cuda : add ELU support (#14657)

ggml-ci

.devops/intel.Dockerfile

@@ -49,19 +49,23 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-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
+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
 
+ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.devops/musa.Dockerfile

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc3.1.1
+ARG MUSA_VERSION=rc4.0.1
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
 
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
@@ -21,21 +21,14 @@ RUN apt-get update && \
     libcurl4-openssl-dev \
     libgomp1
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
-
-RUN pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt
-
 WORKDIR /app
 
 COPY . .
 
-# Use the default MUSA archs if not specified
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/tools.sh

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

.editorconfig

@@ -48,3 +48,7 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[vendor/miniaudio/miniaudio.h]
+trim_trailing_whitespace = unset
+insert_final_newline = unset

.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, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         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, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true

.github/labeler.yml

@@ -1,10 +1,4 @@
 # 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:
@@ -86,3 +80,15 @@ 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

@@ -0,0 +1,51 @@
+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

@@ -26,12 +26,12 @@ jobs:
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  libcurl4-openssl-dev:riscv64
+                  g++-14-riscv64-linux-gnu
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
                          -DLLAMA_BUILD_TOOLS=ON \
@@ -72,12 +72,12 @@ jobs:
                   glslc \
                   gcc-14-riscv64-linux-gnu \
                   g++-14-riscv64-linux-gnu \
-                  libvulkan-dev:riscv64 \
-                  libcurl4-openssl-dev:riscv64
+                  libvulkan-dev:riscv64
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -118,12 +118,12 @@ jobs:
                   build-essential \
                   glslc \
                   crossbuild-essential-arm64 \
-                  libvulkan-dev:arm64 \
-                  libcurl4-openssl-dev:arm64
+                  libvulkan-dev:arm64
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -163,12 +163,12 @@ jobs:
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu \
-                  libcurl4-openssl-dev:ppc64el
+                  g++-14-powerpc64le-linux-gnu
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
                          -DLLAMA_BUILD_TOOLS=ON \
@@ -209,12 +209,12 @@ jobs:
                   glslc \
                   gcc-14-powerpc64le-linux-gnu \
                   g++-14-powerpc64le-linux-gnu \
-                  libvulkan-dev:ppc64el \
-                  libcurl4-openssl-dev:ppc64el
+                  libvulkan-dev:ppc64el
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -231,3 +231,116 @@ 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

@@ -5,10 +5,43 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.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',
+      '.github/workflows/build-cmake-pkg.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', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    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'
+    ]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -51,7 +84,8 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_METAL_EMBED_LIBRARY=OFF \
+            -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -306,8 +340,9 @@ 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
+          ctest -L main --verbose --timeout 4200
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
@@ -351,7 +386,7 @@ jobs:
 
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
-    container: mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 
     steps:
       - name: Clone
@@ -477,6 +512,9 @@ 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
 
@@ -627,7 +665,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       fail-fast: false
@@ -677,28 +715,31 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
         include:
-          - 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: '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: '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'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            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'
           - 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
@@ -712,12 +753,6 @@ 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' }}
@@ -733,9 +768,9 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ 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-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -768,6 +803,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -777,6 +814,7 @@ 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
@@ -787,7 +825,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
+        if: ${{ matrix.arch == 'x64' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -839,12 +877,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-2022-cmake-cuda:
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -878,7 +916,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^
@@ -892,7 +930,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -926,7 +964,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -1,4 +1,4 @@
-name: Create Release
+name: Release
 
 on:
   workflow_dispatch: # allows manual triggering
@@ -49,7 +49,8 @@ jobs:
         run: |
           sysctl -a
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -103,7 +104,8 @@ 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_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
@@ -131,8 +133,9 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
+          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
+          # - build: 'arm64'
+          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -159,6 +162,11 @@ 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 $(nproc)
@@ -207,6 +215,11 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$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 $(nproc)
@@ -227,35 +240,17 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
           name: llama-bin-ubuntu-vulkan-x64.zip
 
-  windows:
-    runs-on: windows-latest
-
-    env:
-      OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.309.0
+  windows-cpu:
+    runs-on: windows-2025
 
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -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 -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: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
-          - build: 'cpu-arm64'
-            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
-          - build: 'opencl-adreno-arm64'
-            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'
+          - arch: 'x64'
+          - arch: 'arm64'
 
     steps:
       - name: Clone
-        id: checkout
         uses: actions/checkout@v4
         with:
           fetch-depth: 0
@@ -263,28 +258,87 @@ jobs:
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-${{ matrix.build }}
+          key: windows-latest-cmake-cpu-${{ matrix.arch }}
           variant: ccache
           evict-old-files: 1d
 
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas-x64' }}
+      - name: Install Ninja
         run: |
-          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
-          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/openblas
-          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
-          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+          choco install ninja
+
+      - name: libCURL
+        id: get_libcurl
+        uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
+
+      - name: Build
+        shell: cmd
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          cmake -S . -B build -G "Ninja Multi-Config" ^
+            -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
+            -DGGML_OPENMP=ON ^
+            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        env:
+          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.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
+          name: llama-bin-win-cpu-${{ matrix.arch }}.zip
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      OPENBLAS_VERSION: 0.3.23
+      VULKAN_VERSION: 1.4.313.2
+
+    strategy:
+      matrix:
+        include:
+          - backend: 'vulkan'
+            arch: 'x64'
+            defines: '-DGGML_VULKAN=ON'
+            target: 'ggml-vulkan'
+          - backend: 'opencl-adreno'
+            arch: 'arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+            target: 'ggml-opencl'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
+          variant: ccache
+          evict-old-files: 1d
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
+        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-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -296,7 +350,7 @@ jobs:
 
       - name: Install OpenCL Headers and Libs
         id: install_opencl
-        if: ${{ matrix.build == 'opencl-adreno-arm64' }}
+        if: ${{ matrix.backend == 'opencl-adreno' && matrix.arch == 'arm64' }}
         run: |
           git clone https://github.com/KhronosGroup/OpenCL-Headers
           cd OpenCL-Headers
@@ -314,60 +368,34 @@ jobs:
             -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
           cmake --build build-arm64-release --target install --config release
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -S . -B build ${{ matrix.defines }} `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Add libopenblas.dll
-        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
-          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake -S . -B build ${{ matrix.defines }} -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_BACKEND_DL=ON -DLLAMA_CURL=OFF
+          cmake --build build --config Release --target ${{ matrix.target }}
 
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
+          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
+          path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
+          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Install ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -386,45 +414,30 @@ jobs:
         run: |
           choco install ninja
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
         shell: cmd
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
-            ${{ env.CMAKE_ARGS }}
+            -DLLAMA_CURL=OFF
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
         run: |
@@ -432,16 +445,16 @@ jobs:
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         uses: actions/upload-artifact@v4
         with:
-          path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -451,12 +464,11 @@ jobs:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -469,15 +481,18 @@ jobs:
         run:  |
           scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
-      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
       - name: Build
         id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
+        shell: cmd
+        run: |
+          call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+          cmake -G "Ninja" -B build ^
+            -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx ^
+            -DCMAKE_BUILD_TYPE=Release ^
+            -DGGML_BACKEND_DL=ON -DBUILD_SHARED_LIBS=ON ^
+            -DGGML_CPU=OFF -DGGML_SYCL=ON ^
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-sycl -j
 
       - name: Build the release package
         id: pack_artifacts
@@ -502,27 +517,27 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
 
           echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+          7z a llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Clone rocWMMA repository
         id: clone_rocwmma
@@ -532,7 +547,7 @@ jobs:
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-hip-release
+          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
           evict-old-files: 1d
 
       - name: Install
@@ -550,50 +565,39 @@ jobs:
         run: |
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+            -DLLAMA_CURL=OFF
+          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
 
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -655,14 +659,16 @@ jobs:
     runs-on: ubuntu-latest
 
     needs:
-      - ubuntu-22-cpu
-      - ubuntu-22-vulkan
       - windows
+      - windows-cpu
       - windows-cuda
       - windows-sycl
       - windows-hip
+      - ubuntu-22-cpu
+      - ubuntu-22-vulkan
       - macOS-arm64
       - macOS-x64
+      - ios-xcode-build
 
     steps:
       - name: Clone
@@ -680,10 +686,43 @@ jobs:
         uses: actions/download-artifact@v4
         with:
           path: ./artifact
+          merge-multiple: true
 
       - name: Move artifacts
         id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+        run: |
+          mkdir -p release
+
+          echo "Adding CPU backend files to existing zips..."
+          for arch in x64 arm64; do
+            cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
+            temp_dir=$(mktemp -d)
+            echo "Extracting CPU backend for $arch..."
+            unzip "$cpu_zip" -d "$temp_dir"
+
+            echo "Adding CPU files to $arch zips..."
+            for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
+              if [[ "$target_zip" == "$cpu_zip" ]]; then
+                continue
+              fi
+              echo "Adding CPU backend to $(basename "$target_zip")"
+              realpath_target_zip=$(realpath "$target_zip")
+              (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
+            done
+
+            rm -rf "$temp_dir"
+          done
+
+          echo "Renaming and moving zips to release..."
+          for zip_file in artifact/llama-bin-win-*.zip; do
+            base_name=$(basename "$zip_file" .zip)
+            zip_name="llama-${{ steps.tag.outputs.name }}-${base_name#llama-}.zip"
+            echo "Moving $zip_file to release/$zip_name"
+            mv "$zip_file" "release/$zip_name"
+          done
+
+          echo "Moving other artifacts..."
+          mv -v artifact/*.zip release
 
       - name: Create release
         id: create_release
@@ -702,7 +741,7 @@ jobs:
             const path = require('path');
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
+            for (let file of await fs.readdirSync('./release')) {
               if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
@@ -710,7 +749,7 @@ jobs:
                   repo: context.repo.repo,
                   release_id: release_id,
                   name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                  data: await fs.readFileSync(`./release/${file}`)
                 });
               }
             }

.github/workflows/server.yml

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

.github/workflows/update-ops-docs.yml

@@ -0,0 +1,40 @@
+name: Update Operations Documentation
+
+on:
+    push:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+    pull_request:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+
+jobs:
+    update-ops-docs:
+        runs-on: ubuntu-latest
+
+        steps:
+        - name: Checkout repository
+          uses: actions/checkout@v4
+
+        - name: Set up Python
+          uses: actions/setup-python@v5
+          with:
+              python-version: '3.x'
+
+        - name: Generate operations documentation to temporary file
+          run: |
+              mkdir -p /tmp/ops_check
+              ./scripts/create_ops_docs.py /tmp/ops_check/ops.md
+
+        - name: Check if docs/ops.md matches generated version
+          run: |
+              if ! diff -q docs/ops.md /tmp/ops_check/ops.md; then
+                  echo "Operations documentation (docs/ops.md) is not up to date with the backend CSV files."
+                  echo "To fix: run ./scripts/create_ops_docs.py and commit the updated docs/ops.md along with your changes"
+                  echo "Differences found:"
+                  diff docs/ops.md /tmp/ops_check/ops.md || true
+                  exit 1
+              fi
+              echo "Operations documentation is up to date."

.github/workflows/winget.yml

@@ -0,0 +1,42 @@
+name: Update Winget Package
+
+on:
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    - cron: '28 5 * * *' # Update every day at 5:28 UTC
+
+jobs:
+  update:
+    name: Update Winget Package
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Install cargo binstall
+        uses: cargo-bins/cargo-binstall@268643a6b5ea099f5718ee5cd3ff7dc89a5eb49b
+
+      - name: Install komac
+        run: |
+          cargo binstall komac@2.11.2 -y
+
+      - name: Find latest release
+        id: find_latest_release
+        uses: actions/github-script@v6
+        with:
+          script: |
+            const { data: releases } = await github.rest.repos.listReleases({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+            });
+            console.log("Latest release:", releases[0].tag_name);
+            return releases[0].tag_name;
+
+      - name: Update manifest
+        env:
+          VERSION: ${{ steps.find_latest_release.outputs.result }}
+        run: |
+          echo "Updating manifest..."
+          komac update --version ${{ env.VERSION }} \
+            --urls "https://github.com/ggml-org/llama.cpp/releases/download/${{ env.VERSION }}/llama-${{ env.VERSION }}-bin-win-vulkan-x64.zip" \
+            --token ${{ secrets.WINGET_GITHUB_TOKEN }} \
+            --submit \
+            ggml.llamacpp

.gitmodules

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

CMakeLists.txt

@@ -89,6 +89,14 @@ 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})
@@ -112,7 +120,6 @@ 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)
@@ -155,10 +162,17 @@ 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
 #
@@ -199,10 +213,6 @@ 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")

CMakePresets.json

@@ -55,6 +55,17 @@
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
         }
     },
+    {
+        "name": "x64-linux-gcc", "hidden": true,
+        "cacheVariables": {
+            "CMAKE_C_COMPILER": "gcc",
+            "CMAKE_CXX_COMPILER": "g++"
+        }
+    },
+    { "name": "x64-linux-gcc-debug", "inherits": [ "base", "x64-linux-gcc", "debug" ] },
+    { "name": "x64-linux-gcc-release", "inherits": [ "base", "x64-linux-gcc", "release" ] },
+    { "name": "x64-linux-gcc-reldbg", "inherits": [ "base", "x64-linux-gcc", "reldbg" ] },
+    { "name": "x64-linux-gcc+static-release", "inherits": [ "base", "x64-linux-gcc", "release", "static" ] },
 
     { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },

Makefile

@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
 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/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/repack.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

@@ -3,11 +3,12 @@
 ![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) / [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)
+[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
 
-Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
+LLM inference in C/C++
 
 ## Recent API changes
 
@@ -16,11 +17,9 @@ 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
+- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
+- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - 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
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -28,6 +27,30 @@ 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
@@ -37,7 +60,7 @@ range of hardware - locally and in the cloud.
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
-- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
@@ -110,6 +133,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
+- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 
 #### Multimodal
 
@@ -130,6 +154,7 @@ 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)
@@ -229,6 +254,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
+
 ## Supported backends
 
 | Backend | Target devices |
@@ -237,7 +263,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
 | [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](docs/build.md#musa) | Moore Threads MTT GPU |
+| [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
@@ -245,16 +271,6 @@ 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`:
@@ -262,7 +278,11 @@ 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]`.
+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
+```
 
 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/`.
 
@@ -580,3 +600,4 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
 - [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
 - [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
+- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain

build-xcframework.sh

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

ci/README.md

@@ -54,7 +54,7 @@ docker run --privileged -it \
     -v $HOME/llama.cpp/ci-cache:/ci-cache \
     -v $HOME/llama.cpp/ci-results:/ci-results \
     -v $PWD:/ws -w /ws \
-    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+    mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 ```
 
 Inside the container, execute the following commands:

ci/run.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # sample usage:
 #
@@ -39,14 +39,27 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 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 -DCMAKE_CUDA_ARCHITECTURES=native"
+    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
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -766,7 +779,7 @@ function gg_run_rerank_tiny {
     model_f16="${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?</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
+    (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
 
     # sample output
     # rerank score 0:    0.029

common/CMakeLists.txt

@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,36 +18,26 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
         endif()
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
+        # For build-info.cpp below
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${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()
 
-# 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" ${GIT_INDEX}
-    VERBATIM
-)
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+
 set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
+add_library(${TARGET} OBJECT ${OUTPUT_FILE})
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
@@ -58,19 +48,20 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
+    chat-parser.cpp
+    chat-parser.h
     chat.cpp
     chat.h
     common.cpp
     common.h
     console.cpp
     console.h
+    json-partial.cpp
+    json-partial.h
     json-schema-to-grammar.cpp
-    json.hpp
     llguidance.cpp
     log.cpp
     log.h
-    minja/chat-template.hpp
-    minja/minja.hpp
     ngram-cache.cpp
     ngram-cache.h
     regex-partial.cpp
@@ -95,8 +86,7 @@ if (LLAMA_CURL)
     endif()
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
 endif ()
 
 if (LLAMA_LLGUIDANCE)
@@ -121,13 +111,13 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.20 (+ fix to build on GCC 15):
-        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
+        # v1.0.1:
+        GIT_TAG d795912fedc7d393de740177ea9ea761e7905774
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE
         CONFIGURE_COMMAND ""
-        BUILD_COMMAND cargo build --release
+        BUILD_COMMAND cargo build --release --package llguidance
         INSTALL_COMMAND ""
         BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
         UPDATE_COMMAND ""
@@ -143,7 +133,7 @@ if (LLAMA_LLGUIDANCE)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
 endif ()
 
-target_include_directories(${TARGET} PUBLIC .)
+target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
 

common/arg.cpp

@@ -1,10 +1,11 @@
-#include "gguf.h" // for reading GGUF splits
 #include "arg.h"
 
+#include "chat.h"
 #include "common.h"
+#include "gguf.h" // for reading GGUF splits
+#include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
-#include "chat.h"
 
 // fix problem with std::min and std::max
 #if defined(_WIN32)
@@ -15,6 +16,9 @@
 #include <windows.h>
 #endif
 
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
 #include <climits>
 #include <cstdarg>
@@ -34,12 +38,10 @@
 #include <future>
 #endif
 
-#include "json-schema-to-grammar.h"
-
 using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_SERVER,
 };
 
@@ -242,7 +244,56 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
 }
 
 // download one single file from remote URL to local path
-static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) {
+static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token, bool offline) {
+    // Check if the file already exists locally
+    auto file_exists = std::filesystem::exists(path);
+
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
+    std::string etag;
+    std::string last_modified;
+
+    if (file_exists) {
+        if (offline) {
+            LOG_INF("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
+            return true; // skip verification/downloading
+        }
+        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+        std::ifstream metadata_in(metadata_path);
+        if (metadata_in.good()) {
+            try {
+                metadata_in >> metadata;
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
+                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                    etag = metadata.at("etag");
+                }
+                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                    last_modified = metadata.at("lastModified");
+                }
+            } catch (const nlohmann::json::exception & e) {
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+            }
+        }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
+    } else {
+        if (offline) {
+            LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
+            return false;
+        }
+        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
+    }
+
+    // Send a HEAD request to retrieve the etag and last-modified headers
+    struct common_load_model_from_url_headers {
+        std::string etag;
+        std::string last_modified;
+    };
+
+    common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist
+
     // Initialize libcurl
     curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
     curl_slist_ptr http_headers;
@@ -269,91 +320,47 @@ static bool common_download_file_single(const std::string & url, const std::stri
     curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
 #endif
 
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
+    typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
+    auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
+        common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
 
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
-    std::string etag;
-    std::string last_modified;
+        static std::regex header_regex("([^:]+): (.*)\r\n");
+        static std::regex etag_regex("ETag", std::regex_constants::icase);
+        static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
 
-    if (file_exists) {
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+        std::string header(buffer, n_items);
+        std::smatch match;
+        if (std::regex_match(header, match, header_regex)) {
+            const std::string & key = match[1];
+            const std::string & value = match[2];
+            if (std::regex_match(key, match, etag_regex)) {
+                headers->etag = value;
+            } else if (std::regex_match(key, match, last_modified_regex)) {
+                headers->last_modified = value;
             }
         }
-        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
-    } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
-    }
-
-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
+        return n_items;
     };
 
-    common_load_model_from_url_headers headers;
-    bool head_request_ok = false;
-    bool should_download = !file_exists; // by default, we should download if the file does not exist
+    curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
+    curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-    // get ETag to see if the remote file has changed
-    {
-        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
+    // we only allow retrying once for HEAD requests
+    // this is for the use case of using running offline (no internet), retrying can be annoying
+    bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
+    if (!was_perform_successful) {
+        head_request_ok = false;
+    }
 
-            static std::regex header_regex("([^:]+): (.*)\r\n");
-            static std::regex etag_regex("ETag", std::regex_constants::icase);
-            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-            std::string header(buffer, n_items);
-            std::smatch match;
-            if (std::regex_match(header, match, header_regex)) {
-                const std::string & key = match[1];
-                const std::string & value = match[2];
-                if (std::regex_match(key, match, etag_regex)) {
-                    headers->etag = value;
-                } else if (std::regex_match(key, match, last_modified_regex)) {
-                    headers->last_modified = value;
-                }
-            }
-            return n_items;
-        };
-
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-        // we only allow retrying once for HEAD requests
-        // this is for the use case of using running offline (no internet), retrying can be annoying
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
-        if (!was_perform_successful) {
-            head_request_ok = false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code == 200) {
-            head_request_ok = true;
-        } else {
-            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-            head_request_ok = false;
-        }
+    long http_code = 0;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+    if (http_code == 200) {
+        head_request_ok = true;
+    } else {
+        LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        head_request_ok = false;
     }
 
     // if head_request_ok is false, we don't have the etag or last-modified headers
@@ -460,12 +467,12 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
 // download multiple files from remote URLs to local paths
 // the input is a vector of pairs <url, path>
-static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token) {
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token, bool offline) {
     // Prepare download in parallel
     std::vector<std::future<bool>> futures_download;
     for (auto const & item : urls) {
-        futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair<std::string, std::string> & it) -> bool {
-            return common_download_file_single(it.first, it.second, bearer_token);
+        futures_download.push_back(std::async(std::launch::async, [bearer_token, offline](const std::pair<std::string, std::string> & it) -> bool {
+            return common_download_file_single(it.first, it.second, bearer_token, offline);
         }, item));
     }
 
@@ -481,14 +488,15 @@ static bool common_download_file_multiple(const std::vector<std::pair<std::strin
 
 static bool common_download_model(
         const common_params_model & model,
-        const std::string & bearer_token) {
+        const std::string & bearer_token,
+        bool offline) {
     // Basic validation of the model.url
     if (model.url.empty()) {
         LOG_ERR("%s: invalid model url\n", __func__);
         return false;
     }
 
-    if (!common_download_file_single(model.url, model.path, bearer_token)) {
+    if (!common_download_file_single(model.url, model.path, bearer_token, offline)) {
         return false;
     }
 
@@ -547,7 +555,7 @@ static bool common_download_model(
         }
 
         // Download in parallel
-        common_download_file_multiple(urls, bearer_token);
+        common_download_file_multiple(urls, bearer_token, offline);
     }
 
     return true;
@@ -608,7 +616,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
  *
  * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
  */
-static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token) {
+static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token, bool offline) {
     auto parts = string_split<std::string>(hf_repo_with_tag, ':');
     std::string tag = parts.size() > 1 ? parts.back() : "latest";
     std::string hf_repo = parts[0];
@@ -638,20 +646,25 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
     long res_code = 0;
     std::string res_str;
     bool use_cache = false;
-    try {
-        auto res = common_remote_get_content(url, params);
-        res_code = res.first;
-        res_str = std::string(res.second.data(), res.second.size());
-    } catch (const std::exception & e) {
-        LOG_WRN("error: failed to get manifest: %s\n", e.what());
-        LOG_WRN("try reading from cache\n");
-        // try to read from cache
+    if (!offline) {
         try {
+            auto res = common_remote_get_content(url, params);
+            res_code = res.first;
+            res_str = std::string(res.second.data(), res.second.size());
+        } catch (const std::exception & e) {
+            LOG_WRN("error: failed to get manifest at %s: %s\n", url.c_str(), e.what());
+        }
+    }
+    if (res_code == 0) {
+        if (std::filesystem::exists(cached_response_path)) {
+            LOG_WRN("trying to read manifest from cache: %s\n", cached_response_path.c_str());
             res_str = read_file(cached_response_path);
             res_code = 200;
             use_cache = true;
-        } catch (const std::exception & e) {
-            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        } else {
+            throw std::runtime_error(
+                offline ? "error: failed to get manifest (offline mode)"
+                : "error: failed to get manifest (check your internet connection)");
         }
     }
     std::string ggufFile;
@@ -698,24 +711,25 @@ bool common_has_curl() {
     return false;
 }
 
-static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
+static bool common_download_file_single(const std::string &, const std::string &, const std::string &, bool) {
     LOG_ERR("error: built without CURL, cannot download model from internet\n");
     return false;
 }
 
-static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &) {
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &, bool) {
     LOG_ERR("error: built without CURL, cannot download model from the internet\n");
     return false;
 }
 
 static bool common_download_model(
         const common_params_model &,
-        const std::string &) {
+        const std::string &,
+        bool) {
     LOG_ERR("error: built without CURL, cannot download model from the internet\n");
     return false;
 }
 
-static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) {
+static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool) {
     LOG_ERR("error: built without CURL, cannot download model from the internet\n");
     return {};
 }
@@ -742,7 +756,8 @@ struct handle_model_result {
 static handle_model_result common_params_handle_model(
         struct common_params_model & model,
         const std::string & bearer_token,
-        const std::string & model_path_default) {
+        const std::string & model_path_default,
+        bool offline) {
     handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
@@ -750,7 +765,7 @@ static handle_model_result common_params_handle_model(
             // short-hand to avoid specifying --hf-file -> default it to --model
             if (model.hf_file.empty()) {
                 if (model.path.empty()) {
-                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token);
+                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token, offline);
                     if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
                         exit(1); // built without CURL, error message already printed
                     }
@@ -791,7 +806,7 @@ static handle_model_result common_params_handle_model(
 
     // then, download it if needed
     if (!model.url.empty()) {
-        bool ok = common_download_model(model, bearer_token);
+        bool ok = common_download_model(model, bearer_token, offline);
         if (!ok) {
             LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
             exit(1);
@@ -934,7 +949,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
 
     // handle model and download
     {
-        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH);
+        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH, params.offline);
         if (params.no_mmproj) {
             params.mmproj = {};
         } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
@@ -944,12 +959,12 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         // only download mmproj if the current example is using it
         for (auto & ex : mmproj_examples) {
             if (ctx_arg.ex == ex) {
-                common_params_handle_model(params.mmproj,    params.hf_token, "");
+                common_params_handle_model(params.mmproj,    params.hf_token, "", params.offline);
                 break;
             }
         }
-        common_params_handle_model(params.speculative.model, params.hf_token, "");
-        common_params_handle_model(params.vocoder.model,     params.hf_token, "");
+        common_params_handle_model(params.speculative.model, params.hf_token, "", params.offline);
+        common_params_handle_model(params.vocoder.model,     params.hf_token, "", params.offline);
     }
 
     if (params.escape) {
@@ -973,10 +988,6 @@ 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",
@@ -1333,9 +1344,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ));
     add_opt(common_arg(
         {"--prio"}, "N",
-        string_format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority),
+        string_format("set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: %d)\n", params.cpuparams.priority),
         [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
+            if (prio < GGML_SCHED_PRIO_LOW || prio > GGML_SCHED_PRIO_REALTIME) {
                 throw std::invalid_argument("invalid value");
             }
             params.cpuparams.priority = (enum ggml_sched_priority) prio;
@@ -1445,6 +1456,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_keep = value;
         }
     ));
+    add_opt(common_arg(
+        {"--swa-full"},
+        string_format("use full-size SWA cache (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)", params.swa_full ? "true" : "false"),
+        [](common_params & params) {
+            params.swa_full = true;
+        }
+    ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1670,7 +1689,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2057,13 +2076,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.grp_attn_w = value;
         }
     ).set_env("LLAMA_ARG_GRP_ATTN_W").set_examples({LLAMA_EXAMPLE_MAIN}));
-    add_opt(common_arg(
-        {"-dkvc", "--dump-kv-cache"},
-        "verbose print of the KV cache",
-        [](common_params & params) {
-            params.dump_kv_cache = true;
-        }
-    ));
     add_opt(common_arg(
         {"-nkvo", "--no-kv-offload"},
         "disable KV offload",
@@ -2232,12 +2244,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
-        {"--image"}, "FILE",
-        "path to an image file. use with multimodal models. Specify multiple times for batching",
+        {"--image", "--audio"}, "FILE",
+        "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",
         [](common_params & params, const std::string & value) {
             params.image.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples({LLAMA_EXAMPLE_MTMD}));
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
@@ -2694,6 +2706,13 @@ 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()),
@@ -2715,6 +2734,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.public_path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
+    add_opt(common_arg(
+        {"--api-prefix"}, "PREFIX",
+        string_format("prefix path the server serves from, without the trailing slash (default: %s)", params.api_prefix.c_str()),
+        [](common_params & params, const std::string & value) {
+            params.api_prefix = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
     add_opt(common_arg(
         {"--no-webui"},
         string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
@@ -2731,9 +2757,10 @@ 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)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
         [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -2774,6 +2801,16 @@ 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),
@@ -2847,15 +2884,25 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
-        "reasoning format (default: deepseek; allowed values: deepseek, none)\n"
-        "controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).\n"
-        "only supported for non-streamed responses",
+        "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
+        "- none: leaves thoughts unparsed in `message.content`\n"
+        "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
+        "(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 { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
+    add_opt(common_arg(
+        {"--reasoning-budget"}, "N",
+        "controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)",
+        [](common_params & params, int value) {
+            if (value != 0 && value != -1) { throw std::invalid_argument("invalid value"); }
+            params.reasoning_budget = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK_BUDGET"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
@@ -2867,7 +2914,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(
@@ -2954,7 +3001,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             /**/ if (value == "jsonl") { params.batched_bench_output_jsonl = true; }
             else if (value == "md") { params.batched_bench_output_jsonl = false; }
-            else { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(
@@ -2986,6 +3033,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_verbosity_thold(INT_MAX);
         }
     ));
+    add_opt(common_arg(
+        {"--offline"},
+        "Offline mode: forces use of cache, prevents network access",
+        [](common_params & params) {
+            params.offline = true;
+        }
+    ).set_env("LLAMA_OFFLINE"));
     add_opt(common_arg(
         {"-lv", "--verbosity", "--log-verbosity"}, "N",
         "Set the verbosity threshold. Messages with a higher verbosity will be ignored.",
@@ -3180,6 +3234,32 @@ 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 = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -0,0 +1,385 @@
+#include "chat-parser.h"
+#include "common.h"
+#include "log.h"
+#include "regex-partial.h"
+
+#include <optional>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+    : input_(input), is_partial_(is_partial), syntax_(syntax)
+{
+    result_.role = "assistant";
+
+    while (true) {
+        std::string id = std::to_string(std::rand());
+        if (input.find(id) == std::string::npos) {
+            healing_marker_ = id;
+            break;
+        }
+    }
+}
+
+std::string common_chat_msg_parser::str(const common_string_range & rng) const {
+    GGML_ASSERT(rng.begin <= rng.end);
+    return input_.substr(rng.begin, rng.end - rng.begin);
+}
+
+void common_chat_msg_parser::add_content(const std::string &content) {
+    result_.content += content;
+}
+
+void common_chat_msg_parser::add_reasoning_content(const std::string &reasoning_content) {
+    result_.reasoning_content += reasoning_content;
+}
+
+bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::string & id, const std::string & arguments) {
+    if (name.empty()) {
+        return false;
+    }
+
+    common_chat_tool_call tool_call;
+    tool_call.name = name;
+    tool_call.arguments = arguments;
+    tool_call.id = id;
+
+    // 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) {
+    std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
+    std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
+    std::string arguments = tool_call.contains("arguments") ? tool_call.at("arguments") : "";
+    return add_tool_call(name, id, arguments);
+}
+
+bool common_chat_msg_parser::add_tool_calls(const json & arr) {
+    for (const auto & item : arr) {
+        if (!add_tool_call(item)) {
+            return false;
+        }
+    }
+    return true;
+}
+void common_chat_msg_parser::finish() {
+    if (!is_partial_ && pos_ != input_.size()) {
+        throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));
+    }
+}
+
+bool common_chat_msg_parser::consume_spaces() {
+    const auto length = input_.size();
+    auto consumed = false;
+    while (pos_ < length && std::isspace(input_[pos_])) {
+        ++pos_;
+        consumed = true;
+    }
+    return consumed;
+}
+
+bool common_chat_msg_parser::try_consume_literal(const std::string & literal) {
+    auto pos = pos_;
+    for (auto i = 0u; i < literal.size(); ++i) {
+        if (pos >= input_.size()) {
+            return false;
+        }
+        if (input_[pos] != literal[i]) {
+            return false;
+        }
+        ++pos;
+    }
+    pos_ = pos;
+    return true;
+}
+
+std::optional<common_chat_msg_parser::find_regex_result>  common_chat_msg_parser::try_find_literal(const std::string & literal) {
+    auto idx = input_.find(literal, pos_);
+    if (idx != std::string::npos) {
+        find_regex_result res;
+        res.prelude = input_.substr(pos_, idx - pos_);
+        auto end = idx + literal.size();
+        res.groups.emplace_back(common_string_range{idx, end});
+        move_to(end);
+        return res;
+    }
+    if (is_partial_) {
+        idx = string_find_partial_stop(input_, literal);
+        if (idx != std::string::npos && idx >= pos_) {
+            find_regex_result res;
+            res.prelude = input_.substr(pos_, idx - pos_);
+            auto end = input_.size();
+            res.groups.emplace_back(common_string_range{idx, end});
+            move_to(end);
+            return res;
+        }
+    }
+    return std::nullopt;
+}
+
+void common_chat_msg_parser::consume_literal(const std::string & literal) {
+    if (!try_consume_literal(literal)) {
+        throw common_chat_msg_partial_exception(literal);
+    }
+}
+
+bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
+    auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
+        auto stripped_reasoning = string_strip(reasoning);
+        if (stripped_reasoning.empty()) {
+            return;
+        }
+        if (syntax_.reasoning_in_content) {
+            add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "<think>" : start_think);
+            add_content(stripped_reasoning);
+            if (closed) {
+                add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
+            }
+        } else {
+            add_reasoning_content(stripped_reasoning);
+        }
+    };
+    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
+        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
+            if (auto res = try_find_literal(end_think)) {
+                handle_reasoning(res->prelude, /* closed */ true);
+                consume_spaces();
+                return true;
+            }
+            auto rest = consume_rest();
+            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);
+            // }
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string common_chat_msg_parser::consume_rest() {
+    auto rest = input_.substr(pos_);
+    pos_ = input_.size();
+    return rest;
+}
+
+// Tries to find the regex, consumes it (pos right after it) and gives the prelude (right before it) and the groups to the callback.
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from, bool add_prelude_to_content) {
+    auto m = regex.search(input_, from == std::string::npos ? pos_ : from);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
+    pos_ = m.groups[0].end;
+
+    if (add_prelude_to_content) {
+        add_content(prelude);
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    return find_regex_result{prelude, m.groups};
+}
+
+common_chat_msg_parser::find_regex_result common_chat_msg_parser::consume_regex(const common_regex & regex) {
+    if (auto result = try_consume_regex(regex)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception(regex.str());
+}
+
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_consume_regex(const common_regex & regex) {
+    auto m = regex.search(input_, pos_);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    if (m.groups[0].begin != pos_) {
+        // Didn't match at the current position.
+        return std::nullopt;
+    }
+    pos_ = m.groups[0].end;
+
+    return find_regex_result {
+        /* .prelude = */ "",
+        m.groups,
+    };
+}
+
+std::optional<common_json> common_chat_msg_parser::try_consume_json() {
+    auto it = input_.cbegin() + pos_;
+    const auto end = input_.cend();
+    common_json result;
+    if (!common_json_parse(it, end, healing_marker_, result)) {
+        return std::nullopt;
+    }
+    pos_ = std::distance(input_.cbegin(), it);
+    if (result.healing_marker.marker.empty()) {
+        // No healing marker, just return the parsed json
+        return result;
+    }
+    if (!is_partial()) {
+        throw common_chat_msg_partial_exception("JSON");
+    }
+    return result;
+}
+
+common_json common_chat_msg_parser::consume_json() {
+    if (auto result = try_consume_json()) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+common_chat_msg_parser::consume_json_result common_chat_msg_parser::consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    if (auto result = try_consume_json_with_dumped_args(args_paths, content_paths)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parser::try_consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    auto partial = try_consume_json();
+    if (!partial) {
+        return std::nullopt;
+    }
+    auto is_arguments_path = [&](const std::vector<std::string> & path) {
+        return std::find(args_paths.begin(), args_paths.end(), path) != args_paths.end();
+    };
+    auto is_content_path = [&](const std::vector<std::string> & path) {
+        return std::find(content_paths.begin(), content_paths.end(), path) != content_paths.end();
+    };
+
+    if (partial->healing_marker.marker.empty()) {
+        if (args_paths.empty()) {
+            // No arguments to dump, and JSON was parsed fully.
+            return consume_json_result {
+                partial->json,
+                /* .is_partial = */ false,
+            };
+        }
+        if (is_arguments_path({})) {
+            // Entire JSON is the arguments and was parsed fully.
+            return consume_json_result {
+                partial->json.dump(),
+                /* .is_partial = */ false,
+            };
+        }
+    }
+
+    LOG_DBG("Parsed partial JSON: %s (json_healing_marker: %s)\n", partial->json.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+
+    auto found_healing_marker = false;
+    std::vector<std::string> path;
+    std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
+        if (is_arguments_path(path)) {
+            auto arguments = j.dump();
+            if (is_partial() && !partial->healing_marker.marker.empty()) {
+                auto idx = arguments.find(partial->healing_marker.json_dump_marker);
+                if (idx != std::string::npos) {
+                    arguments.resize(idx);
+                    found_healing_marker = true;
+                }
+                if (arguments == "\"") {
+                    // This happens because of completing `:"$magic` after `"arguments"`
+                    arguments = "";
+                }
+            }
+            return arguments;
+        }
+        if (is_content_path(path)) {
+            if (!j.is_string()) {
+                throw std::runtime_error("Content path must be a string");
+            }
+            std::string str = j;
+            auto idx = str.find(partial->healing_marker.marker); // not using json_dump_marker as we're inside a string
+            if (idx != std::string::npos) {
+                str.resize(idx);
+                found_healing_marker = true;
+            }
+            return str;
+        }
+        if (j.is_object()) {
+            auto obj = json::object();
+            for (const auto & p : j.items()) {
+                const auto & key = p.key();
+                const auto & value = p.value();
+                const std::string key_str = key; // NOLINT
+                auto idx = key_str.find(healing_marker_);
+                if (idx != std::string::npos) {
+                    found_healing_marker = true;
+                    break;
+                }
+                path.push_back(key_str);
+                if (value.is_string()) {
+                    const std::string value_str = value;
+                    if (value_str.find(healing_marker_) != std::string::npos) {
+                        found_healing_marker = true;
+                        if (is_content_path(path)) {
+                            if (partial->healing_marker.marker == partial->healing_marker.json_dump_marker) {
+                                // The healing occurred inside the string: good. Otherwise we just ditch the entire key/value pair.
+                                obj[key] = remove_unsupported_healings_and_dump_args(value);
+                            }
+                        }
+                        break;
+                    }
+                    obj[key] = value;
+                } else {
+                    obj[key] = remove_unsupported_healings_and_dump_args(value);
+                }
+                path.pop_back();
+            }
+            return obj;
+        }
+        if (j.is_array()) {
+            auto arr = json::array();
+            for (const auto & value : j) {
+                if (value.is_string()) {
+                    std::string str = value;
+                    auto idx = str.find(healing_marker_);
+                    if (idx != std::string::npos) {
+                        // Don't heal array values that aren't in the arguments.
+                        found_healing_marker = true;
+                        break;
+                    }
+                }
+                arr.push_back(remove_unsupported_healings_and_dump_args(value));
+            }
+            return arr;
+        }
+        return j;
+    };
+
+    auto cleaned = remove_unsupported_healings_and_dump_args(partial->json);
+    LOG_DBG("Cleaned up JSON %s to %s (json_healing_marker : '%s')\n", partial->json.dump().c_str(), cleaned.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+    return consume_json_result {
+        cleaned,
+        /* .is_partial = */ found_healing_marker,
+    };
+}
+
+void common_chat_msg_parser::clear_tools() {
+    result_.tool_calls.clear();
+}

common/chat-parser.h

@@ -0,0 +1,120 @@
+#pragma once
+
+#include "chat.h"
+#include "json-partial.h"
+#include "regex-partial.h"
+
+#include <nlohmann/json.hpp>
+
+#include <optional>
+#include <string>
+#include <vector>
+
+class common_chat_msg_partial_exception : public std::runtime_error {
+  public:
+    common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
+};
+
+class common_chat_msg_parser {
+    std::string input_;
+    bool is_partial_;
+    common_chat_syntax syntax_;
+    std::string healing_marker_;
+
+    size_t pos_ = 0;
+    common_chat_msg result_;
+
+  public:
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    const std::string & input() const { return input_; }
+    size_t pos() const { return pos_; }
+    const std::string & healing_marker() const { return healing_marker_; }
+    const bool & is_partial() const { return is_partial_; }
+    const common_chat_msg & result() const { return result_; }
+    const common_chat_syntax & syntax() const { return syntax_; }
+
+    void move_to(size_t pos) {
+        if (pos > input_.size()) {
+            throw std::runtime_error("Invalid position!");
+        }
+        pos_ = pos;
+    }
+    void move_back(size_t n) {
+        if (pos_ < n) {
+            throw std::runtime_error("Can't move back that far!");
+        }
+        pos_ -= n;
+    }
+
+    // Get the substring of the input at the given range
+    std::string str(const common_string_range & rng) const;
+
+    // Appends to the result.content field
+    void add_content(const std::string & content);
+
+    // Appends to the result.reasoning_content field
+    void add_reasoning_content(const std::string & reasoning_content);
+
+    // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
+    bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
+
+    // Adds a tool call using the "name", "id" and "arguments" fields of the json object
+    bool add_tool_call(const nlohmann::ordered_json & tool_call);
+
+    // Adds an array of tool calls using their "name", "id" and "arguments" fields.
+    bool add_tool_calls(const nlohmann::ordered_json & arr);
+
+    void finish();
+
+    bool consume_spaces();
+
+    void consume_literal(const std::string & literal);
+
+    bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
+
+    std::string consume_rest();
+
+    struct find_regex_result {
+        std::string prelude;
+        std::vector<common_string_range> groups;
+    };
+
+    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
+
+    bool try_consume_literal(const std::string & literal);
+
+    std::optional<find_regex_result> try_find_literal(const std::string & literal);
+
+    find_regex_result consume_regex(const common_regex & regex);
+
+    std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
+
+    std::optional<common_json> try_consume_json();
+    common_json consume_json();
+
+    struct consume_json_result {
+        nlohmann::ordered_json value;
+        bool is_partial;
+    };
+
+    /*
+        Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
+
+        By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
+        e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
+
+        But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
+        - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
+        - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
+    */
+    consume_json_result consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+    std::optional<consume_json_result> try_consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+
+    void clear_tools();
+};

common/chat.h

@@ -3,9 +3,11 @@
 #pragma once
 
 #include "common.h"
+#include <functional>
 #include <chrono>
 #include <string>
 #include <vector>
+#include <map>
 
 struct common_chat_templates;
 
@@ -13,11 +15,19 @@ struct common_chat_tool_call {
     std::string name;
     std::string arguments;
     std::string id;
+
+    bool operator==(const common_chat_tool_call & other) const {
+        return name == other.name && arguments == other.arguments && id == other.id;
+    }
 };
 
 struct common_chat_msg_content_part {
     std::string type;
     std::string text;
+
+    bool operator==(const common_chat_msg_content_part & other) const {
+        return type == other.type && text == other.text;
+    }
 };
 
 struct common_chat_msg {
@@ -28,6 +38,51 @@ struct common_chat_msg {
     std::string reasoning_content;
     std::string tool_name;
     std::string tool_call_id;
+
+    template <class T> T to_json_oaicompat() const;
+
+    bool empty() const {
+        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
+    }
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+        for (auto i = 0u; i < tool_calls.size(); i++) {
+            if (ids_cache.size() <= i) {
+                auto id = tool_calls[i].id;
+                if (id.empty()) {
+                    id = gen_tool_call_id();
+                }
+                ids_cache.push_back(id);
+            }
+            tool_calls[i].id = ids_cache[i];
+        }
+    }
+    bool operator==(const common_chat_msg & other) const {
+        return role == other.role
+            && content == other.content
+            && content_parts == other.content_parts
+            && tool_calls == other.tool_calls
+            && reasoning_content == other.reasoning_content
+            && tool_name == other.tool_name
+            && tool_call_id == other.tool_call_id;
+    }
+    bool operator!=(const common_chat_msg & other) const {
+        return !(*this == other);
+    }
+};
+
+struct common_chat_msg_diff {
+    std::string reasoning_content_delta;
+    std::string content_delta;
+    size_t tool_call_index = std::string::npos;
+    common_chat_tool_call tool_call_delta;
+
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg);
+
+    bool operator==(const common_chat_msg_diff & other) const {
+        return content_delta == other.content_delta
+        && tool_call_index == other.tool_call_index
+        && tool_call_delta == other.tool_call_delta;
+    }
 };
 
 struct common_chat_tool {
@@ -49,14 +104,11 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_LLAMA_3_X,
     COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
     COMMON_CHAT_FORMAT_DEEPSEEK_R1,
-    COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
-    COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
-    COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
@@ -71,8 +123,10 @@ struct common_chat_templates_inputs {
     std::vector<common_chat_tool> tools;
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
-    bool extract_reasoning     = true;
+    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 {
@@ -80,11 +134,21 @@ struct common_chat_params {
     std::string                         prompt;
     std::string                         grammar;
     bool                                grammar_lazy = false;
+    bool                                thinking_forced_open = false;
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
 };
 
+struct common_chat_syntax {
+    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
+    bool                     reasoning_in_content  = false;
+    bool                     thinking_forced_open  = false;
+    bool                     parse_tool_calls      = true;
+};
+
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
 
@@ -121,8 +185,9 @@ std::string common_chat_format_example(
     const struct common_chat_templates * tmpls,
     bool use_jinja);
 
-std::string               common_chat_format_name(common_chat_format format);
-common_chat_msg           common_chat_parse(      const std::string & input, common_chat_format format);
+const char*               common_chat_format_name(common_chat_format format);
+const char*               common_reasoning_format_name(common_reasoning_format format);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
@@ -135,3 +200,5 @@ template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common
 // T can be std::string containing JSON or nlohmann::ordered_json
 template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
 template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+
+template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);

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

@@ -1,24 +0,0 @@
-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

@@ -203,6 +203,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     DWORD p = NORMAL_PRIORITY_CLASS;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = BELOW_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
         case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
@@ -228,6 +229,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     int p = 0;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p =  5;  break;
         case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
         case GGML_SCHED_PRIO_HIGH:     p = -10; break;
@@ -464,7 +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, "\\$0");
+    return std::regex_replace(s, special_chars, "\\$&");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -704,11 +706,17 @@ 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);
@@ -765,6 +773,9 @@ 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
@@ -782,9 +793,16 @@ 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();
 
-        const bool success = CreateDirectoryW(test, NULL);
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -798,8 +816,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
-
-        pos_slash += 1;
     }
 
     return true;
@@ -849,7 +865,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -895,31 +911,6 @@ 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;
-        }
-
-        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
-            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);
@@ -929,7 +920,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -961,6 +952,35 @@ 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;
@@ -1036,7 +1056,7 @@ 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_kv_self_clear(lctx);
+        llama_memory_clear(llama_get_memory(lctx), true);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
@@ -1102,6 +1122,9 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
     }
 
+    mparams.progress_callback           = params.load_progress_callback;
+    mparams.progress_callback_user_data = params.load_progress_callback_user_data;
+
     return mparams;
 }
 
@@ -1133,11 +1156,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
-
-    if (params.reranking) {
-        cparams.embeddings    = true;
-        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
-    }
+    cparams.swa_full          = params.swa_full;
 
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
@@ -1271,6 +1290,9 @@ 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);
@@ -1325,81 +1347,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
     return text;
 }
 
-//
-// KV cache utils
-//
-
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        int seq_count = 0;
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) { seq_count++; }
-        }
-        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    std::unordered_map<llama_seq_id, size_t> seqs;
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] < 0) { continue; }
-            if (seqs.find(cs_curr[j]) == seqs.end()) {
-                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                const size_t sz = seqs.size();
-                seqs[cs_curr[j]] = sz;
-            }
-        }
-        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-    }
-
-    printf("=== Sequence legend: ");
-    for (const auto & it : seqs) {
-        printf("%zu=%d, ", it.second, it.first);
-    }
-    printf("'+'=other sequence ids");
-
-    c_curr = view.cells;
-    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) {
-                const auto & it = seqs.find(cs_curr[j]);
-                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
-            } else {
-                putchar('.');
-            }
-        }
-        putchar(' ');
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
 //
 // Embedding utils
 //

common/common.h

@@ -8,6 +8,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <map>
 #include <sstream>
 
 #ifdef _WIN32
@@ -76,7 +77,7 @@ enum llama_example {
     LLAMA_EXAMPLE_SERVER,
     LLAMA_EXAMPLE_CVECTOR_GENERATOR,
     LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
@@ -115,7 +116,7 @@ enum common_grammar_trigger_type {
     COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
     COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
     COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
 };
 
 struct common_grammar_trigger {
@@ -199,6 +200,9 @@ 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;
 
@@ -215,7 +219,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    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.
 };
 
 struct common_params {
@@ -291,6 +296,7 @@ struct common_params {
     int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
+    bool    offline                    = false;
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -323,13 +329,13 @@ struct common_params {
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
@@ -353,7 +359,7 @@ 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
-    bool reranking         = false; // enable reranking support on server
+    std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
@@ -364,10 +370,12 @@ struct common_params {
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
+    std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
@@ -375,6 +383,8 @@ 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;
@@ -428,6 +438,11 @@ struct common_params {
 
     // common params
     std::string out_file; // output filename for all example programs
+    // optional callback for model loading progress and cancellation:
+    // called with a progress value between 0.0 and 1.0.
+    // return false from callback to abort model loading or true to continue
+    llama_progress_callback load_progress_callback = NULL;
+    void *                  load_progress_callback_user_data = NULL;
 };
 
 // call once at the start of a program if it uses libcommon
@@ -616,16 +631,6 @@ std::string common_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// Dump the KV cache view showing individual sequences in each cell (long output).
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
 //
 // Embedding utils
 //

common/json-partial.cpp

@@ -0,0 +1,256 @@
+#include "json-partial.h"
+
+#include "log.h"
+
+#include <nlohmann/json.hpp>
+
+#include <string>
+
+using json = nlohmann::ordered_json;
+
+enum common_json_stack_element_type {
+    COMMON_JSON_STACK_ELEMENT_OBJECT,
+    COMMON_JSON_STACK_ELEMENT_KEY,
+    COMMON_JSON_STACK_ELEMENT_ARRAY,
+};
+
+struct common_json_stack_element {
+    common_json_stack_element_type type;
+    std::string key;
+};
+
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    std::string::const_iterator it = input.begin();
+    const auto end = input.end();
+    return common_json_parse(it, end, healing_marker, out);
+}
+
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    // // https://json.nlohmann.me/features/parsing/sax_interface/
+    struct json_error_locator : public nlohmann::json_sax<json> {
+        std::size_t position;
+        bool found_error;
+        std::string last_token;
+        std::string exception_message;
+        std::vector<common_json_stack_element> stack;
+
+        json_error_locator() : position(0), found_error(false) {}
+
+        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
+            this->position = position - 1;
+            this->found_error = true;
+            this->last_token = last_token;
+            this->exception_message = ex.what();
+            return false;
+        }
+        void close_value() {
+            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
+                stack.pop_back();
+            }
+        }
+        bool null() override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool boolean(bool) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_integer(number_integer_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_unsigned(number_unsigned_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_float(number_float_t, const string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool string(string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool binary(binary_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool start_object(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
+            return true;
+        }
+        bool end_object() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+        bool key(string_t & key) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
+            return true;
+        }
+        bool start_array(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
+            return true;
+        }
+        bool end_array() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+    };
+    json_error_locator err_loc;
+    auto start = it;
+    json::sax_parse(it, end, &err_loc);
+
+    if (err_loc.found_error) {
+        it = start;
+        auto temptative_end = it + err_loc.position;
+        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
+
+        auto input = std::string(it, temptative_end);
+        try {
+            out.json = json::parse(input);
+            // out.json = json::parse(it, temptative_end);
+            it = temptative_end;
+            return true;
+        } catch (const std::exception & ex) {
+            // No, needs healing.
+            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
+        }
+        auto can_parse = [](const std::string & str) {
+            try {
+                auto _ = json::parse(str); // NOLINT
+                return true;
+            } catch (const std::exception &) {
+                return false;
+            }
+        };
+        if (!healing_marker.empty() && !err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
+            if (last_non_sp_pos == std::string::npos) {
+                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+            }
+            auto last_non_sp_char = str[last_non_sp_pos];
+            // Used to detect stops on a number, which may not be complete.
+            auto was_maybe_number = [&]() {
+                if (!str.empty() && std::isspace(str.back())) {
+                    return false;
+                }
+                return std::isdigit(last_non_sp_char) ||
+                    last_non_sp_char == '.' ||
+                    last_non_sp_char == 'e' ||
+                    last_non_sp_char == 'E' ||
+                    last_non_sp_char == '-';
+            };
+
+            std::string closing;
+            for (size_t i = err_loc.stack.size(); i > 0; i--) {
+                auto & el = err_loc.stack[i - 1];
+                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                    closing += "}";
+                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                    closing += "]";
+                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
+                    throw std::runtime_error("Unexpected stack element type");
+                }
+            }
+
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
+
+            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
+                // We're inside an object value
+                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
+                    // Was about to create an object value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + ": 1" + closing)) {
+                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
+                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
+                    // Was about to create an object
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an object value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an object value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else {
+                    // find last :
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+                    }
+                    // Cutting back to opening : for object value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
+                    // Was about to create an array value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an array value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an array value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
+                    // Had just finished a value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
+                } else {
+                    auto last_pos = str.find_last_of("[,");
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
+                    }
+                    // Cutting back to last [ or , for array value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
+                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\": 1" + closing)) {
+                    // Was inside an object key string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
+                    // Was inside an object key string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else {
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+                    }
+                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else {
+                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+            }
+            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
+        // TODO: handle unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
+        // fprintf(stderr, "Closing: TODO\n");
+        return false;
+    }
+    out.json = json::parse(it, end);
+    it = end;
+    return true;
+}

common/json-partial.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include <nlohmann/json.hpp>
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

common/json-schema-to-grammar.cpp

@@ -1,8 +1,9 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
-#include <fstream>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -40,49 +41,6 @@ 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();
@@ -111,14 +69,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         }
         out << "}";
     };
-    std::function<void(const string_view &, const string_view &)> uniform_range =
-        [&](const string_view & from, const string_view & to) {
+    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
+        [&](const std::string_view & from, const std::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).str() << "\"";
+                out << "\"" << from.substr(0, i) << "\"";
             }
             if (i < from.length() && i < to.length()) {
                 if (i > 0) {

common/json-schema-to-grammar.h

@@ -1,9 +1,9 @@
 #pragma once
 
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
+#include <nlohmann/json_fwd.hpp>
+
+#include <functional>
+#include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);

common/sampling.cpp

@@ -161,7 +161,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
-        std::vector<std::string> patterns_at_start;
+        std::vector<std::string> trigger_patterns;
         std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
@@ -173,10 +173,13 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
-                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
                 {
-                    const auto & pattern = trigger.value;
-                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
+                    patterns_anywhere.push_back(trigger.value);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
+                {
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -190,10 +193,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        std::vector<std::string> trigger_patterns;
-        if (!patterns_at_start.empty()) {
-            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
-        }
         if (!patterns_anywhere.empty()) {
             trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
         }

common/speculative.cpp

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

convert_hf_to_gguf_update.py

@@ -1,28 +1,6 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 
-# This script downloads the tokenizer models of the specified models from Huggingface and
-# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
-#
-# This is necessary in order to analyze the type of pre-tokenizer used by the model and
-# provide the necessary information to llama.cpp via the GGUF header in order to implement
-# the same pre-tokenizer.
-#
-# ref: https://github.com/ggml-org/llama.cpp/pull/6920
-#
-# Instructions:
-#
-# - Add a new model to the "models" list
-# - Run the script with your huggingface token:
-#
-#   python3 convert_hf_to_gguf_update.py <huggingface_token>
-#
-# - The convert_hf_to_gguf.py script will have had its get_vocab_base_pre() function updated
-# - Update llama.cpp with the new pre-tokenizer if necessary
-#
-# TODO: generate tokenizer tests for llama.cpp
-#
-
 import logging
 import os
 import pathlib
@@ -32,6 +10,7 @@ import requests
 import sys
 import json
 import shutil
+import argparse
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -41,6 +20,11 @@ logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger("convert_hf_to_gguf_update")
 sess = requests.Session()
 
+convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
+convert_py = convert_py_pth.read_text(encoding="utf-8")
+hf_token_pth = pathlib.Path.home() / ".cache" / "huggingface" / "token"
+hf_token = hf_token_pth.read_text(encoding="utf-8").strip() if hf_token_pth.exists() else None
+
 
 class TOKENIZER_TYPE(IntEnum):
     SPM = auto()
@@ -49,20 +33,49 @@ class TOKENIZER_TYPE(IntEnum):
     UGM = auto()
 
 
+DOC_STRING = """
+This script downloads the tokenizer models of the specified models from Huggingface and
+generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
+
+/!\\ It is intended to be used by contributors and is not meant to be run by end users
+
+This is necessary in order to analyze the type of pre-tokenizer used by the model and
+provide the necessary information to llama.cpp via the GGUF header in order to implement
+the same pre-tokenizer.
+
+ref: https://github.com/ggml-org/llama.cpp/pull/6920
+
+Instructions:
+
+- Add a new model to the "models" list
+- Run the script with your huggingface token
+    By default, token will be read from ~/.cache/huggingface/token
+- The convert_hf_to_gguf.py script will have had its get_vocab_base_pre() function updated
+- Update llama.cpp with the new pre-tokenizer if necessary
+"""
+# TODO: generate tokenizer tests for llama.cpp
+
+parser = argparse.ArgumentParser(description=DOC_STRING, formatter_class=argparse.RawTextHelpFormatter)
+parser.add_argument(
+    "--full", action="store_true",
+    help="download full list of models - make sure you have access to all of them",
+)
+parser.add_argument(
+    "hf_token",
+    help="optional HF token",
+    nargs="?",
+)
+args = parser.parse_args()
+hf_token = args.hf_token if args.hf_token is not None else hf_token
+
+if hf_token is None:
+    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
+    sys.exit(1)
+
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
 CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
 
-if len(sys.argv) == 2:
-    token = sys.argv[1]
-    if not token.startswith("hf_"):
-        logger.info("Huggingface token seems invalid")
-        logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
-        sys.exit(1)
-else:
-    logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
-    sys.exit(1)
-
 # TODO: add models here, base models preferred
 models = [
     {"name": "llama-spm",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
@@ -103,7 +116,6 @@ models = [
     {"name": "exaone",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
     {"name": "phi-2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
     {"name": "chameleon",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
-    {"name": "minerva-7b",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", },
     {"name": "roberta-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"},
     {"name": "gigachat",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct"},
     {"name": "megrez",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
@@ -114,9 +126,26 @@ models = [
     {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
-    {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
+    {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
+    {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
+    {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
+]
+
+# some models are known to be broken upstream, so we will skip them as exceptions
+pre_computed_hashes = [
+    # chatglm-bpe has 2 hashes, why?
+    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
+    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
+    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
+    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
+    # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
 ]
 
 
@@ -169,9 +198,29 @@ def download_model(model):
             if os.path.isfile(save_path):
                 logger.info(f"{name}: File {save_path} already exists - skipping")
                 continue
-            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+            download_file_with_auth(f"{repo}/resolve/main/{file}", hf_token, save_path)
 
 
+# get list of existing models and chkhsh from the convert_hf_to_gguf.py file
+# returns mapping res --> chkhsh
+def get_existing_models(convert_py):
+    pattern = r'if chkhsh == "([a-f0-9]{64})":\s*\n\s*.*\s*res = "([^"]+)"'
+    matches = re.findall(pattern, convert_py)
+    output = {}
+    for chkhsh, res in matches:
+        output[res] = chkhsh
+    return output
+
+
+existing_models = {}
+all_models = models.copy()
+if not args.full:
+    # Filter out models that already exist in convert_hf_to_gguf.py
+    existing_models = get_existing_models(convert_py)
+    all_models = models.copy()
+    models = [model for model in all_models if model["name"] not in existing_models]
+
+logging.info(f"Downloading {len(models)} models...")
 for model in models:
     try:
         download_model(model)
@@ -182,9 +231,10 @@ for model in models:
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in models:
+for model in [*all_models, *pre_computed_hashes]:
     name = model["name"]
     tokt = model["tokt"]
+    chkhsh = model.get("chkhsh")
 
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
@@ -195,35 +245,44 @@ for model in models:
         continue
 
     # create the tokenizer
-    try:
-        if name == "t5":
-            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
-        else:
-            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
-        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
-        continue  # Skip to the next model if the tokenizer can't be loaded
+    if chkhsh is not None:
+        # if the model has a pre-computed hash, use it
+        logger.info(f"Using pre-computed hash for model {name}: {chkhsh}")
+    elif name in existing_models:
+        # if the model already exists in convert_hf_to_gguf.py, skip compute hash
+        chkhsh = existing_models[name]
+    else:
+        # otherwise, compute the hash of the tokenizer
+        try:
+            logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
+            if name == "t5":
+                tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+        except OSError as e:
+            logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+            continue  # Skip to the next model if the tokenizer can't be loaded
 
-    chktok = tokenizer.encode(CHK_TXT)
-    chkhsh = sha256(str(chktok).encode()).hexdigest()
+        chktok = tokenizer.encode(CHK_TXT)
+        chkhsh = sha256(str(chktok).encode()).hexdigest()
 
-    logger.info(f"model: {name}")
-    logger.info(f"tokt: {tokt}")
-    logger.info(f"repo: {model['repo']}")
-    logger.info(f"chktok: {chktok}")
-    logger.info(f"chkhsh: {chkhsh}")
+        logger.info(f"model: {name}")
+        logger.info(f"tokt: {tokt}")
+        logger.info(f"repo: {model['repo']}")
+        logger.info(f"chktok: {chktok}")
+        logger.info(f"chkhsh: {chkhsh}")
 
-    # print the "pre_tokenizer" content from the tokenizer.json
-    with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
-        cfg = json.load(f)
-        normalizer = cfg["normalizer"]
-        logger.info("normalizer: " + json.dumps(normalizer, indent=4))
-        pre_tokenizer = cfg["pre_tokenizer"]
-        logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
-        if "ignore_merges" in cfg["model"]:
-            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
+        # print the "pre_tokenizer" content from the tokenizer.json
+        with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
+            cfg = json.load(f)
+            normalizer = cfg["normalizer"]
+            logger.info("normalizer: " + json.dumps(normalizer, indent=4))
+            pre_tokenizer = cfg["pre_tokenizer"]
+            logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+            if "ignore_merges" in cfg["model"]:
+                logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
 
-    logger.info("")
+        logger.info("")
 
     src_ifs += f"        if chkhsh == \"{chkhsh}\":\n"
     src_ifs += f"            # ref: {model['repo']}\n"
@@ -271,8 +330,6 @@ src_func = f"""
         return res
 """
 
-convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
-convert_py = convert_py_pth.read_text(encoding="utf-8")
 convert_py = re.sub(
     r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
     lambda m: m.group(1) + src_func + m.group(3),
@@ -288,7 +345,7 @@ logger.info("+++ convert_hf_to_gguf.py was updated")
 
 tests = [
     "ied 4 ½ months",
-    "Führer",
+    "Äpfel",
     "",
     " ",
     "  ",
@@ -367,6 +424,10 @@ for model in models:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 
+    if not os.path.exists(f"models/ggml-vocab-{name}.gguf"):
+        logger.info(f"Skip vocab files for model {name}, no GGUF file found")
+        continue
+
     with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
         for text in tests:
             f.write(f"{text}")

docs/backend/CANN.md

@@ -8,6 +8,7 @@
  - [DataType Supports](#datatype-supports)
  - [Docker](#docker)
  - [Linux](#linux)
+ - [Environment variable setup](#environment-variable-setup)
  - [TODO](#todo)
 
 
@@ -56,60 +57,82 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 
 ## Model Supports
 
-| Model Name                  | FP16  | Q8_0 | Q4_0 |
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
 |:----------------------------|:-----:|:----:|:----:|
-| AquilaChat2-7B              |   √   |   √  |   √  |
-| Baichuan-7b                 |   √   |   √  |   √  |
-| Baichuan2-7B-Chat           |   √   |   √  |   √  |
-| bitnet_b1_58-large          |   √   |   √  |   √  |
-| bloom-560m                  |   √   |   x  |   √  |
-| bloomz-alpaca-560m          |   √   |   x  |   √  |
-| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
-| chatglm3-6B                 |   x   |   x  |   x  |
-| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
-| CodeShell-7B                |   √   |   √  |   √  |
-| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
-| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
-| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
-| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
-| falcon-7b-instruct          |   √   |   √  |   √  |
-| flan-t5-large               |   √   |   √  |   √  |
-| gemma-2-9b-it               |   √   |   √  |   √  |
-| glm-4-9B                    |   x   |   x  |   x  |
-| gpt2                        |   √   |   √  |   √  |
-| Gpt2-163M                   |   √   |   √  |   √  |
-| granite-3B-code-instruct    |   √   |   √  |   √  |
+| Llama-2                     |   √   |   √  |   √  |
+| Llama-3                     |   √   |   √  |   √  |
+| Mistral-7B                  |   √   |   √  |   √  |
+| Mistral MOE                 |   √   |   √  |   √  |
+| DBRX                        |   -   |   -  |   -  |
+| Falcon                      |   √   |   √  |   √  |
+| Chinese LLaMA/Alpaca        |   √   |   √  |   √  |
+| Vigogne(French)             |   √   |   √  |   √  |
+| BERT                        |   x   |   x  |   x  |
+| Koala                       |   √   |   √  |   √  |
+| Baichuan                    |   √   |   √  |   √  |
+| Aquila 1 & 2                |   √   |   √  |   √  |
+| Starcoder models            |   √   |   √  |   √  |
+| Refact                      |   √   |   √  |   √  |
+| MPT                         |   √   |   √  |   √  |
+| Bloom                       |   √   |   √  |   √  |
+| Yi models                   |   √   |   √  |   √  |
+| stablelm models             |   √   |   √  |   √  |
+| DeepSeek models             |   x   |   x  |   x  |
+| Qwen models                 |   √   |   √  |   √  |
+| PLaMo-13B                   |   √   |   √  |   √  |
+| Phi models                  |   √   |   √  |   √  |
+| PhiMoE                      |   √   |   √  |   √  |
+| GPT-2                       |   √   |   √  |   √  |
+| Orion                       |   √   |   √  |   √  |
+| InternlLM2                  |   √   |   √  |   √  |
+| CodeShell                   |   √   |   √  |   √  |
+| Gemma                       |   √   |   √  |   √  |
+| Mamba                       |   √   |   √  |   √  |
+| Xverse                      |   √   |   √  |   √  |
+| command-r models            |   √   |   √  |   √  |
+| Grok-1                      |   -   |   -  |   -  |
+| SEA-LION                    |   √   |   √  |   √  |
 | GritLM-7B                   |   √   |   √  |   √  |
-| internlm2_5-7b-chat         |   √   |   √  |   √  |
-| koala-7B-HF                 |   √   |   √  |   √  |
-| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
-| Llama-3-Smaug-8B            |   √   |   √  |   √  |
-| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
-| Llama3-8B                   |   √   |   √  |   √  |
-| Llama3-8b-chinese           |   √   |   √  |   √  |
-| mamba-130m-hf               |   √   |   √  |   √  |
-| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
-| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
-| mpt-7B                      |   √   |   √  |   √  |
-| OLMo-1B-hf                  |   √   |   √  |   √  |
-| OpenELM-3B-Instruct         |   √   |   √  |   √  |
-| Orion-14b-base              |   √   |   √  |   √  |
-| phi1                        |   x   |   x  |   x  |
-| phi2                        |   x   |   x  |   x  |
-| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
-| plamo-13b                   |   √   |   √  |   √  |
-| pythia-70M                  |   x   |   x  |   x  |
-| Qwen-7B                     |   √   |   √  |   √  |
-| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
-| Refact-1_6B-fim             |   √   |   √  |   √  |
-| SmolLM-135M                 |   √   |   √  |   √  |
-| stablelm-zephyr             |   x   |   x  |   x  |
-| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
-| starcoderbase-1b            |   √   |   √  |   √  |
-| starcoder2-3b               |   √   |   √  |   √  |
-| vigogne-7b-chat             |   √   |   √  |   √  |
-| xverse-7b-chat              |   √   |   √  |   √  |
-| Yi-6b-Chat                  |   √   |   √  |   √  |
+| OLMo                        |   √   |   √  |   √  |
+| OLMo 2                      |   √   |   √  |   √  |
+| OLMoE                       |   √   |   √  |   √  |
+| Granite models              |   √   |   √  |   √  |
+| GPT-NeoX                    |   √   |   √  |   √  |
+| Pythia                      |   √   |   √  |   √  |
+| Snowflake-Arctic MoE        |   -   |   -  |   -  |
+| Smaug                       |   √   |   √  |   √  |
+| Poro 34B                    |   √   |   √  |   √  |
+| Bitnet b1.58 models         |   √   |   x  |   x  |
+| Flan-T5                     |   √   |   √  |   √  |
+| Open Elm models             |   x   |   √  |   √  |
+| chatGLM3-6B + ChatGLM4-9b +  GLMEdge-1.5b + GLMEdge-4b    |   √   |   √  |   √  |
+| GLM-4-0414                  |   √   |   √  |   √  |
+| SmolLM                      |   √   |   √  |   √  |
+| EXAONE-3.0-7.8B-Instruct    |   √   |   √  |   √  |
+| FalconMamba Models          |   √   |   √  |   √  |
+| Jais Models                 |   -   |   x  |   x  |
+| Bielik-11B-v2.3             |   √   |   √  |   √  |
+| RWKV-6                      |   -   |   √  |   √  |
+| QRWKV-6                     |   √   |   √  |   √  |
+| GigaChat-20B-A3B            |   x   |   x  |   x  |
+| Trillion-7B-preview         |   √   |   √  |   √  |
+| Ling models                 |   √   |   √  |   √  |
+
+
+**Multimodal**
+| Model Name                  | FP16  | Q4_0 | Q8_0 |
+|:----------------------------|:-----:|:----:|:----:|
+| LLaVA 1.5 models, LLaVA 1.6 models      |   x   |   x  |   x  |
+|  BakLLaVA                   |   √   |   √  |   √  |
+|  Obsidian                   |   √   |   -  |   -  |
+|  ShareGPT4V                 |   x   |   -  |   -  |
+|  MobileVLM 1.7B/3B models   |   -   |   -  |   -  |
+|  Yi-VL                      |   -   |   -  |   -  |
+|  Mini CPM                   |   √   |   √  |   √  |
+|  Moondream                  |   √   |   √  |   √  |
+|  Bunny                      |   √   |   -  |   -  |
+|  GLM-EDGE                   |   √   |   √  |   √  |
+|  Qwen2-VL                   |   √   |   √  |   √  |
 
 
 
@@ -258,6 +281,34 @@ cmake --build build --config release
 ### **GitHub contribution**:
 Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay.
 
+## Updates
+### Basic Flash Attention Support
+The basic FA kernel with aclnnops has been added in aclnn_ops.cpp.
+Currently, the FA only supports the cases with FP16 KV tensors and NO logit softcap.
+Since the aclnn interface for flash attention cannot support the logit softcap, we will only update the quantized version in the future.
+
+Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang@pku.edu.cn), Ruiyang Ma (ruiyang@stu.pku.edu.cn), and Guojie Luo (gluo@pku.edu.cn).
+
+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

@@ -17,25 +17,25 @@
 
 **SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.
 
-**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to Intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
-- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
 ### Llama.cpp + SYCL
 
-The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
+The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
+SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
 
 ## Recommended Release
 
-The SYCL backend would be broken by some PRs due to no online CI.
-
-The following release is verified with good quality:
+The following releases are verified and recommended:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
+|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |ArcB580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
@@ -106,15 +106,14 @@ SYCL backend supports Intel GPU Family:
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
-| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake    |
-| Intel iGPU                    | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750, B580         |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
+| Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
 
 *Notes:*
 
 - **Memory**
   - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
-
   - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
 
 - **Execution Unit (EU)**
@@ -138,9 +137,11 @@ Note: AMD GPU support is highly experimental and is incompatible with F16.
 Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
 
 ## Docker
-The docker build option is currently limited to *intel GPU* targets.
+
+The docker build option is currently limited to *Intel GPU* targets.
 
 ### Build image
+
 ```sh
 # Using FP16
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
@@ -148,9 +149,10 @@ docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f
 
 *Notes*:
 
-To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.
+To build in default FP32 *(Slower than FP16 alternative)*, set `--build-arg="GGML_SYCL_F16=OFF"` in the previous command.
 
 You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
+Check the [documentation for Docker](../docker.md) to see the available images.
 
 ### Run container
 
@@ -250,7 +252,7 @@ sycl-ls
 
 - **Intel GPU**
 
-When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
+When targeting an intel GPU, the user should expect one or more devices among the available SYCL devices. Please make sure that at least one GPU is present via `sycl-ls`, for instance `[level_zero:gpu]` in the sample output below:
 
 ```
 [opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
@@ -282,7 +284,7 @@ For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
 
 #### Intel GPU
 
-```
+```sh
 ./examples/sycl/build.sh
 ```
 
@@ -351,7 +353,7 @@ cmake --build build --config Release -j -v
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -398,11 +400,15 @@ Choose one of following methods to run.
 
 ```sh
 ./examples/sycl/run-llama2.sh 0
+# OR
+./examples/sycl/run-llama3.sh 0
 ```
 - Use multiple devices:
 
 ```sh
 ./examples/sycl/run-llama2.sh
+# OR
+./examples/sycl/run-llama3.sh
 ```
 
 2. Command line
@@ -425,13 +431,13 @@ Examples:
 - Use device 0:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer
 ```
 
 *Notes:*
@@ -452,7 +458,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 1. Install GPU driver
 
-Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
 2. Install Visual Studio
 
@@ -629,7 +635,7 @@ Once it is completed, final results will be in **build/Release/bin**
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -648,7 +654,7 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow
 build\bin\llama-ls-sycl-device.exe
 ```
 
-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *Intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -658,13 +664,14 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+
 #### Choose level-zero devices
 
 |Chosen Device ID|Setting|
 |-|-|
-|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|0|Default option. You may also want to `set ONEAPI_DEVICE_SELECTOR="level_zero:0"`|
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|
 
 #### Execute
 
@@ -673,7 +680,13 @@ Choose one of following methods to run.
 1. Script
 
 ```
-examples\sycl\win-run-llama2.bat
+examples\sycl\win-run-llama-2.bat
+```
+
+or
+
+```
+examples\sycl\win-run-llama-3.bat
 ```
 
 2. Command line
@@ -697,13 +710,13 @@ Examples:
 - Use device 0:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer
 ```
 
 
@@ -714,7 +727,9 @@ Note:
 ```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
+
 Or
+
 ```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
@@ -726,21 +741,23 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
 | GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
 | GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
+1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+
 #### Runtime
 
 | 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 based on Intel GPU type, to compare the performance increase |
+| 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_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 |
@@ -752,7 +769,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Q&A
 
-- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+- Error:  `error while loading shared libraries: libsycl.so: cannot open shared object file: No such file or directory`.
 
   - Potential cause: Unavailable oneAPI installation or not set ENV variables.
   - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
@@ -781,18 +798,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   It's same for other projects including llama.cpp SYCL backend.
 
-- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+- `Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)`, `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 3503030272 Bytes of memory on device`, or `failed to allocate SYCL0 buffer`
 
-  Device Memory is not enough.
+  You are running out of Device Memory.
 
   |Reason|Solution|
   |-|-|
-  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
-  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
+  | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
+  | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|
 
 ### **GitHub contribution**:
-Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
+Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 
 ## TODO
 
-- NA
+- Review ZES_ENABLE_SYSMAN: https://github.com/intel/compute-runtime/blob/master/programmers-guide/SYSMAN.md#support-and-limitations

docs/build-s390x.md

@@ -0,0 +1,246 @@
+> [!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

@@ -1,5 +1,9 @@
 # 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
@@ -63,6 +67,7 @@ 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
 
@@ -552,6 +557,10 @@ 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/development/HOWTO-add-model.md

@@ -83,20 +83,22 @@ NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the conv
 
 ### 2. Define the model architecture in `llama.cpp`
 
-The model params and tensors layout must be defined in `llama.cpp`:
-1. Define a new `llm_arch`
-2. Define the tensors layout in `LLM_TENSOR_NAMES`
-3. Add any non-standard metadata in `llm_load_hparams`
-4. Create the tensors for inference in `llm_load_tensors`
-5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
+The model params and tensors layout must be defined in `llama.cpp` source files:
+1. Define a new `llm_arch` enum value in `src/llama-arch.h`.
+2. In `src/llama-arch.cpp`:
+    - Add the architecture name to the `LLM_ARCH_NAMES` map.
+    - Add the tensor mappings to the `LLM_TENSOR_NAMES` map.
+3. Add any non-standard metadata loading in the `llama_model_loader` constructor in `src/llama-model-loader.cpp`.
+4. If the model has a RoPE operation, add a case for the architecture in `llama_model_rope_type` function in `src/llama-model.cpp`.
 
 NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorch` dimensions.
 
 ### 3. Build the GGML graph implementation
 
-This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
-
-Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
+This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `src/llama-model.cpp`.
+Create a new struct that inherits from `llm_graph_context` and implement the graph-building logic in its constructor.
+Have a look at existing implementations like `llm_build_llama`, `llm_build_dbrx` or `llm_build_bert`.
+Then, in the `llama_model::build_graph` method, add a case for your architecture to instantiate your new graph-building struct.
 
 Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
 

docs/docker.md

@@ -22,6 +22,12 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `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).
 
@@ -104,7 +110,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
 
 The defaults are:
 
-- `MUSA_VERSION` set to `rc3.1.1`
+- `MUSA_VERSION` set to `rc4.0.1`
 
 The resulting images, are essentially the same as the non-MUSA images:
 

docs/function-calling.md

@@ -2,7 +2,6 @@
 
 [chat.h](../common/chat.h) (https://github.com/ggml-org/llama.cpp/pull/9639) adds support for [OpenAI-style function calling](https://platform.openai.com/docs/guides/function-calling) and is used in:
 - `llama-server` when started w/ `--jinja` flag
-- `llama-cli` (WIP: https://github.com/ggml-org/llama.cpp/pull/11556)
 
 ## Universal support w/ Native & Generic handlers
 
@@ -12,7 +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 (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
+  - Qwen 2.5 Coder
   - Mistral Nemo
   - Firefunction v2
   - Command R7B
@@ -325,36 +324,65 @@ To get the official template from original HuggingFace repos, you can use [scrip
 > [!TIP]
 > If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills)
 
+> [!CAUTION]
+> Beware of extreme KV quantizations (e.g. `-ctk q4_0`), they can substantially degrade the model's tool calling performance.
+
 Test in CLI (or with any library / software that can use OpenAI-compatible API backends):
 
 ```bash
 curl http://localhost:8080/v1/chat/completions -d '{
-"model": "gpt-3.5-turbo",
-"tools": [
-    {
-    "type":"function",
-    "function":{
-        "name":"python",
-        "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
-        "parameters":{
-        "type":"object",
-        "properties":{
-            "code":{
-            "type":"string",
-            "description":"The code to run in the ipython interpreter."
+    "model": "gpt-3.5-turbo",
+    "tools": [
+        {
+        "type":"function",
+        "function":{
+            "name":"python",
+            "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
+            "parameters":{
+            "type":"object",
+            "properties":{
+                "code":{
+                "type":"string",
+                "description":"The code to run in the ipython interpreter."
+                }
+            },
+            "required":["code"]
             }
-        },
-        "required":["code"]
         }
-    }
-    }
-],
-"messages": [
-    {
-    "role": "user",
-    "content": "Print a hello world message with python."
-    }
-]
+        }
+    ],
+    "messages": [
+        {
+        "role": "user",
+        "content": "Print a hello world message with python."
+        }
+    ]
+}'
+
+
+curl http://localhost:8080/v1/chat/completions -d '{
+    "model": "gpt-3.5-turbo",
+    "messages": [
+        {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."},
+        {"role": "user", "content": "What is the weather in Istanbul?"}
+    ],
+    "tools": [{
+        "type":"function",
+        "function":{
+            "name":"get_current_weather",
+            "description":"Get the current weather in a given location",
+            "parameters":{
+                "type":"object",
+                "properties":{
+                    "location":{
+                        "type":"string",
+                        "description":"The city and country/state, e.g. `San Francisco, CA`, or `Paris, France`"
+                    }
+                },
+                "required":["location"]
+            }
+        }
+    }]
 }'
 ```
 

docs/install.md

@@ -1,28 +1,42 @@
 # Install pre-built version of llama.cpp
 
-## Homebrew
+| Install via | Windows | Mac | Linux |
+|-------------|---------|-----|-------|
+| Winget      | ✅      |      |      |
+| Homebrew    |         | ✅   | ✅   |
+| MacPorts    |         | ✅   |      |
+| Nix         |         | ✅   | ✅   |
 
-On Mac and Linux, the homebrew package manager can be used via
+## 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)
 
 ```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
+## MacPorts (Mac)
 
 ```sh
 sudo port install llama.cpp
 ```
-see also: https://ports.macports.org/port/llama.cpp/details/
 
-## Nix
+See also: https://ports.macports.org/port/llama.cpp/details/
 
-On Mac and Linux, the Nix package manager can be used via
+## Nix (Mac and Linux)
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
+
 For flake enabled installs.
 
 Or
@@ -34,13 +48,3 @@ 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

@@ -4,7 +4,9 @@ llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools
 - [llama-mtmd-cli](../tools/mtmd/README.md)
 - [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
 
-To enable it, can use use one of the 2 methods below:
+Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
+
+To enable it, you can use one of the 2 methods below:
 
 - Use `-hf` option with a supported model (see a list of pre-quantized model below)
     - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
@@ -31,12 +33,14 @@ llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
 
 ## Pre-quantized models
 
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/ggml-org
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/collections/ggml-org/multimodal-ggufs-68244e01ff1f39e5bebeeedc
 
 Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
 
 NOTE: some models may require large context window, for example: `-c 8192`
 
+**Vision models**:
+
 ```sh
 # Gemma 3
 (tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
@@ -77,4 +81,33 @@ NOTE: some models may require large context window, for example: `-c 8192`
 
 # Llama 4 Scout
 (tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
+
+# Moondream2 20250414 version
+(tool_name) -hf ggml-org/moondream2-20250414-GGUF
+
 ```
+
+**Audio models**:
+
+```sh
+# Ultravox 0.5
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF
+
+# Qwen2-Audio and SeaLLM-Audio
+# note: no pre-quantized GGUF this model, as they have very poor result
+# ref: https://github.com/ggml-org/llama.cpp/pull/13760
+```
+
+**Mixed modalities**:
+
+```sh
+# Qwen2.5 Omni
+# Capabilities: audio input, vision input
+(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

docs/ops.md

@@ -0,0 +1,95 @@
+# GGML Operations
+
+List of GGML operations and backend support status.
+
+Legend:
+- ✅ Fully supported by this backend
+- 🟡 Partially supported by this backend
+- ❌ Not supported by this backend
+
+| Operation | BLAS | CPU | CUDA | Metal |
+|-----------|------|------|------|------|
+|                              ABS | ❌ | ✅ | 🟡 | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ |
+|                              ADD | ❌ | ✅ | ✅ | 🟡 |
+|                             ADD1 | ❌ | ✅ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ |
+|                            CLAMP | ❌ | ✅ | ✅ | 🟡 |
+|                           CONCAT | ❌ | ✅ | 🟡 | ✅ |
+|                             CONT | ❌ | ✅ | 🟡 | ✅ |
+|                       CONV_2D_DW | ❌ | ✅ | ✅ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ |
+|                CONV_TRANSPOSE_2D | ❌ | ✅ | ✅ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | 🟡 |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ✅ | ✅ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | 🟡 |
+|                              DIV | ❌ | ✅ | ✅ | 🟡 |
+|                              DUP | ❌ | ✅ | 🟡 | 🟡 |
+|                              ELU | ❌ | ✅ | ❌ | 🟡 |
+|                              EXP | ❌ | ✅ | 🟡 | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | ✅ | 🟡 | 🟡 |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | 🟡 |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | 🟡 |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | 🟡 |
+|                             GELU | ❌ | ✅ | 🟡 | 🟡 |
+|                         GELU_ERF | ❌ | ✅ | 🟡 | 🟡 |
+|                       GELU_QUICK | ❌ | ✅ | 🟡 | 🟡 |
+|                         GET_ROWS | ❌ | ✅ | 🟡 | ✅ |
+|                    GET_ROWS_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ |
+|                      HARDSIGMOID | ❌ | ✅ | 🟡 | ❌ |
+|                        HARDSWISH | ❌ | ✅ | 🟡 | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | 🟡 |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ |
+|                              LOG | ❌ | ✅ | ✅ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ |
+|                              MUL | ❌ | ✅ | ✅ | 🟡 |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | ✅ | ✅ | ✅ |
+|                              NEG | ❌ | ✅ | 🟡 | 🟡 |
+|                             NORM | ❌ | ✅ | ✅ | 🟡 |
+|                   OPT_STEP_ADAMW | ❌ | ✅ | ✅ | ❌ |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ❌ | ✅ |
+|                          POOL_2D | ❌ | ✅ | ✅ | ✅ |
+|                            REGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             RELU | ❌ | ✅ | 🟡 | 🟡 |
+|                           REPEAT | ❌ | ✅ | 🟡 | ✅ |
+|                      REPEAT_BACK | ❌ | ✅ | ✅ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | 🟡 |
+|                    RMS_NORM_BACK | ❌ | ✅ | ✅ | ❌ |
+|                     RMS_NORM_MUL | ❌ | ✅ | ✅ | ✅ |
+|                             ROPE | ❌ | ✅ | ✅ | ✅ |
+|                        ROPE_BACK | ❌ | ✅ | ✅ | ❌ |
+|                        RWKV_WKV6 | ❌ | ✅ | ✅ | ✅ |
+|                        RWKV_WKV7 | ❌ | ✅ | ✅ | ✅ |
+|                            SCALE | ❌ | ✅ | ✅ | ✅ |
+|                              SET | ❌ | ✅ | ❌ | ✅ |
+|                         SET_ROWS | ❌ | 🟡 | ❌ | 🟡 |
+|                              SGN | ❌ | ✅ | 🟡 | ❌ |
+|                          SIGMOID | ❌ | ✅ | 🟡 | 🟡 |
+|                             SILU | ❌ | ✅ | 🟡 | 🟡 |
+|                        SILU_BACK | ❌ | ✅ | ✅ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | 🟡 |
+|                         SOFT_MAX | ❌ | ✅ | ✅ | ✅ |
+|                    SOFT_MAX_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | 🟡 |
+|                             SQRT | ❌ | ✅ | ✅ | 🟡 |
+|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ |
+|                         SSM_SCAN | ❌ | ✅ | ✅ | ✅ |
+|                             STEP | ❌ | ✅ | 🟡 | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | 🟡 |
+|                              SUM | ❌ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ |
+|                           SWIGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             TANH | ❌ | ✅ | 🟡 | 🟡 |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ |
+|                          UPSCALE | ❌ | ✅ | ✅ | 🟡 |

examples/Miku.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env 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_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/chat-13B.sh

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

examples/chat-persistent.sh

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

examples/chat-vicuna.sh

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

examples/chat.sh

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

examples/embedding/embedding.cpp

@@ -37,12 +37,12 @@ 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_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_encode(ctx, batch) < 0) {
-        LOG_ERR("%s : failed to encode\n", __func__);
+    if (llama_decode(ctx, batch) < 0) {
+        LOG_ERR("%s : failed to process\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -133,10 +133,36 @@ 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) {
-        auto inp = common_tokenize(ctx, prompt, true, true);
+        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);
+        }
         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);
@@ -145,11 +171,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP
+    // check if the last token is SEP/EOS
     // 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)) {
-            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
+        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__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
@@ -236,9 +262,24 @@ 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++) {
-                // 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]);
+                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());
+                    }
+                }
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding

examples/eval-callback/eval-callback.cpp

@@ -55,6 +55,8 @@ 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) {
@@ -134,6 +136,11 @@ 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

@@ -41,12 +41,11 @@ 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 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_self_clear(ctx);
-        llama_set_embeddings(ctx, true);
+        llama_memory_clear(llama_get_memory(ctx), true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -102,8 +101,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_self_clear(ctx);
-    llama_set_embeddings(ctx, false);
+    llama_memory_clear(llama_get_memory(ctx), true);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -166,6 +164,8 @@ 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);
@@ -213,6 +213,8 @@ 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 @@
-#!/bin/bash
+#!/usr/bin/env 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_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         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_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
         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_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         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_kv_self_clear(reinterpret_cast<llama_context *>(context));
+    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
 }

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

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

examples/lookahead/lookahead.cpp

@@ -50,8 +50,6 @@ int main(int argc, char ** argv) {
     const int N = 5;  // n-gram size
     const int G = 15; // max verification n-grams
 
-    const bool dump_kv_cache = params.dump_kv_cache;
-
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -62,6 +60,8 @@ 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 +96,7 @@ 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_kv_self_seq_cp(ctx, 0, s, -1, -1);
+        llama_memory_seq_cp(mem, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -152,9 +152,6 @@ int main(int argc, char ** argv) {
     // here we keep adding new n-grams as we go
     ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);
 
-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
-
     const auto t_dec_start = ggml_time_us();
 
     // sample first token
@@ -172,12 +169,6 @@ int main(int argc, char ** argv) {
     }
 
     while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
         //
         // Example for W = 5, N = 4, G = 2:
@@ -438,17 +429,17 @@ 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_kv_self_seq_rm(ctx, -1, n_past, -1);
+        llama_memory_seq_rm(mem, -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_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);
+            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);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+                llama_memory_seq_cp(mem, 0, s, -1, -1);
             }
         }
     }
@@ -473,8 +464,6 @@ int main(int argc, char ** argv) {
 
     common_sampler_free(smpl);
 
-    llama_kv_cache_view_free(&kvc_view);
-
     llama_batch_free(batch);
 
     llama_backend_free();

examples/lookup/lookup.cpp

@@ -24,8 +24,6 @@ int main(int argc, char ** argv){
     // max. number of additional tokens to draft if match is found
     const int n_draft = params.speculative.n_max;
 
-    const bool dump_kv_cache = params.dump_kv_cache;
-
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -110,18 +108,9 @@ int main(int argc, char ** argv){
 
     llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
 
-    // debug
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
-
     const auto t_dec_start = ggml_time_us();
 
     while (true) {
-        // debug
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         // print current draft sequence
         LOG_DBG("drafted %s\n", string_from(ctx, draft).c_str());
 
@@ -192,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
+        llama_memory_seq_rm(llama_get_memory(ctx), 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/parallel/README.md

@@ -4,7 +4,7 @@ Simplified simulation of serving incoming requests in parallel
 
 ## Example
 
-Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of 10 junk questions (`-j 10`) followed by the actual question.
+Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of up to 10 junk questions (`--junk 10`) followed by the actual question.
 
 ```bash
 llama-parallel -m model.gguf -np 8 -ns 128 --top-k 1 -pps --junk 10 -c 16384

examples/parallel/parallel.cpp

@@ -158,7 +158,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.n_predict = 128;
-    params.n_junk = 0;
+    params.n_junk = 1;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
@@ -178,13 +178,11 @@ int main(int argc, char ** argv) {
     // insert new requests as soon as the previous one is done
     const bool cont_batching = params.cont_batching;
 
-    const bool dump_kv_cache = params.dump_kv_cache;
-
     // is the system prompt shared in the cache
     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 = params.n_junk;
+    const int32_t n_junk = std::max(1, params.n_junk);
 
     // init llama.cpp
     llama_backend_init();
@@ -196,6 +194,8 @@ 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
@@ -241,8 +241,6 @@ int main(int argc, char ** argv) {
     int32_t n_total_gen    = 0;
     int32_t n_cache_miss   = 0;
 
-    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
-
     const auto t_main_start = ggml_time_us();
 
     LOG_INF("%s: Simulating parallel requests from clients:\n", __func__);
@@ -263,7 +261,7 @@ 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_kv_self_seq_cp(ctx, 0, i, -1, -1);
+            llama_memory_seq_cp(mem, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -272,11 +270,6 @@ int main(int argc, char ** argv) {
     LOG_INF("Processing requests ...\n\n");
 
     while (true) {
-        if (dump_kv_cache) {
-            llama_kv_cache_view_update(ctx, &kvc_view);
-            common_kv_cache_dump_view_seqs(kvc_view, 40);
-        }
-
         common_batch_clear(batch);
 
         // decode any currently ongoing sequences
@@ -295,9 +288,9 @@ 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_kv_self_seq_rm(ctx, i, -1, -1);
+                llama_memory_seq_rm(mem, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                llama_memory_seq_cp(mem, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -324,7 +317,10 @@ int main(int argc, char ** argv) {
                     } else {
                         client.prompt += k_system;
                     }
-                    for (int i = 0; i < n_junk; ++i) {
+
+                    const int n_junk_cur = rand() % n_junk;
+
+                    for (int i = 0; i < n_junk_cur; ++i) {
                         const int r = rand() % k_questions.size();
                         client.prompt += "User:\n" + k_questions[r] + "\nAssistant:\n " + k_answers[r] + "\n";
                     }
@@ -349,7 +345,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_INF("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur);
 
                     g_seq_id += 1;
 
@@ -368,7 +364,9 @@ int main(int argc, char ** argv) {
         // process in chunks of params.n_batch
         int32_t n_batch = params.n_batch;
 
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+        int32_t i_next = 0;
+
+        for (int32_t i = 0; i < batch.n_tokens; i = i_next) {
             // experiment: process in powers of 2
             //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
             //    n_batch /= 2;
@@ -376,7 +374,7 @@ int main(int argc, char ** argv) {
             //    continue;
             //}
 
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
 
             llama_batch batch_view = {
                 n_tokens,
@@ -396,19 +394,24 @@ int main(int argc, char ** argv) {
                     return 1;
                 }
 
-                LOG_ERR("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+                LOG_WRN("%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;
@@ -446,8 +449,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
+                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,6 +126,8 @@ 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) {
@@ -133,11 +135,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            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);
+            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);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -167,12 +168,10 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, 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);
+        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);
 
-        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -198,12 +197,10 @@ 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_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);
+            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);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
     }
 

examples/reason-act.sh

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

examples/retrieval/retrieval.cpp

@@ -81,14 +81,14 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+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_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), false);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
     if (llama_decode(ctx, batch) < 0) {
-        LOG_ERR("%s : failed to decode\n", __func__);
+        LOG_ERR("%s : failed to process\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -233,7 +233,7 @@ int main(int argc, char ** argv) {
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
             float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd);
+            batch_process(ctx, batch, out, s, n_embd);
             common_batch_clear(batch);
             p += s;
             s = 0;
@@ -246,7 +246,7 @@ int main(int argc, char ** argv) {
 
     // final batch
     float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd);
+    batch_process(ctx, batch, out, s, n_embd);
 
     // save embeddings to chunks
     for (int i = 0; i < n_chunks; i++) {
@@ -267,7 +267,7 @@ int main(int argc, char ** argv) {
         batch_add_seq(query_batch, query_tokens, 0);
 
         std::vector<float> query_emb(n_embd, 0);
-        batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
+        batch_process(ctx, query_batch, query_emb.data(), 1, n_embd);
 
         common_batch_clear(query_batch);
 

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_kv_self_clear(ctx3);
+        llama_memory_clear(llama_get_memory(ctx3), true);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/server-llama2-13B.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env 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_kv_self_used_cells(ctx) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,15 +113,16 @@ 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_kv_self_used_cells(ctx);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");
                 exit(0);
             }
 
-            if (llama_decode(ctx, batch)) {
-                GGML_ABORT("failed to decode\n");
+            int ret = llama_decode(ctx, batch);
+            if (ret != 0) {
+                GGML_ABORT("failed to decode, ret = %d\n", ret);
             }
 
             // 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_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -142,6 +142,8 @@ 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;
@@ -420,14 +422,14 @@ 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_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_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_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);
+                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);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +446,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_memory_seq_rm(mem_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);
 
@@ -503,8 +505,8 @@ 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_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
+                        llama_memory_seq_cp(mem_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) {
@@ -585,9 +587,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_self_seq_keep(ctx_tgt, 0);
+            llama_memory_seq_keep(mem_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_memory_seq_cp(mem_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 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
@@ -12,16 +12,16 @@ source /opt/intel/oneapi/setvars.sh
 
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
 MODEL_FILE=models/llama-2-7b.Q4_0.gguf
-NGL=33
-CONEXT=4096
+NGL=99
+CONTEXT=4096
 
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/run-llama3.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+#  MIT license
+#  Copyright (C) 2025 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+# If you want more control, DPC++ Allows selecting a specific device through the
+# following environment variable
+#export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
+NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
+CONTEXT=4096
+
+if [ $# -gt 0 ]; then
+    GGML_SYCL_DEVICE=$1
+    echo "Using $GGML_SYCL_DEVICE as the main GPU"
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+else
+    #use multiple GPUs with same max compute units
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -c ${CONTEXT}
+fi

examples/sycl/win-run-llama2.bat

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0

examples/sycl/win-run-llama3.bat

@@ -0,0 +1,9 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+
+.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -e -ngl 99

examples/training/README.md

@@ -10,8 +10,8 @@ Proof of concept:
 
 ``` sh
 export model_name=llama_3.2-1b && export quantization=f32
-./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
-./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+./build/bin/llama-finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/llama-perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
 ```
 
 The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

examples/ts-type-to-grammar.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env 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_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_CPU_REPACK       "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})
@@ -129,14 +129,16 @@ option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 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 (WIN32)
+if (MINGW)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
@@ -171,15 +173,14 @@ 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)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
-option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 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)
@@ -264,7 +265,6 @@ 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
@@ -358,6 +358,13 @@ 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)
@@ -367,6 +374,8 @@ 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
     )
@@ -386,4 +395,46 @@ 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

@@ -24,3 +24,27 @@ function(ggml_get_flags CCID CCVER)
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
+
+function(ggml_get_system_arch)
+    if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+        CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+        set(GGML_SYSTEM_ARCH "ARM" PARENT_SCOPE)
+    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
+            CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
+        set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
+        set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+        set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+        set(GGML_SYSTEM_ARCH "riscv64" PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+        set(GGML_SYSTEM_ARCH "s390x" PARENT_SCOPE)
+    else()
+        set(GGML_SYSTEM_ARCH "UNKNOWN" PARENT_SCOPE)
+    endif()
+endfunction()

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);
+    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);
 
     // 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

@@ -101,6 +101,7 @@ 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);
 
@@ -133,6 +134,7 @@ 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

@@ -1,50 +0,0 @@
-#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

@@ -314,6 +314,13 @@
 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, ...);
 
@@ -470,6 +477,7 @@ 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,
@@ -481,14 +489,16 @@ 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,
         GGML_OP_POOL_2D,
         GGML_OP_POOL_2D_BACK,
-        GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_UPSCALE,
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
+        GGML_OP_ROLL,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -518,6 +528,8 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
 
+        GGML_OP_GLU,
+
         GGML_OP_COUNT,
     };
 
@@ -536,10 +548,21 @@ extern "C" {
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
+        GGML_UNARY_OP_GELU_ERF,
 
         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,
@@ -625,6 +648,9 @@ 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);
@@ -655,6 +681,7 @@ 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);
@@ -685,6 +712,9 @@ 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);
 
@@ -756,6 +786,7 @@ 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);
@@ -934,6 +965,15 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // repeat a to the specified shape
+    GGML_API struct ggml_tensor * ggml_repeat_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+                       int64_t    ne0,
+                       int64_t    ne1,
+                       int64_t    ne2,
+                       int64_t    ne3);
+
     // sums repetitions in a into shape of b
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
@@ -1024,6 +1064,16 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // GELU using erf (error function) when possible
+    // some backends may fallback to approximation based on Abramowitz and Stegun formula
+    GGML_API struct ggml_tensor * ggml_gelu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_erf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu_quick(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1065,6 +1115,89 @@ 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,
@@ -1164,6 +1297,19 @@ extern "C" {
             struct ggml_tensor  * a,
             float                 s);
 
+    // x = s * a + b
+    GGML_API struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    GGML_API struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
             struct ggml_context * ctx,
@@ -1354,6 +1500,23 @@ 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);
@@ -1391,8 +1554,14 @@ 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,
@@ -1702,6 +1871,17 @@ 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,
@@ -1744,6 +1924,12 @@ 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
@@ -1756,14 +1942,26 @@ extern "C" {
 
     // interpolate
     // interpolate scale to specified dimensions
-    GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_DEPRECATED(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);
+            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...]
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1781,6 +1979,17 @@ 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]
@@ -1815,11 +2024,17 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // 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 !!
+    // 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
+    //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1858,7 +2073,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -2075,9 +2291,6 @@ 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);
 
@@ -2161,6 +2374,7 @@ extern "C" {
 
     // scheduling priorities
     enum ggml_sched_priority {
+        GGML_SCHED_PRIO_LOW = -1,
         GGML_SCHED_PRIO_NORMAL,
         GGML_SCHED_PRIO_MEDIUM,
         GGML_SCHED_PRIO_HIGH,

ggml/src/CMakeLists.txt

@@ -109,6 +109,8 @@ if (MSVC)
 else ()
     set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
+ggml_get_system_arch()
+message(STATUS "GGML_SYSTEM_ARCH: ${GGML_SYSTEM_ARCH}")
 
 if (NOT MSVC)
     if (GGML_STATIC)
@@ -123,7 +125,6 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 
@@ -194,6 +195,7 @@ add_library(ggml-base
             ../include/ggml-opt.h
             ../include/gguf.h
             ggml.c
+            ggml.cpp
             ggml-alloc.c
             ggml-backend.cpp
             ggml-opt.cpp
@@ -210,6 +212,7 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
+add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -224,6 +227,7 @@ 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})
@@ -266,17 +270,27 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
+    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 ${ARGN})
-        set(GGML_${feat} ON)
-    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()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -286,17 +300,62 @@ 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()
-    ggml_add_cpu_backend_variant(x64)
-    ggml_add_cpu_backend_variant(sse42        SSE42)
-    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
-    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
-    if (NOT MSVC)
-        # 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)
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        ggml_add_cpu_backend_variant(x64)
+        ggml_add_cpu_backend_variant(sse42        SSE42)
+        ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+        ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+        ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+        ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+        ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+        if (NOT MSVC)
+            # 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}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")
@@ -306,7 +365,6 @@ 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

@@ -61,14 +61,13 @@
 #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;
@@ -91,6 +90,8 @@ 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
@@ -184,9 +185,6 @@ 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
@@ -570,7 +568,6 @@ 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

@@ -817,8 +817,9 @@ 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]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(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)]);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1340,7 +1341,10 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
-        ggml_backend_sched_synchronize(sched);
+        // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
+        for (int i = 0; i < sched->n_backends; i++) {
+            ggml_backend_synchronize(sched->backends[i]);
+        }
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
@@ -1564,7 +1568,6 @@ bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgra
 
     ggml_backend_sched_split_graph(sched, graph);
 
-
     if (!ggml_backend_sched_alloc_splits(sched)) {
         return false;
     }
@@ -1598,6 +1601,12 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
     for (int i = 0; i < sched->n_backends; i++) {
         ggml_backend_synchronize(sched->backends[i]);
     }
+    if (!sched->is_alloc) {
+        // if the graph is not already allocated, always use copy 0 after a synchronization
+        // this ensures that during generation the same copy is used every time,
+        // which avoids changes in the graph that could cause CUDA or other graphs to be disabled
+        sched->cur_copy = 0;
+    }
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
@@ -1818,7 +1827,7 @@ 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) {
+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) {
     struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
     if (copy.buffer == NULL) {
         return false;
@@ -1829,28 +1838,45 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     assert(g1->n_nodes == g2->n_nodes);
 
-    for (int i = 0; i < g1->n_nodes; i++) {
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
+    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);
 
-        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;
+        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;
+            }
         }
+        GGML_ASSERT(test_node_idx != -1);
 
-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
+        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;
+            }
         }
     }
-
     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(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(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")
 endif()

ggml/src/ggml-cann/CMakeLists.txt

@@ -30,6 +30,7 @@ string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
 string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
 set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
 string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
+message(STATUS "CANN: SOC_VERSION =  ${SOC_VERSION}")
 
 if (CANN_INSTALL_DIR)
     # Only Support Linux.

ggml/src/ggml-cann/acl_tensor.cpp

@@ -31,6 +31,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_FLOAT;
         case GGML_TYPE_F16:
             return ACL_FLOAT16;
+        case GGML_TYPE_BF16:
+            return ACL_BF16;
         case GGML_TYPE_I8:
             return ACL_INT8;
         case GGML_TYPE_I16:

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,7 +65,9 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v2.h>
+#include <aclnnop/aclnn_grouped_matmul_v3.h>
+#include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_zero.h>
 #include <float.h>
 
 #include <cmath>
@@ -74,11 +76,13 @@
 #include <vector>
 
 #include "ggml-impl.h"
+#include "ggml.h"
 
 #define GGML_COMMON_DECL_C
 
 #include "../ggml-common.h"
 
+
 void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclTensor ** acl_src0,
                  aclTensor ** acl_src1, aclTensor ** acl_dst) {
     GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_can_repeat(src1, src0));
@@ -801,10 +805,11 @@ 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);
 }
 
 /**
@@ -2651,6 +2656,67 @@ 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;
@@ -2697,14 +2763,10 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         }
     }
 
-    // GroupedMatmulV2 required tensor_list.size < 128
     size_t GROUP_SIZE = 128;
-    std::vector<std::vector<aclTensor*>> src0_tensor_vec_vec;
-    std::vector<std::vector<aclTensor*>> src1_tensor_vec_vec;
-    std::vector<std::vector<aclTensor*>> dst_tensor_vec_vec;
-
-    // split and call GroupedMatmulV2
+    // GroupedMatmulV3 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
+        // split and call GroupedMatmulV3
         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);
@@ -2714,7 +2776,7 @@ 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, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, 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);
@@ -2722,6 +2784,133 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     return;
 }
 
+/**
+ * @brief Performs expert-specific matrix multiplication (MoE) with
+ * quantized precision using the CANN backend.
+ *
+ * This function executes a matrix multiplication operation tailored for
+ * Mixture of Experts (MoE) models, where the input tensor is multiplied
+ * with expert-specific quantized weight matrices. It leverages the CANN
+ * backend to perform efficient low-precision computations and stores the
+ * quantized result in the destination tensor `dst`.
+ *
+ * Quantization techniques reduce memory footprint and improve performance
+ * by using lower-bit representations (e.g., int8) instead of floating-point.
+ * This function is designed to work with such formats and may incorporate
+ * optimizations like identity-based fast paths or routing masks for sparse
+ * expert selection.
+ *
+ * @param ctx The context for executing CANN backend operations.
+ * @param dst The destination tensor where the quantized MoE multiplication result
+ * will be stored.
+ *
+ * @note This function assumes quantized data types and is designed for
+ * MoE architectures with potential sparse expert routing.
+ */
+static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    // TODO: Use aclnnGroupedMatMul
+    //dst   [M, K, N, 1]
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // copy index from npu to cpu
+    int64_t n_as = ne02; // A
+    int64_t n_ids = ids->ne[0]; // K
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
+        ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    const enum ggml_type type = dst->src[0]->type;
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    } else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    } else {
+        GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = weight_elem_size;
+    src0_row.nb[1] = weight_elem_size * ne00;
+    src0_row.nb[2] = weight_elem_size * ne00;
+    src0_row.nb[3] = weight_elem_size * ne00;
+    size_t weight_stride = ne00 * ne01 * weight_elem_size;
+    size_t weight_size = weight_stride * ne02 * ne03;
+
+    // scale [D, M, 1, 1] -> scale && permute
+    size_t scale_elem_size = sizeof(uint16_t);
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
+
+    // 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
+    ggml_cann_pool_alloc weight_allocator(ctx.pool());
+    void* weight_buffer = weight_allocator.alloc(nb02);
+    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*weight_stride;
+            void* scale_tmp_ptr = src0_original + weight_size + i02*scale_stride;
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            // mem cpy
+            ggml_cann_async_memcpy(ctx, weight_buffer, src0_tmp_ptr, weight_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+            void* scale_buffer = (char*)weight_buffer + weight_stride;
+            ggml_cann_async_memcpy(ctx, scale_buffer, scale_tmp_ptr, scale_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+
+            src0_row.data = weight_buffer;
+            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;
+}
+
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const enum ggml_type type = dst->src[0]->type;
     switch (type) {
@@ -2729,8 +2918,339 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16:
             ggml_cann_mul_mat_id_fp(ctx, dst);
             break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            ggml_cann_mul_mat_id_quant(ctx, dst);
+            break;
         default:
             GGML_ABORT("Unsupported type for mul_mat_id");
             break;
     }
 }
+
+void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
+
+    ggml_tensor* src0 = dst->src[0]; // q, fp32
+    ggml_tensor* src1 = dst->src[1]; // k, fp16
+    ggml_tensor* src2 = dst->src[2]; // v, fp16
+    ggml_tensor* src3 = dst->src[3]; // mask, fp16
+
+    float maxBias = 0.0f;
+    float scaleValue = 1.0f;
+    float logitSoftcap = 0.0f;
+    memcpy(&scaleValue,    (float*)dst->op_params + 0, sizeof(float));
+    memcpy(&maxBias,       (float*)dst->op_params + 1, sizeof(float));
+    memcpy(&logitSoftcap,  (float*)dst->op_params + 2, sizeof(float));
+
+    if(logitSoftcap == 0.0f){
+        size_t faElemSize = sizeof(uint16_t);
+        auto   faDataType = ACL_FLOAT16; //ACL_BF16;
+
+        aclTensor* acl_src0_f16_tensor = nullptr;
+        aclTensor* acl_src1_f16_tensor = nullptr;
+        aclTensor* acl_src2_f16_tensor = nullptr;
+        aclTensor* acl_dst_f16_tensor  = nullptr;
+
+        // Step 1: cast the src0 (Query) to fp16 if needed
+        ggml_cann_pool_alloc src0_f16_allocator(ctx.pool());
+        void* src0_f16_buffer = nullptr;
+
+        if(ggml_cann_type_mapping(src0->type) != faDataType){
+            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0);
+            src0_f16_buffer = src0_f16_allocator.alloc(
+                                    ggml_nelements(src0) * faElemSize);
+
+            int64_t* src0_f16_ne = src0->ne;
+            size_t   src0_f16_nb[GGML_MAX_DIMS];
+            src0_f16_nb[0] = sizeof(uint16_t);
+            for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                src0_f16_nb[i] = src0_f16_nb[i - 1] * src0_f16_ne[i - 1];
+            }
+
+            acl_src0_f16_tensor = ggml_cann_create_tensor(
+                src0_f16_buffer, faDataType, faElemSize,
+                src0_f16_ne, src0_f16_nb, GGML_MAX_DIMS
+            );
+            aclnn_cast(ctx, acl_src0_f32_tensor, acl_src0_f16_tensor, faDataType);
+            ggml_cann_release_resources(ctx, acl_src0_f32_tensor);
+        }else{
+            acl_src0_f16_tensor = ggml_cann_create_tensor(src0);
+        }
+
+        // Step 2: create the acl tensors for src1 (Key), src2 (Value),
+        //         and the direct output from FusedInferAttention
+
+        acl_src1_f16_tensor = ggml_cann_create_tensor(src1);
+        acl_src2_f16_tensor = ggml_cann_create_tensor(src2);
+
+        ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
+        void* out_f16_buffer = out_f16_allocator.alloc(
+                                    ggml_nelements(dst) * faElemSize);
+
+        int64_t* out_f16_ne = src0->ne;
+        size_t out_f16_nb[GGML_MAX_DIMS];
+        out_f16_nb[0] = faElemSize;
+        for(int i = 1; i < GGML_MAX_DIMS; ++i){
+            out_f16_nb[i] = out_f16_nb[i - 1] * out_f16_ne[i - 1];
+        }
+
+        acl_dst_f16_tensor = ggml_cann_create_tensor(
+            out_f16_buffer, faDataType, faElemSize,
+            out_f16_ne, out_f16_nb, GGML_MAX_DIMS
+        );
+
+        // Step 3: create the PSEShift tensor if needed
+        //         this tensor is considered as mask (f16) in the llama.cpp
+
+        aclTensor* bcast_pse_tensor = nullptr;
+        int64_t bcast_pse_ne[GGML_MAX_DIMS];
+        size_t bcast_pse_nb[GGML_MAX_DIMS];
+        ggml_cann_pool_alloc bcast_pse_allocator(ctx.pool());
+        void* bcast_pse_buffer = nullptr;
+
+        if(src3 != nullptr){
+            bcast_pse_buffer = bcast_pse_allocator.alloc(
+                            ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t));
+
+            if(src0->ne[1] > 1){
+                // Case 1: broadcast pse for prefill stage with multiple head
+                aclTensor* acl_mask_f16_tensor = ggml_cann_create_tensor(src3);
+                bcast_pse_ne[0] = src3->ne[0];
+                bcast_pse_ne[1] = src3->ne[1];
+                bcast_pse_ne[2] = src0->ne[2];
+                bcast_pse_ne[3] = src3->ne[3];
+
+                bcast_pse_nb[0] = sizeof(uint16_t);
+                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
+                }
+
+                bcast_pse_tensor = ggml_cann_create_tensor(
+                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+
+                int64_t repeats[] = {1, src0->ne[2], 1, 1};
+                aclnn_repeat(ctx, acl_mask_f16_tensor, bcast_pse_tensor, repeats);
+
+                ggml_cann_release_resources(ctx, acl_mask_f16_tensor);
+            }else{
+                // Case 2: trunc the first row and broadcast pse for decode stage with multiple head
+                int64_t trunc_pse_ne[GGML_MAX_DIMS] = {src3->ne[0], src0->ne[1], src3->ne[2], src3->ne[3]};
+                size_t* trunc_pse_nb = src3->nb;
+
+                aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
+                    src3->data, ACL_FLOAT16, sizeof(uint16_t),
+                    trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS);
+
+                bcast_pse_ne[0] = src3->ne[0];
+                bcast_pse_ne[1] = src0->ne[1];
+                bcast_pse_ne[2] = src0->ne[2];
+                bcast_pse_ne[3] = src3->ne[3];
+
+                bcast_pse_nb[0] = sizeof(uint16_t);
+                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
+                }
+
+                bcast_pse_tensor = ggml_cann_create_tensor(
+                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+
+                int64_t repeats[] = {1, src0->ne[2], 1, 1};
+                aclnn_repeat(ctx, acl_mask_f16_trunc_tensor, bcast_pse_tensor, repeats);
+
+                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
+            }
+
+            // Compute the slope if needed. Derived from ggml_cann_softmax().
+            if(maxBias != 0.0f){
+                // alibi
+                const int64_t ne2_ne3 = src0->ne[2] * src0->ne[3];
+                const int64_t n_head = src0->ne[2];
+                const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
+                float m0 = powf(2.0f, -(maxBias) / n_heads_log2_floor);
+                float m1 = powf(2.0f, -(maxBias / 2.0f) / n_heads_log2_floor);
+                // init arange
+                ggml_cann_pool_alloc arange_allocator(ctx.pool(),
+                                                    ne2_ne3 * faElemSize);
+                void* tmp_arange_buffer = arange_allocator.get();
+
+                // arange1: [1, ..., n_heads_log2_floor+1)
+                float start = 1;
+                float stop = n_heads_log2_floor + 1;
+                float step = 1;
+                int64_t n_elements_arange = n_heads_log2_floor;
+
+                int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
+                size_t tmp_arange1_nb[] = {faElemSize};
+                aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
+                    tmp_arange_buffer, faDataType, faElemSize,
+                    tmp_arange1_ne, tmp_arange1_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+                aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
+
+                aclTensor* tmp_arange2_tensor = nullptr;
+                if (n_heads_log2_floor < ne2_ne3) {
+                    // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
+                    start = 1;
+                    stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
+                    step = 2;
+                    n_elements_arange = ne2_ne3 - n_heads_log2_floor;
+                    int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+                    size_t tmp_arange2_nb[] = {faElemSize};
+
+                    aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
+                        (char*)tmp_arange_buffer +
+                            n_heads_log2_floor * faElemSize,
+                        faDataType, faElemSize,
+                        tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                    aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
+                                n_elements_arange);
+                }
+
+                // init mk_base
+                ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
+                                                    ne2_ne3 * faElemSize);
+                void* tmp_mk_base_buffer = mk_base_allocator.get();
+                int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
+                size_t tmp_mk_base1_nb[] = {faElemSize};
+                aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_base1_ne, tmp_mk_base1_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+                aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
+
+                aclTensor* tmp_mk_base2_tensor = nullptr;
+                if (n_heads_log2_floor < ne2_ne3) {
+                    int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+                    size_t tmp_mk_base2_nb[] = {faElemSize};
+                    aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
+                        (char*)tmp_mk_base_buffer +
+                            n_heads_log2_floor * faElemSize,
+                        faDataType, faElemSize,
+                        tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                    aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
+                }
+
+                // init mk
+                int64_t tmp_mk_base_ne[] = {ne2_ne3};
+                size_t tmp_mk_base_nb[] = {faElemSize};
+                aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_base_ne, tmp_mk_base_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
+                    tmp_arange_buffer, faDataType, faElemSize,
+                    tmp_mk_base_ne, tmp_mk_base_nb,
+                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+                aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
+
+                // reshape mk
+                int64_t tmp_mk_ne[] = {1, 1, src0->ne[2], src0->ne[3]};
+                size_t tmp_mk_nb[GGML_MAX_DIMS];
+                tmp_mk_nb[0] = faElemSize;
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                    tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
+                }
+                aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
+                    tmp_mk_base_buffer, faDataType, faElemSize,
+                    tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
+                    ACL_FORMAT_ND);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, tmp_mk_tensor);
+
+                ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
+                    tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
+                    tmp_arange_tensor, tmp_mk_tensor);
+            }
+        }
+
+        // Step 4: set the inputs for FusedInferAttention.
+        int kvTensorNum = 1;
+        aclTensor* acl_q_tensor = acl_src0_f16_tensor;
+        aclTensor* acl_k_tensors[] = {acl_src1_f16_tensor};
+        aclTensor* acl_v_tensors[] = {acl_src2_f16_tensor};
+        auto acl_k_tensor_list = aclCreateTensorList(acl_k_tensors, kvTensorNum);
+        auto acl_v_tensor_list = aclCreateTensorList(acl_v_tensors, kvTensorNum);
+
+        int64_t numHeads = src0->ne[2]; // N
+        int64_t numKeyValueHeads = src1->ne[2];
+        // double  scaleValue = 1 / sqrt(src0->ne[0]); // 1/sqrt(d)
+        int64_t preTokens = 65535;
+        int64_t nextTokens = 65535;
+        char layout[5] = {'B', 'N', 'S', 'D', 0};
+        int64_t sparseMode = 0;
+        int64_t innerPrecise = (src0->ne[1] == 1) ? 0 : 2;
+        int64_t blockSize = 0;
+        int64_t antiquantMode = 0;
+        bool softmaxLseFlag = false;
+        int64_t keyAntiquantMode = 0;
+        int64_t valueAntiquantMode = 0;
+
+        // Step 5: launch the FusedInferAttentionScoreV2 kernel.
+        // Refer to https://gitee.com/ascend/cann-ops-adv/blob/master/docs/FusedInferAttentionScoreV2.md
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, FusedInferAttentionScoreV2,
+            acl_q_tensor, acl_k_tensor_list, acl_v_tensor_list, // q, k, v
+            bcast_pse_tensor, nullptr, // pse, mask
+            nullptr, nullptr, // actSeqLen, actSeqLenkv
+            nullptr, nullptr, // deqScale1, quantScale1
+            nullptr, nullptr, nullptr, // deqScale2, quantScale2, quantOffset2
+            nullptr, nullptr, // antiquantScale, antiquantOffset
+            nullptr, // blockTable
+            nullptr, nullptr, // qPadSize, kvPadSize
+            nullptr, nullptr, // kAntiquantScale, kAntiQuantOffset
+            nullptr, nullptr, // vAntiquantScale, vAntiQuantOffset
+            nullptr, nullptr, nullptr, // kSharedPrefix, vSharedPrefix, actSharedLen
+            numHeads, scaleValue, // heads, scaleValue
+            preTokens, nextTokens, // preTokens, nextTokens
+            layout, // inputLayout
+            numKeyValueHeads, // numKVHeads
+            sparseMode, innerPrecise, // sparseMode, innerPrecise
+            blockSize, antiquantMode, // blockSize, antiquantMode
+            softmaxLseFlag, // softmaxLseFlag
+            keyAntiquantMode, valueAntiquantMode, // keyAntiqMode, valueAntiqMode
+            acl_dst_f16_tensor, // attentionOut
+            nullptr // softmaxLse
+        );
+
+        // Step 6: post-processing, permute and cast to f32
+
+        int64_t new_dim[] = {0, 2, 1, 3};
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+
+        if(ggml_cann_type_mapping(dst->type) != faDataType){
+            ggml_cann_pool_alloc perm_out_f16_allocator(ctx.pool());
+            perm_out_f16_allocator.alloc(ggml_nelements(dst) * faElemSize);
+            void* perm_out_f16_buffer = perm_out_f16_allocator.get();
+
+            int64_t* perm_out_f16_ne = dst->ne;
+            size_t  perm_out_f16_nb[GGML_MAX_DIMS];
+            perm_out_f16_nb[0] = faElemSize;
+            for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                perm_out_f16_nb[i] = perm_out_f16_nb[i - 1] * perm_out_f16_ne[i - 1];
+            }
+            aclTensor* acl_perm_out_f16_tensor = ggml_cann_create_tensor(
+                perm_out_f16_buffer, faDataType, faElemSize,
+                perm_out_f16_ne, perm_out_f16_nb, GGML_MAX_DIMS);
+            aclnn_permute(ctx, acl_dst_f16_tensor, acl_perm_out_f16_tensor, new_dim, GGML_MAX_DIMS);
+            aclnn_cast(ctx,
+                acl_perm_out_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
+            ggml_cann_release_resources(ctx, acl_perm_out_f16_tensor);
+        }else{
+            // only need to permute
+            aclnn_permute(ctx, acl_dst_f16_tensor, acl_dst_tensor, new_dim, GGML_MAX_DIMS);
+        }
+        ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
+                                         acl_src1_f16_tensor,
+                                         acl_src2_f16_tensor,
+                                         acl_dst_f16_tensor,
+                                         acl_dst_tensor);
+        if(src3 != nullptr){
+            ggml_cann_release_resources(ctx, bcast_pse_tensor);
+        }
+    }else{
+        GGML_ABORT("Function is not implemented.");
+    }
+}