Update tools/llama-bench/llama-bench.cpp

Co-authored-by: Diego Devesa <slarengh@gmail.com>

.devops/intel.Dockerfile

@@ -49,23 +49,19 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update && \
-    apt-get install -y \
-        git \
-        python3 \
-        python3-pip \
-        python3-venv && \
-    python3 -m venv /opt/venv && \
-    . /opt/venv/bin/activate && \
-    pip install --upgrade pip setuptools wheel && \
-    pip install -r requirements.txt && \
-    apt autoremove -y && \
-    apt clean -y && \
-    rm -rf /tmp/* /var/tmp/* && \
-    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
-    find /var/cache -type f -delete
+RUN apt-get update \
+    && apt-get install -y \
+    git \
+    python3 \
+    python3-pip \
+    && pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt \
+    && apt autoremove -y \
+    && apt clean -y \
+    && rm -rf /tmp/* /var/tmp/* \
+    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
+    && find /var/cache -type f -delete
 
-ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.devops/tools.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 # Read the first argument into a variable

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
     validations:
       required: true

.github/labeler.yml

@@ -1,4 +1,10 @@
 # https://github.com/actions/labeler
+Kompute:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-kompute.h
+            - ggml/src/ggml-kompute/**
+            - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
@@ -80,15 +86,3 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
-
-Ascend NPU:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-cann.h
-            - ggml/src/ggml-cann/**
-            - docs/backend/CANN.md
-OpenCL:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-opencl.h
-            - ggml/src/ggml-opencl/**

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

@@ -1,51 +0,0 @@
-name: Build relocatable cmake package
-on:
-  workflow_dispatch:
-  workflow_call:
-
-jobs:
-  linux:
-    runs-on: ubuntu-24.04
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y build-essential tcl
-
-      - name: Build
-        run: |
-          PREFIX="$(pwd)"/inst
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
-                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release
-          cmake --install build --prefix "$PREFIX" --config Release
-
-          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
-          set build(version) "0.0.$build(number)"
-
-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
-
-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
-                  exit 1
-              }
-          }
-          puts "success."
-          EOF
-
-          cd examples/simple-cmake-pkg
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build

.github/workflows/build-linux-cross.yml

@@ -231,116 +231,3 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
-
-  debian-13-loongarch64-cpu-cross:
-    runs-on: ubuntu-24.04
-    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup LoongArch
-        run: |
-          rm -f /etc/apt/sources.list.d/*
-          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
-          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
-          EOF
-          ( echo 'quiet "true";'; \
-            echo 'APT::Get::Assume-Yes "true";'; \
-            echo 'APT::Install-Recommends "false";'; \
-            echo 'Acquire::Check-Valid-Until "false";'; \
-            echo 'Acquire::Retries "5";'; \
-          ) > /etc/apt/apt.conf.d/99snapshot-repos
-
-          apt-get update
-          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
-          dpkg --add-architecture loong64
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
-          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
-          EOF
-
-          apt-get update || true    ;# Prevent failure due to missing URLs.
-
-          apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-loongarch64-linux-gnu \
-                  g++-14-loongarch64-linux-gnu
-
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
-                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
-
-          cmake --build build --config Release -j $(nproc)
-
-  debian-13-loongarch64-vulkan-cross:
-    runs-on: ubuntu-24.04
-    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup LoongArch
-        run: |
-          rm -f /etc/apt/sources.list.d/*
-          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
-          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
-          EOF
-          ( echo 'quiet "true";'; \
-            echo 'APT::Get::Assume-Yes "true";'; \
-            echo 'APT::Install-Recommends "false";'; \
-            echo 'Acquire::Check-Valid-Until "false";'; \
-            echo 'Acquire::Retries "5";'; \
-          ) > /etc/apt/apt.conf.d/99snapshot-repos
-
-          apt-get update
-          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
-          dpkg --add-architecture loong64
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
-          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
-          EOF
-
-          apt-get update || true    ;# Prevent failure due to missing URLs.
-
-          apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  gcc-14-loongarch64-linux-gnu \
-                  g++-14-loongarch64-linux-gnu \
-                  libvulkan-dev:loong64
-
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
-                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
-
-          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -5,43 +5,10 @@ on:
   push:
     branches:
       - master
-    paths: [
-      '.github/workflows/build.yml',
-      '.github/workflows/build-linux-cross.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.swift',
-      '**/*.m',
-      '**/*.metal',
-      '**/*.comp'
-    ]
-
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build.yml',
-      '.github/workflows/build-linux-cross.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.swift',
-      '**/*.m',
-      '**/*.metal',
-      '**/*.comp'
-    ]
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -84,8 +51,7 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=OFF \
-            -DGGML_METAL_SHADER_DEBUG=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -340,7 +306,6 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -512,9 +477,6 @@ jobs:
   build-linux-cross:
     uses: ./.github/workflows/build-linux-cross.yml
 
-  build-cmake-pkg:
-    uses: ./.github/workflows/build-cmake-pkg.yml
-
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -665,7 +627,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     strategy:
       fail-fast: false
@@ -715,31 +677,28 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64 (static)'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
+          - build: 'cpu-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'
           - build: 'openblas-x64'
-            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: 'vulkan-x64'
-            arch: 'x64'
-            defines: '-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'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
-            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
-            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+         # - build: 'kompute-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_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
       - name: Clone
@@ -753,6 +712,12 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
+      - name: Clone Kompute submodule
+        id: clone_kompute
+        if: ${{ matrix.build == 'kompute-x64' }}
+        run: |
+          git submodule update --init ggml/src/ggml-kompute/kompute
+
       - name: Download OpenBLAS
         id: get_openblas
         if: ${{ matrix.build == 'openblas-x64' }}
@@ -768,9 +733,9 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
         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"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$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"
@@ -803,8 +768,6 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -814,7 +777,6 @@ jobs:
           cmake -S . -B build ${{ matrix.defines }} `
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
@@ -825,7 +787,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.arch == 'x64' }}
+        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -877,12 +839,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2022-cmake-cuda:
-    runs-on: windows-2022
+  windows-2019-cmake-cuda:
+    runs-on: windows-2019
 
     strategy:
       matrix:
-        cuda: ['12.4']
+        cuda: ['12.4', '11.7']
 
     steps:
       - name: Clone
@@ -916,7 +878,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^
@@ -930,7 +892,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     defaults:
       run:
@@ -964,7 +926,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -49,8 +49,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='@loader_path' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -104,8 +103,7 @@ jobs:
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='@loader_path' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
@@ -133,9 +131,8 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
-          # - build: 'arm64'
-          #   os: ubuntu-22.04-arm
+          - build: 'arm64'
+            os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -162,11 +159,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_NATIVE=OFF \
-            -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -215,11 +207,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_NATIVE=OFF \
-            -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -241,7 +228,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -277,7 +264,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DGGML_NATIVE=OFF ^
@@ -294,7 +281,7 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -304,11 +291,11 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
@@ -338,7 +325,7 @@ jobs:
         id: get_vulkan
         if: ${{ matrix.backend == '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"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$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 $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -386,11 +373,11 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2022
+    runs-on: windows-2019
 
     strategy:
       matrix:
-        cuda: ['12.4']
+        cuda: ['12.4', '11.7']
 
     steps:
       - name: Clone
@@ -418,7 +405,7 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^
@@ -454,7 +441,7 @@ jobs:
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     defaults:
       run:
@@ -526,7 +513,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     strategy:
       matrix:

.github/workflows/server.yml

@@ -180,7 +180,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2022
+    runs-on: windows-2019
 
     steps:
       - name: Clone

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "kompute"]
+	path = ggml/src/ggml-kompute/kompute
+	url = https://github.com/nomic-ai/kompute.git

CMakeLists.txt

@@ -89,14 +89,6 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
-if (NOT DEFINED LLAMA_BUILD_NUMBER)
-    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-endif()
-if (NOT DEFINED LLAMA_BUILD_COMMIT)
-    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-endif()
-set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
-
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -120,6 +112,7 @@ endfunction()
 
 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
+llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
@@ -162,17 +155,10 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
-    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
-    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
-if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
-endif()
-
 #
 # build the library
 #
@@ -213,6 +199,10 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
+set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")

Makefile

@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
+	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
 endif
 
 # warnings
@@ -970,7 +970,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/repack.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \

README.md

@@ -3,10 +3,9 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![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)
 [![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)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
@@ -18,6 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
+- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
 - A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
@@ -28,30 +28,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
-## Quick start
-
-Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
-
-- Install `llama.cpp` using [brew, nix or winget](docs/install.md)
-- Run with Docker - see our [Docker documentation](docs/docker.md)
-- Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
-- Build from source by cloning this repository - check out [our build guide](docs/build.md)
-
-Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
-
-Example command:
-
-```sh
-# Use a local model file
-llama-cli -m my_model.gguf
-
-# Or download and run a model directly from Hugging Face
-llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
-
-# Launch OpenAI-compatible API server
-llama-server -hf ggml-org/gemma-3-1b-it-GGUF
-```
-
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -154,7 +130,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Bindings</summary>
 
-- Python: [ddh0/easy-llama](https://github.com/ddh0/easy-llama)
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
@@ -254,7 +229,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
-
 ## Supported backends
 
 | Backend | Target devices |
@@ -271,6 +245,16 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
+## Building the project
+
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
+
+- Clone this repository and build locally, see [how to build](docs/build.md)
+- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
+- Use a Docker image, see [documentation for Docker](docs/docker.md)
+- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
+
 ## Obtaining and quantizing models
 
 The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
@@ -278,11 +262,7 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
-
-```sh
-llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
-```
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
 
 By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
 

build-xcframework.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #
 # Options
 IOS_MIN_OS_VERSION=16.4

ci/run.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -39,27 +39,14 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
-
-    if command -v nvidia-smi >/dev/null 2>&1; then
-        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
-        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
-            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
-        else
-            echo "Warning: Using fallback CUDA architectures"
-            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
-        fi
-    else
-        echo "Error: nvidia-smi not found, cannot build with CUDA"
-        exit 1
-    fi
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -779,7 +766,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/CMakeLists.txt

@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
-    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,26 +18,36 @@ if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
+            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
         endif()
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        # For build-info.cpp below
-        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
 endif()
 
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
-configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
 set(TARGET build_info)
-add_library(${TARGET} OBJECT ${OUTPUT_FILE})
+add_library(${TARGET} OBJECT build-info.cpp)
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

common/arg.cpp

@@ -988,6 +988,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
+    if (params.reranking && params.embedding) {
+        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
+    }
+
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -2706,13 +2710,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.embd_sep = value;
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
-    add_opt(common_arg(
-        {"--cls-separator"}, "STRING",
-        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
-        [](common_params & params, const std::string & value) {
-            params.cls_sep = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
         string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -2750,10 +2747,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
         [](common_params & params) {
-            params.embedding = true;
-            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
+            params.reranking = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -2794,16 +2790,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ssl_file_cert = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
-    add_opt(common_arg(
-        {"--chat-template-kwargs"}, "STRING",
-        string_format("sets additional params for the json template parser"),
-        [](common_params & params, const std::string &  value) {
-            auto parsed = json::parse(value);
-            for (const auto & item : parsed.items()) {
-                params.default_template_kwargs[item.key()] = item.value().dump();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -2883,7 +2869,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
-            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
             else { throw std::invalid_argument("invalid value"); }
         }
@@ -3227,32 +3212,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
-    add_opt(common_arg(
-        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for K for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_k)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_k = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
-    add_opt(common_arg(
-        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for V for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_v)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_v = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));
 
     add_opt(common_arg(
         {"-mv", "--model-vocoder"}, "FNAME",

common/build-info.cpp.in

@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,7 +49,6 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
-
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
@@ -155,10 +154,9 @@ bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think
             if (!rest.empty()) {
                 handle_reasoning(rest, /* closed */ !is_partial());
             }
-            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
-            // if (!syntax_.thinking_forced_open) {
-            //     throw common_chat_msg_partial_exception(end_think);
-            // }
+            if (!syntax_.thinking_forced_open) {
+                throw common_chat_msg_partial_exception(end_think);
+            }
             return true;
         }
     }
@@ -379,7 +377,3 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
-
-void common_chat_msg_parser::clear_tools() {
-    result_.tool_calls.clear();
-}

common/chat-parser.h

@@ -115,6 +115,4 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
-
-    void clear_tools();
 };

common/chat.cpp

@@ -17,8 +17,6 @@
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
-
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -84,10 +82,10 @@ json common_chat_msg::to_json_oaicompat() const
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg) {
     std::vector<common_chat_msg_diff> diffs;
-    if (previous_msg.reasoning_content != new_msg.reasoning_content) {
-        auto & diff = diffs.emplace_back();
-        diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, new_msg.reasoning_content);
-    }
+    // if (previous_msg.reasoning_content != current.reasoning_content) {
+    //     auto & diff = diffs.emplace_back();
+    //     diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, current.reasoning_content);
+    // }
     if (previous_msg.content != new_msg.content) {
         auto & diff = diffs.emplace_back();
         diff.content_delta = string_diff(previous_msg.content, new_msg.content);
@@ -142,7 +140,6 @@ struct templates_params {
     bool add_generation_prompt = true;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
-    json extra_context;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -388,9 +385,9 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
 
 template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
-    if (!diff.reasoning_content_delta.empty()) {
-        delta["reasoning_content"] = diff.reasoning_content_delta;
-    }
+    // if (!diff.reasoning_content_delta.empty()) {
+    //     delta["reasoning_content"] = msg.reasoning_content;
+    // }
     if (!diff.content_delta.empty()) {
         delta["content"] = diff.content_delta;
     }
@@ -601,7 +598,6 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
-        case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
@@ -723,23 +719,16 @@ static void foreach_function(const json & tools, const std::function<void(const
 
 static std::string apply(
     const common_chat_template & tmpl,
-    const struct templates_params & inputs,
-    const std::optional<json> & messages_override = std::nullopt,
-    const std::optional<json> & tools_override = std::nullopt,
-    const std::optional<json> & additional_context = std::nullopt)
+    const nlohmann::ordered_json & messages,
+    const nlohmann::ordered_json & tools,
+    bool add_generation_prompt,
+    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
 {
     minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
-    if (tools_override) {
-        tmpl_inputs.tools = *tools_override;
-    } else {
-        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
-    }
-    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
-    tmpl_inputs.extra_context = inputs.extra_context;
-    if (additional_context) {
-        tmpl_inputs.extra_context.merge_patch(*additional_context);
-    }
+    tmpl_inputs.messages = messages;
+    tmpl_inputs.tools = tools;
+    tmpl_inputs.add_generation_prompt = add_generation_prompt;
+    tmpl_inputs.extra_context = extra_context;
     // TODO: add flag to control date/time, if only for testing purposes.
     // tmpl_inputs.now = std::chrono::system_clock::now();
 
@@ -838,7 +827,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
         inputs.messages,
         "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
 
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
+    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_GENERIC;
     return data;
 }
@@ -914,7 +903,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.preserved_tokens = {
         "[TOOL_CALLS]",
     };
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
@@ -944,7 +933,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             adjusted_messages.push_back(msg);
         }
     }
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
+    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
         if (!inputs.enable_thinking) {
@@ -1132,7 +1121,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
         {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1197,7 +1186,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs);
+    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
 
     // Hacks to fix the official (broken) prompt.
     // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1292,7 +1281,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
+    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
@@ -1348,7 +1337,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
     // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1475,7 +1464,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
 
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     // TODO: if (has_raw_python)
     return data;
 }
@@ -1508,15 +1497,14 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
-    json extra_context = json {
+    json additional_context = {
         {"enable_thinking", inputs.enable_thinking},
     };
-    extra_context.update(inputs.extra_context);
 
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!extra_context["enable_thinking"]) {
+        if (!inputs.enable_thinking) {
             data.prompt += "</think>";
         } else {
             data.thinking_forced_open = true;
@@ -1702,7 +1690,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     data.grammar_lazy = false;
     if (!inputs.json_schema.is_null()) {
@@ -1733,12 +1721,6 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
-
-    params.extra_context = json::object();
-    for (auto el : inputs.chat_template_kwargs) {
-        params.extra_context[el.first] = json::parse(el.second);
-    }
-
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }
@@ -1855,7 +1837,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     if (res < 0) {
         // if the custom "tmpl" is not supported, we throw an error
         // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-        throw std::runtime_error("this custom template is not supported, try using --jinja");
+        throw std::runtime_error("this custom template is not supported");
     }
 
     // if it turns out that our buffer is too small, we resize it
@@ -1938,9 +1920,7 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            builder.clear_tools();
-            builder.move_to(0);
-            common_chat_parse_content_only(builder);
+            throw std::runtime_error(ex.what());
         }
     }
     auto msg = builder.result();

common/chat.h

@@ -7,7 +7,6 @@
 #include <chrono>
 #include <string>
 #include <vector>
-#include <map>
 
 struct common_chat_templates;
 
@@ -71,7 +70,7 @@ struct common_chat_msg {
 };
 
 struct common_chat_msg_diff {
-    std::string reasoning_content_delta;
+    // std::string reasoning_content_delta;
     std::string content_delta;
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;
@@ -126,7 +125,6 @@ struct common_chat_templates_inputs {
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
-    std::map<std::string, std::string> chat_template_kwargs;
 };
 
 struct common_chat_params {

common/cmake/build-info-gen-cpp.cmake

@@ -0,0 +1,24 @@
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
+
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+    endif()
+else()
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+endif()

common/common.cpp

@@ -466,7 +466,7 @@ size_t string_find_partial_stop(const std::string_view & str, const std::string_
 
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$&");
+    return std::regex_replace(s, special_chars, "\\$0");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -706,17 +706,11 @@ bool fs_validate_filename(const std::string & filename) {
         // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
-
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif
 
         filename_utf32 = converter.from_bytes(filename);
@@ -773,9 +767,6 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
-#include <iostream>
-
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -793,16 +784,9 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // process path from front to back, procedurally creating directories
     while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
         const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
 
-        pos_slash += 1;
-
-        // skip the drive letter, in some systems it can return an access denied error
-        if (subpath.length() == 2 && subpath[1] == ':') {
-            continue;
-        }
-
-        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
-
+        const bool success = CreateDirectoryW(test, NULL);
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -816,6 +800,8 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
+
+        pos_slash += 1;
     }
 
     return true;
@@ -911,6 +897,34 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -920,7 +934,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -952,35 +966,6 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
-    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1056,7 +1041,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_memory_clear(llama_get_memory(lctx), true);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
@@ -1158,6 +1143,11 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
 
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }
+
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
 
@@ -1290,9 +1280,6 @@ std::vector<llama_token> common_tokenize(
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
     n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-    if (n_tokens == std::numeric_limits<int32_t>::min()) {
-        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
-    }
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);

common/common.h

@@ -8,7 +8,6 @@
 #include <string>
 #include <string_view>
 #include <vector>
-#include <map>
 #include <sstream>
 
 #ifdef _WIN32
@@ -200,9 +199,6 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
-    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
-    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
-
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -219,8 +215,7 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
-    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
+    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };
 
 struct common_params {
@@ -359,7 +354,7 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    std::string cls_sep    = "\t";  // separator of classification sequences
+    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
@@ -382,8 +377,6 @@ struct common_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
-    std::map<std::string, std::string> default_template_kwargs;
-
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
     bool endpoint_slots   = false;

common/json-schema-to-grammar.cpp

@@ -41,6 +41,49 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
+/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
+class string_view {
+    const std::string & _str;
+    const size_t _start;
+    const size_t _end;
+public:
+    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
+
+    size_t size() const {
+        return _end - _start;
+    }
+
+    size_t length() const {
+        return size();
+    }
+
+    operator std::string() const {
+        return str();
+    }
+
+    std::string str() const {
+        return _str.substr(_start, _end - _start);
+    }
+
+    string_view substr(size_t pos, size_t len = std::string::npos) const {
+        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
+    }
+
+    char operator[](size_t pos) const {
+        auto index = _start + pos;
+        if (index >= _end) {
+            throw std::out_of_range("string_view index out of range");
+        }
+        return _str[_start + pos];
+    }
+
+    bool operator==(const string_view & other) const {
+        std::string this_str = *this;
+        std::string other_str = other;
+        return this_str == other_str;
+    }
+};
+
 static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
     auto has_min = min_value != std::numeric_limits<int>::min();
     auto has_max = max_value != std::numeric_limits<int>::max();
@@ -69,14 +112,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         }
         out << "}";
     };
-    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
-        [&](const std::string_view & from, const std::string_view & to) {
+    std::function<void(const string_view &, const string_view &)> uniform_range =
+        [&](const string_view & from, const string_view & to) {
             size_t i = 0;
             while (i < from.length() && i < to.length() && from[i] == to[i]) {
                 i++;
             }
             if (i > 0) {
-                out << "\"" << from.substr(0, i) << "\"";
+                out << "\"" << from.substr(0, i).str() << "\"";
             }
             if (i < from.length() && i < to.length()) {
                 if (i > 0) {

common/speculative.cpp

@@ -144,8 +144,6 @@ llama_tokens common_speculative_gen_draft(
     auto & smpl   = spec->smpl;
     auto & prompt = spec->prompt;
 
-    auto * mem = llama_get_memory(ctx);
-
     int reuse_i = 0;
     int reuse_n = 0;
 
@@ -175,7 +173,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_memory_clear(mem, false);
+        llama_kv_self_clear(ctx);
 
         prompt.clear();
     } else {
@@ -194,14 +192,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_memory_seq_rm (mem, 0, 0, reuse_i);
-            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
+            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_memory_seq_rm (mem, 0, reuse_n, -1);
+            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -310,8 +310,6 @@ class ModelBase:
                             gguf.MODEL_TENSOR.POSNET_NORM2,
                             gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
                             gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
-                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
-                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -322,11 +320,7 @@ class ModelBase:
                     self.match_model_tensor_name(new_name, key, bid)
                     for key in (
                         gguf.MODEL_TENSOR.TOKEN_EMBD,
-                        gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
                         gguf.MODEL_TENSOR.OUTPUT,
-                        gguf.MODEL_TENSOR.ALTUP_ROUTER,
-                        gguf.MODEL_TENSOR.LAUREL_L,
-                        gguf.MODEL_TENSOR.LAUREL_R,
                     )
                 ):
                     if self.ftype in (
@@ -525,7 +519,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -927,20 +921,13 @@ class TextModel(ModelBase):
         tokenizer = SentencePieceProcessor()
         tokenizer.LoadFromFile(str(tokenizer_path))
 
-        vocab_size = self.find_hparam([
-            "vocab_size_per_layer_input", # gemma3n
-            "vocab_size",
-        ], optional=True) or tokenizer.vocab_size()
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
         toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
-            if token_id >= vocab_size:
-                logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}')
-                break
-
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.GetScore(token_id)
@@ -1911,7 +1898,9 @@ class LlamaModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -1993,8 +1982,7 @@ class LlamaModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -2029,20 +2017,6 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("ArceeForCausalLM")
-class ArceeModel(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.ARCEE
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self._try_set_pooling_type()
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
-
-
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -2158,6 +2132,7 @@ class Llama4Model(LlamaModel):
 
     def set_vocab(self):
         self._set_vocab_gpt2()
+        self.gguf_writer.add_add_bos_token(True)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2206,7 +2181,7 @@ class Llama4VisionModel(MmprojModel):
                 name += ".weight"
             if "multi_modal_projector.linear_1" in name:
                 # despite the name with number postfix, this is a single fully connected layer
-                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)]
+                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)]
             return [(self.map_tensor_name(name), data_torch)]
         return []
 
@@ -2329,7 +2304,9 @@ class DeciModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -2369,8 +2346,7 @@ class DeciModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -2743,52 +2719,6 @@ class Qwen2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Ernie4_5_ForCausalLM")
-class Ernie4_5Model(TextModel):
-    model_arch = gguf.MODEL_ARCH.ERNIE4_5
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        num_heads = self.hparams["num_attention_heads"]
-        num_kv_heads = self.hparams["num_key_value_heads"]
-        head_dim = self.hparams["head_dim"]
-
-        if "ernie." in name:
-            name = name.replace("ernie.", "model.")
-        # split the qkv weights
-        # qkv_proj shape: [(num_heads + 2 * num_kv_heads) * head_dim, hidden_size]
-        if "qkv_proj" in name:
-            name_q = name.replace("qkv_proj.weight", "q_proj.weight")
-            name_k = name.replace("qkv_proj.weight", "k_proj.weight")
-            name_v = name.replace("qkv_proj.weight", "v_proj.weight")
-            total_q_dim = num_heads * head_dim
-            total_k_dim = num_kv_heads * head_dim
-            total_v_dim = num_kv_heads * head_dim
-            q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)
-            return [
-                (self.map_tensor_name(name_q), q_proj_weight),
-                (self.map_tensor_name(name_k), k_proj_weight),
-                (self.map_tensor_name(name_v), v_proj_weight)
-            ]
-        # split the up_gate_proj into gate and up
-        # up_gate_proj shape: [2 * intermediate_size, hidden_size]
-        if "up_gate_proj" in name:
-            name_up = name.replace("up_gate_proj.weight", "up_proj.weight")
-            name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")
-            dim_half = data_torch.shape[0] // 2
-            gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)
-            return [
-                (self.map_tensor_name(name_gate), gate_proj_weight),
-                (self.map_tensor_name(name_up), up_proj_weight)
-            ]
-        return [(self.map_tensor_name(name), data_torch)]
-
-
 @ModelBase.register(
     "Qwen2VLModel",
     "Qwen2VLForConditionalGeneration",
@@ -3734,7 +3664,9 @@ class InternLM3Model(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -3777,7 +3709,8 @@ class BertModel(TextModel):
         self._try_set_pooling_type()
 
         if self.cls_out_labels:
-            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
+            key_name = gguf.Keys.Classifier.OUTPUT_LABELS.format(arch = gguf.MODEL_ARCH_NAMES[self.model_arch])
+            self.gguf_writer.add_array(key_name, [v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3881,7 +3814,7 @@ class BertModel(TextModel):
             remove_whitespaces = tokenizer.clean_up_tokenization_spaces
             precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])
 
-            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)
+            vocab_size = self.hparams.get("vocab_size", tokenizer.vocab_size)
         else:
             sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
             sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
@@ -3894,7 +3827,7 @@ class BertModel(TextModel):
             tokenizer = SentencePieceProcessor()
             tokenizer.LoadFromFile(str(tokenizer_path))
 
-            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size())
+            vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
@@ -3924,26 +3857,33 @@ class BertModel(TextModel):
             unk_token = tokenizer_config_json.get("unk_token")
             unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))
 
-            for token_id in range(tokenizer.vocab_size):
+            for token_id in range(vocab_size):
                 piece = tokenizer._convert_id_to_token(token_id)
-                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:
-                    text = piece.encode("utf-8")
-                    score = tokenizer_json["model"]["vocab"][token_id][1]
+                text = piece.encode("utf-8")
+                score = tokenizer_json["model"]["vocab"][token_id][1]
 
-                    toktype = SentencePieceTokenTypes.NORMAL
-                    if token_id == unk_token_id:
-                        toktype = SentencePieceTokenTypes.UNKNOWN
-                    elif token_id in tokenizer.all_special_ids:
-                        toktype = SentencePieceTokenTypes.CONTROL
-                    elif token_id in added_vocab.values():
-                        toktype = SentencePieceTokenTypes.USER_DEFINED
-                    # No reliable way to detect this, but jina doesn't have any
-                    # elif tokenizer.IsByte(token_id):
-                    #     toktype = SentencePieceTokenTypes.BYTE
+                toktype = SentencePieceTokenTypes.NORMAL
+                if token_id == unk_token_id:
+                    toktype = SentencePieceTokenTypes.UNKNOWN
+                elif token_id in tokenizer.all_special_ids:
+                    toktype = SentencePieceTokenTypes.CONTROL
+                elif token_id in added_vocab.values():
+                    toktype = SentencePieceTokenTypes.USER_DEFINED
+                # No reliable way to detect this, but jina doesn't have any
+                # elif tokenizer.IsByte(token_id):
+                #     toktype = SentencePieceTokenTypes.BYTE
 
-                    tokens[token_id] = text
-                    scores[token_id] = score
-                    toktypes[token_id] = toktype
+                tokens[token_id] = text
+                scores[token_id] = score
+                toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
 
         if isinstance(tokenizer, SentencePieceProcessor):
             # realign tokens (see HF tokenizer code)
@@ -3956,12 +3896,6 @@ class BertModel(TextModel):
                 SentencePieceTokenTypes.UNKNOWN,
             ] + toktypes[3:-1]
 
-            if self.model_arch == gguf.MODEL_ARCH.NOMIC_BERT_MOE:
-                # Add mask token missing from sentencepiece.bpe.model
-                tokens[250001] = b'<mask>'
-                scores[250001] = 0.0
-                toktypes[250001] = SentencePieceTokenTypes.CONTROL
-
         self.gguf_writer.add_tokenizer_model("t5")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
@@ -3976,6 +3910,9 @@ class BertModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
 
 @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
 class DistilBertModel(BertModel):
@@ -4017,6 +3954,8 @@ class RobertaModel(BertModel):
         bpe_tok_path = self.dir_model / "tokenizer.json"
         if bpe_tok_path.exists():
             self._set_vocab_gpt2()
+            self.gguf_writer.add_add_bos_token(True)
+            self.gguf_writer.add_add_eos_token(True)
 
             # we need this to validate the size of the token_type embeddings
             # though currently we are passing all zeros to the token_type embeddings
@@ -4122,34 +4061,6 @@ class NomicBertModel(BertModel):
         raise ValueError(f"unknown tokenizer: {toktyp}")
 
 
-@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
-class NeoBert(BertModel):
-    model_arch = gguf.MODEL_ARCH.NEO_BERT
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-
-        # NeoBERT uses 2/3 of the intermediate size as feed forward length
-        self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
-        self.gguf_writer.add_rope_freq_base(10000.0)  # default value for NeoBERT
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
-
-        f_rms_eps = self.hparams.get("norm_eps", 1e-6)  # default value for NeoBERT
-        self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-        logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
-
-        self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
-
-    def modify_tensors(self, data_torch, name, bid):
-        if name.startswith("decoder."):
-            return []
-
-        if name.startswith("model."):
-            name = name[6:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
@@ -4276,7 +4187,6 @@ class Gemma2Model(TextModel):
 @ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
 class Gemma3Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GEMMA3
-    norm_shift = 1.0  # Gemma3RMSNorm adds 1.0 to the norm value
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -4298,8 +4208,9 @@ class Gemma3Model(TextModel):
         self.gguf_writer.add_value_length(hparams.get("head_dim", 256))
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
-        # attn_logit_softcapping is removed in Gemma3
+        # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3
         assert hparams.get("attn_logit_softcapping") is None
+        assert hparams.get("final_logit_softcapping") is None
         self.gguf_writer.add_sliding_window(hparams["sliding_window"])
         self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
         if hparams.get("rope_scaling") is not None:
@@ -4311,7 +4222,7 @@ class Gemma3Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-        if "language_model." in name:
+        if name.startswith("language_model."):
             name = name.replace("language_model.", "")
 
         elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
@@ -4326,9 +4237,8 @@ class Gemma3Model(TextModel):
 
         # ref code in Gemma3RMSNorm
         # output = output * (1.0 + self.weight.float())
-        # note: this is not the case on gemma3n
         if name.endswith("norm.weight"):
-            data_torch = data_torch + self.norm_shift
+            data_torch = data_torch + 1
 
         return [(self.map_tensor_name(name), data_torch)]
 
@@ -4385,101 +4295,6 @@ class Gemma3VisionModel(MmprojModel):
         return [] # skip other tensors
 
 
-@ModelBase.register("Gemma3nForConditionalGeneration")
-class Gemma3NModel(Gemma3Model):
-    model_arch = gguf.MODEL_ARCH.GEMMA3N
-    norm_shift = 0.0 # same value with Gemma3p5RMSNorm scale_shift on python code
-
-    _altup_proj: list[Tensor] = []
-    _altup_unembd: list[Tensor] = []
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        assert self.hparams["altup_num_inputs"] == 4, "Current conversion only supports 4 altup inputs"
-        self._altup_proj = [
-            torch.Tensor(), # to be replaced
-            torch.Tensor(), # to be replaced
-            torch.Tensor(), # to be replaced
-        ]
-        self._altup_unembd = [
-            torch.Tensor(), # to be replaced
-            torch.Tensor(), # to be replaced
-            torch.Tensor(), # to be replaced
-        ]
-
-    def set_vocab(self):
-        super().set_vocab()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_altup_active_idx(self.hparams["altup_active_idx"])
-        self.gguf_writer.add_altup_num_inputs(self.hparams["altup_num_inputs"])
-        self.gguf_writer.add_embedding_length_per_layer_input(self.hparams["hidden_size_per_layer_input"])
-        self.gguf_writer.add_shared_kv_layers(self.hparams["num_kv_shared_layers"])
-
-        activation_sparsity_scale = []
-        for s in self.hparams["activation_sparsity_pattern"]:
-            normal_dist = torch.distributions.normal.Normal(0, 1)
-            std_multiplier = normal_dist.icdf(torch.tensor(s, dtype=torch.float32))
-            activation_sparsity_scale.append(std_multiplier.item())
-        self.gguf_writer.add_activation_sparsity_scale(activation_sparsity_scale)
-
-        sliding_window_pattern = []
-        for t in self.hparams["layer_types"]:
-            sliding_window_pattern.append(t == "sliding_attention")
-        self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
-
-    def _stack_matrices(self, matrices: list[Tensor]) -> Tensor | None:
-        has_all = all(m.numel() > 0 for m in matrices)
-        if not has_all:
-            return None
-        else:
-            return torch.stack(matrices, dim=0)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if name.endswith("_scale"):
-            name = name + ".weight"
-
-        # TODO: implement self.prediction_coefs.weight.clamp_(...)
-
-        if "language_model." not in name:
-            return [] # skip non-language model tensors
-
-        if "altup_unembed_projections" in name:
-            data_torch = data_torch.to(device="cpu")
-            if ".0." in name:
-                self._altup_unembd[0] = data_torch
-            elif ".1." in name:
-                self._altup_unembd[1] = data_torch
-            elif ".2." in name:
-                self._altup_unembd[2] = data_torch
-            else:
-                raise ValueError(f"Unknown name: {name}")
-            out = self._stack_matrices(self._altup_unembd)
-            if out is not None:
-                return [(self.map_tensor_name("model.altup_unembed_projections.weight"), out)]
-            else:
-                return []
-
-        if "altup_projections" in name:
-            data_torch = data_torch.to(device="cpu")
-            if ".0." in name:
-                self._altup_proj[0] = data_torch
-            elif ".1." in name:
-                self._altup_proj[1] = data_torch
-            elif ".2." in name:
-                self._altup_proj[2] = data_torch
-            else:
-                raise ValueError(f"Unknown name: {name}")
-            out = self._stack_matrices(self._altup_proj)
-            if out is not None:
-                return [(self.map_tensor_name("model.altup_projections.weight"), out)]
-            else:
-                return []
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @ModelBase.register("Starcoder2ForCausalLM")
 class StarCoder2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.STARCODER2
@@ -4778,14 +4593,6 @@ class ARwkv7Model(Rwkv7Model):
 class MambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
-    def __init__(self, dir_model: Path, *args, **kwargs):
-        # Avoid using AutoConfig for hparams
-        hparams = kwargs.pop("hparams", None)
-        if hparams is None:
-            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-                hparams = json.load(f)
-        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
-
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
         # Round vocab size to next multiple of 8
@@ -4860,100 +4667,6 @@ class MambaModel(TextModel):
         return [(new_name, data_torch)]
 
 
-@ModelBase.register("Mamba2ForCausalLM")
-class Mamba2Model(TextModel):
-    model_arch = gguf.MODEL_ARCH.MAMBA2
-
-    def __init__(self, dir_model: Path, *args, **kwargs):
-        # Avoid using AutoConfig for hparams
-        # It wrongly assumes all Mamba2 models are Mamba-Codestral-7B-v0.1
-        hparams = kwargs.pop("hparams", None)
-        if hparams is None:
-            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-                hparams = json.load(f)
-        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
-
-    def set_vocab(self):
-        vocab_size = self.hparams["vocab_size"]
-        # Round vocab size to next multiple of 16
-        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
-        # pad using ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
-        self.hparams["vocab_size"] = vocab_size
-
-        if (self.dir_model / "tokenizer.model").is_file():
-            self._set_vocab_sentencepiece()
-        elif (self.dir_model / "tokenizer.model.v3").is_file():
-            # mamba-codestral
-            raise NotImplementedError(f"Please rename {self.dir_model / 'tokenizer.model.v3'} to {self.dir_model / 'tokenizer.model'}")
-        elif (self.dir_model / "tokenizer.json").is_file():
-            self._set_vocab_gpt2()
-        else:
-            # Use the GPT-NeoX tokenizer when no tokenizer files are present
-            self._set_vocab_builtin("gpt-neox", vocab_size)
-
-    def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
-        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
-        head_dim = self.find_hparam(["head_dim"],                    optional=True) or 64
-        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
-
-        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
-        # Fail early for models which don't have a block expansion factor of 2
-        # TODO: does this really matter?
-        assert d_inner == 2 * d_model
-        assert d_inner % head_dim == 0
-
-        self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
-        self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
-        self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.block_count)
-        self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
-        self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
-        self.gguf_writer.add_ssm_group_count(n_group)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-
-        if name.startswith("model.backbone") or name.startswith("model.lm_head"):
-            # map Mamba-Codestral-7B-v0.1 tensor names to the names used by Mamba-2
-            name = name.removeprefix("model.")
-
-        if name.endswith(".dt_bias"):
-            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
-
-        new_name = self.map_tensor_name(name)
-
-        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
-            data_torch = data_torch.squeeze()
-        elif any(self.match_model_tensor_name(new_name, t, bid, suffix="") for t in [
-            gguf.MODEL_TENSOR.SSM_A,
-            gguf.MODEL_TENSOR.SSM_D,
-        ]):
-            # unsqueeze A to use similar shape semantics as Mamba-1
-            # (D is also unsqueezed, but for more straightforward broadcast internally)
-            data_torch = data_torch.reshape((*data_torch.shape, 1))
-        elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
-            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-            d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-            n_group = self.hparams.get("n_groups", 1)
-            data_torch = data_torch.reshape((n_group, d_inner // n_group))
-
-        if name.endswith(".A_log"):
-            logger.debug("A_log --> A ==> " + new_name)
-            data_torch = -torch.exp(data_torch)
-
-        yield (new_name, data_torch)
-
-
 @ModelBase.register("CohereForCausalLM")
 class CommandR2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.COMMAND_R
@@ -5087,6 +4800,25 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.intermediate_size = self.hparams["intermediate_size"]
+
+    def get_tensors(self):
+        for name, data in super().get_tensors():
+            if 'gated_layer' in name:
+                d1 = data[:self.intermediate_size, :]
+                name1 = name.replace('gated_layers', 'gated_layers_w')
+                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
+                d2 = data[self.intermediate_size:, :]
+                name2 = name.replace('gated_layers', 'gated_layers_v')
+                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
+                yield name1, d1
+                yield name2, d2
+                continue
+
+            yield name, data
+
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -5099,6 +4831,16 @@ class JinaBertV2Model(BertModel):
             self.gguf_writer.add_token_type_count(2)
         else:
             raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "bert.", remove the prefix
+        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+        if name.startswith("bert."):
+            name = name[5:]
+
+        return super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("OpenELMForCausalLM")
@@ -5340,7 +5082,9 @@ class DeepseekModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
@@ -5544,34 +5288,6 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("Dots1ForCausalLM")
-class Dots1Model(Qwen2MoeModel):
-    model_arch = gguf.MODEL_ARCH.DOTS1
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
-        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
-        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
-
-        if self.hparams["scoring_func"] == "noaux_tc":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
-        if name.endswith("e_score_correction_bias"):
-            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
-        if "shared_experts" in name:
-            return [(self.map_tensor_name(name), data_torch)]
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @ModelBase.register("PLMForCausalLM")
 class PLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PLM
@@ -5700,6 +5416,9 @@ class T5Model(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5837,6 +5556,9 @@ class T5EncoderModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -6224,8 +5946,7 @@ class ExaoneModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -6337,8 +6058,7 @@ class BailingMoeModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if (rope_dim := hparams.get("head_dim")) is None:
-            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
         rope_scaling = self.hparams.get("rope_scaling") or {}
@@ -6370,8 +6090,7 @@ class BailingMoeModel(TextModel):
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
         n_embd = self.hparams["hidden_size"]
-        if (head_dim := self.hparams.get("head_dim")) is None:
-            head_dim = n_embd // n_head
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
 
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
 
@@ -6632,8 +6351,8 @@ def parse_args() -> argparse.Namespace:
         help="model is executed on big endian machine",
     )
     parser.add_argument(
-        "model", type=str,
-        help="directory containing model file or huggingface repository ID (if --remote)",
+        "model", type=Path,
+        help="directory containing model file",
         nargs="?",
     )
     parser.add_argument(
@@ -6714,20 +6433,12 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # maybe we should fallback to text model's arch in that case, since not many models have both
     text_config = hparams.get("text_config", {})
     vision_config = hparams.get("vision_config", {})
-    arch = None
-    if (arches := hparams.get("architectures")) is not None and len(arches) > 0:
-        arch = arches[0]
-    elif "ssm_cfg" in hparams:
-        # For non-hf Mamba and Mamba2 models
-        arch = hparams["ssm_cfg"].get("layer", "Mamba") + "ForCausalLM"
-
+    arch = hparams["architectures"][0]
     # if "architectures" is found in the sub-config, use that instead
     if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
         arch = text_config["architectures"][0]
     elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
         arch = vision_config["architectures"][0]
-    if arch is None:
-        raise ValueError("Failed to detect model architecture")
     return arch
 
 
@@ -6744,20 +6455,18 @@ def main() -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
+    dir_model = args.model
+
     if args.remote:
-        hf_repo_id = args.model
         from huggingface_hub import snapshot_download
         local_dir = snapshot_download(
-            repo_id=hf_repo_id,
+            repo_id=str(dir_model),
             allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
-    else:
-        hf_repo_id = None
-        dir_model = Path(args.model)
 
     if not dir_model.is_dir():
-        logger.error(f'Error: {dir_model} is not a directory')
+        logger.error(f'Error: {args.model} is not a directory')
         sys.exit(1)
 
     ftype_map: dict[str, gguf.LlamaFileType] = {
@@ -6777,9 +6486,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
-    elif hf_repo_id:
+    elif args.remote:
         # if remote, use the model ID as the output file name
-        fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf")
+        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -6808,7 +6517,7 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id)
+                                     remote_hf_model_id=str(args.model) if args.remote else None)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")

docs/backend/CANN.md

@@ -8,7 +8,6 @@
  - [DataType Supports](#datatype-supports)
  - [Docker](#docker)
  - [Linux](#linux)
- - [Environment variable setup](#environment-variable-setup)
  - [TODO](#todo)
 
 
@@ -291,24 +290,5 @@ Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang
 
 We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers from Huawei Technologies Co., Ltd for their help during the code development and pull request.
 
-## Environment variable setup
-
-### GGML_CANN_ASYNC_MODE
-
-Enables asynchronous operator submission. Disabled by default.
-
-### GGML_CANN_MEM_POOL
-
-Specifies the memory pool management strategy:
-
-- vmm: Utilizes a virtual memory manager pool. If hardware support for VMM is unavailable, falls back to the legacy (leg) memory pool.
-
-- prio: Employs a priority queue-based memory pool management.
-- leg: Uses a fixed-size buffer pool.
-
-### GGML_CANN_DISABLE_BUF_POOL_CLEAN
-
-Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
-
 ## TODO
 - Support more models and data types.

docs/backend/SYCL.md

@@ -757,7 +757,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
+| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

docs/build-s390x.md

@@ -1,246 +0,0 @@
-> [!IMPORTANT]
-> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
-
-# Build llama.cpp locally (for s390x)
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
-
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
-
-**To get the code:**
-
-```bash
-git clone https://github.com/ggml-org/llama.cpp
-cd llama.cpp
-```
-
-## CPU Build with BLAS
-
-Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
-
-```bash
-cmake -S . -B build             \
-    -DCMAKE_BUILD_TYPE=Release  \
-    -DGGML_BLAS=ON              \
-    -DGGML_BLAS_VENDOR=OpenBLAS
-
-cmake --build build --config Release -j $(nproc)
-```
-
-**Notes**:
-
--   For faster repeated compilation, install [ccache](https://ccache.dev/)
--   By default, VXE/VXE2 is enabled. To disable it (not recommended):
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_VXE=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
--   By default, NNPA is enabled when available. To disable it (not recommended):
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_NNPA=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
--   For debug builds:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Debug    \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS
-    cmake --build build --config Debug -j $(nproc)
-    ```
-
--   For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DBUILD_SHARED_LIBS=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
-## Getting GGUF Models
-
-All models need to be converted to Big-Endian. You can achieve this in three cases:
-
-1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
-
-    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
-
-    You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
-
-    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
-
-2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
-
-    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
-
-    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile model-name-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        model-directory/
-    ```
-
-    For example,
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile granite-3.3-2b-instruct-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        granite-3.3-2b-instruct/
-    ```
-
-3. **Convert existing GGUF Little-Endian model to Big-Endian**
-
-    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
-
-    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
-
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
-    ```
-
-    For example,
-
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
-    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
-    ```
-
-    **Notes:**
-
-    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
-
-## IBM Accelerators
-
-### 1. SIMD Acceleration
-
-Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
-
-### 2. NNPA Vector Intrinsics Acceleration
-
-Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
-
-### 3. zDNN Accelerator
-
-_Only available in IBM z16 or later system. No direction at the moment._
-
-### 4. Spyre Accelerator
-
-_No direction at the moment._
-
-## Performance Tuning
-
-### 1. Virtualization Setup
-
-It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
-
-Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
-
-### 2. IFL (Core) Count
-
-It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
-
-Note: IFL count does not equate to vCPU count.
-
-### 3. SMT vs NOSMT (Simultaneous Multithreading)
-
-It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
-
-### 4. BLAS vs NOBLAS
-
-IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
-
-## Frequently Asked Questions (FAQ)
-
-1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
-
-    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
-
-    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
-
-2. I'm getting extremely poor performance when running inference on a model
-
-    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
-
-3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
-
-    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
-
-## Getting Help on IBM Z & LinuxONE
-
-1. **Bugs, Feature Requests**
-
-    Please file an issue in llama.cpp and ensure that the title contains "s390x".
-
-2. **Other Questions**
-
-    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
-
-## Appendix A: Hardware Support Matrix
-
-|         | Support | Minimum Compiler Version |
-| ------- | ------- | ------------------------ |
-| IBM z15 | ✅      |                          |
-| IBM z16 | ✅      |                          |
-| IBM z17 | ✅      | GCC 15.1.0               |
-
--   ✅ - supported and verified to run as intended
--   🚫 - unsupported, we are unlikely able to provide support
-
-## Appendix B: SIMD Support Matrix
-
-|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
-| ---------- | ----------- | ---- | ---- | ----- |
-| FP32       | ✅          | ✅   | ❓   | ❓    |
-| FP16       | ✅          | ✅   | ❓   | ❓    |
-| BF16       | 🚫          | 🚫   | ❓   | ❓    |
-| Q4_0       | ✅          | ✅   | ❓   | ❓    |
-| Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
-| Q8_0       | ✅          | ✅   | ❓   | ❓    |
-| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
-| Q3_K       | ✅          | ✅   | ❓   | ❓    |
-| Q4_K       | ✅          | ✅   | ❓   | ❓    |
-| Q5_K       | ✅          | ✅   | ❓   | ❓    |
-| Q6_K       | ✅          | ✅   | ❓   | ❓    |
-| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
-| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
-| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
-| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
-| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
-
--   ✅ - acceleration available
--   🚫 - acceleration unavailable, will still run using scalar implementation
--   ❓ - acceleration unknown, please contribute if you can test it yourself

docs/build.md

@@ -1,9 +1,5 @@
 # Build llama.cpp locally
 
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
-
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
-
 **To get the Code:**
 
 ```bash
@@ -67,7 +63,6 @@ cmake --build build --config Release
       cmake --preset x64-windows-llvm-release
       cmake --build build-x64-windows-llvm-release
       ```
-- Curl usage is enabled by default and can be turned off with `-DLLAMA_CURL=OFF`. Otherwise you need to install development libraries for libcurl.
 
 ## BLAS Build
 
@@ -557,10 +552,6 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
-## IBM Z & LinuxONE
-
-To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
-
 ## Notes about GPU-accelerated backends
 
 The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.

docs/docker.md

@@ -25,9 +25,6 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 

docs/function-calling.md

@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
   - Functionary v3.1 / v3.2
   - Hermes 2/3, Qwen 2.5
-  - Qwen 2.5 Coder
+  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/install.md

@@ -1,42 +1,28 @@
 # Install pre-built version of llama.cpp
 
-| Install via | Windows | Mac | Linux |
-|-------------|---------|-----|-------|
-| Winget      | ✅      |      |      |
-| Homebrew    |         | ✅   | ✅   |
-| MacPorts    |         | ✅   |      |
-| Nix         |         | ✅   | ✅   |
+## Homebrew
 
-## Winget (Windows)
-
-```sh
-winget install llama.cpp
-```
-
-The package is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/issues/8188
-
-## Homebrew (Mac and Linux)
+On Mac and Linux, the homebrew package manager can be used via
 
 ```sh
 brew install llama.cpp
 ```
-
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
-## MacPorts (Mac)
+## MacPorts
 
 ```sh
 sudo port install llama.cpp
 ```
+see also: https://ports.macports.org/port/llama.cpp/details/
 
-See also: https://ports.macports.org/port/llama.cpp/details/
+## Nix
 
-## Nix (Mac and Linux)
+On Mac and Linux, the Nix package manager can be used via
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
-
 For flake enabled installs.
 
 Or
@@ -48,3 +34,13 @@ nix-env --file '<nixpkgs>' --install --attr llama-cpp
 For non-flake enabled installs.
 
 This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+## Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+
+```sh
+flox install llama-cpp
+```
+
+Flox follows the nixpkgs build of llama.cpp.

docs/multimodal.md

@@ -107,7 +107,3 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
-
-## Finding more models:
-
-GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf

examples/Miku.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
+    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/chat-13B.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -e
 

examples/chat-persistent.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -euo pipefail
 

examples/chat-vicuna.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -e
 

examples/chat.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 #
 # Temporary script - will be removed in the future

examples/embedding/embedding.cpp

@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_memory_clear(llama_get_memory(ctx), true);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -133,36 +133,10 @@ int main(int argc, char ** argv) {
     // max batch size
     const uint64_t n_batch = params.n_batch;
 
-    // get added sep and eos token, if any
-    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
-    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
-
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        std::vector<llama_token> inp;
-
-        // split classification pairs and insert expected separator tokens
-        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
-            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
-            std::string final_prompt;
-
-            for (size_t i = 0; i < pairs.size(); i++) {
-                final_prompt += pairs[i];
-                if (i != pairs.size() - 1) {
-                    if (!added_eos_token.empty()) {
-                        final_prompt += added_eos_token;
-                    }
-                    if (!added_sep_token.empty()) {
-                        final_prompt += added_sep_token;
-                    }
-                }
-            }
-
-            inp = common_tokenize(ctx, final_prompt, true, true);
-        } else {
-            inp = common_tokenize(ctx, prompt, true, true);
-        }
+        auto inp = common_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -171,11 +145,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP/EOS
+    // check if the last token is SEP
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
-            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
+        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
+            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
@@ -262,24 +236,9 @@ int main(int argc, char ** argv) {
                 LOG("\n");
             }
         } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
-            const uint32_t n_cls_out = llama_model_n_cls_out(model);
-            std::vector<std::string> cls_out_labels;
-
-            for (uint32_t i = 0; i < n_cls_out; i++) {
-                const char * label = llama_model_cls_label(model, i);
-                const std::string label_i(label == nullptr ? "" : label);
-                cls_out_labels.emplace_back(label_i.empty() ? std::to_string(i) : label_i);
-            }
-
             for (int j = 0; j < n_embd_count; j++) {
-                for (uint32_t i = 0; i < n_cls_out; i++) {
-                    // NOTE: if you change this log - update the tests in ci/run.sh
-                    if (n_cls_out == 1) {
-                        LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
-                    } else {
-                        LOG("rerank score %d: %8.3f [%s]\n", j, emb[j * n_embd + i], cls_out_labels[i].c_str());
-                    }
-                }
+                // NOTE: if you change this log - update the tests in ci/run.sh
+                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding

examples/eval-callback/eval-callback.cpp

@@ -55,8 +55,6 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                     } else if (type == GGML_TYPE_F32) {
                         v = *(float *) &data[i];
-                    } else if (type == GGML_TYPE_I64) {
-                        v = (float) *(int64_t *) &data[i];
                     } else if (type == GGML_TYPE_I32) {
                         v = (float) *(int32_t *) &data[i];
                     } else if (type == GGML_TYPE_I16) {
@@ -136,11 +134,6 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (tokens.empty()) {
-        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
-        return false;
-    }
-
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;

examples/gritlm/gritlm.cpp

@@ -41,11 +41,12 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, true);
+            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_memory_clear(llama_get_memory(ctx), true);
+        llama_kv_self_clear(ctx);
+        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -101,7 +102,8 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_memory_clear(llama_get_memory(ctx), true);
+    llama_kv_self_clear(ctx);
+    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -164,8 +166,6 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
-    cparams.embeddings = true;
-
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,8 +213,6 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
-    llama_set_embeddings(ctx, false);
-
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {

examples/jeopardy/jeopardy.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_memory_clear(llama_get_memory(context), false);
+        llama_kv_self_clear(context);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_memory_clear(llama_get_memory(context), false);
+        llama_kv_self_clear(context);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_memory_clear(llama_get_memory(context), false);
+        llama_kv_self_clear(context);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
+    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_memory_clear(llama_get_memory(context), false)
+            llama_kv_self_clear(context)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_memory_clear(llama_get_memory(context), false)
+            llama_kv_self_clear(context)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_memory_clear(llama_get_memory(context), false)
+            llama_kv_self_clear(context)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_memory_clear(llama_get_memory(context), true)
+        llama_kv_self_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/lookahead/lookahead.cpp

@@ -60,8 +60,6 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
-    auto * mem = llama_get_memory(ctx);
-
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // Tokenize the prompt
@@ -96,7 +94,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_memory_seq_cp(mem, 0, s, -1, -1);
+        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -429,17 +427,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_memory_seq_rm(mem, -1, n_past, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_memory_seq_keep(mem, seq_id_best);
-            llama_memory_seq_cp  (mem, seq_id_best, 0, -1, -1);
-            llama_memory_seq_rm  (mem, seq_id_best,    -1, -1);
+            llama_kv_self_seq_keep(ctx, seq_id_best);
+            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_memory_seq_cp(mem, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -181,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_memory_seq_rm(llama_get_memory(ctx), 0, n_past, -1);
+        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/parallel/parallel.cpp

@@ -158,7 +158,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.n_predict = 128;
-    params.n_junk = 1;
+    params.n_junk = 0;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {
     const bool is_sp_shared = params.is_pp_shared;
 
     // extra text to insert in each client's prompt in order to make it larger
-    const int32_t n_junk = std::max(1, params.n_junk);
+    const int32_t n_junk = params.n_junk;
 
     // init llama.cpp
     llama_backend_init();
@@ -194,8 +194,6 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
-    auto * mem = llama_get_memory(ctx);
-
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
@@ -261,7 +259,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_memory_seq_cp(mem, 0, i, -1, -1);
+            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -288,9 +286,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_memory_seq_rm(mem, i, -1, -1);
+                llama_kv_self_seq_rm(ctx, i, -1, -1);
                 // but keep the system prompt
-                llama_memory_seq_cp(mem, 0, i, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -364,9 +362,7 @@ int main(int argc, char ** argv) {
         // process in chunks of params.n_batch
         int32_t n_batch = params.n_batch;
 
-        int32_t i_next = 0;
-
-        for (int32_t i = 0; i < batch.n_tokens; i = i_next) {
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
             // experiment: process in powers of 2
             //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
             //    n_batch /= 2;
@@ -374,7 +370,7 @@ int main(int argc, char ** argv) {
             //    continue;
             //}
 
-            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
 
             llama_batch batch_view = {
                 n_tokens,
@@ -394,24 +390,19 @@ int main(int argc, char ** argv) {
                     return 1;
                 }
 
-                LOG_WRN("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+                LOG_ERR("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
 
                 n_cache_miss += 1;
 
                 // retry with half the batch size to try to find a free slot in the KV cache
                 n_batch /= 2;
+                i -= n_batch;
 
                 continue;
             }
 
             LOG_DBG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
 
-            // move the head of the batch forward with the number of tokens we just processed
-            i_next = i + n_tokens;
-
-            // on successful decode, restore the original batch size
-            n_batch = params.n_batch;
-
             for (auto & client : clients) {
                 if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
                     continue;
@@ -449,8 +440,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
-                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
+                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
+                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,8 +126,6 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
-    auto * mem = llama_get_memory(ctx);
-
     // fill the KV cache
     for (int i = 0; i < n_ctx; i += n_batch) {
         if (i > 0 && n_grp > 1) {
@@ -135,10 +133,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_memory_seq_add(mem, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_memory_seq_div(mem, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_update  (ctx);
 
-            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -168,10 +167,12 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
-        llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
+        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+      //llama_kv_self_defrag (ctx);
+        llama_kv_self_update (ctx);
 
-        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
+        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -197,10 +198,12 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
-            llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
+            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+          //llama_kv_self_defrag (ctx);
+            llama_kv_self_update (ctx);
 
-            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/reason-act.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 cd `dirname $0`
 cd ..

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_process(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_memory_clear(llama_get_memory(ctx), false);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/save-load-state/save-load-state.cpp

@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_memory_clear(llama_get_memory(ctx3), true);
+        llama_kv_self_clear(ctx3);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/server-llama2-13B.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -e
 

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1;
+        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,16 +113,15 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
+            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");
                 exit(0);
             }
 
-            int ret = llama_decode(ctx, batch);
-            if (ret != 0) {
-                GGML_ABORT("failed to decode, ret = %d\n", ret);
+            if (llama_decode(ctx, batch)) {
+                GGML_ABORT("failed to decode\n");
             }
 
             // sample the next token

examples/speculative-simple/speculative-simple.cpp

@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
+            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -142,8 +142,6 @@ int main(int argc, char ** argv) {
         }
     }
 
-    auto * mem_tgt = llama_get_memory(ctx_tgt);
-    auto * mem_dft = llama_get_memory(ctx_dft);
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -422,14 +420,14 @@ int main(int argc, char ** argv) {
             {
                 LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
 
-                llama_memory_seq_keep(mem_dft, s_keep);
-                llama_memory_seq_cp  (mem_dft, s_keep, 0, -1, -1);
-                llama_memory_seq_keep(mem_dft, 0);
+                llama_kv_self_seq_keep(ctx_dft, s_keep);
+                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_dft, 0);
 
-                llama_memory_seq_rm  (mem_tgt, s_keep, n_past_tgt, -1);
-                llama_memory_seq_keep(mem_tgt, s_keep);
-                llama_memory_seq_cp  (mem_tgt, s_keep, 0, -1, -1);
-                llama_memory_seq_keep(mem_tgt, 0);
+                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
+                llama_kv_self_seq_keep(ctx_tgt, s_keep);
+                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -446,7 +444,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_memory_seq_rm(mem_dft, 0, n_past_dft, -1);
+            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
             // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode(ctx_dft, batch_dft);
 
@@ -505,8 +503,8 @@ int main(int argc, char ** argv) {
                     if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
-                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
-                        llama_memory_seq_cp(mem_dft, s, n_seq_cur, -1, -1);
+                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
+                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
 
                         // all previous tokens from this branch are now also part of the new branch
                         for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -587,9 +585,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_memory_seq_keep(mem_tgt, 0);
+            llama_kv_self_seq_keep(ctx_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_memory_seq_cp(mem_tgt, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
             }
 
             // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());

examples/sycl/build.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT

examples/sycl/run-llama2.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation

examples/sycl/run-llama3.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 #  MIT license
 #  Copyright (C) 2025 Intel Corporation

examples/ts-type-to-grammar.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #
 # ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
 # python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json

ggml/CMakeLists.txt

@@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
 message(DEBUG "INS_ENB             : ${INS_ENB}")
 
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
-option(GGML_CPU_REPACK       "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
 option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
@@ -131,14 +131,13 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
-option(GGML_NNPA             "ggml: enable nnpa"             ON)
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
 set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
-if (MINGW)
+if (WIN32)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
@@ -173,7 +172,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -181,6 +179,7 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -265,6 +264,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cpp.h
     include/ggml-cuda.h
+    include/ggml-kompute.h
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
@@ -358,13 +358,6 @@ write_basic_package_version_file(
     VERSION ${GGML_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
-target_compile_definitions(ggml-base PRIVATE
-    GGML_VERSION="${GGML_INSTALL_VERSION}"
-    GGML_COMMIT="${GGML_BUILD_COMMIT}"
-)
-message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
-message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")
-
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
@@ -374,8 +367,6 @@ if (MSVC)
         /wd4005  # Macro redefinition
         /wd4244  # Conversion from one type to another type, possible loss of data
         /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
-        /wd4305  # Conversion from 'type1' to 'type2', possible loss of data
-        /wd4566  # Conversion from 'char' to 'wchar_t', possible loss of data
         /wd4996  # Disable POSIX deprecation warnings
         /wd4702  # Unreachable code warnings
     )
@@ -395,46 +386,4 @@ if (MSVC)
     disable_msvc_warnings(ggml-cpu-skylakex)
     disable_msvc_warnings(ggml-cpu-icelake)
     disable_msvc_warnings(ggml-cpu-alderlake)
-
-    if (GGML_BUILD_EXAMPLES)
-        disable_msvc_warnings(common-ggml)
-        disable_msvc_warnings(common)
-
-        disable_msvc_warnings(mnist-common)
-        disable_msvc_warnings(mnist-eval)
-        disable_msvc_warnings(mnist-train)
-
-        disable_msvc_warnings(gpt-2-ctx)
-        disable_msvc_warnings(gpt-2-alloc)
-        disable_msvc_warnings(gpt-2-backend)
-        disable_msvc_warnings(gpt-2-sched)
-        disable_msvc_warnings(gpt-2-quantize)
-        disable_msvc_warnings(gpt-2-batched)
-
-        disable_msvc_warnings(gpt-j)
-        disable_msvc_warnings(gpt-j-quantize)
-
-        disable_msvc_warnings(magika)
-        disable_msvc_warnings(yolov3-tiny)
-        disable_msvc_warnings(sam)
-
-        disable_msvc_warnings(simple-ctx)
-        disable_msvc_warnings(simple-backend)
-    endif()
-
-    if (GGML_BUILD_TESTS)
-        disable_msvc_warnings(test-mul-mat)
-        disable_msvc_warnings(test-arange)
-        disable_msvc_warnings(test-backend-ops)
-        disable_msvc_warnings(test-cont)
-        disable_msvc_warnings(test-conv-transpose)
-        disable_msvc_warnings(test-conv-transpose-1d)
-        disable_msvc_warnings(test-conv1d)
-        disable_msvc_warnings(test-conv2d)
-        disable_msvc_warnings(test-conv2d-dw)
-        disable_msvc_warnings(test-customop)
-        disable_msvc_warnings(test-dup)
-        disable_msvc_warnings(test-opt)
-        disable_msvc_warnings(test-pool)
-    endif ()
 endif()

ggml/cmake/common.cmake

@@ -36,7 +36,8 @@ function(ggml_get_system_arch)
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
         set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
+    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
+            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)

ggml/include/ggml-backend.h

@@ -339,7 +339,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
 
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml/include/ggml-cpu.h

@@ -101,7 +101,6 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
-    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -134,7 +133,6 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
-    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

ggml/include/ggml-kompute.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define GGML_KOMPUTE_MAX_DEVICES 16
+
+struct ggml_vk_device {
+    int index;
+    int type; // same as VkPhysicalDeviceType
+    size_t heapSize;
+    const char * name;
+    const char * vendor;
+    int subgroupSize;
+    uint64_t bufferAlignment;
+    uint64_t maxAlloc;
+};
+
+struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
+bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
+bool ggml_vk_has_vulkan(void);
+bool ggml_vk_has_device(void);
+struct ggml_vk_device ggml_vk_current_device(void);
+
+//
+// backend API
+//
+
+// forward declaration
+typedef struct ggml_backend * ggml_backend_t;
+
+GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
+
+GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml.h

@@ -314,13 +314,6 @@
 extern "C" {
 #endif
 
-    // Function type used in fatal error callbacks
-    typedef void (*ggml_abort_callback_t)(const char * error_message);
-
-    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
-    // Returns the old callback for chaining
-    GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
-
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
@@ -477,7 +470,6 @@ extern "C" {
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
         GGML_OP_GET_ROWS_BACK,
-        GGML_OP_SET_ROWS,
         GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
@@ -489,7 +481,6 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
-        GGML_OP_CONV_2D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -498,7 +489,6 @@ extern "C" {
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
-        GGML_OP_ROLL,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -528,8 +518,6 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
 
-        GGML_OP_GLU,
-
         GGML_OP_COUNT,
     };
 
@@ -553,16 +541,6 @@ extern "C" {
         GGML_UNARY_OP_COUNT,
     };
 
-    enum ggml_glu_op {
-        GGML_GLU_OP_REGLU,
-        GGML_GLU_OP_GEGLU,
-        GGML_GLU_OP_SWIGLU,
-        GGML_GLU_OP_GEGLU_ERF,
-        GGML_GLU_OP_GEGLU_QUICK,
-
-        GGML_GLU_OP_COUNT,
-    };
-
     enum ggml_object_type {
         GGML_OBJECT_TYPE_TENSOR,
         GGML_OBJECT_TYPE_GRAPH,
@@ -648,9 +626,6 @@ extern "C" {
 
     // misc
 
-    GGML_API const char * ggml_version(void);
-    GGML_API const char * ggml_commit(void);
-
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
     GGML_API int64_t ggml_time_ms(void);
     GGML_API int64_t ggml_time_us(void);
@@ -681,7 +656,6 @@ extern "C" {
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
-    GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
     GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
@@ -712,9 +686,6 @@ extern "C" {
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
-    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
-    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
-
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -786,7 +757,6 @@ extern "C" {
     GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
     GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
-    GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1115,89 +1085,6 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // gated linear unit ops
-    // A: n columns, r rows,
-    // result is n / 2 columns, r rows,
-    // expects gate in second half of row, unless swapped is true
-    GGML_API struct ggml_tensor * ggml_glu(
-            struct ggml_context * ctx,
-             struct ggml_tensor * a,
-             enum ggml_glu_op     op,
-             bool                 swapped);
-
-    GGML_API struct ggml_tensor * ggml_reglu(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_reglu_swapped(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu_swapped(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_swiglu(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_swiglu_swapped(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu_erf(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu_quick(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
-    // A: n columns, r rows,
-    // B: n columns, r rows,
-    GGML_API struct ggml_tensor * ggml_glu_split(
-            struct ggml_context * ctx,
-             struct ggml_tensor * a,
-             struct ggml_tensor * b,
-             enum ggml_glu_op     op);
-
-    GGML_API struct ggml_tensor * ggml_reglu_split(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
-    GGML_API struct ggml_tensor * ggml_geglu_split(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
-    GGML_API struct ggml_tensor * ggml_swiglu_split(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
-    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
-    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1487,23 +1374,6 @@ extern "C" {
             struct ggml_tensor  * b,  // row indices
             struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
-    // a TD  [n_embd, ne1,    ne2,    ne3]
-    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
-    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
-    //
-    // undefined behavior if destination rows overlap
-    //
-    // broadcast:
-    //   ne2 % ne11 == 0
-    //   ne3 % ne12 == 0
-    //
-    // return view(a)
-    GGML_API struct ggml_tensor * ggml_set_rows(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,  // destination
-            struct ggml_tensor  * b,  // source
-            struct ggml_tensor  * c); // row indices
-
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
         struct ggml_tensor      * a);
@@ -1541,14 +1411,8 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // a    [ne0, ne01, ne02, ne03]
-    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
-    //
-    // broadcast:
-    //   ne02 % ne12 == 0
-    //   ne03 % ne13 == 0
-    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
+    // mask is optional
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
@@ -1858,17 +1722,6 @@ extern "C" {
             struct ggml_tensor  * b,
             int                   stride);
 
-    GGML_API struct ggml_tensor * ggml_conv_2d_direct(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
-            struct ggml_tensor  * b,   // input data [W, H, C, N]
-            int                   s0,  // stride dimension 0
-            int                   s1,  // stride dimension 1
-            int                   p0,  // padding dimension 0
-            int                   p1,  // padding dimension 1
-            int                   d0,  // dilation dimension 0
-            int                   d1); // dilation dimension 1
-
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -1911,12 +1764,6 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
-
-        GGML_SCALE_MODE_COUNT
-    };
-
-    enum ggml_scale_flag {
-        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
     };
 
     // interpolate
@@ -1929,26 +1776,14 @@ extern "C" {
 
     // interpolate
     // interpolate scale to specified dimensions
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
             int                   ne3,
-            enum ggml_scale_mode  mode),
-        "use ggml_interpolate instead");
-
-    // Up- or downsamples the input to the specified size.
-    // 2D scale modes (eg. bilinear) are applied to the first two dimensions.
-    GGML_API struct ggml_tensor * ggml_interpolate(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int64_t               ne0,
-            int64_t               ne1,
-            int64_t               ne2,
-            int64_t               ne3,
-            uint32_t              mode); // ggml_scale_mode [ | ggml_scale_flag...]
+            enum ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1966,17 +1801,6 @@ extern "C" {
             int                   p0,
             int                   p1);
 
-    // Move tensor elements by an offset given for each dimension. Elements that
-    // are shifted beyond the last position are wrapped around to the beginning.
-    GGML_API struct ggml_tensor * ggml_roll(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   shift0,
-            int                   shift1,
-            int                   shift2,
-            int                   shift3);
-
-
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]
@@ -2011,17 +1835,11 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
-    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
-    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
-    //
-    // broadcast:
-    //   n_head % n_head_kv == 0
-    //   n_head % ne32      == 0
-    //   ne3    % ne33      == 0
-    //
+    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -2060,8 +1878,7 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C,
-            struct ggml_tensor  * ids);
+            struct ggml_tensor  * C);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -2278,6 +2095,9 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
     GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
+    GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 

ggml/src/CMakeLists.txt

@@ -125,6 +125,7 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 
@@ -195,7 +196,6 @@ add_library(ggml-base
             ../include/ggml-opt.h
             ../include/gguf.h
             ggml.c
-            ggml.cpp
             ggml-alloc.c
             ggml-backend.cpp
             ggml-opt.cpp
@@ -212,7 +212,6 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
-add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -227,7 +226,6 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
-        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})
@@ -270,27 +268,17 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    if (GGML_SYSTEM_ARCH STREQUAL "x86")
-        foreach (feat NATIVE
-                      SSE42
-                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                      AMX_TILE AMX_INT8 AMX_BF16)
-            set(GGML_${feat} OFF)
-        endforeach()
+    foreach (feat NATIVE
+                  SSE42
+                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                  AMX_TILE AMX_INT8 AMX_BF16)
+        set(GGML_${feat} OFF)
+    endforeach()
 
-        foreach (feat ${ARGN})
-            set(GGML_${feat} ON)
-        endforeach()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
-        foreach (feat ${ARGN})
-            set(GGML_INTERNAL_${feat} ON)
-        endforeach()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
-        foreach (feat ${ARGN})
-            set(GGML_INTERNAL_${feat} ON)
-        endforeach()
-    endif()
+    foreach (feat ${ARGN})
+        set(GGML_${feat} ON)
+    endforeach()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -300,8 +288,6 @@ ggml_add_backend(CPU)
 if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
-    elseif (GGML_CPU_ARM_ARCH)
-        message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
@@ -315,47 +301,8 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
-    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            # Many of these features are optional so we build versions with popular
-            # combinations and name the backends based on the version they were
-            # first released with
-            ggml_add_cpu_backend_variant(armv8.0_1)
-            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
-            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
-            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
-            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
-            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
-            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
-        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
-            # Android-specific backends with SoC-compatible feature sets
-            ggml_add_cpu_backend_variant(android_armv8.0_1)
-            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
-            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
-        elseif (APPLE)
-            ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
-            ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)
-            ggml_add_cpu_backend_variant(apple_m4             DOTPROD MATMUL_INT8 NOSVE SME)
-        else()
-            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
-        endif()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(power0)
-            ggml_add_cpu_backend_variant(power7_1       POWER7)
-            ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)
-            ggml_add_cpu_backend_variant(power8_1       POWER8)
-            ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
-            ggml_add_cpu_backend_variant(power9         POWER9  VSX)
-            ggml_add_cpu_backend_variant(power10        POWER10 VSX)
-            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
-        else()
-            message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
-        endif()
     else()
-        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")
@@ -365,6 +312,7 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
+ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)

ggml/src/ggml-backend-reg.cpp

@@ -61,13 +61,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_KOMPUTE
+#include "ggml-kompute.h"
+#endif
+
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
 
 namespace fs = std::filesystem;
@@ -90,8 +91,6 @@ static std::string path_str(const fs::path & path) {
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif
 
 #ifdef _WIN32
@@ -185,6 +184,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
+#ifdef GGML_USE_KOMPUTE
+        register_backend(ggml_backend_kompute_reg());
+#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -568,6 +570,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
+    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);

ggml/src/ggml-backend.cpp

@@ -817,9 +817,8 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
-                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1827,7 +1826,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
     ggml_free(copy.ctx_unallocated);
 }
 
-bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node) {
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
     struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
     if (copy.buffer == NULL) {
         return false;
@@ -1838,45 +1837,28 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     assert(g1->n_nodes == g2->n_nodes);
 
-    if (test_node != nullptr) {
-        // Compute the whole graph and only test the output for a specific tensor
-        ggml_backend_graph_compute(backend1, g1);
-        ggml_backend_graph_compute(backend2, g2);
+    for (int i = 0; i < g1->n_nodes; i++) {
+        struct ggml_tensor * t1 = g1->nodes[i];
+        struct ggml_tensor * t2 = g2->nodes[i];
 
-        int test_node_idx = -1;
-        for (int i = 0; i < g1->n_nodes; i++) {
-            struct ggml_tensor * t1 = g1->nodes[i];
-            if (t1 == test_node) {
-                test_node_idx = i;
-                break;
-            }
+        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+        ggml_backend_graph_compute(backend1, &g1v);
+        ggml_backend_graph_compute(backend2, &g2v);
+
+        if (ggml_is_view_op(t1->op)) {
+            continue;
         }
-        GGML_ASSERT(test_node_idx != -1);
 
-        callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
-    } else {
-        for (int i = 0; i < g1->n_nodes; i++) {
-            struct ggml_tensor * t1 = g1->nodes[i];
-            struct ggml_tensor * t2 = g2->nodes[i];
-
-            assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-            struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-            struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-            ggml_backend_graph_compute(backend1, &g1v);
-            ggml_backend_graph_compute(backend2, &g2v);
-
-            if (ggml_is_view_op(t1->op)) {
-                continue;
-            }
-
-            // compare results, calculate rms etc
-            if (!callback(i, t1, t2, user_data)) {
-                break;
-            }
+        // compare results, calculate rms etc
+        if (!callback(i, t1, t2, user_data)) {
+            break;
         }
     }
+
     ggml_backend_graph_copy_free(copy);
 
     return true;

ggml/src/ggml-blas/CMakeLists.txt

@@ -81,7 +81,7 @@ if (BLAS_FOUND)
     target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
     target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
 else()
-    message(FATAL_ERROR "BLAS not found, please refer to "
-                        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
-                        " to set correct GGML_BLAS_VENDOR")
+    message(ERROR "BLAS not found, please refer to "
+                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                  " to set correct GGML_BLAS_VENDOR")
 endif()

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,9 +65,8 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v3.h>
+#include <aclnnop/aclnn_grouped_matmul_v2.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
-#include <aclnnop/aclnn_zero.h>
 #include <float.h>
 
 #include <cmath>
@@ -805,11 +804,10 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
         nb[i] = nb[i - 1] * ne[i - 1];
     }
 
+    ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
     aclTensor* zero =
         ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
     return zero;
-    GGML_UNUSED(n_bytes);
 }
 
 /**
@@ -2656,67 +2654,6 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
     }
 
-#ifdef ASCEND_310P
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row = *dst;
-
-    if (src0->type == GGML_TYPE_F16) {
-        src0_row.type = GGML_TYPE_F32;
-    }
-
-    // src0_row [D, M, 1, 1] weight without permute
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[0] = ori_src0_nb[0];
-    src0_row.nb[1] = ori_src0_nb[1];
-    src0_row.nb[2] = ori_src0_nb[1];
-    src0_row.nb[3] = ori_src0_nb[1];
-
-    // src1_row [D, 1, 1, 1] -> input
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    // dst_row [M, 1, 1, 1] -> out
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    //create weight for one row
-    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-        for (int64_t id = 0; id < n_ids; id++) {
-            // expert index
-            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-            GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-            // If B = 1 (broadcast), always use 0; otherwise, use id.
-            int64_t i11 = (ne11 == 1 ? 0 : id);
-            int64_t i12 = iid1;
-
-            int64_t i1 = id;
-            int64_t i2 = i12;
-
-            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
-            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
-            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
-
-            src0_row.data = src0_tmp_ptr;
-            src1_row.data = src1_tmp_ptr;
-            dst_row.data = dst_tmp_ptr;
-            dst_row.src[0] = &src0_row;
-            dst_row.src[1] = &src1_row;
-
-            ggml_cann_mul_mat(ctx, &dst_row);
-        }
-    }
-    return;
-#endif
-
     std::vector<aclTensor*> src0_tensor_vec;
     std::vector<aclTensor*> src1_tensor_vec;
     std::vector<aclTensor*> dst_tensor_vec;
@@ -2764,9 +2701,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     }
 
     size_t GROUP_SIZE = 128;
-    // GroupedMatmulV3 required tensor_list.size < 128
+    // GroupedMatmulV2 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV3
+        // split and call GroupedMatmulV2
         size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
         std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
         std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2776,7 +2713,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
         aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
             nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
 
         ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);

ggml/src/ggml-cann/common.h

@@ -37,7 +37,6 @@
 #include <thread>
 #include <unistd.h>
 #include <functional>
-#include <optional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -104,9 +103,6 @@ const ggml_cann_device_info& ggml_cann_info();
 void ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
-std::optional<std::string> get_env(const std::string& name);
-bool parse_bool(const std::string& value);
-
 /**
  * @brief Abstract base class for memory pools used by CANN.
  */
@@ -358,8 +354,7 @@ struct ggml_backend_cann_context {
         : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
-
-        async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -31,8 +31,6 @@
 #include <mutex>
 #include <queue>
 #include <chrono>
-#include <unordered_set>
-#include <optional>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -95,26 +93,6 @@ int32_t ggml_cann_get_device() {
     return id;
 }
 
-/**
- * @brief Get the value of the specified environment variable (name).
- *        if not empty, return a std::string object
- */
-std::optional<std::string> get_env(const std::string& name) {
-    const char* val = std::getenv(name.c_str());
-    if (!val) return std::nullopt;
-    std::string res = std::string(val);
-    std::transform(res.begin(), res.end(), res.begin(), ::tolower);
-    return res;
-}
-
-/**
- * @brief Verify whether the environment variable is a valid value.
- */
-bool parse_bool(const std::string& value) {
-    std::unordered_set<std::string> valid_values = {"on", "1", "yes", "y", "enable", "true"};
-    return valid_values.find(value) != valid_values.end();
-}
-
 /**
  * @brief Initialize the CANN device information.
  *
@@ -236,7 +214,7 @@ struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf_prio(int device) : device(device) {
-        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
     }
 
     /**
@@ -432,7 +410,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf(int device) : device(device) {
-        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
     }
 
     /**
@@ -753,18 +731,16 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    std::string mem_pool_type = get_env("GGML_CANN_MEM_POOL").value_or("");
-
-    if (mem_pool_type == "prio") {
-        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
-        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
-    }
-
-    if (ggml_cann_info().devices[device].vmm && mem_pool_type != "leg") {
+    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
+    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
         GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
         return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
     }
-
+    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
+    if (enable_buf_prio) {
+        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
+    }
     GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
@@ -2086,12 +2062,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
-        case GGML_OP_SET_ROWS:
-            {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
-            } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@@ -2193,6 +2163,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
@@ -2210,10 +2181,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
-        case GGML_OP_SOFT_MAX:
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2236,8 +2203,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-common.h

@@ -1074,10 +1074,6 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
-GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
-    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
-GGML_TABLE_END()
-
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,17 +1,3 @@
-function(ggml_add_cpu_backend_features cpu_name arch)
-    # The feature detection code is compiled as a separate target so that
-    # it can be built without the architecture flags
-    # Since multiple variants of the CPU backend may be included in the same
-    # build, using set_source_files_properties() to set the arch flags is not possible
-    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
-    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
-    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
-    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
-    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
-    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
-endfunction()
-
 function(ggml_add_cpu_backend_variant_impl tag_name)
     if (tag_name)
         set(GGML_CPU_NAME ggml-cpu-${tag_name})
@@ -24,14 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/repack.cpp
-        ggml-cpu/repack.h
-        ggml-cpu/hbm.cpp
-        ggml-cpu/hbm.h
-        ggml-cpu/quants.c
-        ggml-cpu/quants.h
-        ggml-cpu/traits.cpp
-        ggml-cpu/traits.h
+        ggml-cpu/ggml-cpu-aarch64.cpp
+        ggml-cpu/ggml-cpu-aarch64.h
+        ggml-cpu/ggml-cpu-hbm.cpp
+        ggml-cpu/ggml-cpu-hbm.h
+        ggml-cpu/ggml-cpu-quants.c
+        ggml-cpu/ggml-cpu-quants.h
+        ggml-cpu/ggml-cpu-traits.cpp
+        ggml-cpu/ggml-cpu-traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -98,11 +84,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (GGML_SYSTEM_ARCH STREQUAL "ARM")
         message(STATUS "ARM detected")
-        list(APPEND GGML_CPU_SOURCES
-            ggml-cpu/arch/arm/quants.c
-            ggml-cpu/arch/arm/repack.cpp
-            )
-
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
         else()
@@ -157,49 +138,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
-                elseif(GGML_CPU_ALL_VARIANTS)
-                    # Begin with the lowest baseline
-                    set(ARM_MCPU "armv8-a")
-                    set(ARCH_TAGS "")
-                    set(ARCH_DEFINITIONS "")
-
-                    # When a feature is selected, bump the MCPU to the first
-                    # version that supported it
-                    if (GGML_INTERNAL_DOTPROD)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
-                    endif()
-                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
-                    endif()
-                    if (GGML_INTERNAL_SVE)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
-                    endif()
-                    if (GGML_INTERNAL_MATMUL_INT8)
-                        set(ARM_MCPU "armv8.6-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
-                    endif()
-                    if (GGML_INTERNAL_SVE2)
-                        set(ARM_MCPU "armv8.6-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
-                    endif()
-                    if (GGML_INTERNAL_NOSVE)
-                        set(ARCH_TAGS "${ARCH_TAGS}+nosve")
-                    endif()
-                    if (GGML_INTERNAL_SME)
-                        set(ARM_MCPU "armv9.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
-                    endif()
-                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
-                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                 endif()
             endif()
 
@@ -229,11 +167,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
-        list(APPEND GGML_CPU_SOURCES
-            ggml-cpu/arch/x86/quants.c
-            ggml-cpu/arch/x86/repack.cpp
-            )
-
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -363,11 +296,21 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
                 message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
             endif()
-            ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
+
+            # The feature detection code is compiled as a separate target so that
+            # it can be built without the architecture flags
+            # Since multiple variants of the CPU backend may be included in the same
+            # build, using set_source_files_properties() to set the arch flags is not possible
+            set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
+            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
+            target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
+            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+            set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")
-        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
         if (GGML_NATIVE)
             if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
                 file(READ "/proc/cpuinfo" POWER10_M)
@@ -375,8 +318,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
             endif()
 
-            string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
-            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
@@ -388,27 +330,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
             endif()
-        elseif(GGML_CPU_ALL_VARIANTS)
-            # Begin with the lowest baseline
-            set(ARCH_DEFINITIONS "")
-
-            # When a feature is selected, bump the MCPU to the first
-            # version that supported it
-            foreach(PVER RANGE 7 11)
-                if(DEFINED GGML_INTERNAL_POWER${PVER})
-                    set(POWERPC_MCPU "power${PVER}")
-                    list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
-                endif()
-            endforeach()
-            if (GGML_INTERNAL_VSX)
-                list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
-                list(APPEND ARCH_FLAGS -mvsx)
-            endif()
-
-            if (DEFINED POWERPC_MCPU)
-                list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
-            endif()
-            ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
         else()
             if (GGML_CPU_POWERPC_CPUTYPE)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
@@ -416,8 +337,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
         message(STATUS "loongarch64 detected")
-        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
-
         list(APPEND ARCH_FLAGS -march=loongarch64)
         if (GGML_LASX)
             list(APPEND ARCH_FLAGS -mlasx)
@@ -427,10 +346,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
         message(STATUS "riscv64 detected")
-        list(APPEND GGML_CPU_SOURCES
-            ggml-cpu/arch/riscv/quants.c
-            ggml-cpu/arch/riscv/repack.cpp
-            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -442,13 +357,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
-        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
         string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
-            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -465,25 +378,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (GGML_VXE)
-            message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
-            list(APPEND ARCH_DEFINITIONS GGML_VXE)
         endif()
-
-        if (GGML_NNPA)
-            message(STATUS "NNPA enabled")
-            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
-        endif()
-    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
-        message(STATUS "Wasm detected")
-        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
-        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
+        message(STATUS "Unknown architecture")
     endif()
 
-    if (GGML_CPU_REPACK)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
+    if (GGML_CPU_AARCH64)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
     endif()
 
     if (GGML_CPU_KLEIDIAI)
@@ -494,9 +396,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -589,9 +491,4 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (EMSCRIPTEN)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
-
-    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
-        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
-        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
-    endif()
 endfunction()

ggml/src/ggml-cpu/amx/amx.cpp

@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "traits.h"
+#include "ggml-cpu-traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,8 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "simd-mappings.h"
-#include "quants.h"
+#include "ggml-cpu-quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>
@@ -454,7 +453,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
 
         // Quantize these floats
         const float iscale = 127.f / amax;
-        y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
+        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
         const float id = ( amax != 0.0f ) ? iscale : 0.f;
         const __m512 vscale = _mm512_set1_ps(id);
 
@@ -1091,7 +1090,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1114,8 +1113,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
         const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
-            const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1138,7 +1137,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1438,7 +1437,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b
@@ -1499,8 +1498,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
-                vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
             }
 
             // load b
@@ -1572,7 +1571,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     va[k] = _mm512_add_epi8(va[k], off);
                 }
-                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b

ggml/src/ggml-cpu/arch-fallback.h

@@ -1,184 +0,0 @@
-#pragma once
-
-// Rename `_generic` functions if no native implementation is available.
-// This effectively selects the generic implementation.
-
-#if defined(GGML_CPU_GENERIC)
-// quants.c
-#define quantize_row_q8_0_generic quantize_row_q8_0
-#define quantize_row_q8_1_generic quantize_row_q8_1
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
-#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
-#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
-#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
-// repack.cpp
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__POWERPC__) || defined(__powerpc__)
-// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__loongarch64)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__riscv)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__s390x__)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__wasm__)
-// quants.c
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#endif

ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp

@@ -1,94 +0,0 @@
-#include "ggml-backend-impl.h"
-
-#if defined(__aarch64__)
-
-#if defined(__linux__)
-#include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
-#endif
-
-struct aarch64_features {
-    // has_neon not needed, aarch64 has NEON guaranteed
-    bool has_dotprod     = false;
-    bool has_fp16_va     = false;
-    bool has_sve         = false;
-    bool has_sve2        = false;
-    bool has_i8mm        = false;
-    bool has_sme         = false;
-
-    aarch64_features() {
-#if defined(__linux__)
-        uint32_t hwcap = getauxval(AT_HWCAP);
-        uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-        has_fp16_va = !!(hwcap & HWCAP_FPHP);
-        has_sve     = !!(hwcap & HWCAP_SVE);
-        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
-        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-        has_sme     = !!(hwcap2 & HWCAP2_SME);
-#elif defined(__APPLE__)
-        int oldp = 0;
-        size_t size = sizeof(oldp);
-
-        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
-            has_dotprod = static_cast<bool>(oldp);
-        }
-
-        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
-            has_i8mm = static_cast<bool>(oldp);
-        }
-
-        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
-            has_sme = static_cast<bool>(oldp);
-        }
-
-        // Apple apparently does not implement SVE yet
-#endif
-    }
-};
-
-static int ggml_backend_cpu_aarch64_score() {
-    int score = 1;
-    aarch64_features af;
-
-#ifdef GGML_USE_DOTPROD
-    if (!af.has_dotprod) { return 0; }
-    score += 1<<1;
-#endif
-#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
-    if (!af.has_fp16_va) { return 0; }
-    score += 1<<2;
-#endif
-#ifdef GGML_USE_SVE
-    if (!af.has_sve) { return 0; }
-    score += 1<<3;
-#endif
-#ifdef GGML_USE_MATMUL_INT8
-    if (!af.has_i8mm) { return 0; }
-    score += 1<<4;
-#endif
-#ifdef GGML_USE_SVE2
-    if (!af.has_sve2) { return 0; }
-    score += 1<<5;
-#endif
-#ifdef GGML_USE_SME
-    if (!af.has_sme) { return 0; }
-    score += 1<<6;
-#endif
-
-    return score;
-}
-
-GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
-
-# endif // defined(__aarch64__)

ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp

@@ -1,82 +0,0 @@
-# include "ggml-backend-impl.h"
-
-#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
-
-#if defined(__linux__)
-#include <sys/auxv.h>
-#endif
-
-#include <string>
-
-struct powerpc_features {
-    std::string platform = "";
-    int power_version    = -1;
-
-    bool has_vsx         = false;
-
-    powerpc_features() {
-#if defined(__linux__)
-        unsigned long auxval = getauxval(AT_PLATFORM);
-        if (auxval) {
-            platform = std::string(reinterpret_cast<const char*>(auxval));
-            // TBD: Do systems exist that return this in uppercase?
-            if (platform.substr(0, 5) == "power") {
-                // Extractt a numeric suffix, if one exists
-                int vpos = -1;
-                for (int i = platform.length() - 1; i >= 0; i--) {
-                    if (std::isdigit(platform[i])) {
-                        vpos = i;
-                    } else {
-                        break;
-                    }
-                }
-                if (vpos > -1) {
-                    power_version = std::stoi(platform.substr(vpos));
-                }
-            }
-        }
-#endif
-        if (power_version >= 9) {
-            has_vsx = true;
-        }
-    }
-};
-
-static int ggml_backend_cpu_powerpc_score() {
-    int score = 1;
-    powerpc_features pf;
-
-// Platform scores
-#if defined(GGML_USE_POWER7)
-    if (pf.power_version < 7) { return 0; }
-    score += 1<<1;
-#endif
-#if defined(GGML_USE_POWER8)
-    if (pf.power_version < 8) { return 0; }
-    score += 1<<2;
-#endif
-#if defined(GGML_USE_POWER9)
-    if (pf.power_version < 9) { return 0; }
-    score += 1<<3;
-#endif
-#if defined(GGML_USE_POWER10)
-    if (pf.power_version < 10) { return 0; }
-    score += 1<<4;
-#endif
-#if defined(GGML_USE_POWER11)
-    if (pf.power_version < 11) { return 0; }
-    score += 1<<5;
-#endif
-
-// Feature scores
-#if defined(GGML_USE_VSX)
-    if (!pf.has_vsx) { return 0; }
-    score += 1<<6;
-#endif
-
-    return score;
-}
-
-GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
-
-#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -1,397 +0,0 @@
-#define GGML_COMMON_IMPL_CPP
-#define GGML_COMMON_DECL_CPP
-#include "ggml-common.h"
-#include "ggml-backend-impl.h"
-
-#include "ggml-impl.h"
-#include "ggml-cpu.h"
-#include "ggml-cpu-impl.h"
-#include "simd-mappings.h"
-#include "traits.h"
-
-#include <cmath>
-#include <cstring>
-#include <cassert>
-#include <cstdlib> // for qsort
-#include <cstdio>  // for GGML_ASSERT
-
-#define GGML_CPU_CLANG_WORKAROUND
-#include "../../repack.h"
-
-#if defined(__GNUC__)
-#pragma GCC diagnostic ignored "-Woverlength-strings"
-#endif
-
-#define UNUSED GGML_UNUSED
-
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
-
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
-
-#if defined __riscv_v
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-            for (int l = 0; l < nb; l++) {
-                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
-                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
-                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
-                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
-                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
-                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
-                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
-                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
-                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
-
-                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
-                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                // vector version needs Zvfhmin extension
-                const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                const float b_scales[8] = {
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
-                };
-                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
-                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
-            }
-            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
-        }
-        return;
-    }
-
-#endif
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
-}
-
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
-
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
-
-#if defined __riscv_v
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                for (int l = 0; l < nb; l++) {
-                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                    // vector version needs Zvfhmin extension
-                    const float a_scales[4] = {
-                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
-                    };
-                    const float b_scales[8] = {
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
-                    };
-                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-
-                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
-                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
-                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
-                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l0;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l0 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
-                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
-                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
-                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
-                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l1;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l1 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
-                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
-                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
-                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
-                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l2;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l2 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
-                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
-                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
-                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
-                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l3;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l3 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
-                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
-                    }
-                }
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
-            }
-        }
-
-        return;
-    }
-
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
-}

ggml/src/ggml-cpu/common.h

@@ -1,10 +1,9 @@
 #pragma once
 
 #include "ggml.h"
-#include "traits.h"
+#include "ggml-cpu-traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
-#include "simd-mappings.h"
 
 #ifdef __cplusplus
 
@@ -13,11 +12,11 @@
 // convenience functions/macros for use in template calls
 // note: these won't be required after the 'traits' lookup table is used.
 static inline ggml_fp16_t f32_to_f16(float x) {
-    return GGML_CPU_FP32_TO_FP16(x);
+    return GGML_FP32_TO_FP16(x);
 }
 
 static inline float f16_to_f32(ggml_fp16_t x) {
-    return GGML_CPU_FP16_TO_FP32(x);
+    return GGML_FP16_TO_FP32(x);
 }
 
 static inline ggml_bf16_t f32_to_bf16(float x) {

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "ggml-cpu-traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

ggml/src/ggml-cpu/ggml-cpu-hbm.cpp

@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "hbm.h"
+#include "ggml-cpu-hbm.h"
 
 // buffer type HBM
 

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -62,17 +62,11 @@ struct ggml_compute_params {
 #if defined(__s390x__) && defined(__VEC__)
 #ifndef __VXE__
 #define __VXE__
-#endif  // __VXE__
+#endif
 #ifndef __VXE2__
 #define __VXE2__
-#endif  // __VXE2__
-#endif  // __s390x__ && __VEC__
-
-#if defined(__s390x__) && defined(GGML_NNPA)
-#ifndef __NNPA__
-#define __NNPA__
-#endif  // __NNPA__
-#endif  // __s390x__ && GGML_NNPA
+#endif
+#endif
 
 #if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
@@ -377,7 +371,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
 #endif
 
-typedef signed   char char8x16_t  __attribute__((vector_size(16)));
+typedef signed char char8x16_t __attribute__((vector_size(16)));
 typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
 
 typedef int8_t  int8x16_t __attribute__((vector_size(16)));
@@ -388,10 +382,10 @@ typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
 typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
 typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
 
-typedef float  float32x4_t  __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute__((vector_size(16)));
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute((vector_size(16)));
 
-typedef signed   long long long64x2_t  __attribute__((vector_size(16)));
+typedef signed long long long64x2_t __attribute((vector_size(16)));
 typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
 
 typedef struct ggml_uint8x16x2_t {
@@ -509,9 +503,6 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
-void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
-int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
-
 #ifdef __cplusplus
 }
 #endif