias, single caller moved to init() * ui: Merge device, theme and viewport into a single deviceStore All three are reactive browser-environment signals, now exposed as one class store: deviceStore.isMobile, deviceStore.isIOSDevice / isIOSSafari / isWKWebView / isStandalone and deviceStore.systemTheme.isDark. The systemTheme name disambiguates the OS preference from the user theme preference in settingsStore. Drops the unused viewport export (only isMobile was consumed). * ui: Merge build info into version store One VersionStore class with build (llama.cpp build number from build.json) and frontend (PWA version from _app/version.json), matching the class pattern of the other stores. * ui: Colocate context gauge popup state with its components The gauge popup state is local UI state shared only by the ChatFormContextGauge subtree, so it lives next to its consumers instead of the app-scope stores barrel.	2026-08-18 16:37:26 +02:00
Niklas Wenzel	04b569142d	common: share thread pools when `n_threads` differ (#27138 )	2026-08-18 16:23:43 +02:00
Ed Addario	0596704284	quant : Optimise memory usage by evicting weights after processing each layer (#22877 ) * Evict weights from memory after processing each layer * Revert changes * Move unmap to libllama * Unmap weights offloaded to backend * Change member's constness * Remove unmap weights offloaded to backend	2026-08-18 16:22:32 +02:00
Titaniumtown	0882c7bc89	sycl: honor GGML_HINT_SRC0_IS_HADAMARD (#27298 ) Kernel is a port of `ggml-cuda/fwht.cu` (us/run, median): ``` m x n x k GEMM FWHT speedup 64 x 1 x 64 10.20 2.93 3.48x 64 x 2048 x 64 10.75 2.71 3.97x 128 x 1 x 128 10.33 2.88 3.59x 128 x 32 x 128 9.20 2.77 3.33x 128 x 2048 x 128 16.46 2.76 5.95x 256 x 1 x 256 10.19 2.77 3.68x 256 x 2048 x 256 16.69 3.41 4.89x 512 x 2048 x 512 54.16 12.89 4.20x ```	2026-08-18 21:21:25 +08:00
Thiago Padilha	afd439df1f	unicode : include '~' in collapsed symbol class (#26972 ) The collapsed \p{S} class was missing '~', which split " ~" into separate pre-tokens and prevented the Ġ~ BPE merge used by DeepSeek V4. This caused re-tokenized prompts to diverge from sampled tokens and broke KV cache reuse. Assisted-by: Codex	2026-08-18 15:15:22 +02:00
Georgi Gerganov	169e4a7ff2	readme : update status badges + regen AUTHORS (#27317 ) * readme : update status badges * authors : regen	2026-08-18 14:35:04 +03:00
Zijun Yu	9d77fa1725	ci : Update OpenVINO to 2026.3, skip nemotron-h rollback test (#27292 ) * update to ov-2026.3, update device drivers * ci: skip nemotron-h rollback test on OpenVINO The OpenVINO backend does not support SSM_SCAN, so the Nemotron-H recurrent state rollback graph is split and cannot preserve the recurrent cache output shape. Keep the test enabled for other backends and retain the qwen35 OpenVINO rollback coverage. --------- Co-authored-by: ravi9 <ravi.panchumarthy@intel.com>	2026-08-18 12:02:22 +02:00
shivamkumard-ctrl	c0296022f3	ci: add Windows ARM64 CUDA support to the manual workflow (#27300 ) - Add a CUDA 13.4 ARM64 matrix entry. - Build only ggml-cuda for x64 and ARM64.	2026-08-18 11:55:45 +02:00
BlackFoil	7acdbb1f19	mtmd: fix LFM2 image tiling threshold (#27057 ) * mtmd: fix LFM2 image tiling threshold * refactor testing * fix * fix on windows --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co>	2026-08-18 11:11:19 +02:00
Georgi Gerganov	1511ce3bc3	sync : ggml	2026-08-18 11:30:03 +03:00
Georgi Gerganov	da786dc23e	ggml : bump version to 0.20.2 (ggml/1589)	2026-08-18 11:30:03 +03:00
Georgi Gerganov	27e345b574	build : fix xcframework + cmake clean-up (#27304 ) * xcframework : fix build * mtmd : remove unused include path * vendor : use vendor::hash alias target in cmake CMake reserves "::" in target names for imported/alias targets, so the real target keeps the name vendor-hash and a vendor::hash ALIAS target is added. Consumers (mtmd, llama-gguf-hash) now link against the namespaced alias. Assisted-by: pi:llama.cpp/Qwen3.8-27B * vendor : add cmake targets for all vendored libs with vendor:: aliases Add INTERFACE targets for the header-only vendor libs (miniaudio, nlohmann, sheredom, stb) and ALIAS targets named vendor::<lib> for all of them, including cpp-httplib and hash. Each exposes the vendor/ root so includes are namespaced, e.g. <nlohmann/json.hpp>. Consolidate the per-lib add_subdirectory calls into a single add_subdirectory(vendor), keeping the cpp-httplib gate on LLAMA_BUILD_COMMON. Consumers (llama-common, mtmd) now link the aliases instead of relying on raw vendor/ include paths. hash: consumers now include via "hash/hash.h"; the vendor/hash dir is kept as a PRIVATE include so the synced upstream sources compile unmodified. Assisted-by: pi:llama.cpp/Qwen3.8-27B * readme : use foo/bar names in acknowledgements Assisted-by: pi:llama.cpp/Qwen3.8-27B * ocd : fix valign	2026-08-18 11:16:51 +03:00
Georgi Gerganov	8b86400975	ci : create pre-release with change log and nightly link in make-release (#27302 ) * ci : create pre-release with change log and nightly link in make-release After pushing the tag, create a pre-release using ggml-org/action-create-release. The release description is generated by scripts/make-release-desc.sh: the change log between the current and previous version (one line per commit), a link to the corresponding nightly build when it exists, and a note that semantic versioning is still work in progress. Assisted-by: pi:llama.cpp/Qwen3.8-27B * cmake : bump version to 0.1.2 Assisted-by: pi:llama.cpp/Qwen3.8-27B * ci : find the nightly tag by commit in make-release-desc.sh The nightly release is guaranteed by the release checks to point at HEAD, so instead of reconstructing its name (commit count, branch, hash) just pick the b* tag pointing at HEAD. This also drops the RELEASE_BRANCH env var from the workflow. Assisted-by: pi:llama.cpp/Qwen3.8-27B * ci : resolve the release commit from the version tag in make-release-desc.sh The change log and nightly lookup now use the commit the version tag points at (HEAD when the tag does not exist), instead of always HEAD. This makes the script usable locally for older versions, e.g. ./scripts/make-release-desc.sh v0.1.1. The tag is resolved to a SHA first, since --points-at does not peel annotated tags. Assisted-by: pi:llama.cpp/Qwen3.8-27B * ci : normalize the version argument in make-release-desc.sh Accept the version with or without the leading v (0.1.1 == v0.1.1) and reject anything else, instead of silently treating a bare version as a non-existent tag name. Assisted-by: pi:llama.cpp/Qwen3.8-27B * cont : clean-up	2026-08-18 10:14:41 +03:00
ynankani	25ae3a9b33	CUDA: MMVQ nwarps=8 for bs=1 for dense models on DGX Spark (#26843 ) * CUDA: MMVQ nwarps=8 for bs=1 for dense models on DGX Spark Signed-off-by: ynankani <ynankani@nvidia.com> * skip moe experts and allow others based on k geometry (allow only small idle tail) Signed-off-by: ynankani <ynankani@nvidia.com> * rename MMVQ DGX Spark params to GB10 and fix MSVC constexpr lambda capture Signed-off-by: ynankani <ynankani@nvidia.com> --------- Signed-off-by: ynankani <ynankani@nvidia.com>	2026-08-18 09:45:53 +05:30
Aleksander Grygier	01818e4956	ui: enforce alphabetical enum member ordering (#27272 )	2026-08-17 23:52:00 +02:00
Aleksander Grygier	0021a77de0	ui: Refactor Built-In Tools naming (Server/Browser) (#27271 ) * server: rename built-in tools to server tools * ui: rename built-in tools to server/browser tools	2026-08-17 22:23:22 +02:00
Eve	058df671b2	ci: more optimizations (#26983 ) * replace rpc job with cpu * remove vulkan cache * move windows to build vulkan	2026-08-17 19:59:40 +00:00
Pascal	087f94d82e	doc: document MCP stdio servers and CORS defaults in the server README [no release] [no ci] (#26847 ) * doc: document MCP stdio servers and CORS defaults in the server README The MCP arguments were listed but nothing explained what an MCP server is or how to declare one. Cover the stdio transport, the config keys, the tool naming, and add a POSIX shell echo server as a minimal example. Also document the CORS behavior: the default reflected origin, the switch to localhost once tools are enabled, and the recommended setting per deployment. * doc: drop the inline MCP shell example from the server README The example parsed JSON-RPC by hand and sat in a page people copy paste from, into servers spawned with the privileges of llama-server. Point to the specification instead. Link the pull request that introduced the feature, and keep a short mcp.json snippet so the table of configuration keys has a declaration to refer to.	2026-08-17 21:32:15 +02:00
Xuan-Son Nguyen	533b18257b	server: save processed mtmd chunks as placeholder (#27278 )	2026-08-17 21:29:50 +02:00
Xuan-Son Nguyen	ed1c3a20f5	mtmd: use sha256 for input hashing (#27274 ) * mtmd: use sha256 for input hashing * void conflict with boringssl	2026-08-17 21:20:12 +02:00
Sigbjørn Skjæret	d8df12ebc4	vocab : support integer tokenizer scores (#27260 )	2026-08-17 20:13:15 +02:00
Tarek Dakhran	b75ecd1971	mtmd : skip thumbnail for non-tiled LFM2 images (#27246 ) * mtmd : skip thumbnail for non-tiled LFM2 images Matches HF behavior. * Update tools/mtmd/mtmd-image.cpp --------- Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>	2026-08-17 19:41:32 +02:00