speculative : adapt to new llama API

ggml-ci

.clang-tidy

@@ -13,7 +13,6 @@ Checks: >
     -readability-magic-numbers,
     -readability-uppercase-literal-suffix,
     -readability-simplify-boolean-expr,
-    -readability-math-missing-parentheses,
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,

.devops/cpu.Dockerfile

@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -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_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/intel.Dockerfile

@@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -1,4 +1,4 @@
-ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10
+ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
 
 FROM ascendai/cann:$ASCEND_VERSION AS build
 
@@ -6,7 +6,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN yum install -y gcc g++ cmake make libcurl-devel
+RUN yum install -y gcc g++ cmake make
 ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
 ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
 ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}

.devops/musa.Dockerfile

@@ -35,7 +35,7 @@ COPY . .
 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_CURL=ON -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 -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/rocm.Dockerfile

@@ -17,8 +17,8 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # gfx906 is deprecated
 #check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
 
-ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
-#ARG ROCM_DOCKER_ARCH=gfx1100
+#ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
+ARG ROCM_DOCKER_ARCH=gfx1100
 
 # Set nvcc architectured
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .
 
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -16,7 +16,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

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

@@ -1,25 +0,0 @@
-name: 'Windows - Setup CURL'
-description: 'Composite action, to be reused in other workflow'
-inputs:
-  curl_version:
-    description: 'CURL version'
-    required: false
-    default: '8.6.0_6'
-outputs:
-  curl_path:
-    description: "Path to the downloaded libcurl"
-    value: ${{ steps.get_libcurl.outputs.curl_path }}
-
-runs:
-  using: "composite"
-  steps:
-    - name: libCURL
-      id: get_libcurl
-      shell: powershell
-      env:
-        CURL_VERSION: ${{ inputs.curl_version }}
-      run: |
-        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
-        mkdir $env:RUNNER_TEMP/libcurl
-        tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
-        echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

.github/workflows/bench.yml.disabled

@@ -104,6 +104,7 @@ jobs:
           cmake -B build \
               -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
               -DLLAMA_CUBLAS=ON \
               -DCUDAToolkit_ROOT=/usr/local/cuda \
               -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \

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

@@ -1,124 +0,0 @@
-name: Build on Linux using cross-compiler
-on:
-  workflow_dispatch:
-  workflow_call:
-
-jobs:
-  ubuntu-latest-riscv64-cpu-cross:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  libcurl4-openssl-dev:riscv64
-
-      - name: Build
-        run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-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)
-
-  ubuntu-latest-riscv64-vulkan-cross:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  libvulkan-dev:riscv64 \
-                  libcurl4-openssl-dev:riscv64
-
-      - name: Build
-        run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-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)
-
-  ubuntu-latest-arm64-vulkan-cross:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Setup Arm64
-        run: |
-          sudo dpkg --add-architecture arm64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  crossbuild-essential-arm64 \
-                  libvulkan-dev:arm64 \
-                  libcurl4-openssl-dev:arm64
-
-      - name: Build
-        run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
-                         -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
-                         -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-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

@@ -10,7 +10,7 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
   pull_request:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
@@ -54,7 +54,6 @@ jobs:
         continue-on-error: true
         run: |
           brew update
-          brew install curl
 
       - name: Build
         id: cmake_build
@@ -63,6 +62,7 @@ jobs:
           cmake -B build \
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DGGML_RPC=ON
@@ -92,6 +92,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -122,7 +123,6 @@ jobs:
         continue-on-error: true
         run: |
           brew update
-          brew install curl
 
       - name: Build
         id: cmake_build
@@ -133,6 +133,7 @@ jobs:
           cmake -B build \
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@@ -161,6 +162,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -205,6 +207,7 @@ jobs:
         run: |
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -243,6 +246,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -277,7 +281,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Build
         id: cmake_build
@@ -318,7 +322,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Build
         id: cmake_build
@@ -356,7 +360,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Build
         id: cmake_build
@@ -393,7 +397,7 @@ jobs:
           wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
           sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk
 
       - name: Build
         id: cmake_build
@@ -427,6 +431,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -449,7 +454,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -525,7 +530,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
 
       - name: install oneAPI MKL library
         shell: bash
@@ -573,7 +578,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
 
       - name: install oneAPI MKL library
         shell: bash
@@ -601,10 +606,6 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-# Disabled for now due to sporadic issue syncing.
-#  build-linux-cross:
-#    uses: ./.github/workflows/build-linux-cross.yml
-
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -632,7 +633,6 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -668,7 +668,6 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -677,36 +676,6 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-  macOS-latest-cmake-visionos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=visionOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
   macOS-latest-swift:
     runs-on: macos-latest
 
@@ -738,7 +707,6 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_CURL=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -806,7 +774,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.304.1
 
     strategy:
       matrix:
@@ -899,17 +867,10 @@ 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
-
       - 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"
+          cmake -S . -B build ${{ matrix.defines }}
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add libopenblas.dll
@@ -969,10 +930,9 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
+          Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
+          Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -998,7 +958,7 @@ jobs:
             DEBIAN_FRONTEND: noninteractive
           run: |
               apt update
-              apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
+              apt install -y cmake build-essential ninja-build libgomp1 git
 
         - name: ccache
           uses: hendrikmuhs/ccache-action@v1.2.16
@@ -1100,23 +1060,16 @@ 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"
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
+            -DGGML_RPC=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1137,10 +1090,7 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        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-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -1195,8 +1145,6 @@ 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
@@ -1282,14 +1230,8 @@ jobs:
           key: ${{ github.job }}
           evict-old-files: 1d
 
-      - 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}"
@@ -1300,11 +1242,9 @@ jobs:
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_RPC=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+            -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
-  # TODO: reuse windows-latest-cmake-hip instead of duplicating this job
   windows-latest-cmake-hip-release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
     runs-on: windows-latest
@@ -1346,14 +1286,8 @@ 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}"
@@ -1365,8 +1299,7 @@ jobs:
             -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
-            -DGGML_RPC=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+            -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
@@ -1388,10 +1321,7 @@ jobs:
 
       - 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\*
 
       - name: Upload artifacts
@@ -1416,7 +1346,6 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_CURL=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -1767,17 +1696,16 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
-        shell: bash -el {0}
+       shell: bash -el {0}
+    runs-on: ubuntu-24.04-arm
     strategy:
       matrix:
-        arch: [x86, aarch64]
         cann:
-          - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
+          - '8.0.rc3.beta1-910b-openeuler22.03-py3.10'
         device:
           - 'ascend910b3'
         build:
           - 'Release'
-    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     container: ascendai/cann:${{ matrix.cann }}
     steps:
       - name: Checkout
@@ -1786,7 +1714,7 @@ jobs:
       - name: Dependencies
         run: |
           yum update -y
-          yum install -y git gcc gcc-c++ make cmake libcurl-devel
+          yum install -y git gcc gcc-c++ make cmake
 
       - name: Build
         run: |

.github/workflows/docker.yml

@@ -36,13 +36,13 @@ jobs:
       matrix:
         config:
           # Multi-stage build
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
-          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
-          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
+          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

.github/workflows/server.yml

@@ -129,6 +129,7 @@ jobs:
           cmake -B build \
               -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
               -DGGML_OPENMP=OFF ;
@@ -141,6 +142,7 @@ jobs:
           cmake -B build \
               -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
@@ -152,6 +154,7 @@ jobs:
           cmake -B build \
               -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
@@ -192,14 +195,17 @@ jobs:
 
       - name: libCURL
         id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
+        env:
+          CURL_VERSION: 8.6.0_6
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+          mkdir $env:RUNNER_TEMP/libcurl
+          tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
 
       - name: Build
         id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
+          cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -215,10 +221,8 @@ jobs:
 
       - name: Copy Libcurl
         id: prepare_libcurl
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
 
       - name: Tests
         id: server_integration_tests

CMakeLists.txt

@@ -81,7 +81,7 @@ option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 
 # 3rd party libs
-option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
+option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 # Required for relocatable CMake package
@@ -168,11 +168,6 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #
 
-if (NOT LLAMA_BUILD_COMMON)
-    message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL")
-    set(LLAMA_CURL OFF)
-endif()
-
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
 endif()
@@ -247,20 +242,3 @@ configure_file(cmake/llama.pc.in
 
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
-    message(STATUS "Running inside GitHub Actions - copying license files")
-
-    # Copy all files from licenses/ to build/bin/
-    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
-    foreach(LICENSE_FILE ${LICENSE_FILES})
-        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
-        configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY)
-    endforeach()
-endif()
-

Makefile

@@ -780,6 +780,10 @@ ifdef GGML_HIP
 
 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
+ifdef GGML_HIP_UMA
+	MK_CPPFLAGS += -DGGML_HIP_UMA
+endif # GGML_HIP_UMA
+
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas

README.md

@@ -9,6 +9,13 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
+> [!IMPORTANT]
+> New `llama.cpp` package location: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pkgs/container/llama.cpp)
+>
+> Update your container URLs to: `ghcr.io/ggml-org/llama.cpp`
+>
+> More info: https://github.com/ggml-org/llama.cpp/discussions/11801
+
 ## Recent API changes
 
 - [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
@@ -16,9 +23,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **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` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
-- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
+- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
+- **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
@@ -98,7 +104,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
 - [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
-- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e)
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
@@ -107,8 +112,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [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)
 
 #### Multimodal
 
@@ -242,7 +245,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |
 
 ## Building the project
 
@@ -261,9 +263,7 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
-
-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/`.
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from Hugging Face by using this CLI argument: `-hf <user>/<model>[:quant]`
 
 After downloading a model, use the CLI tools to run it locally - see below.
 
@@ -528,35 +528,6 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-## XCFramework
-The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
-and macOS. It can be used in Swift projects without the need to compile the
-library from source. For example:
-```swift
-// swift-tools-version: 5.10
-// The swift-tools-version declares the minimum version of Swift required to build this package.
-
-import PackageDescription
-
-let package = Package(
-    name: "MyLlamaPackage",
-    targets: [
-        .executableTarget(
-            name: "MyLlamaPackage",
-            dependencies: [
-                "LlamaFramework"
-            ]),
-        .binaryTarget(
-            name: "LlamaFramework",
-            url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
-            checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
-        )
-    ]
-)
-```
-The above example is using an intermediate build `b5046` of the library. This can be modified
-to use a different version by changing the URL and checksum.
-
 ## Completions
 Command-line completion is available for some environments.
 

SECURITY.md

@@ -40,8 +40,7 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks
 
 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
-* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
+* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value
 * Encrypt your data if sending it over the network.
 
 ### Multi-Tenant environments

build-xcframework.sh

@@ -41,11 +41,6 @@ COMMON_CMAKE_ARGS=(
     -DGGML_OPENMP=${GGML_OPENMP}
 )
 
-XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
-MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
-MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
-echo "Detected Xcode version: $XCODE_VERSION"
-
 check_required_tool() {
     local tool=$1
     local install_message=$2
@@ -330,28 +325,21 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if command -v xcrun vtool &>/dev/null; then
+        if command -v vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
-                    xcrun vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
+                    vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
                         -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                     ;;
                 "visionos")
                     echo "Marking binary as a framework binary for visionOS..."
-                    if [[ "$MAJOR_VERSION" -gt 16 ]] || [[ "$MAJOR_VERSION" -eq 16 && "$MINOR_VERSION" -gt 2 ]]; then
-                        echo "Xcode version greater than 16.2, using visionOS."
-                        VISION_OS_BUILD_VERSION="visionos"
-                    else
-                        echo "Xcode version less than or equal to 16.2, using xros."
-                        VISION_OS_BUILD_VERSION="xros"
-                    fi
-                    xcrun vtool -set-build-version ${VISION_OS_BUILD_VERSION} ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
+                    vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
                         -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                     ;;
                 "tvos")
                     echo "Marking binary as a framework binary for tvOS..."
-                    xcrun vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
+                    vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
                         -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                     ;;
             esac
@@ -411,7 +399,6 @@ cmake -B build-ios-sim -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-ios-sim --config Release -- -quiet
 
@@ -424,7 +411,6 @@ cmake -B build-ios-device -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-ios-device --config Release -- -quiet
 
@@ -435,7 +421,6 @@ cmake -B build-macos -G Xcode \
     -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-macos --config Release -- -quiet
 
@@ -447,9 +432,8 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos --config Release -- -quiet
 
@@ -461,9 +445,8 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
 
@@ -479,7 +462,6 @@ cmake -B build-tvos-sim -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-tvos-sim --config Release -- -quiet
 
@@ -494,7 +476,6 @@ cmake -B build-tvos-device -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-tvos-device --config Release -- -quiet
 

ci/README.md

@@ -26,43 +26,4 @@ GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 # with SYCL support
 source /opt/intel/oneapi/setvars.sh
 GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-
-# with MUSA support
-GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```
-
-## Running MUSA CI in a Docker Container
-
-Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
-
-### 1. Create a local directory to store cached models, configuration files and venv:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-cache
-```
-
-### 2. Create a local directory to store CI run results:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-results
-```
-
-### 3. Start a Docker container and run the CI:
-
-```bash
-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
-```
-
-Inside the container, execute the following commands:
-
-```bash
-apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
-git config --global --add safe.directory /ws
-GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
-```
-
-This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

ci/run.sh

@@ -16,9 +16,6 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with MUSA support
-# GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -39,7 +36,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
@@ -55,24 +52,13 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
-    # Use only main GPU
-    export ONEAPI_DEVICE_SELECTOR="level_zero:0"
-    # Enable sysman for correct memory reporting
-    export ZES_ENABLE_SYSMAN=1
-    # to circumvent precision issues on CPY operations
-    export SYCL_PROGRAM_COMPILE_OPTIONS="-cl-fp32-correctly-rounded-divide-sqrt"
+
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
-
-if [ ! -z ${GG_BUILD_MUSA} ]; then
-    # Use qy1 by default (MTT S80)
-    MUSA_ARCH=${MUSA_ARCH:-21}
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
-fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -822,7 +808,7 @@ export LLAMA_LOG_PREFIX=1
 export LLAMA_LOG_TIMESTAMPS=1
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
@@ -840,10 +826,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-if [ -z ${GG_BUILD_SYCL} ]; then
-    # SYCL build breaks with debug build flags
-    test $ret -eq 0 && gg_run ctest_debug
-fi
+
+test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -851,9 +835,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run test_scripts_debug
-        fi
+        test $ret -eq 0 && gg_run test_scripts_debug
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -864,9 +846,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
             test $ret -eq 0 && gg_run pythia_2_8b
             #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run ctest_with_model_debug
-        fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi

common/CMakeLists.txt

@@ -85,10 +85,7 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info)
 
 # Use curl to download model url
 if (LLAMA_CURL)
-    find_package(CURL)
-    if (NOT CURL_FOUND)
-        message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF")
-    endif()
+    find_package(CURL REQUIRED)
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
     find_library(CURL_LIBRARY curl REQUIRED)
@@ -117,8 +114,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.10:
-        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
+        # v0.6.12:
+        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/arg.h

@@ -78,12 +78,3 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 
 // function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
-bool common_has_curl();
-
-struct common_remote_params {
-    std::vector<std::string> headers;
-    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
-    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
-};
-// get remote file content, returns <http_code, raw_response_body>
-std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

common/chat.cpp

@@ -1622,7 +1622,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
-    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null() && params.tools.is_array() && params.json_schema.is_null()) {
+    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_hermes_2_pro(tmpl, params);
     }
 

common/common.cpp

@@ -7,6 +7,9 @@
 
 #include "common.h"
 #include "log.h"
+// Change JSON_ASSERT from assert() to GGML_ASSERT:
+#define JSON_ASSERT GGML_ASSERT
+#include "json.hpp"
 #include "llama.h"
 
 #include <algorithm>
@@ -48,11 +51,47 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #endif
+#if defined(LLAMA_USE_CURL)
+#include <curl/curl.h>
+#include <curl/easy.h>
+#include <future>
+#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#if defined(LLAMA_USE_CURL)
+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
+#else
+#include <sys/syslimits.h>
+#endif
+#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
+
+//
+// CURL utils
+//
+
+using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
+
+// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
+struct curl_slist_ptr {
+    struct curl_slist * ptr = nullptr;
+    ~curl_slist_ptr() {
+        if (ptr) {
+            curl_slist_free_all(ptr);
+        }
+    }
+};
+#endif // LLAMA_USE_CURL
+
+using json = nlohmann::ordered_json;
+
 //
 // CPU utils
 //
@@ -543,41 +582,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
     return buf.str();
 }
 
-std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch) {
-    std::stringstream buf;
-
-    buf << "[ ";
-
-    bool first = true;
-    for (int i = 0; i < batch.n_tokens; ++i) {
-        if (!first) {
-            buf << ", ";
-        } else {
-            first = false;
-        }
-
-        auto detokenized = common_token_to_piece(ctx, batch.token[i]);
-
-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
-        buf << "\n"          << std::to_string(i)
-            << ", token '"   << detokenized << "'"
-            << ", pos "      << std::to_string(batch.pos[i])
-            << ", n_seq_id " << std::to_string(batch.n_seq_id[i])
-            << ", seq_id "   << std::to_string(batch.seq_id[i][0])
-            << ", logits "   << std::to_string(batch.logits[i]);
-    }
-
-    buf << " ]";
-
-    return buf.str();
-}
-
 void string_process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -830,7 +834,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
+#ifdef __linux__
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -840,9 +844,7 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#else
-#  error Unknown architecture
-#endif
+#endif // __linux__
         cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
     }
@@ -863,14 +865,22 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
+    llama_model * model = nullptr;
+
+    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
+        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
+    } else if (!params.model_url.empty()) {
+        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
+    } else {
+        model = llama_model_load_from_file(params.model.c_str(), mparams);
+    }
+
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
         return iparams;
     }
 
@@ -905,7 +915,7 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_model_free(model);
         return iparams;
     }
@@ -1006,7 +1016,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (llama_model_has_encoder(model)) {
-            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
+            llama_batch_ext_ptr batch(llama_batch_ext_init_from_text(tmp.data(), tmp.size(), 0, 0, true));
+            llama_encode_ext(lctx, batch.get());
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
             if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
                 decoder_start_token_id = bos;
@@ -1015,7 +1026,8 @@ struct common_init_result common_init_from_params(common_params & params) {
             tmp.push_back(decoder_start_token_id);
         }
         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_batch_ext_ptr batch(llama_batch_ext_init_from_text(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0, true));
+            llama_decode_ext(lctx, batch.get());
         }
         llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
@@ -1029,19 +1041,6 @@ struct common_init_result common_init_from_params(common_params & params) {
     return iparams;
 }
 
-std::string get_model_endpoint() {
-    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
-    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
-    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
-    std::string model_endpoint = "https://huggingface.co/";
-    if (endpoint_env) {
-        model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') model_endpoint += '/';
-    }
-    return model_endpoint;
-}
-
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     llama_clear_adapter_lora(ctx);
     for (auto & la : lora) {
@@ -1057,18 +1056,15 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (!params.devices.empty()) {
         mparams.devices = params.devices.data();
     }
-
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
-
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
-
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1076,13 +1072,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
-    if (params.tensor_buft_overrides.empty()) {
-        mparams.tensor_buft_overrides = NULL;
-    } else {
-        GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
-        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
-    }
-
     return mparams;
 }
 
@@ -1142,14 +1131,461 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     return tpp;
 }
 
+#ifdef LLAMA_USE_CURL
+
+#define CURL_MAX_RETRY 3
+#define CURL_RETRY_DELAY_SECONDS 2
+
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+    int remaining_attempts = max_attempts;
+
+    while (remaining_attempts > 0) {
+        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+
+        CURLcode res = curl_easy_perform(curl);
+        if (res == CURLE_OK) {
+            return true;
+        }
+
+        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
+        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
+
+        remaining_attempts--;
+        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+    }
+
+    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+
+    return false;
+}
+
+static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
+    // Initialize libcurl
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    if (!curl) {
+        LOG_ERR("%s: error initializing libcurl\n", __func__);
+        return false;
+    }
+
+    bool force_download = false;
+
+    // Set the URL, allow to follow http redirection
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+
+    // Check if hf-token or bearer-token was specified
+    if (!hf_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + hf_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    }
+
+#if defined(_WIN32)
+    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
+    //   operating system. Currently implemented under MS-Windows.
+    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);
+
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    nlohmann::json metadata;
+    std::string etag;
+    std::string last_modified;
+
+    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_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
+                if (metadata.contains("url") && metadata.at("url").is_string()) {
+                    auto previous_url = metadata.at("url").get<std::string>();
+                    if (previous_url != url) {
+                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
+                        return false;
+                    }
+                }
+                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());
+                return false;
+            }
+        }
+    } 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;
+    };
+
+    common_load_model_from_url_headers headers;
+
+    {
+        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;
+
+            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);
+
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code != 200) {
+            // HEAD not supported, we don't know if the file has changed
+            // force trigger downloading
+            force_download = true;
+            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        }
+    }
+
+    bool should_download = !file_exists || force_download;
+    if (!should_download) {
+        if (!etag.empty() && etag != headers.etag) {
+            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
+            should_download = true;
+        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
+            should_download = true;
+        }
+    }
+    if (should_download) {
+        std::string path_temporary = path + ".downloadInProgress";
+        if (file_exists) {
+            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
+            if (remove(path.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+                return false;
+            }
+        }
+
+        // Set the output file
+
+        struct FILE_deleter {
+            void operator()(FILE * f) const {
+                fclose(f);
+            }
+        };
+
+        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
+        if (!outfile) {
+            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
+            return false;
+        }
+
+        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
+        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
+            return fwrite(data, size, nmemb, (FILE *)fd);
+        };
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
+
+        //  display download progress
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
+
+        // helper function to hide password in URL
+        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
+            std::size_t protocol_pos = url.find("://");
+            if (protocol_pos == std::string::npos) {
+                return url;  // Malformed URL
+            }
+
+            std::size_t at_pos = url.find('@', protocol_pos + 3);
+            if (at_pos == std::string::npos) {
+                return url;  // No password in URL
+            }
+
+            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
+        };
+
+        // start the download
+        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code < 200 || http_code >= 400) {
+            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
+            return false;
+        }
+
+        // Causes file to be closed explicitly here before we rename it.
+        outfile.reset();
+
+        // Write the updated JSON metadata file.
+        metadata.update({
+            {"url", url},
+            {"etag", headers.etag},
+            {"lastModified", headers.last_modified}
+        });
+        std::ofstream(metadata_path) << metadata.dump(4);
+        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+
+        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
+            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+            return false;
+        }
+    }
+
+    return true;
+}
+
+struct llama_model * common_load_model_from_url(
+        const std::string & model_url,
+        const std::string & local_path,
+        const std::string & hf_token,
+        const struct llama_model_params & params) {
+    // Basic validation of the model_url
+    if (model_url.empty()) {
+        LOG_ERR("%s: invalid model_url\n", __func__);
+        return NULL;
+    }
+
+    if (!common_download_file(model_url, local_path, hf_token)) {
+        return NULL;
+    }
+
+    // check for additional GGUFs split to download
+    int n_split = 0;
+    {
+        struct gguf_init_params gguf_params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ NULL,
+        };
+        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
+        if (!ctx_gguf) {
+            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
+            return NULL;
+        }
+
+        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+        if (key_n_split >= 0) {
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+        }
+
+        gguf_free(ctx_gguf);
+    }
+
+    if (n_split > 1) {
+        char split_prefix[PATH_MAX] = {0};
+        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+
+        // Verify the first split file format
+        // and extract split URL and PATH prefixes
+        {
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
+                return NULL;
+            }
+
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
+                return NULL;
+            }
+        }
+
+        // Prepare download in parallel
+        std::vector<std::future<bool>> futures_download;
+        for (int idx = 1; idx < n_split; idx++) {
+            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
+                char split_path[PATH_MAX] = {0};
+                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
+
+                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
+
+                return common_download_file(split_url, split_path, hf_token);
+            }, idx));
+        }
+
+        // Wait for all downloads to complete
+        for (auto & f : futures_download) {
+            if (!f.get()) {
+                return NULL;
+            }
+        }
+    }
+
+    return llama_model_load_from_file(local_path.c_str(), params);
+}
+
+struct llama_model * common_load_model_from_hf(
+        const std::string & repo,
+        const std::string & remote_path,
+        const std::string & local_path,
+        const std::string & hf_token,
+        const struct llama_model_params & params) {
+    // construct hugging face model url:
+    //
+    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
+    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
+    //
+    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //
+
+    std::string model_url = "https://huggingface.co/";
+    model_url += repo;
+    model_url += "/resolve/main/";
+    model_url += remote_path;
+
+    return common_load_model_from_url(model_url, local_path, hf_token, params);
+}
+
+/**
+ * Allow getting the HF file from the HF repo with tag (like ollama), for example:
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
+ * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
+ *
+ * Return pair of <repo, file> (with "repo" already having tag removed)
+ *
+ * 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.
+ */
+std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
+    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];
+    if (string_split<std::string>(hf_repo, '/').size() != 2) {
+        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
+    }
+
+    // fetch model info from Hugging Face Hub API
+    json model_info;
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::string res_str;
+    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (!hf_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + hf_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+    }
+    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        throw std::runtime_error("error: cannot make GET request to HF API");
+    }
+
+    long res_code;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+    if (res_code == 200) {
+        model_info = json::parse(res_str);
+    } else if (res_code == 401) {
+        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
+    } else {
+        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
+    }
+
+    // check response
+    if (!model_info.contains("ggufFile")) {
+        throw std::runtime_error("error: model does not have ggufFile");
+    }
+    json & gguf_file = model_info.at("ggufFile");
+    if (!gguf_file.contains("rfilename")) {
+        throw std::runtime_error("error: ggufFile does not have rfilename");
+    }
+
+    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
+}
+
+#else
+
+struct llama_model * common_load_model_from_url(
+        const std::string & /*model_url*/,
+        const std::string & /*local_path*/,
+        const std::string & /*hf_token*/,
+        const struct llama_model_params & /*params*/) {
+    LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
+    return nullptr;
+}
+
+struct llama_model * common_load_model_from_hf(
+        const std::string & /*repo*/,
+        const std::string & /*remote_path*/,
+        const std::string & /*local_path*/,
+        const std::string & /*hf_token*/,
+        const struct llama_model_params & /*params*/) {
+    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
+    return nullptr;
+}
+
+std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
+    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
+    return std::make_pair("", "");
+}
+
+#endif // LLAMA_USE_CURL
+
 //
 // Batch utils
 //
 
+// DEPRECATED
 void common_batch_clear(struct llama_batch & batch) {
     batch.n_tokens = 0;
 }
 
+// DEPRECATED
 void common_batch_add(
                  struct llama_batch & batch,
                         llama_token   id,
@@ -1565,3 +2001,26 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+template <>
+json common_grammar_trigger::to_json() const {
+    json out {
+        {"type", (int) type},
+        {"value", value},
+    };
+    if (type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+        out["token"] = (int) token;
+    }
+    return out;
+}
+
+template <>
+common_grammar_trigger common_grammar_trigger::from_json(const json & in) {
+    common_grammar_trigger out;
+    out.type = (common_grammar_trigger_type) in.at("type").get<int>();
+    out.value = in.at("value").get<std::string>();
+    if (out.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+        out.token = (llama_token) in.at("token").get<int>();
+    }
+    return out;
+}

common/common.h

@@ -121,6 +121,10 @@ struct common_grammar_trigger {
     common_grammar_trigger_type type;
     std::string value;
     llama_token token = LLAMA_TOKEN_NULL;
+
+    // T can only be nlohmann::ordered_json
+    template <class T> T to_json() const;
+    template <class T> static common_grammar_trigger from_json(const T & in);
 };
 
 // sampling parameters
@@ -180,13 +184,6 @@ struct common_params_sampling {
     std::string print() const;
 };
 
-struct common_params_model {
-    std::string path    = ""; // model local path                                           // NOLINT
-    std::string url     = ""; // model url to download                                      // NOLINT
-    std::string hf_repo = ""; // HF repo                                                    // NOLINT
-    std::string hf_file = ""; // HF file                                                    // NOLINT
-};
-
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
@@ -200,11 +197,19 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    struct common_params_model model;
+    std::string hf_repo = ""; // HF repo                                                     // NOLINT
+    std::string hf_file = ""; // HF file                                                     // NOLINT
+
+    std::string model = "";     // draft model for speculative decoding                      // NOLINT
+    std::string model_url = ""; // model url to download                                     // NOLINT
 };
 
 struct common_params_vocoder {
-    struct common_params_model model;
+    std::string hf_repo = ""; // HF repo                                                     // NOLINT
+    std::string hf_file = ""; // HF file                                                     // NOLINT
+
+    std::string model     = ""; // model path                                                // NOLINT
+    std::string model_url = ""; // model url to download                                     // NOLINT
 
     std::string speaker_file = ""; // speaker file path                                      // NOLINT
 
@@ -262,10 +267,12 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    struct common_params_model model;
-
+    std::string model                = ""; // model path                                                    // NOLINT
     std::string model_alias          = ""; // model alias                                                   // NOLINT
+    std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
+    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
+    std::string hf_file              = ""; // HF file                                                       // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
@@ -279,7 +286,6 @@ struct common_params {
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
     std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
@@ -341,9 +347,7 @@ struct common_params {
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
     // multimodal models (see examples/llava)
-    struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;     // use GPU for multimodal model
-    bool no_mmproj = false;         // explicitly disable multimodal model
+    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -512,7 +516,6 @@ void string_process_escapes(std::string & input);
 std::string string_from(bool value);
 std::string string_from(const std::vector<int> & values);
 std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
-std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
 
 //
 // Filesystem utils
@@ -542,17 +545,34 @@ struct llama_model_params     common_model_params_to_llama  (      common_params
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
+struct llama_model * common_load_model_from_url(
+    const std::string & model_url,
+    const std::string & local_path,
+    const std::string & hf_token,
+    const struct llama_model_params & params);
+
+struct llama_model * common_load_model_from_hf(
+    const std::string & repo,
+    const std::string & remote_path,
+    const std::string & local_path,
+    const std::string & hf_token,
+    const struct llama_model_params & params);
+
+std::pair<std::string, std::string> common_get_hf_file(
+    const std::string & hf_repo_with_tag,
+    const std::string & hf_token);
+
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
-std::string                   get_model_endpoint();
-
 //
 // Batch utils
 //
 
+// DEPRECATED
 void common_batch_clear(struct llama_batch & batch);
 
+// DEPRECATED
 void common_batch_add(
                  struct llama_batch & batch,
                         llama_token   id,
@@ -560,6 +580,66 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
+// convenient wrapper around llama_batch_ext, to provide a way to get embeddings positions
+// this is meant to be temporary
+struct common_batch {
+    llama_batch_ext_ptr batch;
+    struct batch_token {
+        llama_token  token;
+        llama_seq_id seq_id; // only support single seq for now
+        bool         logits;
+    };
+    std::vector<batch_token> tokens;
+    int n_outputs = 0;
+    common_batch() = default;
+    common_batch(int32_t n_tokens, int32_t n_seq_max) {
+        batch.reset(llama_batch_ext_init(n_tokens, n_seq_max));
+        tokens.reserve(n_tokens);
+    }
+    void clear() {
+        llama_batch_ext_clear(batch.get());
+        tokens.clear();
+    }
+    void add_text(llama_token token, llama_pos pos, llama_seq_id seq_id, bool logits) {
+        llama_batch_ext_add_text(batch.get(), token, pos, &seq_id, 1, logits);
+        tokens.push_back({token, seq_id, logits});
+        if (logits) {
+            n_outputs++;
+        }
+    }
+    void add_text_multi_seq(llama_token token, llama_pos pos, std::vector<llama_seq_id> seq_ids, bool logits) {
+        llama_batch_ext_add_text(batch.get(), token, pos, seq_ids.data(), seq_ids.size(), logits);
+        tokens.push_back({token, seq_ids[0], logits});
+        if (logits) {
+            n_outputs++;
+        }
+    }
+    void set_logits_last() {
+        if (!tokens.empty()) {
+            llama_batch_ext_set_output_last(batch.get());
+            tokens.back().logits = true;
+        }
+    }
+    int32_t get_n_tokens() const {
+        return (int32_t)tokens.size();
+    }
+    llama_batch_ext * get() {
+        return batch.get();
+    }
+    common_batch get_view(int32_t offset, int32_t n_tokens) {
+        common_batch view;
+        view.batch = llama_batch_ext_ptr(llama_batch_ext_get_view(batch.get(), offset, n_tokens));
+        view.tokens.reserve(n_tokens);
+        for (int32_t i = 0; i < n_tokens; i++) {
+            view.tokens.push_back(tokens[offset + i]);
+            if (tokens[offset + i].logits) {
+                view.n_outputs++;
+            }
+        }
+        return view;
+    }
+};
+
 //
 // Token utils
 //

common/json-schema-to-grammar.cpp

@@ -16,9 +16,6 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
-    if (max_items == 0) {
-        return "";
-    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

common/llguidance.cpp

@@ -11,24 +11,25 @@ struct llama_sampler_llg {
     std::string         grammar_kind;
     std::string         grammar_data;
     LlgTokenizer *      tokenizer;
-    LlgMatcher *        grammar;
+    LlgConstraint *     grammar;
+    LlgMaskResult       llg_res;
+    bool                has_llg_res;
 };
 
-static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                          const char * grammar_data) {
+static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                             const char * grammar_data) {
     LlgConstraintInit cinit;
     llg_constraint_init_set_defaults(&cinit, tokenizer);
     const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
     if (log_level && *log_level) {
         cinit.log_stderr_level = atoi(log_level);
     }
-    auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data);
-    if (llg_matcher_get_error(c)) {
-        LOG_ERR("llg error: %s\n", llg_matcher_get_error(c));
-        llg_free_matcher(c);
+    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
+    if (llg_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_get_error(c));
+        llg_free_constraint(c);
         return nullptr;
     }
-
     return c;
 }
 
@@ -39,29 +40,39 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
 static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        llg_matcher_consume_token(ctx->grammar, token);
+        LlgCommitResult res;
+        llg_commit_token(ctx->grammar, token, &res);
+        ctx->has_llg_res = false;
     }
 }
 
 static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        const uint32_t * mask = llg_matcher_get_mask(ctx->grammar);
-        if (mask == nullptr) {
-            if (llg_matcher_compute_mask(ctx->grammar) == 0) {
-                mask = llg_matcher_get_mask(ctx->grammar);
+        if (!ctx->has_llg_res) {
+            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
+                ctx->has_llg_res = true;
             } else {
-                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
-                llg_free_matcher(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
+                llg_free_constraint(ctx->grammar);
                 ctx->grammar = nullptr;
-                return;
             }
         }
-
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            auto token = cur_p->data[i].id;
-            if ((mask[token / 32] & (1 << (token % 32))) == 0) {
-                cur_p->data[i].logit = -INFINITY;
+        if (ctx->has_llg_res) {
+            if (ctx->llg_res.is_stop) {
+                for (size_t i = 0; i < cur_p->size; ++i) {
+                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
+                        cur_p->data[i].logit = -INFINITY;
+                    }
+                }
+            } else {
+                const uint32_t * mask = ctx->llg_res.sample_mask;
+                for (size_t i = 0; i < cur_p->size; ++i) {
+                    auto token = cur_p->data[i].id;
+                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                        cur_p->data[i].logit = -INFINITY;
+                    }
+                }
             }
         }
     }
@@ -69,9 +80,14 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array
 
 static void llama_sampler_llg_reset(llama_sampler * smpl) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
-    if (ctx->grammar) {
-        llg_matcher_reset(ctx->grammar);
+    if (!ctx->grammar) {
+        return;
     }
+
+    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
+    llg_free_constraint(ctx->grammar);
+    ctx->grammar     = grammar_new;
+    ctx->has_llg_res = false;
 }
 
 static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
@@ -86,7 +102,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
         if (ctx->grammar) {
             result_ctx->grammar_kind = ctx->grammar_kind;
             result_ctx->grammar_data = ctx->grammar_data;
-            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
+            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
             result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
         }
     }
@@ -98,7 +114,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_llg *) smpl->ctx;
 
     if (ctx->grammar) {
-        llg_free_matcher(ctx->grammar);
+        llg_free_constraint(ctx->grammar);
         llg_free_tokenizer(ctx->tokenizer);
     }
 
@@ -223,11 +239,9 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ grammar_data,
             /* .tokenizer    = */ tokenizer,
             /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
+            /* .llg_res      = */ {},
+            /* .has_llg_res  = */ false,
         };
-        if (ctx->grammar) {
-            GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 ==
-                        llg_matcher_get_mask_byte_size(ctx->grammar));
-        }
     } else {
         *ctx = {
             /* .vocab        = */ vocab,
@@ -235,12 +249,15 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ {},
             /* .tokenizer    = */ nullptr,
             /* .grammar      = */ nullptr,
+            /* .llg_res      = */ {},
+            /* .has_llg_res  = */ false,
         };
     }
 
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx);
+        /* .ctx   = */ ctx
+    );
 }
 
 #else

common/minja/chat-template.hpp

@@ -9,19 +9,10 @@
 #pragma once
 
 #include "minja.hpp"
-
-#include <chrono>
-#include <cstddef>
-#include <cstdio>
-#include <exception>
-#include <iomanip>
-#include <memory>
-#include <sstream>
+#include <json.hpp>
 #include <string>
 #include <vector>
 
-#include <json.hpp>
-
 using json = nlohmann::ordered_json;
 
 namespace minja {
@@ -434,7 +425,7 @@ class chat_template {
                         auto obj = json {
                             {"tool_calls", tool_calls},
                         };
-                        if (!content.is_null() && !content.empty()) {
+                        if (!content.is_null() && content != "") {
                             obj["content"] = content;
                         }
                         message["content"] = obj.dump(2);
@@ -444,12 +435,13 @@ class chat_template {
                 if (polyfill_tool_responses && role == "tool") {
                     message["role"] = "user";
                     auto obj = json {
-                        {"tool_response", json::object()},
+                        {"tool_response", {
+                            {"content", message.at("content")},
+                        }},
                     };
                     if (message.contains("name")) {
-                        obj["tool_response"]["tool"] = message.at("name");
+                        obj["tool_response"]["name"] = message.at("name");
                     }
-                    obj["tool_response"]["content"] = message.at("content");
                     if (message.contains("tool_call_id")) {
                         obj["tool_response"]["tool_call_id"] = message.at("tool_call_id");
                     }
@@ -518,7 +510,7 @@ class chat_template {
     static nlohmann::ordered_json add_system(const nlohmann::ordered_json & messages, const std::string & system_prompt) {
         json messages_with_system = messages;
 
-        if (!messages_with_system.empty() && messages_with_system[0].at("role") == "system") {
+        if (messages_with_system.size() > 0 && messages_with_system[0].at("role") == "system") {
             std::string existing_system = messages_with_system.at(0).at("content");
             messages_with_system[0] = json {
                 {"role", "system"},

common/minja/minja.hpp

@@ -8,26 +8,14 @@
 // SPDX-License-Identifier: MIT
 #pragma once
 
-#include <algorithm>
-#include <cctype>
-#include <cstddef>
-#include <cmath>
-#include <exception>
-#include <functional>
 #include <iostream>
-#include <iterator>
-#include <limits>
-#include <map>
-#include <memory>
-#include <regex>
-#include <sstream>
 #include <string>
-#include <stdexcept>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
 #include <vector>
-
+#include <regex>
+#include <memory>
+#include <stdexcept>
+#include <sstream>
+#include <unordered_set>
 #include <json.hpp>
 
 using json = nlohmann::ordered_json;
@@ -743,51 +731,51 @@ public:
 
 struct TextTemplateToken : public TemplateToken {
     std::string text;
-    TextTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, loc, pre, post), text(t) {}
+    TextTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, location, pre, post), text(t) {}
 };
 
 struct ExpressionTemplateToken : public TemplateToken {
     std::shared_ptr<Expression> expr;
-    ExpressionTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && e) : TemplateToken(Type::Expression, loc, pre, post), expr(std::move(e)) {}
+    ExpressionTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && e) : TemplateToken(Type::Expression, location, pre, post), expr(std::move(e)) {}
 };
 
 struct IfTemplateToken : public TemplateToken {
     std::shared_ptr<Expression> condition;
-    IfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::If, loc, pre, post), condition(std::move(c)) {}
+    IfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::If, location, pre, post), condition(std::move(c)) {}
 };
 
 struct ElifTemplateToken : public TemplateToken {
     std::shared_ptr<Expression> condition;
-    ElifTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::Elif, loc, pre, post), condition(std::move(c)) {}
+    ElifTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::Elif, location, pre, post), condition(std::move(c)) {}
 };
 
 struct ElseTemplateToken : public TemplateToken {
-    ElseTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, loc, pre, post) {}
+    ElseTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, location, pre, post) {}
 };
 
 struct EndIfTemplateToken : public TemplateToken {
-    EndIfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, loc, pre, post) {}
+    EndIfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, location, pre, post) {}
 };
 
 struct MacroTemplateToken : public TemplateToken {
     std::shared_ptr<VariableExpr> name;
     Expression::Parameters params;
-    MacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p)
-      : TemplateToken(Type::Macro, loc, pre, post), name(std::move(n)), params(std::move(p)) {}
+    MacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p)
+      : TemplateToken(Type::Macro, location, pre, post), name(std::move(n)), params(std::move(p)) {}
 };
 
 struct EndMacroTemplateToken : public TemplateToken {
-    EndMacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, loc, pre, post) {}
+    EndMacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, location, pre, post) {}
 };
 
 struct FilterTemplateToken : public TemplateToken {
     std::shared_ptr<Expression> filter;
-    FilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && filter)
-      : TemplateToken(Type::Filter, loc, pre, post), filter(std::move(filter)) {}
+    FilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && filter)
+      : TemplateToken(Type::Filter, location, pre, post), filter(std::move(filter)) {}
 };
 
 struct EndFilterTemplateToken : public TemplateToken {
-    EndFilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, loc, pre, post) {}
+    EndFilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, location, pre, post) {}
 };
 
 struct ForTemplateToken : public TemplateToken {
@@ -795,38 +783,38 @@ struct ForTemplateToken : public TemplateToken {
     std::shared_ptr<Expression> iterable;
     std::shared_ptr<Expression> condition;
     bool recursive;
-    ForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::vector<std::string> & vns, std::shared_ptr<Expression> && iter,
+    ForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::vector<std::string> & vns, std::shared_ptr<Expression> && iter,
       std::shared_ptr<Expression> && c, bool r)
-      : TemplateToken(Type::For, loc, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {}
+      : TemplateToken(Type::For, location, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {}
 };
 
 struct EndForTemplateToken : public TemplateToken {
-    EndForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, loc, pre, post) {}
+    EndForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, location, pre, post) {}
 };
 
 struct GenerationTemplateToken : public TemplateToken {
-    GenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, loc, pre, post) {}
+    GenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, location, pre, post) {}
 };
 
 struct EndGenerationTemplateToken : public TemplateToken {
-    EndGenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, loc, pre, post) {}
+    EndGenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, location, pre, post) {}
 };
 
 struct SetTemplateToken : public TemplateToken {
     std::string ns;
     std::vector<std::string> var_names;
     std::shared_ptr<Expression> value;
-    SetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
-      : TemplateToken(Type::Set, loc, pre, post), ns(ns), var_names(vns), value(std::move(v)) {}
+    SetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
+      : TemplateToken(Type::Set, location, pre, post), ns(ns), var_names(vns), value(std::move(v)) {}
 };
 
 struct EndSetTemplateToken : public TemplateToken {
-    EndSetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, loc, pre, post) {}
+    EndSetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, location, pre, post) {}
 };
 
 struct CommentTemplateToken : public TemplateToken {
     std::string text;
-    CommentTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, loc, pre, post), text(t) {}
+    CommentTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, location, pre, post), text(t) {}
 };
 
 enum class LoopControlType { Break, Continue };
@@ -842,7 +830,7 @@ public:
 
 struct LoopControlTemplateToken : public TemplateToken {
     LoopControlType control_type;
-    LoopControlTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, loc, pre, post), control_type(control_type) {}
+    LoopControlTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, location, pre, post), control_type(control_type) {}
 };
 
 class TemplateNode {
@@ -880,8 +868,8 @@ public:
 class SequenceNode : public TemplateNode {
     std::vector<std::shared_ptr<TemplateNode>> children;
 public:
-    SequenceNode(const Location & loc, std::vector<std::shared_ptr<TemplateNode>> && c)
-      : TemplateNode(loc), children(std::move(c)) {}
+    SequenceNode(const Location & location, std::vector<std::shared_ptr<TemplateNode>> && c)
+      : TemplateNode(location), children(std::move(c)) {}
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
         for (const auto& child : children) child->render(out, context);
     }
@@ -890,7 +878,7 @@ public:
 class TextNode : public TemplateNode {
     std::string text;
 public:
-    TextNode(const Location & loc, const std::string& t) : TemplateNode(loc), text(t) {}
+    TextNode(const Location & location, const std::string& t) : TemplateNode(location), text(t) {}
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> &) const override {
       out << text;
     }
@@ -899,7 +887,7 @@ public:
 class ExpressionNode : public TemplateNode {
     std::shared_ptr<Expression> expr;
 public:
-    ExpressionNode(const Location & loc, std::shared_ptr<Expression> && e) : TemplateNode(loc), expr(std::move(e)) {}
+    ExpressionNode(const Location & location, std::shared_ptr<Expression> && e) : TemplateNode(location), expr(std::move(e)) {}
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
       if (!expr) throw std::runtime_error("ExpressionNode.expr is null");
       auto result = expr->evaluate(context);
@@ -916,8 +904,8 @@ public:
 class IfNode : public TemplateNode {
     std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> cascade;
 public:
-    IfNode(const Location & loc, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> && c)
-        : TemplateNode(loc), cascade(std::move(c)) {}
+    IfNode(const Location & location, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> && c)
+        : TemplateNode(location), cascade(std::move(c)) {}
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
       for (const auto& branch : cascade) {
           auto enter_branch = true;
@@ -936,7 +924,7 @@ public:
 class LoopControlNode : public TemplateNode {
     LoopControlType control_type_;
   public:
-    LoopControlNode(const Location & loc, LoopControlType control_type) : TemplateNode(loc), control_type_(control_type) {}
+    LoopControlNode(const Location & location, LoopControlType control_type) : TemplateNode(location), control_type_(control_type) {}
     void do_render(std::ostringstream &, const std::shared_ptr<Context> &) const override {
       throw LoopControlException(control_type_);
     }
@@ -950,9 +938,9 @@ class ForNode : public TemplateNode {
     bool recursive;
     std::shared_ptr<TemplateNode> else_body;
 public:
-    ForNode(const Location & loc, std::vector<std::string> && var_names, std::shared_ptr<Expression> && iterable,
+    ForNode(const Location & location, std::vector<std::string> && var_names, std::shared_ptr<Expression> && iterable,
       std::shared_ptr<Expression> && condition, std::shared_ptr<TemplateNode> && body, bool recursive, std::shared_ptr<TemplateNode> && else_body)
-            : TemplateNode(loc), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {}
+            : TemplateNode(location), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {}
 
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
       // https://jinja.palletsprojects.com/en/3.0.x/templates/#for
@@ -1037,8 +1025,8 @@ class MacroNode : public TemplateNode {
     std::shared_ptr<TemplateNode> body;
     std::unordered_map<std::string, size_t> named_param_positions;
 public:
-    MacroNode(const Location & loc, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p, std::shared_ptr<TemplateNode> && b)
-        : TemplateNode(loc), name(std::move(n)), params(std::move(p)), body(std::move(b)) {
+    MacroNode(const Location & location, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p, std::shared_ptr<TemplateNode> && b)
+        : TemplateNode(location), name(std::move(n)), params(std::move(p)), body(std::move(b)) {
         for (size_t i = 0; i < params.size(); ++i) {
           const auto & name = params[i].first;
           if (!name.empty()) {
@@ -1084,8 +1072,8 @@ class FilterNode : public TemplateNode {
     std::shared_ptr<TemplateNode> body;
 
 public:
-    FilterNode(const Location & loc, std::shared_ptr<Expression> && f, std::shared_ptr<TemplateNode> && b)
-        : TemplateNode(loc), filter(std::move(f)), body(std::move(b)) {}
+    FilterNode(const Location & location, std::shared_ptr<Expression> && f, std::shared_ptr<TemplateNode> && b)
+        : TemplateNode(location), filter(std::move(f)), body(std::move(b)) {}
 
     void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
         if (!filter) throw std::runtime_error("FilterNode.filter is null");
@@ -1107,8 +1095,8 @@ class SetNode : public TemplateNode {
     std::vector<std::string> var_names;
     std::shared_ptr<Expression> value;
 public:
-    SetNode(const Location & loc, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
-        : TemplateNode(loc), ns(ns), var_names(vns), value(std::move(v)) {}
+    SetNode(const Location & location, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
+        : TemplateNode(location), ns(ns), var_names(vns), value(std::move(v)) {}
     void do_render(std::ostringstream &, const std::shared_ptr<Context> & context) const override {
       if (!value) throw std::runtime_error("SetNode.value is null");
       if (!ns.empty()) {
@@ -1130,8 +1118,8 @@ class SetTemplateNode : public TemplateNode {
     std::string name;
     std::shared_ptr<TemplateNode> template_value;
 public:
-    SetTemplateNode(const Location & loc, const std::string & name, std::shared_ptr<TemplateNode> && tv)
-        : TemplateNode(loc), name(name), template_value(std::move(tv)) {}
+    SetTemplateNode(const Location & location, const std::string & name, std::shared_ptr<TemplateNode> && tv)
+        : TemplateNode(location), name(name), template_value(std::move(tv)) {}
     void do_render(std::ostringstream &, const std::shared_ptr<Context> & context) const override {
       if (!template_value) throw std::runtime_error("SetTemplateNode.template_value is null");
       Value value { template_value->render(context) };
@@ -1144,8 +1132,8 @@ class IfExpr : public Expression {
     std::shared_ptr<Expression> then_expr;
     std::shared_ptr<Expression> else_expr;
 public:
-    IfExpr(const Location & loc, std::shared_ptr<Expression> && c, std::shared_ptr<Expression> && t, std::shared_ptr<Expression> && e)
-        : Expression(loc), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {}
+    IfExpr(const Location & location, std::shared_ptr<Expression> && c, std::shared_ptr<Expression> && t, std::shared_ptr<Expression> && e)
+        : Expression(location), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
       if (!condition) throw std::runtime_error("IfExpr.condition is null");
       if (!then_expr) throw std::runtime_error("IfExpr.then_expr is null");
@@ -1162,16 +1150,16 @@ public:
 class LiteralExpr : public Expression {
     Value value;
 public:
-    LiteralExpr(const Location & loc, const Value& v)
-      : Expression(loc), value(v) {}
+    LiteralExpr(const Location & location, const Value& v)
+      : Expression(location), value(v) {}
     Value do_evaluate(const std::shared_ptr<Context> &) const override { return value; }
 };
 
 class ArrayExpr : public Expression {
     std::vector<std::shared_ptr<Expression>> elements;
 public:
-    ArrayExpr(const Location & loc, std::vector<std::shared_ptr<Expression>> && e)
-      : Expression(loc), elements(std::move(e)) {}
+    ArrayExpr(const Location & location, std::vector<std::shared_ptr<Expression>> && e)
+      : Expression(location), elements(std::move(e)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         auto result = Value::array();
         for (const auto& e : elements) {
@@ -1185,8 +1173,8 @@ public:
 class DictExpr : public Expression {
     std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> elements;
 public:
-    DictExpr(const Location & loc, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> && e)
-      : Expression(loc), elements(std::move(e)) {}
+    DictExpr(const Location & location, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> && e)
+      : Expression(location), elements(std::move(e)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         auto result = Value::object();
         for (const auto& [key, value] : elements) {
@@ -1201,8 +1189,8 @@ public:
 class SliceExpr : public Expression {
 public:
     std::shared_ptr<Expression> start, end;
-    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
-      : Expression(loc), start(std::move(s)), end(std::move(e)) {}
+    SliceExpr(const Location & location, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
+      : Expression(location), start(std::move(s)), end(std::move(e)) {}
     Value do_evaluate(const std::shared_ptr<Context> &) const override {
         throw std::runtime_error("SliceExpr not implemented");
     }
@@ -1212,8 +1200,8 @@ class SubscriptExpr : public Expression {
     std::shared_ptr<Expression> base;
     std::shared_ptr<Expression> index;
 public:
-    SubscriptExpr(const Location & loc, std::shared_ptr<Expression> && b, std::shared_ptr<Expression> && i)
-        : Expression(loc), base(std::move(b)), index(std::move(i)) {}
+    SubscriptExpr(const Location & location, std::shared_ptr<Expression> && b, std::shared_ptr<Expression> && i)
+        : Expression(location), base(std::move(b)), index(std::move(i)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!base) throw std::runtime_error("SubscriptExpr.base is null");
         if (!index) throw std::runtime_error("SubscriptExpr.index is null");
@@ -1255,8 +1243,8 @@ public:
     enum class Op { Plus, Minus, LogicalNot, Expansion, ExpansionDict };
     std::shared_ptr<Expression> expr;
     Op op;
-    UnaryOpExpr(const Location & loc, std::shared_ptr<Expression> && e, Op o)
-      : Expression(loc), expr(std::move(e)), op(o) {}
+    UnaryOpExpr(const Location & location, std::shared_ptr<Expression> && e, Op o)
+      : Expression(location), expr(std::move(e)), op(o) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!expr) throw std::runtime_error("UnaryOpExpr.expr is null");
         auto e = expr->evaluate(context);
@@ -1281,8 +1269,8 @@ private:
     std::shared_ptr<Expression> right;
     Op op;
 public:
-    BinaryOpExpr(const Location & loc, std::shared_ptr<Expression> && l, std::shared_ptr<Expression> && r, Op o)
-        : Expression(loc), left(std::move(l)), right(std::move(r)), op(o) {}
+    BinaryOpExpr(const Location & location, std::shared_ptr<Expression> && l, std::shared_ptr<Expression> && r, Op o)
+        : Expression(location), left(std::move(l)), right(std::move(r)), op(o) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!left) throw std::runtime_error("BinaryOpExpr.left is null");
         if (!right) throw std::runtime_error("BinaryOpExpr.right is null");
@@ -1439,8 +1427,8 @@ class MethodCallExpr : public Expression {
     std::shared_ptr<VariableExpr> method;
     ArgumentsExpression args;
 public:
-    MethodCallExpr(const Location & loc, std::shared_ptr<Expression> && obj, std::shared_ptr<VariableExpr> && m, ArgumentsExpression && a)
-        : Expression(loc), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {}
+    MethodCallExpr(const Location & location, std::shared_ptr<Expression> && obj, std::shared_ptr<VariableExpr> && m, ArgumentsExpression && a)
+        : Expression(location), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!object) throw std::runtime_error("MethodCallExpr.object is null");
         if (!method) throw std::runtime_error("MethodCallExpr.method is null");
@@ -1538,8 +1526,8 @@ class CallExpr : public Expression {
 public:
     std::shared_ptr<Expression> object;
     ArgumentsExpression args;
-    CallExpr(const Location & loc, std::shared_ptr<Expression> && obj, ArgumentsExpression && a)
-        : Expression(loc), object(std::move(obj)), args(std::move(a)) {}
+    CallExpr(const Location & location, std::shared_ptr<Expression> && obj, ArgumentsExpression && a)
+        : Expression(location), object(std::move(obj)), args(std::move(a)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!object) throw std::runtime_error("CallExpr.object is null");
         auto obj = object->evaluate(context);
@@ -1554,8 +1542,8 @@ public:
 class FilterExpr : public Expression {
     std::vector<std::shared_ptr<Expression>> parts;
 public:
-    FilterExpr(const Location & loc, std::vector<std::shared_ptr<Expression>> && p)
-      : Expression(loc), parts(std::move(p)) {}
+    FilterExpr(const Location & location, std::vector<std::shared_ptr<Expression>> && p)
+      : Expression(location), parts(std::move(p)) {}
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         Value result;
         bool first = true;
@@ -2472,7 +2460,7 @@ private:
                 static std::regex leading_space_regex(R"(^\s+)");
                 text = std::regex_replace(text, leading_space_regex, "");
               } else if (options.trim_blocks && (it - 1) != begin && !dynamic_cast<ExpressionTemplateToken*>((*(it - 2)).get())) {
-                if (!text.empty() && text[0] == '\n') {
+                if (text.length() > 0 && text[0] == '\n') {
                   text.erase(0, 1);
                 }
               }
@@ -2550,7 +2538,7 @@ public:
         TemplateTokenIterator begin = tokens.begin();
         auto it = begin;
         TemplateTokenIterator end = tokens.end();
-        return parser.parseTemplate(begin, it, end, /* fully= */ true);
+        return parser.parseTemplate(begin, it, end, /* full= */ true);
     }
 };
 
@@ -2589,7 +2577,7 @@ inline std::shared_ptr<Context> Context::builtins() {
     throw std::runtime_error(args.at("message").get<std::string>());
   }));
   globals.set("tojson", simple_function("tojson", { "value", "indent" }, [](const std::shared_ptr<Context> &, Value & args) {
-    return Value(args.at("value").dump(args.get<int64_t>("indent", -1), /* to_json= */ true));
+    return Value(args.at("value").dump(args.get<int64_t>("indent", -1), /* tojson= */ true));
   }));
   globals.set("items", simple_function("items", { "object" }, [](const std::shared_ptr<Context> &, Value & args) {
     auto items = Value::array();
@@ -2611,25 +2599,21 @@ inline std::shared_ptr<Context> Context::builtins() {
   globals.set("last", simple_function("last", { "items" }, [](const std::shared_ptr<Context> &, Value & args) {
     auto items = args.at("items");
     if (!items.is_array()) throw std::runtime_error("object is not a list");
-    if (items.empty()) return Value();
+    if (items.size() == 0) return Value();
     return items.at(items.size() - 1);
   }));
   globals.set("trim", simple_function("trim", { "text" }, [](const std::shared_ptr<Context> &, Value & args) {
     auto & text = args.at("text");
     return text.is_null() ? text : Value(strip(text.get<std::string>()));
   }));
-  auto char_transform_function = [](const std::string & name, const std::function<char(char)> & fn) {
-    return simple_function(name, { "text" }, [=](const std::shared_ptr<Context> &, Value & args) {
-      auto text = args.at("text");
-      if (text.is_null()) return text;
-      std::string res;
-      auto str = text.get<std::string>();
-      std::transform(str.begin(), str.end(), std::back_inserter(res), fn);
-      return Value(res);
-    });
-  };
-  globals.set("lower", char_transform_function("lower", ::tolower));
-  globals.set("upper", char_transform_function("upper", ::toupper));
+  globals.set("lower", simple_function("lower", { "text" }, [](const std::shared_ptr<Context> &, Value & args) {
+    auto text = args.at("text");
+    if (text.is_null()) return text;
+    std::string res;
+    auto str = text.get<std::string>();
+    std::transform(str.begin(), str.end(), std::back_inserter(res), ::tolower);
+    return Value(res);
+  }));
   globals.set("default", Value::callable([=](const std::shared_ptr<Context> &, ArgumentsValue & args) {
     args.expectArgs("default", {2, 3}, {0, 1});
     auto & value = args.args[0];
@@ -2759,17 +2743,12 @@ inline std::shared_ptr<Context> Context::builtins() {
     return Value::callable([=](const std::shared_ptr<Context> & context, ArgumentsValue & args) {
       args.expectArgs(is_select ? "select" : "reject", {2, (std::numeric_limits<size_t>::max)()}, {0, 0});
       auto & items = args.args[0];
-      if (items.is_null()) {
+      if (items.is_null())
         return Value::array();
-      }
-      if (!items.is_array()) {
-        throw std::runtime_error("object is not iterable: " + items.dump());
-      }
+      if (!items.is_array()) throw std::runtime_error("object is not iterable: " + items.dump());
 
       auto filter_fn = context->get(args.args[1]);
-      if (filter_fn.is_null()) {
-        throw std::runtime_error("Undefined filter: " + args.args[1].dump());
-      }
+      if (filter_fn.is_null()) throw std::runtime_error("Undefined filter: " + args.args[1].dump());
 
       auto filter_args = Value::array();
       for (size_t i = 2, n = args.args.size(); i < n; i++) {
@@ -2891,25 +2870,20 @@ inline std::shared_ptr<Context> Context::builtins() {
         auto v = arg.get<int64_t>();
         startEndStep[i] = v;
         param_set[i] = true;
+        }
       }
-    }
-    for (auto & [name, value] : args.kwargs) {
-      size_t i;
-      if (name == "start") {
-        i = 0;
-      } else if (name == "end") {
-        i = 1;
-      } else if (name == "step") {
-        i = 2;
-      } else {
-        throw std::runtime_error("Unknown argument " + name + " for function range");
-      }
+      for (auto & [name, value] : args.kwargs) {
+        size_t i;
+        if (name == "start") i = 0;
+        else if (name == "end") i = 1;
+        else if (name == "step") i = 2;
+        else throw std::runtime_error("Unknown argument " + name + " for function range");
 
-      if (param_set[i]) {
-        throw std::runtime_error("Duplicate argument " + name + " for function range");
-      }
-      startEndStep[i] = value.get<int64_t>();
-      param_set[i] = true;
+        if (param_set[i]) {
+          throw std::runtime_error("Duplicate argument " + name + " for function range");
+        }
+        startEndStep[i] = value.get<int64_t>();
+        param_set[i] = true;
     }
     if (!param_set[1]) {
       throw std::runtime_error("Missing required argument 'end' for function range");

common/sampling.cpp

@@ -208,9 +208,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                                                         trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                         trigger_tokens.data(), trigger_tokens.size())
              :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
-        if (!grmr) {
-            return nullptr;
-        }
     }
 
     auto * result = new common_sampler {

common/speculative.cpp

@@ -14,7 +14,7 @@ struct common_speculative {
     struct llama_context * ctx;
     struct common_sampler * smpl;
 
-    llama_batch batch;
+    llama_batch_ext_ptr batch;
     llama_tokens prompt;
 };
 
@@ -23,7 +23,7 @@ struct common_speculative * common_speculative_init(
     auto * result = new common_speculative {
         /* .ctx    = */ ctx_dft,
         /* .smpl   = */ nullptr,
-        /* .batch  = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .batch  = */ llama_batch_ext_ptr(llama_batch_ext_init(llama_n_batch(ctx_dft), 1)),
         /* .prompt = */ {},
     };
 
@@ -69,8 +69,6 @@ void common_speculative_free(struct common_speculative * spec) {
 
     common_sampler_free(spec->smpl);
 
-    llama_batch_free(spec->batch);
-
     delete spec;
 }
 
@@ -151,6 +149,8 @@ llama_tokens common_speculative_gen_draft(
 
     const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
 
+    const llama_seq_id seq_id = 0;
+
     // reuse as much as possible from the old draft context
     // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
     for (int i = 0; i < (int) prompt.size(); ++i) {
@@ -206,40 +206,40 @@ llama_tokens common_speculative_gen_draft(
     }
 
     // prepare a batch to evaluate any new tokens in the prompt
-    common_batch_clear(batch);
+    llama_batch_ext_clear(batch.get());
 
     for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
         //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
-        common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
+        llama_batch_ext_add_text(batch.get(), prompt_tgt[i], i - i_start, &seq_id, 1, false);
 
         prompt.push_back(prompt_tgt[i]);
     }
 
     // we should rarely end-up here during normal decoding
-    if (batch.n_tokens > 0) {
+    if (llama_batch_ext_get_n_tokens(batch.get()) > 0) {
         //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
 
-        llama_decode(ctx, batch);
+        llama_decode_ext(ctx, batch.get());
     }
 
     const llama_pos n_past = prompt.size();
 
     LOG_DBG("%s: n_past = %d\n", __func__, n_past);
 
-    common_batch_clear(batch);
-    common_batch_add  (batch, id_last, n_past, { 0 }, true);
+    llama_batch_ext_clear(batch.get());
+    llama_batch_ext_add_text(batch.get(), id_last, n_past, &seq_id, 1, true);
 
     prompt.push_back(id_last);
 
     //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
 
-    llama_decode(ctx, batch);
+    llama_decode_ext(ctx, batch.get());
 
     common_sampler_reset(smpl);
 
     // sample n_draft tokens from the draft model
     for (int i = 0; i < params.n_draft; ++i) {
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch.get());
 
         common_sampler_sample(smpl, ctx, 0, true);
 
@@ -266,10 +266,10 @@ llama_tokens common_speculative_gen_draft(
             break;
         }
 
-        common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
+        llama_batch_ext_add_text(batch.get(), id, n_past + i + 1, &seq_id, 1, true);
 
         // evaluate the drafted tokens on the draft model
-        llama_decode(ctx, batch);
+        llama_decode_ext(ctx, batch.get());
 
         prompt.push_back(id);
     }

convert_hf_to_gguf_update.py

@@ -110,12 +110,6 @@ models = [
     {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
-    {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
-    {"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", },
 ]
 
 

convert_lora_to_gguf.py

@@ -24,7 +24,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 # reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import LazyTorchTensor, ModelBase
+from convert_hf_to_gguf import LazyTorchTensor, Model
 
 logger = logging.getLogger("lora-to-gguf")
 
@@ -340,11 +340,11 @@ if __name__ == '__main__':
             sys.exit(1)
     else:
         logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = ModelBase.load_hparams(dir_base_model)
+        hparams = Model.load_hparams(dir_base_model)
 
     with torch.inference_mode():
         try:
-            model_class = ModelBase.from_model_architecture(hparams["architectures"][0])
+            model_class = Model.from_model_architecture(hparams["architectures"][0])
         except NotImplementedError:
             logger.error(f"Model {hparams['architectures'][0]} is not supported")
             sys.exit(1)

docs/backend/OPENCL.md

@@ -145,13 +145,8 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
 * Clang 19
 * Ninja
 * Visual Studio 2022
-* Powershell 7
 
-Visual Studio provides necessary headers and libraries although it is not directly used for building.
-Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
-
-Powershell 7 is used for the following commands.
-If an older version of Powershell is used, these commands may not work as they are.
+Powershell is used for the following instructions.
 
 ### I. Setup Environment
 
@@ -201,9 +196,10 @@ ninja
 
 ## Known Issues
 
-- Currently OpenCL backend does not work on Adreno 6xx GPUs.
+- Qwen2.5 0.5B model produces gibberish output with Adreno kernels.
 
 ## TODO
 
+- Fix Qwen2.5 0.5B
 - Optimization for Q6_K
 - Support and optimization for Q4_K

docs/backend/SYCL.md

@@ -20,7 +20,7 @@
 **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 Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
 - **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.
 
@@ -227,19 +227,30 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 **oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
 
-**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
+
+**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
 
 ```sh
-git clone https://github.com/oneapi-src/oneDNN.git
-cd oneDNN
-cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-cmake --build build-nvidia --config Release
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
+cmake --build buildWithCublas --config Release
 ```
 
 - **Adding support to AMD GPUs**
 
 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
 
+**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
+
+```sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+# Find your HIPTARGET with rocminfo, under the key 'Name:'
+cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
+cmake --build buildWithrocBLAS --config Release
+```
+
 3. **Verify installation and environment**
 
 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -302,39 +313,37 @@ cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -
 cmake --build build --config Release -j -v
 ```
 
-It is possible to come across some precision issues when running tests that stem from using faster
-instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS`
-as `-cl-fp32-correctly-rounded-divide-sqrt`
-
 #### Nvidia GPU
 
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
 ```sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 # Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
 
 # build all binary
 cmake --build build --config Release -j -v
 ```
 
-It is possible to come across some precision issues when running tests that stem from using faster
-instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler.
-
 #### AMD GPU
 
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
 ```sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
+
 # Build LLAMA with rocBLAS acceleration through SYCL
 
 ## AMD
@@ -425,13 +434,13 @@ 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 -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
 ```
 
 - 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 -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
 ```
 
 *Notes:*
@@ -475,12 +484,6 @@ b. Enable oneAPI running environment:
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-- if you are using Powershell, enable the runtime environment with the following:
-
-```
-cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'
-```
-
 c. Verify installation
 
 In the oneAPI command line, run the following to print the available SYCL devices:
@@ -511,13 +514,13 @@ You could download the release package for Windows directly, which including bin
 
 Choose one of following methods to build from source code.
 
-#### 1. Script
+1. Script
 
 ```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-#### 2. CMake
+2. CMake
 
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
@@ -546,84 +549,13 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-#### 3. Visual Studio
+3. Visual Studio
 
-You have two options to use Visual Studio to build llama.cpp:
-- As CMake Project using CMake presets.
-- Creating a Visual Studio solution to handle the project.
-
-**Note**:
-
-All following commands are executed in PowerShell.
-
-##### - Open as a CMake Project
-
-You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets:
-
-- `x64-windows-sycl-release`
-
-- `x64-windows-sycl-debug`
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
-- For a minimal experimental setup, you can build only the inference executable using:
 
-    ```Powershell
-    cmake --build build --config Release -j --target llama-cli
-    ```
-
-##### - Generating a Visual Studio Solution
-
-You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file.
-
-If you want to use the Intel C++ Compiler for the entire `llama.cpp` project, run the following command:
-
-```Powershell
-cmake -B build -G "Visual Studio 17 2022" -T "Intel C++ Compiler 2025" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release
-```
-
-If you prefer to use the Intel C++ Compiler only for `ggml-sycl`, ensure that `ggml` and its backend libraries are built as shared libraries ( i.e. `-DBUILD_SHARED_LIBRARIES=ON`, this is default behaviour):
-
-```Powershell
-cmake -B build -G "Visual Studio 17 2022" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release \
-      -DSYCL_INCLUDE_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\include" \
-      -DSYCL_LIBRARY_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\lib"
-```
-
-If successful the build files have been written to: *path/to/llama.cpp/build*
-Open the project file **build/llama.cpp.sln** with Visual Studio.
-
-Once the Visual Studio solution is created, follow these steps:
-
-1. Open the solution in Visual Studio.
-
-2. Right-click on `ggml-sycl` and select **Properties**.
-
-3. In the left column, expand **C/C++** and select **DPC++**.
-
-4. In the right panel, find **Enable SYCL Offload** and set it to `Yes`.
-
-5. Apply the changes and save.
-
-
-*Navigation Path:*
-
-```
-Properties -> C/C++ -> DPC++ -> Enable SYCL Offload (Yes)
-```
-
-Now, you can build `llama.cpp` with the SYCL backend as a Visual Studio project.
-To do it from menu: `Build -> Build Solution`.
-Once it is completed, final results will be in **build/Release/bin**
-
-*Additional Note*
-
-- You can avoid specifying `SYCL_INCLUDE_DIR` and `SYCL_LIBRARY_DIR` in the CMake command by setting the environment variables:
-
-    - `SYCL_INCLUDE_DIR_HINT`
-
-    - `SYCL_LIBRARY_DIR_HINT`
-
-- Above instruction has been tested with Visual Studio 17 Community edition and oneAPI 2025.0. We expect them to work also with future version if the instructions are adapted accordingly.
+- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
 
 ### III. Run the inference
 
@@ -697,13 +629,13 @@ 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 -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
 ```
 
 - 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 -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
 ```
 
 

docs/build.md

@@ -132,14 +132,12 @@ You may find the official downloads here: [NVIDIA developer site](https://develo
 
 
 #### Compile and run inside a Fedora Toolbox Container
-We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
+We also have a [guide](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
 
 **Recommended for:**
-- ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
-  - (there are no supported CUDA packages for these systems)
-- ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads).
-  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system)
-- ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean.
+
+- ***Particularly*** *convenient* for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
+- Toolbox is installed by default: [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde).
 - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download)
 
 
@@ -191,7 +189,7 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
@@ -218,7 +216,6 @@ By default, all supported compute capabilities are enabled. To customize this be
 
 ```bash
 cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
-cmake --build build --config Release
 ```
 
 This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.
@@ -259,6 +256,8 @@ You can download it from your Linux distro's package manager or from here: [ROCm
       cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
+  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
+  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
@@ -294,10 +293,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
 
-### Unified Memory
-
-On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
 ## Vulkan
 
 **Windows**
@@ -438,116 +433,6 @@ llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
 
 For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
 
-## Arm® KleidiAI™
-KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
-
-To enable KleidiAI, go to the llama.cpp directory and build using CMake
-```bash
-cmake -B build -DGGML_CPU_KLEIDIAI=ON
-cmake --build build --config Release
-```
-You can verify that KleidiAI is being used by running
-```bash
-./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
-```
-If KleidiAI is enabled, the ouput will contain a line similar to:
-```
-load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
-```
-KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
-
-Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
-
-## OpenCL
-
-This provides GPU acceleration through OpenCL on recent Adreno GPU.
-More information about OpenCL backend can be found in [OPENCL.md](./backend/OPENCL.md) for more information.
-
-### Android
-
-Assume NDK is available in `$ANDROID_NDK`. First, install OpenCL headers and ICD loader library if not available,
-
-```sh
-mkdir -p ~/dev/llm
-cd ~/dev/llm
-
-git clone https://github.com/KhronosGroup/OpenCL-Headers && \
-cd OpenCL-Headers && \
-cp -r CL $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-
-cd ~/dev/llm
-
-git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \
-cd OpenCL-ICD-Loader && \
-mkdir build_ndk && cd build_ndk && \
-cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
-  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
-  -DOPENCL_ICD_LOADER_HEADERS_DIR=$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
-  -DANDROID_ABI=arm64-v8a \
-  -DANDROID_PLATFORM=24 \
-  -DANDROID_STL=c++_shared && \
-ninja && \
-cp libOpenCL.so $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-```
-
-Then build llama.cpp with OpenCL enabled,
-
-```sh
-cd ~/dev/llm
-
-git clone https://github.com/ggml-org/llama.cpp && \
-cd llama.cpp && \
-mkdir build-android && cd build-android
-
-cmake .. -G Ninja \
-  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
-  -DANDROID_ABI=arm64-v8a \
-  -DANDROID_PLATFORM=android-28 \
-  -DBUILD_SHARED_LIBS=OFF \
-  -DGGML_OPENCL=ON
-
-ninja
-```
-
-### Windows Arm64
-
-First, install OpenCL headers and ICD loader library if not available,
-
-```powershell
-mkdir -p ~/dev/llm
-
-cd ~/dev/llm
-git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers
-mkdir build && cd build
-cmake .. -G Ninja `
-  -DBUILD_TESTING=OFF `
-  -DOPENCL_HEADERS_BUILD_TESTING=OFF `
-  -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
-  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
-cmake --build . --target install
-
-cd ~/dev/llm
-git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader
-mkdir build && cd build
-cmake .. -G Ninja `
-  -DCMAKE_BUILD_TYPE=Release `
-  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
-  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
-cmake --build . --target install
-```
-
-Then build llama.cpp with OpenCL enabled,
-
-```powershell
-cmake .. -G Ninja `
-  -DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" `
-  -DCMAKE_BUILD_TYPE=Release `
-  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
-  -DBUILD_SHARED_LIBS=OFF `
-  -DGGML_OPENCL=ON
-ninja
-```
-
 ## Android
 
 To read documentation for how to build on Android, [click here](./android.md)

docs/cuda-fedora.md

@@ -14,7 +14,9 @@ In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox
 - [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
 - [Installing Essential Development Tools](#installing-essential-development-tools)
 - [Adding the CUDA Repository](#adding-the-cuda-repository)
-- [Installing Nvidia Driver Libraries](#installing-nvidia-driver-libraries)
+- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
+- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
+- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
 - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
 - [Configuring the Environment](#configuring-the-environment)
 - [Verifying the Installation](#verifying-the-installation)
@@ -65,7 +67,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for
    sudo dnf distro-sync
    ```
 
-2. **Install **Vim** the default text editor (Optional):**
+2. **Install the Default Text Editor (Optional):**
 
    ```bash
    sudo dnf install vim-default-editor --allowerasing
@@ -95,48 +97,36 @@ After adding the repository, synchronize the package manager again:
 sudo dnf distro-sync
 ```
 
-## Installing Nvidia Driver Libraries
+## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
 
-First, we need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go):
+We need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go).
 
 ```bash
 ls -la /usr/lib64/libcuda.so.1
 ```
 
-### If *`libcuda.so.1`* is missing:
-
-```
-ls: cannot access '/usr/lib64/libcuda.so.1': No such file or directory
-```
-
 **Explanation:**
-The host dose not supply the CUDA drivers, **install them now:**
 
-#### Install the Nvidia Driver Libraries on Guest:
+- `nvidia-driver-libs` and `nvidia-driver-cuda-libs` contains necessary NVIDIA driver libraries required by CUDA,
+  on hosts with NVIDIA drivers installed the Fedora Container will supply the host libraries.
+
+### Install Nvidia Driver Libraries on Guest (if `libcuda.so.1` was NOT found).
 
 ```bash
-sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
+sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
 ```
 
-### If *`libcuda.so.1`* exists:
-```
-lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07
-```
+### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found).
 
-**Explanation:**
-The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:**
+If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
 
-#### Update the Toolbox RPM Database to include the Host-Supplied Libraries:
-
-Note: we do not actually install the libraries, we just update the DB so that the guest system knows they are supplied by the host.
-
-##### 1. Download `nvidia-` parts that are supplied by the host RPM's (with dependencies)
+#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies)
 
 ```bash
-sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
+sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs
 ```
 
-##### 2. Update the RPM database to assume the installation of these packages.
+#### 2. Update the RPM database to assume the installation of these packages.
 
 ```bash
 sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
@@ -144,26 +134,23 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
 
 **Note:**
 
-- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere.
+- The `--justdb` option only updates the RPM database, without touching the filesystem.
 
-##### Check that the RPM Database has been correctly updated:
-
-**Note:** This is the same command as in the *"Install the Nvidia Driver Libraries on Guest"* for if *`libcuda.so.1`* was missing.
+#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
 
+After manually installing the dependencies, run:
 
 ```bash
-sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
+sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
 ```
 
-*(this time it will not install anything, as the database things that these packages are already installed)*
+You should receive a message indicating the package is already installed:
 
 ```
 Updating and loading repositories:
 Repositories loaded.
-Package "nvidia-driver-cuda-3:570.124.06-1.fc41.x86_64" is already installed.
-Package "nvidia-driver-libs-3:570.124.06-1.fc41.x86_64" is already installed.
-Package "nvidia-driver-cuda-libs-3:570.124.06-1.fc41.x86_64" is already installed.
-Package "nvidia-persistenced-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-libs-3:570.86.10-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-libs-3:570.86.10-1.fc41.x86_64" is already installed.
 
 Nothing to do.
 ```
@@ -220,9 +207,9 @@ You should see output similar to:
 ```
 nvcc: NVIDIA (R) Cuda compiler driver
 Copyright (c) 2005-2025 NVIDIA Corporation
-Built on Fri_Feb_21_20:23:50_PST_2025
-Cuda compilation tools, release 12.8, V12.8.93
-Build cuda_12.8.r12.8/compiler.35583870_0
+Built on Wed_Jan_15_19:20:09_PST_2025
+Cuda compilation tools, release 12.8, V12.8.61
+Build cuda_12.8.r12.8/compiler.35404655_0
 ```
 
 This output confirms that the CUDA compiler is accessible and indicates the installed version.

docs/install.md

@@ -9,13 +9,6 @@ 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
-
-```sh
-sudo port install llama.cpp
-```
-see also: https://ports.macports.org/port/llama.cpp/details/
-
 ## Nix
 
 On Mac and Linux, the Nix package manager can be used via

docs/multimodal/gemma3.md

@@ -1,51 +0,0 @@
-# Gemma 3 vision
-
-> [!IMPORTANT]
->
-> This is very experimental, only used for demo purpose.
-
-## Quick started
-
-You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)'s Hugging Face account
-
-```bash
-# build
-cmake -B build
-cmake --build build --target llama-mtmd-cli
-
-# alternatively, install from brew (MacOS)
-brew install llama.cpp
-
-# run it
-llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
-
-# note: 1B model does not support vision
-```
-
-## How to get mmproj.gguf?
-
-Simply to add `--mmproj` in when converting model via `convert_hf_to_gguf.py`:
-
-```bash
-cd gemma-3-4b-it
-python ../llama.cpp/convert_hf_to_gguf.py --outfile model.gguf --outtype f16 --mmproj .
-# output file: mmproj-model.gguf
-```
-
-## How to run it?
-
-What you need:
-- The text model GGUF, can be converted using `convert_hf_to_gguf.py`
-- The mmproj file from step above
-- An image file
-
-```bash
-# build
-cmake -B build
-cmake --build build --target llama-mtmd-cli
-
-# run it
-./build/bin/llama-mtmd-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
-```

docs/multimodal/llava.md

@@ -1,143 +0,0 @@
-# LLaVA
-
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
-as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
-
-The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
-and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
-models are available.
-For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
-
-After API is confirmed, more models will be supported / uploaded.
-
-## Usage
-Build the `llama-mtmd-cli` binary.
-
-After building, run: `./llama-mtmd-cli` to see the usage. For example:
-
-```sh
-./llama-mtmd-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf \
-    --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf \
-    --chat-template vicuna
-```
-
-**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
-**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
-
-## LLaVA 1.5
-
-1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
-
-```sh
-git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
-
-git clone https://huggingface.co/openai/clip-vit-large-patch14-336
-```
-
-2. Install the required Python packages:
-
-```sh
-pip install -r examples/llava/requirements.txt
-```
-
-3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
-
-```sh
-python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
-```
-
-4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
-
-```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
-```
-
-5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
-
-```sh
-python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown
-```
-
-Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
-
-## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-
-2) Install the required Python packages:
-
-```sh
-pip install -r examples/llava/requirements.txt
-```
-
-3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
-- you will find a llava.projector and a llava.clip file in your model directory
-
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-5) Create the visual gguf model:
-```console
-python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
-- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-6) Then convert the model to gguf format:
-```console
-python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-7) And finally we can run the llava cli using the 1.6 model version:
-```console
-./llama-mtmd-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf
-```
-
-**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
-
-**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
-
-**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model.
-
-```python
-import os
-import transformers
-
-model_path = ...
-llm_export_path = ...
-
-tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
-model = transformers.AutoModelForImageTextToText.from_pretrained(model_path)
-
-tokenizer.save_pretrained(llm_export_path)
-model.language_model.save_pretrained(llm_export_path)
-```
-
-Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.
-
-## Chat template
-
-For llava-1.5 and llava-1.6, you need to use `vicuna` chat template. Simply add `--chat-template vicuna` to activate this template.
-
-
-## How to know if you are running in llava-1.5 or llava-1.6 mode
-
-When running llava-cli you will see a visual information right before the prompt is being processed:
-
-**Llava-1.5:**
-`encode_image_with_clip: image embedding created: 576 tokens`
-
-**Llava-1.6 (anything above 576):**
-`encode_image_with_clip: image embedding created: 2880 tokens`
-
-
-Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6

examples/CMakeLists.txt

@@ -21,6 +21,11 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
 
+    if (NOT WIN32)
+        # disabled on Windows because it uses internal functions not exported with LLAMA_API
+        add_subdirectory(gbnf-validator)
+    endif()
+
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -53,6 +58,10 @@ else()
         add_subdirectory(convert-llama2c-to-ggml)
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
+        if (NOT WIN32)
+            # disabled on Windows because it uses internal functions not exported with LLAMA_API
+            add_subdirectory(quantize-stats)
+        endif()
         add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)

examples/batched-bench/batched-bench.cpp

@@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@@ -59,24 +59,17 @@ int main(int argc, char ** argv) {
 
     const int32_t n_kv_max = llama_n_ctx(ctx);
 
-    llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
+    llama_batch_ext * batch = llama_batch_ext_init(n_kv_max, 1);
 
     // decode in batches of ctx_params.n_batch tokens
-    auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+    auto decode_helper = [](llama_context * ctx, llama_batch_ext * batch, int32_t n_batch) {
+        const int32_t n_batch_tokens = llama_batch_ext_get_n_tokens(batch);
+        for (int32_t i = 0; i < (int32_t) n_batch_tokens; i += n_batch) {
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (n_batch_tokens - i));
 
-            llama_batch batch_view = {
-                n_tokens,
-                batch.token    + i,
-                nullptr,
-                batch.pos      + i,
-                batch.n_seq_id + i,
-                batch.seq_id   + i,
-                batch.logits   + i,
-            };
+            llama_batch_ext_ptr batch_view = llama_batch_ext_ptr(llama_batch_ext_get_view(batch, i, n_tokens));
 
-            const int ret = llama_decode(ctx, batch_view);
+            const int ret = llama_decode_ext(ctx, batch_view.get());
             if (ret != 0) {
                 LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
@@ -91,7 +84,8 @@ int main(int argc, char ** argv) {
     // warm up
     {
         for (int i = 0; i < 16; ++i) {
-            common_batch_add(batch, 0, i, { 0 }, false);
+            const llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch, 0, i, &seq_id, 1, false);
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -121,14 +115,14 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                common_batch_clear(batch);
+                llama_batch_ext_clear(batch);
 
                 for (int i = 0; i < pp; ++i) {
                     for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        common_batch_add(batch, 0, i, { j }, false);
+                        llama_batch_ext_add_text(batch, 0, i, &j, 1, false);
                     }
                 }
-                batch.logits[batch.n_tokens - 1] = true;
+                llama_batch_ext_set_output_last(batch);
 
                 const auto t_pp_start = ggml_time_us();
 
@@ -150,10 +144,10 @@ int main(int argc, char ** argv) {
                 const auto t_tg_start = ggml_time_us();
 
                 for (int i = 0; i < tg; ++i) {
-                    common_batch_clear(batch);
+                    llama_batch_ext_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        common_batch_add(batch, 0, pp + i, { j }, true);
+                        llama_batch_ext_add_text(batch, 0, pp + i, &j, 1, true);
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -191,7 +185,7 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
-    llama_batch_free(batch);
+    llama_batch_ext_free(batch);
 
     llama_free(ctx);
     llama_model_free(model);

examples/batched/batched.cpp

@@ -41,7 +41,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n" , __func__);
@@ -102,7 +102,7 @@ int main(int argc, char ** argv) {
 
     // create a llama_batch
     // we use this object to submit token data for decoding
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
+    llama_batch_ext * batch = llama_batch_ext_init(std::max(tokens_list.size(), (size_t) n_parallel), n_parallel);
 
     std::vector<llama_seq_id> seq_ids(n_parallel, 0);
     for (int32_t i = 0; i < n_parallel; ++i) {
@@ -111,12 +111,12 @@ int main(int argc, char ** argv) {
 
     // evaluate the initial prompt
     for (size_t i = 0; i < tokens_list.size(); ++i) {
-        common_batch_add(batch, tokens_list[i], i, seq_ids, false);
+        llama_batch_ext_add_text(batch, tokens_list[i], i, seq_ids.data(), seq_ids.size(), false);
     }
-    GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
+    GGML_ASSERT(llama_batch_ext_get_n_tokens(batch) == (int) tokens_list.size());
 
     if (llama_model_has_encoder(model)) {
-        if (llama_encode(ctx, batch)) {
+        if (llama_encode_ext(ctx, batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
@@ -126,14 +126,14 @@ int main(int argc, char ** argv) {
             decoder_start_token_id = llama_vocab_bos(vocab);
         }
 
-        common_batch_clear(batch);
-        common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
+        llama_batch_ext_clear(batch);
+        llama_batch_ext_add_text(batch, decoder_start_token_id, 0, seq_ids.data(), seq_ids.size(), false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
-    batch.logits[batch.n_tokens - 1] = true;
+    llama_batch_ext_set_output_last(batch);
 
-    if (llama_decode(ctx, batch) != 0) {
+    if (llama_decode_ext(ctx, batch) != 0) {
         LOG_ERR("%s: llama_decode() failed\n", __func__);
         return 1;
     }
@@ -155,16 +155,16 @@ int main(int argc, char ** argv) {
 
     // remember the batch index of the last token for each parallel sequence
     // we need this to determine which logits to sample from
-    std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
+    std::vector<int32_t> i_batch(n_parallel, llama_batch_ext_get_n_tokens(batch) - 1);
 
-    int n_cur    = batch.n_tokens;
+    int n_cur    = llama_batch_ext_get_n_tokens(batch);
     int n_decode = 0;
 
     const auto t_main_start = ggml_time_us();
 
     while (n_cur <= n_predict) {
         // prepare the next batch
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -193,23 +193,23 @@ int main(int argc, char ** argv) {
 
             streams[i] += common_token_to_piece(ctx, new_token_id);
 
-            i_batch[i] = batch.n_tokens;
+            i_batch[i] = llama_batch_ext_get_n_tokens(batch);
 
             // push this new token for next evaluation
-            common_batch_add(batch, new_token_id, n_cur, { i }, true);
+            llama_batch_ext_add_text(batch, new_token_id, n_cur, &i, 1, true);
 
             n_decode += 1;
         }
 
         // all streams are finished
-        if (batch.n_tokens == 0) {
+        if (llama_batch_ext_get_n_tokens(batch) == 0) {
             break;
         }
 
         n_cur += 1;
 
         // evaluate the current batch with the transformer model
-        if (llama_decode(ctx, batch)) {
+        if (llama_decode_ext(ctx, batch)) {
             LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
@@ -234,7 +234,7 @@ int main(int argc, char ** argv) {
 
     fprintf(stderr, "\n");
 
-    llama_batch_free(batch);
+    llama_batch_ext_free(batch);
 
     llama_sampler_free(smpl);
     llama_free(ctx);

examples/cvector-generator/cvector-generator.cpp

@@ -343,7 +343,8 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
     llama_kv_self_clear(ctx);
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
+    auto batch = llama_batch_ext_ptr::init_from_text(tokens.data(), tokens.size(), 0, 0, true);
+    if (llama_decode_ext(ctx, batch.get())) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }

examples/embedding/embedding.cpp

@@ -26,14 +26,14 @@ static std::vector<std::string> split_lines(const std::string & s, const std::st
     return lines;
 }
 
-static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
+static void batch_add_seq(common_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
     size_t n_tokens = tokens.size();
     for (size_t i = 0; i < n_tokens; i++) {
-        common_batch_add(batch, tokens[i], i, { seq_id }, true);
+        batch.add_text(tokens[i], i, seq_id, true);
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+static void batch_decode(llama_context * ctx, common_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
     const struct llama_model * model = llama_get_model(ctx);
 
@@ -41,21 +41,21 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     llama_kv_self_clear(ctx);
 
     // run model
-    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, llama_batch_ext_get_n_tokens(batch.get()), n_seq);
     if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
         // encoder-only model
-        if (llama_encode(ctx, batch) < 0) {
+        if (llama_encode_ext(ctx, batch.get()) < 0) {
             LOG_ERR("%s : failed to encode\n", __func__);
         }
     } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
         // decoder-only model
-        if (llama_decode(ctx, batch) < 0) {
+        if (llama_decode_ext(ctx, batch.get()) < 0) {
             LOG_ERR("%s : failed to decode\n", __func__);
         }
     }
 
-    for (int i = 0; i < batch.n_tokens; i++) {
-        if (!batch.logits[i]) {
+    for (int i = 0; i < llama_batch_ext_get_n_tokens(batch.get()); i++) {
+        if (!batch.tokens[i].logits) {
             continue;
         }
 
@@ -69,8 +69,8 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             GGML_ASSERT(embd != NULL && "failed to get token embeddings");
         } else {
             // try to get sequence embeddings - supported only when pooling_type is not NONE
-            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-            embd_pos = batch.seq_id[i][0];
+            embd = llama_get_embeddings_seq(ctx, batch.tokens[i].seq_id);
+            embd_pos = batch.tokens[i].seq_id;
             GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
         }
 
@@ -89,13 +89,6 @@ int main(int argc, char ** argv) {
     common_init();
 
     params.embedding = true;
-
-    // utilize the full context
-    if (params.n_batch < params.n_ctx) {
-        LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
-        params.n_batch = params.n_ctx;
-    }
-
     // For non-causal models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
 
@@ -141,6 +134,7 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
@@ -177,7 +171,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    struct common_batch batch = common_batch(n_batch, 1);
 
     // count number of embeddings
     int n_embd_count = 0;
@@ -204,12 +198,12 @@ int main(int argc, char ** argv) {
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity
-        if (batch.n_tokens + n_toks > n_batch) {
+        if (batch.get_n_tokens() + n_toks > n_batch) {
             float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.get_n_tokens() : s;
             s = 0;
-            common_batch_clear(batch);
+            batch.clear();
         }
 
         // add to batch
@@ -325,7 +319,6 @@ int main(int argc, char ** argv) {
     llama_perf_context_print(ctx);
 
     // clean up
-    llama_batch_free(batch);
     llama_backend_free();
 
     return 0;

examples/eval-callback/eval-callback.cpp

@@ -134,7 +134,8 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
+    auto batch = llama_batch_ext_ptr::init_from_text(tokens.data(), tokens.size(), 0, 0, true);
+    if (llama_decode_ext(ctx, batch.get())) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;
     }

examples/export-lora/export-lora.cpp

@@ -421,7 +421,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity > 1);
     try {
-        lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/gbnf-validator/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-gbnf-validator)
+add_executable(${TARGET} gbnf-validator.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gbnf-validator/gbnf-validator.cpp

@@ -1,5 +1,5 @@
-#include "../src/unicode.h"
-#include "../src/llama-grammar.h"
+#include "unicode.h"
+#include "llama-grammar.h"
 
 #include <cstdio>
 #include <cstdlib>

examples/gguf-split/gguf-split.cpp

@@ -408,6 +408,8 @@ static void gguf_merge(const split_params & split_params) {
         exit(EXIT_FAILURE);
     }
 
+    std::ofstream fout(split_params.output.c_str(), std::ios::binary);
+    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
 
     auto * ctx_out = gguf_init_empty();
 
@@ -451,6 +453,7 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
+                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -463,6 +466,7 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
+                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -475,6 +479,7 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
+                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -495,11 +500,9 @@ static void gguf_merge(const split_params & split_params) {
 
         fprintf(stderr, "\033[3Ddone\n");
     }
-    std::ofstream fout;
-    if (!split_params.dry_run) {
-        fout.open(split_params.output.c_str(), std::ios::binary);
-        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
-        // placeholder for the meta data
+
+    // placeholder for the meta data
+    {
         auto meta_size = gguf_get_meta_size(ctx_out);
         ::zeros(fout, meta_size);
     }
@@ -515,9 +518,7 @@ static void gguf_merge(const split_params & split_params) {
                 ggml_free(ctx_metas[i]);
             }
             gguf_free(ctx_out);
-            if (!split_params.dry_run) {
-                fout.close();
-            }
+            fout.close();
             exit(EXIT_FAILURE);
         }
         fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
@@ -539,11 +540,10 @@ static void gguf_merge(const split_params & split_params) {
             auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
             f_input.seekg(offset);
             f_input.read((char *)read_data.data(), n_bytes);
-            if (!split_params.dry_run) {
-                // write tensor data + padding
-                fout.write((const char *)read_data.data(), n_bytes);
-                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
-            }
+
+            // write tensor data + padding
+            fout.write((const char *)read_data.data(), n_bytes);
+            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
         }
 
         gguf_free(ctx_gguf);
@@ -552,15 +552,16 @@ static void gguf_merge(const split_params & split_params) {
         fprintf(stderr, "\033[3Ddone\n");
     }
 
-    if (!split_params.dry_run) {
+    {
         // go back to beginning of file and write the updated metadata
         fout.seekp(0);
         std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
         gguf_get_meta_data(ctx_out, data.data());
         fout.write((const char *)data.data(), data.size());
+
         fout.close();
+        gguf_free(ctx_out);
     }
-    gguf_free(ctx_out);
 
     fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
             __func__, split_params.output.c_str(), n_split, total_tensors);

examples/gritlm/gritlm.cpp

@@ -13,10 +13,10 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     const llama_model * model = llama_get_model(ctx);
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
+    llama_batch_ext * batch = llama_batch_ext_init(llama_n_batch(ctx), 1);
 
     for (uint64_t i = 0; i < sentences.size(); i++) {
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch);
 
         const std::string input_string = instruction + sentences[i];
 
@@ -41,7 +41,8 @@ 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);
+            const llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch, inputs[j], j, &seq_id, 1 , j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
@@ -50,7 +51,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         llama_set_causal_attn(ctx, false);
 
         // run model
-        llama_decode(ctx, batch);
+        llama_decode_ext(ctx, batch);
 
         // get embedding dimensions
         uint64_t n_embd = llama_model_n_embd(model);
@@ -89,7 +90,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 #endif
     }
 
-    llama_batch_free(batch);
+    llama_batch_ext_free(batch);
 
     return result;
 }
@@ -106,25 +107,26 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
-    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
+    llama_batch_ext * bat = llama_batch_ext_init(llama_n_batch(ctx), 1);
 
     std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
-        common_batch_clear(bat);
+        llama_batch_ext_clear(bat);
         {
             const int32_t n_inputs = inputs.size();
 
             for (int32_t i = 0; i < n_inputs; i++) {
-                common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
+                const llama_seq_id seq_id = 0;
+                llama_batch_ext_add_text(bat, inputs[i], i_current_token++, &seq_id, 1, i == n_inputs - 1);
             }
         }
         inputs.clear();
 
-        llama_decode(ctx, bat);
+        llama_decode_ext(ctx, bat);
 
-        llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
+        llama_token token = llama_sampler_sample(smpl, ctx, llama_batch_ext_get_n_tokens(bat) - 1);
 
         if (token == eos_token) {
             break;
@@ -145,7 +147,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
         std::printf("\n");
     }
 
-    llama_batch_free(bat);
+    llama_batch_ext_free(bat);
 
     return result;
 }
@@ -168,7 +170,7 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
 
     // create generation context
     llama_context * ctx = llama_init_from_model(model, cparams);

examples/imatrix/imatrix.cpp

@@ -497,7 +497,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_self_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+        llama_batch_ext * batch = llama_batch_ext_init(n_batch, 1);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -511,14 +511,15 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_vocab_bos(vocab);
             }
 
-            common_batch_clear(batch);
+            llama_batch_ext_clear(batch);
             for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+                const llama_seq_id seq_id = 0;
+                llama_batch_ext_add_text(batch, tokens[batch_start + i], j*n_batch + i, &seq_id, 1, true);
             }
 
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode_ext(ctx, batch)) {
                 LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
+                llama_batch_ext_free(batch);
                 return false;
             }
 
@@ -531,7 +532,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
             }
         }
 
-        llama_batch_free(batch);
+        llama_batch_ext_free(batch);
 
         const auto t_end = std::chrono::high_resolution_clock::now();
 

examples/infill/infill.cpp

@@ -353,7 +353,8 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                auto batch = llama_batch_ext_ptr::init_from_text(&embd[i], n_eval, n_past, 0, true);
+                if (llama_decode_ext(ctx, batch.get())) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }

examples/json_schema_to_grammar.py

@@ -10,9 +10,6 @@ from typing import Any, List, Optional, Set, Tuple, Union
 
 def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
 
-    if max_items == 0:
-        return ""
-
     if min_items == 0 and max_items == 1:
         return f'{item_rule}?'
 

examples/llama-bench/README.md

@@ -28,7 +28,6 @@ options:
   -p, --n-prompt <n>                        (default: 512)
   -n, --n-gen <n>                           (default: 128)
   -pg <pp,tg>                               (default: )
-  -d, --n-depth <n>                         (default: 0)
   -b, --batch-size <n>                      (default: 2048)
   -ub, --ubatch-size <n>                    (default: 512)
   -ctk, --cache-type-k <t>                  (default: f16)
@@ -67,8 +66,6 @@ With the exception of `-r`, `-o` and `-v`, all options can be specified multiple
 
 Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
 
-Using the `-d <n>` option, each test can be run at a specified context depth, prefilling the KV cache with `<n>` tokens.
-
 For a description of the other options, see the [main example](../main/README.md).
 
 Note:
@@ -151,19 +148,6 @@ $ ./llama-bench -ngl 10,20,30,31,32,33,34,35
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
 
-### Different prefilled context
-
-```
-$ ./llama-bench -d 0,512
-```
-
-| model                          |       size |     params | backend    | ngl |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
-| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           pp512 |      7340.20 ± 23.45 |
-| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           tg128 |        120.60 ± 0.59 |
-| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    pp512 @ d512 |      6425.91 ± 18.88 |
-| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    tg128 @ d512 |        116.71 ± 0.60 |
-
 ## Output formats
 
 By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
@@ -186,9 +170,9 @@ $ ./llama-bench -o csv
 ```
 
 ```csv
-build_commit,build_number,cpu_info,gpu_info,backends,model_filename,model_type,model_size,model_n_params,n_batch,n_ubatch,n_threads,cpu_mask,cpu_strict,poll,type_k,type_v,n_gpu_layers,split_mode,main_gpu,no_kv_offload,flash_attn,tensor_split,use_mmap,embeddings,n_prompt,n_gen,n_depth,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
-"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","512","0","0","2025-04-24T11:57:09Z","70285660","982040","7285.676949","100.064434"
-"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","0","128","0","2025-04-24T11:57:10Z","1067431600","3834831","119.915244","0.430617"
+build_commit,build_number,cuda,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
 ```
 
 ### JSON
@@ -200,78 +184,64 @@ $ ./llama-bench -o json
 ```json
 [
   {
-    "build_commit": "8cf427ff",
-    "build_number": 5163,
-    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
-    "gpu_info": "NVIDIA GeForce RTX 4080",
-    "backends": "CUDA",
-    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
-    "model_type": "qwen2 7B Q4_K - Medium",
-    "model_size": 4677120000,
-    "model_n_params": 7615616512,
-    "n_batch": 2048,
-    "n_ubatch": 512,
-    "n_threads": 8,
-    "cpu_mask": "0x0",
-    "cpu_strict": false,
-    "poll": 50,
-    "type_k": "f16",
-    "type_v": "f16",
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
     "n_gpu_layers": 99,
-    "split_mode": "layer",
     "main_gpu": 0,
-    "no_kv_offload": false,
-    "flash_attn": false,
+    "mul_mat_q": true,
     "tensor_split": "0.00",
-    "use_mmap": true,
-    "embeddings": false,
     "n_prompt": 512,
     "n_gen": 0,
-    "n_depth": 0,
-    "test_time": "2025-04-24T11:58:50Z",
-    "avg_ns": 72135640,
-    "stddev_ns": 1453752,
-    "avg_ts": 7100.002165,
-    "stddev_ts": 140.341520,
-    "samples_ns": [ 74601900, 71632900, 71745200, 71952700, 70745500 ],
-    "samples_ts": [ 6863.1, 7147.55, 7136.37, 7115.79, 7237.21 ]
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
   },
   {
-    "build_commit": "8cf427ff",
-    "build_number": 5163,
-    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
-    "gpu_info": "NVIDIA GeForce RTX 4080",
-    "backends": "CUDA",
-    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
-    "model_type": "qwen2 7B Q4_K - Medium",
-    "model_size": 4677120000,
-    "model_n_params": 7615616512,
-    "n_batch": 2048,
-    "n_ubatch": 512,
-    "n_threads": 8,
-    "cpu_mask": "0x0",
-    "cpu_strict": false,
-    "poll": 50,
-    "type_k": "f16",
-    "type_v": "f16",
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
     "n_gpu_layers": 99,
-    "split_mode": "layer",
     "main_gpu": 0,
-    "no_kv_offload": false,
-    "flash_attn": false,
+    "mul_mat_q": true,
     "tensor_split": "0.00",
-    "use_mmap": true,
-    "embeddings": false,
     "n_prompt": 0,
     "n_gen": 128,
-    "n_depth": 0,
-    "test_time": "2025-04-24T11:58:51Z",
-    "avg_ns": 1076767880,
-    "stddev_ns": 9449585,
-    "avg_ts": 118.881588,
-    "stddev_ts": 1.041811,
-    "samples_ns": [ 1075361300, 1065089400, 1071761200, 1081934900, 1089692600 ],
-    "samples_ts": [ 119.03, 120.178, 119.43, 118.307, 117.464 ]
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
   }
 ]
 ```
@@ -284,8 +254,8 @@ $ ./llama-bench -o jsonl
 ```
 
 ```json lines
-{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 512, "n_gen": 0, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 70497220, "stddev_ns": 883196, "avg_ts": 7263.609157, "stddev_ts": 90.940578, "samples_ns": [ 71551000, 71222800, 70364100, 69439100, 69909100 ],"samples_ts": [ 7155.74, 7188.71, 7276.44, 7373.37, 7323.8 ]}
-{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 0, "n_gen": 128, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 1068078400, "stddev_ns": 6279455, "avg_ts": 119.844681, "stddev_ts": 0.699739, "samples_ns": [ 1066331700, 1064864900, 1079042600, 1063328400, 1066824400 ],"samples_ts": [ 120.038, 120.203, 118.624, 120.377, 119.982 ]}
+{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":512,"n_gen":0,"test_time":"2023-09-23T12:09:57Z","avg_ns":212365953,"stddev_ns":985423,"avg_ts":2410.974041,"stddev_ts":11.163766,"samples_ns":[213837238,211635853,212328053,211329715,212698907],"samples_ts":[2394.34,2419.25,2411.36,2422.75,2407.16]}
+{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":0,"n_gen":128,"test_time":"2023-09-23T12:09:59Z","avg_ns":977425219,"stddev_ns":9268593,"avg_ts":130.965708,"stddev_ts":1.238924,"samples_ns":[984472709,974901233,989474741,970729355,967548060],"samples_ts":[130.019,131.295,129.362,131.86,132.293]}
 ```
 
 
@@ -301,32 +271,25 @@ $ ./llama-bench -o sql
 CREATE TABLE IF NOT EXISTS test (
   build_commit TEXT,
   build_number INTEGER,
+  cuda INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
   cpu_info TEXT,
   gpu_info TEXT,
-  backends TEXT,
   model_filename TEXT,
   model_type TEXT,
   model_size INTEGER,
   model_n_params INTEGER,
   n_batch INTEGER,
-  n_ubatch INTEGER,
   n_threads INTEGER,
-  cpu_mask TEXT,
-  cpu_strict INTEGER,
-  poll INTEGER,
-  type_k TEXT,
-  type_v TEXT,
+  f16_kv INTEGER,
   n_gpu_layers INTEGER,
-  split_mode TEXT,
   main_gpu INTEGER,
-  no_kv_offload INTEGER,
-  flash_attn INTEGER,
+  mul_mat_q INTEGER,
   tensor_split TEXT,
-  use_mmap INTEGER,
-  embeddings INTEGER,
   n_prompt INTEGER,
   n_gen INTEGER,
-  n_depth INTEGER,
   test_time TEXT,
   avg_ns INTEGER,
   stddev_ns INTEGER,
@@ -334,6 +297,6 @@ CREATE TABLE IF NOT EXISTS test (
   stddev_ts REAL
 );
 
-INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '512', '0', '0', '2025-04-24T12:00:08Z', '69905000', '519516', '7324.546977', '54.032613');
-INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '0', '128', '0', '2025-04-24T12:00:09Z', '1063608780', '4464130', '120.346696', '0.504647');
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
 ```

examples/llama-bench/llama-bench.cpp

@@ -36,46 +36,6 @@ static uint64_t get_time_ns() {
     return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
 }
 
-static bool tensor_buft_override_equal(const llama_model_tensor_buft_override& a, const llama_model_tensor_buft_override& b) {
-    if (a.pattern != b.pattern) {
-        // cString comparison that may be null
-        if (a.pattern == nullptr || b.pattern == nullptr) {
-            return false;
-        }
-        if (strcmp(a.pattern, b.pattern) != 0) {
-            return false;
-        }
-    }
-    if (a.buft != b.buft) {
-        return false;
-    }
-    return true;
-}
-
-static bool vec_tensor_buft_override_equal(const std::vector<llama_model_tensor_buft_override>& a, const std::vector<llama_model_tensor_buft_override>& b) {
-    if (a.size() != b.size()) {
-        return false;
-    }
-    for (size_t i = 0; i < a.size(); i++) {
-        if (!tensor_buft_override_equal(a[i], b[i])) {
-            return false;
-        }
-    }
-    return true;
-}
-
-static bool vec_vec_tensor_buft_override_equal(const std::vector<std::vector<llama_model_tensor_buft_override>>& a, const std::vector<std::vector<llama_model_tensor_buft_override>>& b) {
-    if (a.size() != b.size()) {
-        return false;
-    }
-    for (size_t i = 0; i < a.size(); i++) {
-        if (!vec_tensor_buft_override_equal(a[i], b[i])) {
-            return false;
-        }
-    }
-    return true;
-}
-
 template <class T> static std::string join(const std::vector<T> & values, const std::string & delim) {
     std::ostringstream str;
     for (size_t i = 0; i < values.size(); i++) {
@@ -200,7 +160,6 @@ struct cmd_params {
     std::vector<int>                 n_prompt;
     std::vector<int>                 n_gen;
     std::vector<std::pair<int, int>> n_pg;
-    std::vector<int>                 n_depth;
     std::vector<int>                 n_batch;
     std::vector<int>                 n_ubatch;
     std::vector<ggml_type>           type_k;
@@ -216,7 +175,6 @@ struct cmd_params {
     std::vector<bool>                no_kv_offload;
     std::vector<bool>                flash_attn;
     std::vector<std::vector<float>>  tensor_split;
-    std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
     ggml_numa_strategy               numa;
@@ -234,7 +192,6 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt             */ { 512 },
     /* n_gen                */ { 128 },
     /* n_pg                 */ {},
-    /* n_depth              */ { 0 },
     /* n_batch              */ { 2048 },
     /* n_ubatch             */ { 512 },
     /* type_k               */ { GGML_TYPE_F16 },
@@ -250,7 +207,6 @@ static const cmd_params cmd_params_defaults = {
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { false },
     /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
-    /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
@@ -274,7 +230,6 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -n, --n-gen <n>                           (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -pg <pp,tg>                               (default: %s)\n",
            join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
-    printf("  -d, --n-depth <n>                         (default: %s)\n", join(cmd_params_defaults.n_depth, ",").c_str());
     printf("  -b, --batch-size <n>                      (default: %s)\n",
            join(cmd_params_defaults.n_batch, ",").c_str());
     printf("  -ub, --ubatch-size <n>                    (default: %s)\n",
@@ -310,7 +265,6 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -embd, --embeddings <0|1>                 (default: %s)\n",
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
-    printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
@@ -412,13 +366,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) });
-        } else if (arg == "-d" || arg == "--n-depth") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_depth.insert(params.n_depth.end(), p.begin(), p.end());
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -610,87 +557,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 params.tensor_split.push_back(tensor_split);
             }
-        } else if (arg == "-ot" || arg == "--override-tensor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto value = argv[i];
-            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
-            if (buft_list.empty()) {
-                // enumerate all the devices and add their buffer types to the list
-                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                    auto * dev = ggml_backend_dev_get(i);
-                    auto * buft = ggml_backend_dev_buffer_type(dev);
-                    if (buft) {
-                        buft_list[ggml_backend_buft_name(buft)] = buft;
-                    }
-                }
-            }
-            auto override_group_span_len = std::strcspn(value, ",");
-            bool last_group = false;
-            do {
-                if (override_group_span_len == 0) {
-                    // Adds an empty override-tensors for an empty span
-                    params.tensor_buft_overrides.push_back({{}});
-                    if (value[override_group_span_len] == '\0') {
-                        value = &value[override_group_span_len];
-                        last_group = true;
-                    } else {
-                        value = &value[override_group_span_len + 1];
-                        override_group_span_len = std::strcspn(value, ",");
-                    }
-                    continue;
-                }
-                // Stamps null terminators into the argv
-                // value for this option to avoid the
-                // memory leak present in the implementation
-                // over in arg.cpp. Acceptable because we
-                // only parse these args once in this program.
-                auto override_group = value;
-                if (value[override_group_span_len] == '\0') {
-                    value = &value[override_group_span_len];
-                    last_group = true;
-                } else {
-                    value[override_group_span_len] = '\0';
-                    value = &value[override_group_span_len + 1];
-                }
-                std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{};
-                auto override_span_len = std::strcspn(override_group, ";");
-                while (override_span_len > 0) {
-                    auto override = override_group;
-                    if (override_group[override_span_len] != '\0') {
-                        override_group[override_span_len] = '\0';
-                        override_group = &override_group[override_span_len + 1];
-                    } else {
-                        override_group = &override_group[override_span_len];
-                    }
-                    auto tensor_name_span_len = std::strcspn(override, "=");
-                    if (tensor_name_span_len >= override_span_len) {
-                        invalid_param = true;
-                        break;
-                    }
-                    override[tensor_name_span_len] = '\0';
-                    auto tensor_name = override;
-                    auto buffer_type = &override[tensor_name_span_len + 1];
-                    if (buft_list.find(buffer_type) == buft_list.end()) {
-                        printf("Available buffer types:\n");
-                        for (const auto & it : buft_list) {
-                            printf("  %s\n", ggml_backend_buft_name(it.second));
-                        }
-                        invalid_param = true;
-                        break;
-                    }
-                    group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)});
-                    override_span_len = std::strcspn(override_group, ";");
-                }
-                if (invalid_param) {
-                    break;
-                }
-                group_tensor_buft_overrides.push_back({nullptr,nullptr});
-                params.tensor_buft_overrides.push_back(group_tensor_buft_overrides);
-                override_group_span_len = std::strcspn(value, ",");
-            } while (!last_group);
         } else if (arg == "-r" || arg == "--repetitions") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -749,9 +615,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_pg.empty()) {
         params.n_pg = cmd_params_defaults.n_pg;
     }
-    if (params.n_depth.empty()) {
-        params.n_depth = cmd_params_defaults.n_depth;
-    }
     if (params.n_batch.empty()) {
         params.n_batch = cmd_params_defaults.n_batch;
     }
@@ -785,9 +648,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.tensor_split.empty()) {
         params.tensor_split = cmd_params_defaults.tensor_split;
     }
-    if (params.tensor_buft_overrides.empty()) {
-        params.tensor_buft_overrides = cmd_params_defaults.tensor_buft_overrides;
-    }
     if (params.use_mmap.empty()) {
         params.use_mmap = cmd_params_defaults.use_mmap;
     }
@@ -814,7 +674,6 @@ struct cmd_params_instance {
     std::string        model;
     int                n_prompt;
     int                n_gen;
-    int                n_depth;
     int                n_batch;
     int                n_ubatch;
     ggml_type          type_k;
@@ -830,7 +689,6 @@ struct cmd_params_instance {
     bool               no_kv_offload;
     bool               flash_attn;
     std::vector<float> tensor_split;
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool               use_mmap;
     bool               embeddings;
 
@@ -875,26 +733,19 @@ struct cmd_params_instance {
         mparams.tensor_split = tensor_split.data();
         mparams.use_mmap     = use_mmap;
 
-        if (tensor_buft_overrides.empty()) {
-            mparams.tensor_buft_overrides = nullptr;
-        } else {
-            GGML_ASSERT(tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
-            mparams.tensor_buft_overrides = tensor_buft_overrides.data();
-        }
-
         return mparams;
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str &&
                split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
-               tensor_split == other.tensor_split && vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides);
+               tensor_split == other.tensor_split;
     }
 
     llama_context_params to_llama_cparams() const {
         llama_context_params cparams = llama_context_default_params();
 
-        cparams.n_ctx       = n_prompt + n_gen + n_depth;
+        cparams.n_ctx       = n_prompt + n_gen;
         cparams.n_batch     = n_batch;
         cparams.n_ubatch    = n_ubatch;
         cparams.type_k      = type_k;
@@ -918,7 +769,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
-    for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
@@ -930,7 +780,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & nt : params.n_threads)
     for (const auto & cm : params.cpu_mask)
     for (const auto & cs : params.cpu_strict)
-    for (const auto & nd : params.n_depth)
     for (const auto & pl : params.poll) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
@@ -940,7 +789,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
-                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -956,7 +804,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
-                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -971,7 +818,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
-                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -987,7 +833,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
-                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -1002,7 +847,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ n_pg.first,
                 /* .n_gen        = */ n_pg.second,
-                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -1018,7 +862,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
-                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -1053,12 +896,10 @@ struct test {
     bool                     no_kv_offload;
     bool                     flash_attn;
     std::vector<float>       tensor_split;
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool                     use_mmap;
     bool                     embeddings;
     int                      n_prompt;
     int                      n_gen;
-    int                      n_depth;
     std::string              test_time;
     std::vector<uint64_t>    samples_ns;
 
@@ -1086,12 +927,10 @@ struct test {
         no_kv_offload  = inst.no_kv_offload;
         flash_attn     = inst.flash_attn;
         tensor_split   = inst.tensor_split;
-        tensor_buft_overrides = inst.tensor_buft_overrides;
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
         n_prompt       = inst.n_prompt;
         n_gen          = inst.n_gen;
-        n_depth        = inst.n_depth;
         // RFC 3339 date-time format
         time_t t       = time(NULL);
         std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
@@ -1134,10 +973,8 @@ struct test {
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "use_mmap",
-            "embeddings",   "n_prompt",     "n_gen",          "n_depth",    "test_time",    "avg_ns",
-            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
-            "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
-            "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
+            "embeddings",   "n_prompt",     "n_gen",          "test_time",  "avg_ns",       "stddev_ns",
+            "avg_ts",       "stddev_ts",
         };
         return fields;
     }
@@ -1147,8 +984,8 @@ struct test {
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" ||
             field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
-            field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" ||
-            field == "avg_ns" || field == "stddev_ns") {
+            field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" ||
+            field == "stddev_ns") {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
@@ -1163,7 +1000,6 @@ struct test {
 
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
-        std::string tensor_buft_overrides_str;
         int         max_nonzero = 0;
         for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
@@ -1178,26 +1014,6 @@ struct test {
                 tensor_split_str += "/";
             }
         }
-        if (tensor_buft_overrides.size() == 1) {
-            // Last element of tensor_buft_overrides is always a null pattern
-            // so if it is only one element long, it must be a null pattern.
-            GGML_ASSERT(tensor_buft_overrides[0].pattern == nullptr);
-            tensor_buft_overrides_str += "none";
-        } else {
-            for (size_t i = 0; i < tensor_buft_overrides.size()-1; i++) {
-                // Last element of tensor_buft_overrides is always a null pattern
-                if (tensor_buft_overrides[i].pattern == nullptr) {
-                    tensor_buft_overrides_str += "none";
-                } else {
-                    tensor_buft_overrides_str += tensor_buft_overrides[i].pattern;
-                    tensor_buft_overrides_str += "=";
-                    tensor_buft_overrides_str += ggml_backend_buft_name(tensor_buft_overrides[i].buft);
-                }
-                if (i + 2 < tensor_buft_overrides.size()) {
-                    tensor_buft_overrides_str += ";";
-                }
-            }
-        }
         std::vector<std::string> values = { build_commit,
                                             std::to_string(build_number),
                                             cpu_info,
@@ -1221,12 +1037,10 @@ struct test {
                                             std::to_string(no_kv_offload),
                                             std::to_string(flash_attn),
                                             tensor_split_str,
-                                            tensor_buft_overrides_str,
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(n_prompt),
                                             std::to_string(n_gen),
-                                            std::to_string(n_depth),
                                             test_time,
                                             std::to_string(avg_ns()),
                                             std::to_string(stdev_ns()),
@@ -1404,7 +1218,7 @@ struct markdown_printer : public printer {
             return 4;
         }
         if (field == "test") {
-            return 15;
+            return 13;
         }
 
         int width = std::max((int) field.length(), 10);
@@ -1440,9 +1254,6 @@ struct markdown_printer : public printer {
         if (field == "tensor_split") {
             return "ts";
         }
-        if (field == "tensor_buft_overrides") {
-            return "ot";
-        }
         return field;
     }
 
@@ -1496,9 +1307,6 @@ struct markdown_printer : public printer {
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.emplace_back("tensor_split");
         }
-        if (params.tensor_buft_overrides.size() > 1 || !vec_vec_tensor_buft_override_equal(params.tensor_buft_overrides, cmd_params_defaults.tensor_buft_overrides)) {
-            fields.emplace_back("tensor_buft_overrides");
-        }
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
@@ -1554,10 +1362,6 @@ struct markdown_printer : public printer {
                 } else {
                     snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
                 }
-                if (t.n_depth > 0) {
-                    int len = strlen(buf);
-                    snprintf(buf + len, sizeof(buf) - len, " @ d%d", t.n_depth);
-                }
                 value = buf;
             } else if (field == "t/s") {
                 snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
@@ -1623,7 +1427,7 @@ struct sql_printer : public printer {
     }
 };
 
-static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
+static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model   = llama_get_model(ctx);
@@ -1640,14 +1444,15 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
-        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
+        auto batch = llama_batch_ext_ptr::init_from_text(tokens.data(), n_tokens, n_past + n_processed, 0, true);
+        llama_decode_ext(ctx, batch.get());
         n_processed += n_tokens;
     }
 
     llama_synchronize(ctx);
 }
 
-static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
+static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model   = llama_get_model(ctx);
@@ -1657,7 +1462,8 @@ static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
     llama_token token = llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
-        llama_decode(ctx, llama_batch_get_one(&token, 1));
+        auto batch = llama_batch_ext_ptr::init_from_text(&token, 1, n_past + i, 0, true);
+        llama_decode_ext(ctx, batch.get());
         llama_synchronize(ctx);
         token = std::rand() % n_vocab;
     }
@@ -1804,26 +1610,18 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
             }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             if (params.progress) {
                 fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
             }
-            test_gen(ctx, 1, t.n_threads);
+            test_gen(ctx, 1, 0, t.n_threads);
         }
 
         for (int i = 0; i < params.reps; i++) {
             llama_kv_self_clear(ctx);
 
-            if (t.n_depth > 0) {
-                if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
-                            i + 1, params.reps);
-                }
-                test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
-            }
-
             uint64_t t_start = get_time_ns();
 
             if (t.n_prompt > 0) {
@@ -1831,14 +1629,14 @@ int main(int argc, char ** argv) {
                     fprintf(stderr, "llama-bench: benchmark %d/%zu: prompt run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
-                test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
+                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
                     fprintf(stderr, "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
-                test_gen(ctx, t.n_gen, t.n_threads);
+                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
 
             uint64_t t_ns = get_time_ns() - t_start;

examples/llama.android/llama/build.gradle.kts

@@ -18,7 +18,6 @@ android {
         }
         externalNativeBuild {
             cmake {
-                arguments += "-DLLAMA_CURL=OFF"
                 arguments += "-DLLAMA_BUILD_COMMON=ON"
                 arguments += "-DGGML_LLAMAFILE=OFF"
                 arguments += "-DCMAKE_BUILD_TYPE=Release"

examples/llava/CMakeLists.txt

@@ -1,5 +1,3 @@
-# llava (legacy)
-
 add_library(llava OBJECT
             llava.cpp
             llava.h
@@ -24,46 +22,27 @@ if (BUILD_SHARED_LIBS)
     install(TARGETS llava_shared LIBRARY)
 endif()
 
-# mtmd
-
-add_library(mtmd OBJECT
-            mtmd.cpp
-            mtmd.h
-            clip.cpp
-            clip.h
-            clip-impl.h
-            )
-
-target_link_libraries(mtmd PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
-
-target_include_directories(mtmd PUBLIC .)
-target_include_directories(mtmd PRIVATE ../..)
-target_include_directories(mtmd PRIVATE ../../common) # for stb_image.h
-
-target_compile_features(mtmd PRIVATE cxx_std_17)
-
-add_library(mtmd_static STATIC $<TARGET_OBJECTS:mtmd>)
-if (BUILD_SHARED_LIBS)
-    set_target_properties(mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(mtmd PRIVATE LLAMA_SHARED LLAMA_BUILD)
-    add_library(mtmd_shared SHARED $<TARGET_OBJECTS:mtmd>)
-    target_link_libraries(mtmd_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
-    install(TARGETS mtmd_shared LIBRARY)
-endif()
-
 if (NOT MSVC)
     target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
-    target_compile_options(mtmd PRIVATE -Wno-cast-qual) # stb_image.h
 endif()
 
 if(TARGET BUILD_INFO)
     add_dependencies(llava BUILD_INFO)
-    add_dependencies(mtmd BUILD_INFO)
 endif()
 
-add_executable(llama-llava-cli    deprecation-warning.cpp)
-add_executable(llama-gemma3-cli   deprecation-warning.cpp)
-add_executable(llama-minicpmv-cli deprecation-warning.cpp)
+set(TARGET llama-llava-cli)
+add_executable(${TARGET} llava-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+set(TARGET llama-minicpmv-cli)
+add_executable(${TARGET} minicpmv-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-qwen2vl-cli)
 add_executable(${TARGET} qwen2vl-cli.cpp)
@@ -72,11 +51,11 @@ install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
-set(TARGET llama-mtmd-cli)
-add_executable(${TARGET} mtmd-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
+set(TARGET llama-gemma3-cli)
+add_executable(${TARGET} gemma3-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-llava-clip-quantize-cli)

examples/llava/MobileVLM-README.md

@@ -9,15 +9,15 @@ The implementation is based on llava, and is compatible with llava and mobileVLM
 Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
 
 ## Usage
+Build with cmake or run `make llama-llava-cli` to build it.
 
-Build the `llama-mtmd-cli` binary.
-
-After building, run: `./llama-mtmd-cli` to see the usage. For example:
+After building, run: `./llama-llava-cli` to see the usage. For example:
 
 ```sh
-./llama-mtmd-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
+./llama-llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
     --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \
-    --chat-template deepseek
+    --image path/to/an/image.jpg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:"
 ```
 
 ## Model conversion
@@ -82,7 +82,7 @@ refer to `android/adb_run.sh`, modify resources' `name` and `path`
 ### case 1
 **input**
 ```sh
-/data/local/tmp/llama-mtmd-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -102,7 +102,7 @@ llama_print_timings:       total time =   34731.93 ms
 ### case 2
 **input**
 ```sh
-/data/local/tmp/llama-mtmd-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -123,10 +123,10 @@ llama_print_timings:       total time =   34570.79 ms
 
 ## Some result on Android with `Snapdragon 778G` chip
 ### MobileVLM-1.7B case
-#### mtmd-cli release-b2005
+#### llava-cli release-b2005
 **input**
 ```sh
-/data/local/tmp/llama-mtmd-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -147,7 +147,7 @@ llama_print_timings: prompt eval time =    8119.49 ms /   191 tokens (   42.51 m
 llama_print_timings:        eval time =    1005.75 ms /    14 runs   (   71.84 ms per token,    13.92 tokens per second)
 llama_print_timings:       total time =   28038.34 ms /   205 tokens
 ```
-#### mtmd-cli latest-version
+#### llava-cli latest-version
 **input**
 
 Just the same as above.
@@ -169,7 +169,7 @@ llama_print_timings:        eval time =   43894.02 ms /    13 runs   ( 3376.46 m
 llama_print_timings:       total time =  865441.76 ms /   204 tokens
 ```
 ### MobileVLM_V2-1.7B case
-#### mtmd-cli release-2005b
+#### llava-cli release-2005b
 **input**
 
 Just the same as above.
@@ -200,7 +200,7 @@ make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
 ### case 1
 **input**
 ```sh
-./llama-mtmd-cli \
+./llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     --image /data/local/tmp/demo.jpeg \
@@ -224,7 +224,7 @@ llama_print_timings:       total time =    1352.63 ms /   252 tokens
 ### case 2
 **input**
 ```sh
-./llama-mtmd-cli \
+./llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:" \

examples/llava/README-gemma3.md

@@ -0,0 +1,30 @@
+# Gemma 3 vision
+
+> [!IMPORTANT]
+>
+> This is very experimental, only used for demo purpose.
+
+## How to get mmproj.gguf?
+
+```bash
+cd gemma-3-4b-it
+python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py .
+
+# output file is mmproj.gguf
+```
+
+## How to run it?
+
+What you need:
+- The text model GGUF, can be converted using `convert_hf_to_gguf.py`
+- The mmproj file from step above
+- An image file
+
+```bash
+# build
+cmake -B build
+cmake --build build --target llama-gemma3-cli
+
+# run it
+./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+```

examples/llava/README-glmedge.md

@@ -3,12 +3,12 @@
 Currently this implementation supports [glm-edge-v-2b](https://huggingface.co/THUDM/glm-edge-v-2b) and [glm-edge-v-5b](https://huggingface.co/THUDM/glm-edge-v-5b).
 
 ## Usage
-Build the `llama-mtmd-cli` binary.
+Build with cmake or run `make llama-llava-cli` to build it.
 
-After building, run: `./llama-mtmd-cli` to see the usage. For example:
+After building, run: `./llama-llava-cli` to see the usage. For example:
 
 ```sh
-./llama-mtmd-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf
+./llama-llava-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf --image img_path/image.jpg -p "<|system|>\n system prompt <image><|user|>\n prompt <|assistant|>\n"
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.

examples/llava/README-granitevision.md

@@ -176,11 +176,15 @@ Note that currently you cannot quantize the visual encoder because granite visio
 
 
 ### 5. Running the Model in Llama cpp
-Build llama cpp normally; you should have a target binary named `llama-mtmd-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
+Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
 
 ```bash
-$ ./build/bin/llama-mtmd-cli -m $LLM_GGUF_PATH \
+$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
     --mmproj $VISUAL_GGUF_PATH \
+    --image ./media/llama0-banner.png \
     -c 16384 \
+    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat does the text in this image say?\n<|assistant|>\n" \
     --temp 0
 ```
+
+Sample output: `The text in the image reads "LLAMA C++ Can it run DOOM Llama?"`

examples/llava/README-minicpmo2.6.md

@@ -40,9 +40,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 Inference on Linux or Mac
 ```bash
-# run in single-turn mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run f16 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run in conversation mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf
+# run quantized int4 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/README-minicpmv2.5.md

@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 Inference on Linux or Mac
 ```bash
-# run in single-turn mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run f16 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run in conversation mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf
+# run quantized int4 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/README-minicpmv2.6.md

@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 Inference on Linux or Mac
 ```bash
-# run in single-turn mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run f16 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run in conversation mode
-./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf
+# run quantized int4 version
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/README.md

@@ -1,75 +1,158 @@
-# Multimodal Support in llama.cpp
+# LLaVA
 
-This directory provides multimodal capabilities for `llama.cpp`. Initially intended as a showcase for running LLaVA models, its scope has expanded significantly over time to include various other vision-capable models. As a result, LLaVA is no longer the only multimodal architecture supported.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
 
-> [!IMPORTANT]
->
-> Multimodal support can be viewed as a sub-project within `llama.cpp`. It is under **very heavy development**, and **breaking changes are expected**.
+The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
+and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
+models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
-The naming and structure related to multimodal support have evolved, which might cause some confusion. Here's a brief timeline to clarify:
+After API is confirmed, more models will be supported / uploaded.
 
-- [#3436](https://github.com/ggml-org/llama.cpp/pull/3436): Initial support for LLaVA 1.5 was added, introducing `llava.cpp` and `clip.cpp`. The `llava-cli` binary was created for model interaction.
-- [#4954](https://github.com/ggml-org/llama.cpp/pull/4954): Support for MobileVLM was added, becoming the second vision model supported. This built upon the existing `llava.cpp`, `clip.cpp`, and `llava-cli` infrastructure.
-- **Expansion & Fragmentation:** Many new models were subsequently added (e.g., [#7599](https://github.com/ggml-org/llama.cpp/pull/7599), [#10361](https://github.com/ggml-org/llama.cpp/pull/10361), [#12344](https://github.com/ggml-org/llama.cpp/pull/12344), and others). However, `llava-cli` lacked support for the increasingly complex chat templates required by these models. This led to the creation of model-specific binaries like `qwen2vl-cli`, `minicpmv-cli`, and `gemma3-cli`. While functional, this proliferation of command-line tools became confusing for users.
-- [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
-- [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
+## Usage
+Build with cmake or run `make llama-llava-cli` to build it.
 
-## Pre-quantized models
-
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
+After building, run: `./llama-llava-cli` to see the usage. For example:
 
 ```sh
-# Gemma 3
-llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
-
-# SmolVLM
-llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
-
-# Pixtral 12B
-llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+./llama-llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
 ```
 
-## How it works and what is `mmproj`?
+**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
+## LLaVA 1.5
 
-This approach keeps the multimodal components distinct from the core `libllama` library. Separating these allows for faster, independent development cycles. While many modern vision models are based on Vision Transformers (ViTs), their specific pre-processing and projection steps can vary significantly. Integrating this diverse complexity directly into `libllama` is currently challenging.
+1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
-Consequently, running a multimodal model typically requires two GGUF files:
-1.  The standard language model file.
-2.  A corresponding **multimodal projector (`mmproj`)** file, which handles the image encoding and projection.
+```sh
+git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
 
-## What is `libmtmd`?
+git clone https://huggingface.co/openai/clip-vit-large-patch14-336
+```
 
-As outlined in the history, `libmtmd` is the modern library designed to replace the original `llava.cpp` implementation for handling multimodal inputs.
+2. Install the required Python packages:
 
-Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advantages:
-- **Unified Interface:** Aims to consolidate interaction for various multimodal models.
-- **Improved UX/DX:** Features a more intuitive API, inspired by the `Processor` class in the Hugging Face `transformers` library.
-- **Flexibility:** Designed to support multiple input types (text, audio, images) while respecting the wide variety of chat templates used by different models.
+```sh
+pip install -r examples/llava/requirements.txt
+```
 
-## How to obtain `mmproj`
+3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
-Multimodal projector (`mmproj`) files are specific to each model architecture. Please refer to the relevant guide for instructions on how to obtain or create them:
+```sh
+python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
+```
 
-- [LLaVA](../../docs/multimodal/llava.md)
-- [MobileVLM](../../docs/multimodal/MobileVLM.md)
-- [GLM-Edge](../../docs/multimodal/glmedge.md)
-- [MiniCPM-V 2.5](../../docs/multimodal/minicpmv2.5.md)
-- [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
-- [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
-- [IBM Granite Vision](../../docs/multimodal/granitevision.md)
-- [Google Gemma 3](../../docs/multimodal/gemma3.md)
+4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
 
-For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
-- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
-- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
-- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
-- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
+```sh
+python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+```
+
+5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
+
+```sh
+python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown
+```
+
+Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
+
+## LLaVA 1.6 gguf conversion
+1) First clone a LLaVA 1.6 model:
+```console
+git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
+```
+
+2) Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+```console
+python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
+```
+- you will find a llava.projector and a llava.clip file in your model directory
+
+4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+```console
+mkdir vit
+cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
+cp ../llava-v1.6-vicuna-7b/llava.projector vit/
+curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
+```
+
+5) Create the visual gguf model:
+```console
+python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+```
+- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
+
+6) Then convert the model to gguf format:
+```console
+python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
+```
+
+7) And finally we can run the llava cli using the 1.6 model version:
+```console
+./llama-llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
+```
+
+**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
+
+**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
+
+**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model.
+
+```python
+import os
+import transformers
+
+model_path = ...
+llm_export_path = ...
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
+model = transformers.AutoModelForImageTextToText.from_pretrained(model_path)
+
+tokenizer.save_pretrained(llm_export_path)
+model.language_model.save_pretrained(llm_export_path)
+```
+
+Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.
+
+## llava-cli templating and llava-1.6 prompting
+
+llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
+For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
+
+**For Mistral and using llava-cli binary:**
+Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
+The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
+
+**For the 34B this should work:**
+Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
+
+
+
+
+## TODO
+
+- [x] Support non-CPU backend for the image encoding part.
+- [ ] Support different sampling methods.
+- [ ] Support more model variants.

examples/llava/android/adb_run.sh

@@ -10,7 +10,7 @@ prompt="A chat between a curious user and an artificial intelligence assistant.
 # prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
 
 program_dir="build_64/bin"
-binName="llama-mtmd-cli"
+binName="llama-llava-cli"
 n_threads=4
 
 

examples/llava/clip-impl.h

@@ -1,347 +0,0 @@
-#include "ggml.h"
-#include "gguf.h"
-#include "clip.h"
-
-#include "clip.h"
-
-#include <climits>
-#include <cstdarg>
-#include <string>
-#include <map>
-#include <sstream>
-#include <vector>
-#include <memory>
-
-// Internal header for clip.cpp
-
-#define KEY_FTYPE               "general.file_type"
-#define KEY_NAME                "general.name"
-#define KEY_DESCRIPTION         "general.description"
-#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
-#define KEY_USE_GELU            "clip.use_gelu"
-#define KEY_USE_SILU            "clip.use_silu"
-#define KEY_N_EMBD              "clip.vision.embedding_length"
-#define KEY_N_FF                "clip.vision.feed_forward_length"
-#define KEY_N_BLOCK             "clip.vision.block_count"
-#define KEY_N_HEAD              "clip.vision.attention.head_count"
-#define KEY_LAYER_NORM_EPS      "clip.vision.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM            "clip.vision.projection_dim"
-#define KEY_IMAGE_SIZE          "clip.vision.image_size"
-#define KEY_PATCH_SIZE          "clip.vision.patch_size"
-#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
-#define KEY_IMAGE_STD           "clip.vision.image_std"
-#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
-#define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
-#define KEY_PROJ_TYPE           "clip.projector_type"
-
-#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
-#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
-
-#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
-#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
-#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
-#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
-#define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
-
-
-//
-// tensor name constants
-//
-
-#define TN_POS_EMBD        "%s.position_embd.weight"
-#define TN_CLASS_EMBD      "v.class_embd"
-#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
-#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
-#define TN_PATCH_BIAS      "v.patch_embd.bias"
-#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
-#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
-#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
-#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
-#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
-#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
-#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
-#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
-#define TN_LN_1            "%s.blk.%d.ln1.%s"
-#define TN_LN_2            "%s.blk.%d.ln2.%s"
-#define TN_LN_PRE          "%s.pre_ln.%s"
-#define TN_LN_POST         "%s.post_ln.%s"
-#define TN_LLAVA_PROJ      "mm.%d.%s"
-#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
-#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
-#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
-#define TN_IMAGE_NEWLINE   "model.image_newline"
-#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
-#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
-#define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
-#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
-
-// mimicpmv
-#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
-#define TN_MINICPMV_QUERY      "resampler.query"
-#define TN_MINICPMV_PROJ       "resampler.proj.weight"
-#define TN_MINICPMV_KV_PROJ    "resampler.kv.weight"
-#define TN_MINICPMV_ATTN       "resampler.attn.%s.%s"
-#define TN_MINICPMV_LN         "resampler.ln_%s.%s"
-
-#define TN_GLM_ADAPER_CONV      "adapter.conv.%s"
-#define TN_GLM_ADAPTER_LINEAR   "adapter.linear.linear.%s"
-#define TN_GLM_ADAPTER_NORM_1   "adapter.linear.norm1.%s"
-#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
-#define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
-#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
-
-enum projector_type {
-    PROJECTOR_TYPE_MLP,
-    PROJECTOR_TYPE_MLP_NORM,
-    PROJECTOR_TYPE_LDP,
-    PROJECTOR_TYPE_LDPV2,
-    PROJECTOR_TYPE_MINICPMV,
-    PROJECTOR_TYPE_GLM_EDGE,
-    PROJECTOR_TYPE_QWEN2VL,
-    PROJECTOR_TYPE_GEMMA3,
-    PROJECTOR_TYPE_IDEFICS3,
-    PROJECTOR_TYPE_PIXTRAL,
-    PROJECTOR_TYPE_QWEN25VL,
-    PROJECTOR_TYPE_UNKNOWN,
-};
-
-static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
-    { PROJECTOR_TYPE_MLP,       "mlp" },
-    { PROJECTOR_TYPE_LDP,       "ldp" },
-    { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
-    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
-    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
-    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
-    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
-    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
-    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
-    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
-};
-
-static projector_type clip_projector_type_from_string(const std::string & str) {
-    for (const auto & pair : PROJECTOR_TYPE_NAMES) {
-        if (pair.second == str) {
-            return pair.first;
-        }
-    }
-    return PROJECTOR_TYPE_UNKNOWN;
-}
-
-// RGB uint8 image
-struct clip_image_u8 {
-    int nx;
-    int ny;
-
-    std::vector<uint8_t> buf;
-};
-
-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
-struct clip_image_f32 {
-    int nx;
-    int ny;
-
-    std::vector<float> buf;
-};
-
-//
-// logging
-//
-
-static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
-    (void) level;
-    (void) user_data;
-    fputs(text, stderr);
-    fflush(stderr);
-}
-
-struct clip_logger_state {
-    ggml_log_level verbosity_thold;
-    ggml_log_callback log_callback;
-    void * log_callback_user_data;
-};
-
-extern struct clip_logger_state g_logger_state;
-
-static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
-    if (format == NULL) {
-        return;
-    }
-    va_list args_copy;
-    va_copy(args_copy, args);
-    char buffer[128];
-    int len = vsnprintf(buffer, 128, format, args);
-    if (len < 128) {
-        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
-    } else {
-        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
-        vsnprintf(buffer2, len + 1, format, args_copy);
-        buffer2[len] = 0;
-        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
-        free(buffer2);
-    }
-    va_end(args_copy);
-}
-
-static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
-    va_list args;
-    va_start(args, format);
-    clip_log_internal_v(level, format, args);
-    va_end(args);
-}
-
-#define LOG_TMPL(level, ...) \
-    do { \
-        if ((level) >= g_logger_state.verbosity_thold) { \
-            clip_log_internal((level), __VA_ARGS__); \
-        } \
-    } while (0)
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
-
-//
-// cpp wrappers
-//
-
-// wrapper for clip_image_size
-struct clip_image_size_deleter {
-    void operator()(clip_image_size * val) { clip_image_size_free(val); }
-};
-typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
-
-// wrapper for clip_image_u8
-struct clip_image_u8_deleter {
-    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
-};
-typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
-
-// wrapper for clip_image_f32
-struct clip_image_f32_deleter {
-    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
-};
-typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
-
-struct clip_image_u8_batch {
-    std::vector<clip_image_u8_ptr> entries;
-};
-
-struct clip_image_f32_batch {
-    std::vector<clip_image_f32_ptr> entries;
-};
-
-//
-// common utils
-//
-
-static std::string string_format(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), buf.size());
-}
-
-static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    if (search.empty()) {
-        return;
-    }
-    std::string builder;
-    builder.reserve(s.length());
-    size_t pos = 0;
-    size_t last_pos = 0;
-    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        builder.append(s, last_pos, pos - last_pos);
-        builder.append(replace);
-        last_pos = pos + search.length();
-    }
-    builder.append(s, last_pos, std::string::npos);
-    s = std::move(builder);
-}
-
-// split string by a `std::string delim` instead of `char delim`
-static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
-    std::vector<std::string> tokens;
-    size_t pos = 0;
-    std::string token;
-    while ((pos = s.find(delimiter)) != std::string::npos) {
-        token = s.substr(0, pos);
-        tokens.push_back(token);
-        s.erase(0, pos + delimiter.length());
-    }
-    tokens.push_back(s);
-    return tokens;
-}
-
-//
-// gguf utils
-//
-
-static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
-    switch (type) {
-        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
-        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
-        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
-        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
-        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
-        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
-        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
-        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
-        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
-        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
-        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
-        default:                return string_format("unknown type %d", type);
-    }
-}
-
-static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
-    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
-
-    switch (type) {
-        case GGUF_TYPE_STRING:
-            return gguf_get_val_str(ctx_gguf, i);
-        case GGUF_TYPE_ARRAY:
-            {
-                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
-                int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
-                std::stringstream ss;
-                ss << "[";
-                for (int j = 0; j < arr_n; j++) {
-                    if (arr_type == GGUF_TYPE_STRING) {
-                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
-                        // escape quotes
-                        string_replace_all(val, "\\", "\\\\");
-                        string_replace_all(val, "\"", "\\\"");
-                        ss << '"' << val << '"';
-                    } else if (arr_type == GGUF_TYPE_ARRAY) {
-                        ss << "???";
-                    } else {
-                        ss << gguf_data_to_str(arr_type, data, j);
-                    }
-                    if (j < arr_n - 1) {
-                        ss << ", ";
-                    }
-                }
-                ss << "]";
-                return ss.str();
-            }
-        default:
-            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
-    }
-}
-
-//
-// API used internally with mtmd
-//
-
-projector_type clip_get_projector_type(const struct clip_ctx * ctx);

examples/llava/clip.h

@@ -1,7 +1,6 @@
 #ifndef CLIP_H
 #define CLIP_H
 
-#include "ggml.h"
 #include <stddef.h>
 #include <stdint.h>
 
@@ -30,13 +29,19 @@ struct clip_image_size {
     int height;
 };
 
-struct clip_image_f32;
-struct clip_image_u8_batch;
-struct clip_image_f32_batch;
+struct clip_image_u8_batch {
+    struct clip_image_u8 * data;
+    size_t size;
+};
+
+struct clip_image_f32_batch {
+    struct clip_image_f32 * data;
+    size_t size;
+};
 
 struct clip_context_params {
     bool use_gpu;
-    enum ggml_log_level verbosity;
+    int verbosity;
 };
 
 // deprecated, use clip_init
@@ -49,9 +54,9 @@ CLIP_API void clip_free(struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
 
-CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
 
 // TODO: should be enum, not string
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
@@ -67,26 +72,15 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
 CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
 
-CLIP_API struct clip_image_size      * clip_image_size_init();
-CLIP_API struct clip_image_u8        * clip_image_u8_init ();
-CLIP_API struct clip_image_f32       * clip_image_f32_init();
-CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
+CLIP_API struct clip_image_size * clip_image_size_init();
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
-// nx, ny are the output image dimensions
-CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
-
-CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
-// use for accessing underlay data of clip_image_f32_batch
-CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
-CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
-CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
-CLIP_API struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
-
 /**
  * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
  * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
@@ -111,8 +105,8 @@ CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
-CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
-CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
+
+CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
 
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 

examples/llava/deprecation-warning.cpp

@@ -1,22 +0,0 @@
-#include <cstdio>
-#include <string>
-
-int main(int argc, char** argv) {
-    std::string filename = "main";
-    if (argc >= 1) {
-        filename = argv[0];
-    }
-
-    // Get only the program name from the full path
-    size_t pos = filename.find_last_of("/\\");
-    if (pos != std::string::npos) {
-        filename = filename.substr(pos+1);
-    }
-
-    fprintf(stdout, "\n");
-    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
-    fprintf(stdout, "Please use 'llama-mtmd-cli' instead.\n");
-    fprintf(stdout, "\n");
-
-    return EXIT_FAILURE;
-}

examples/llava/gemma3-cli.cpp

@@ -0,0 +1,311 @@
+#include "arg.h"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "stb_image.h"
+#include "llama.h"
+#include "llama-cpp.h"
+#include "ggml.h"
+#include "console.h"
+
+#include <vector>
+#include <limits.h>
+#include <inttypes.h>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+static bool g_is_generating = false;
+
+/**
+ * Please note that this is NOT a production-ready stuff.
+ * It is a playground for trying Gemma 3 vision capabilities.
+ * For contributors: please keep this code simple and easy to understand.
+ */
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG(
+        "Experimental CLI for using Gemma 3 vision model\n\n"
+        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
+        "  -m and --mmproj are required\n"
+        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
+        argv[0]
+    );
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (g_is_generating) {
+            g_is_generating = false;
+        } else {
+            console::cleanup();
+            LOG("\nInterrupted by user\n");
+            _exit(130);
+        }
+    }
+}
+#endif
+
+struct gemma3_context {
+    struct clip_ctx    * ctx_clip = NULL;
+    common_init_result   llama_init;
+
+    llama_model       * model;
+    llama_context     * lctx;
+    const llama_vocab * vocab;
+    llama_batch_ext_ptr batch;
+
+    int n_threads    = 1;
+    llama_pos n_past = 0;
+
+    gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) {
+        model = llama_init.model.get();
+        lctx = llama_init.context.get();
+        vocab = llama_model_get_vocab(model);
+        n_threads = params.cpuparams.n_threads;
+        batch.reset(llama_batch_ext_init(params.n_batch, 1));
+        init_clip_model(params);
+    }
+
+    void init_clip_model(common_params & params) {
+        const char * clip_path = params.mmproj.c_str();
+        ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
+    }
+
+    ~gemma3_context() {
+        clip_free(ctx_clip);
+    }
+};
+
+static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
+    llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
+    llama_batch_ext_clear(ctx.batch.get());
+    for (llama_token & t : tokens) {
+        llama_seq_id seq_id = 0;
+        llama_batch_ext_add_text(ctx.batch.get(), t, ctx.n_past++, &seq_id, 1, false);
+    }
+    if (logits_last) {
+        llama_batch_ext_set_output_last(ctx.batch.get());
+    }
+    // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
+    if (llama_decode_ext(ctx.lctx, ctx.batch.get())) {
+        LOG_ERR("Failed to decode text\n");
+        return 1;
+    }
+    return 0;
+}
+
+static int eval_image(gemma3_context & ctx, std::string & fname) {
+    std::vector<float> image_embd_v;
+    int n_embd = llama_model_n_embd(ctx.model);
+    int n_tokens = 256;
+    image_embd_v.resize(n_tokens * n_embd);
+
+    bool ok;
+    struct clip_image_u8 * img_u8 = clip_image_u8_init();
+    ok = clip_image_load_from_file(fname.c_str(), img_u8);
+    if (!ok) {
+        LOG_ERR("Unable to load image %s\n", fname.c_str());
+        clip_image_u8_free(img_u8);
+        return 2; // non-fatal error
+    }
+
+    clip_image_f32_batch batch_f32;
+    ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32);
+    if (!ok) {
+        LOG_ERR("Unable to preprocess image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+
+    int64_t t0 = ggml_time_ms();
+    LOG("Encoding image %s\n", fname.c_str());
+    ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data());
+    if (!ok) {
+        LOG_ERR("Unable to encode image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+    LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+
+    clip_image_f32_batch_free(&batch_f32);
+    clip_image_u8_free(img_u8);
+
+    // decode image embeddings
+    int64_t t1 = ggml_time_ms();
+    eval_text(ctx, "<start_of_image>");
+    llama_set_causal_attn(ctx.lctx, false);
+    llama_batch_ext_ptr batch_img(llama_batch_ext_init_from_embd(image_embd_v.data(), n_tokens, n_embd, ctx.n_past, 0));
+    if (llama_decode_ext(ctx.lctx, batch_img.get())) {
+        LOG_ERR("failed to decode image\n");
+        return 1;
+    }
+    ctx.n_past += n_tokens;
+    llama_set_causal_attn(ctx.lctx, true);
+    eval_text(ctx, "<end_of_image>");
+    LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
+    return 0;
+}
+
+static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
+    for (int i = 0; i < n_predict; i++) {
+        if (i > n_predict || !g_is_generating) {
+            printf("\n");
+            break;
+        }
+
+        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        common_sampler_accept(smpl, token_id, true);
+
+        if (llama_vocab_is_eog(ctx.vocab, token_id)) {
+            printf("\n");
+            break; // end of generation
+        }
+
+        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        fflush(stdout);
+
+        // eval the token
+        llama_batch_ext_clear(ctx.batch.get());
+        llama_seq_id seq_id = 0;
+        llama_batch_ext_add_text(ctx.batch.get(), token_id, ctx.n_past++, &seq_id, 1, true);
+        if (llama_decode_ext(ctx.lctx, ctx.batch.get())) {
+            LOG_ERR("failed to decode token\n");
+            return 1;
+        }
+    }
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+    params.sampling.temp = 0.2; // lower temp by default for better quality
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty()) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    gemma3_context ctx(params);
+    printf("%s: %s\n", __func__, params.model.c_str());
+
+    bool is_single_turn = !params.prompt.empty() && !params.image.empty();
+
+    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
+    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
+
+    // ctrl+C handling
+    {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    }
+
+    if (eval_text(ctx, "<bos>")) {
+        return 1;
+    }
+
+    if (is_single_turn) {
+        g_is_generating = true;
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+        for (auto & fname : params.image) {
+            if (eval_image(ctx, fname)) {
+                return 1;
+            }
+        }
+        if (eval_text(ctx, params.prompt + "<end_of_turn><start_of_turn>model\n", true)) {
+            return 1;
+        }
+        if (generate_response(ctx, smpl, n_predict)) {
+            return 1;
+        }
+
+    } else {
+        LOG("\n Running in chat mode, available commands:");
+        LOG("\n   /image <path>    load an image");
+        LOG("\n   /clear           clear the chat history");
+        LOG("\n   /quit or /exit   exit the program");
+        LOG("\n");
+
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+
+        while (true) {
+            g_is_generating = false;
+            LOG("\n> ");
+            console::set_display(console::user_input);
+            std::string line;
+            console::readline(line, false);
+            console::set_display(console::reset);
+            line = string_strip(line);
+            if (line.empty()) {
+                continue;
+            }
+            if (line == "/quit" || line == "/exit") {
+                break;
+            }
+            if (line == "/clear") {
+                ctx.n_past = 0;
+                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                LOG("Chat history cleared\n\n");
+                continue;
+            }
+            g_is_generating = true;
+            if (line.find("/image") == 0) {
+                std::string image = line.substr(7);
+                int res = eval_image(ctx, image);
+                if (res == 2) {
+                    continue; // image not found
+                }
+                if (res) {
+                    return 1;
+                }
+                continue;
+            }
+            if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
+                return 1;
+            }
+            if (generate_response(ctx, smpl, n_predict)) {
+                return 1;
+            }
+            if (eval_text(ctx, "<end_of_turn><start_of_turn>user\n")) {
+                return 1;
+            }
+        }
+    }
+
+    return 0;
+}

examples/llava/gemma3_convert_encoder_to_gguf.py

@@ -0,0 +1,307 @@
+import gguf
+import argparse
+import logging
+import sys
+import torch
+import json
+import os
+import numpy as np
+from typing import cast, ContextManager, Any, Iterator
+from pathlib import Path
+from torch import Tensor
+
+logger = logging.getLogger("gemma3-mmproj")
+
+
+# (copied from convert_hf_to_gguf.py)
+# tree of lazy tensors
+class LazyTorchTensor(gguf.LazyBase):
+    _tensor_type = torch.Tensor
+    # to keep the type-checker happy
+    dtype: torch.dtype
+    shape: torch.Size
+
+    # only used when converting a torch.Tensor to a np.ndarray
+    _dtype_map: dict[torch.dtype, type] = {
+        torch.float16: np.float16,
+        torch.float32: np.float32,
+    }
+
+    # used for safetensors slices
+    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
+    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
+    _dtype_str_map: dict[str, torch.dtype] = {
+        "F64": torch.float64,
+        "F32": torch.float32,
+        "BF16": torch.bfloat16,
+        "F16": torch.float16,
+        # "U64": torch.uint64,
+        "I64": torch.int64,
+        # "U32": torch.uint32,
+        "I32": torch.int32,
+        # "U16": torch.uint16,
+        "I16": torch.int16,
+        "U8": torch.uint8,
+        "I8": torch.int8,
+        "BOOL": torch.bool,
+        "F8_E4M3": torch.float8_e4m3fn,
+        "F8_E5M2": torch.float8_e5m2,
+    }
+
+    def numpy(self) -> gguf.LazyNumpyTensor:
+        dtype = self._dtype_map[self.dtype]
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
+            args=(self,),
+            func=(lambda s: s.numpy())
+        )
+
+    @classmethod
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
+
+    @classmethod
+    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
+        dtype = cls._dtype_str_map[st_slice.get_dtype()]
+        shape: tuple[int, ...] = tuple(st_slice.get_shape())
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
+        return cast(torch.Tensor, lazy)
+
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.Tensor.numpy:
+            return args[0].numpy()
+
+        return cls._wrap_fn(func)(*args, **kwargs)
+
+
+class Gemma3VisionTower:
+    hparams: dict
+    gguf_writer: gguf.GGUFWriter
+    fname_out: Path
+    ftype: gguf.LlamaFileType
+
+    @staticmethod
+    def load_hparams(dir_model: Path):
+        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
+        part_names: list[str] = []
+        for filename in os.listdir(dir_model):
+            if filename.startswith(prefix) and filename.endswith(suffix):
+                part_names.append(filename)
+        part_names.sort()
+        return part_names
+
+    def __init__(self,
+                 dir_model: Path,
+                 fname_out: Path,
+                 ftype: gguf.LlamaFileType,
+                 is_big_endian: bool,):
+        hparams = Gemma3VisionTower.load_hparams(dir_model)
+        self.hparams = hparams
+        self.fname_out = fname_out
+        self.ftype = ftype
+        endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess)
+
+        text_config = hparams["text_config"]
+        vision_config = hparams["vision_config"]
+
+        assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration"
+        assert text_config is not None
+        assert vision_config is not None
+
+        self.gguf_writer.add_string ("clip.projector_type",              "gemma3")
+        self.gguf_writer.add_bool   ("clip.has_text_encoder",            False)
+        self.gguf_writer.add_bool   ("clip.has_vision_encoder",          True)
+        self.gguf_writer.add_bool   ("clip.has_llava_projector",         False) # legacy
+        self.gguf_writer.add_uint32 ("clip.vision.image_size",           vision_config["image_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.patch_size",           vision_config["patch_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.embedding_length",     vision_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length",  vision_config["intermediate_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.projection_dim",       text_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.block_count",          vision_config["num_hidden_layers"])
+        self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"])
+        self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6))
+        # default values taken from HF tranformers code
+        self.gguf_writer.add_array  ("clip.vision.image_mean", [0.5, 0.5, 0.5])
+        self.gguf_writer.add_array  ("clip.vision.image_std",  [0.5, 0.5, 0.5])
+        self.gguf_writer.add_bool   ("clip.use_gelu", True)
+
+        # load tensors
+        for name, data_torch in self.get_tensors(dir_model):
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+            self.add_tensor(name, data_torch)
+
+    def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]:
+        part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors")
+        tensor_names_from_parts: set[str] = set()
+        for part_name in part_names:
+            logger.info(f"gguf: loading model part '{part_name}'")
+            from safetensors import safe_open
+            ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu"))
+            with ctx as model_part:
+                tensor_names_from_parts.update(model_part.keys())
+
+                for name in model_part.keys():
+                    data = model_part.get_slice(name)
+                    data = LazyTorchTensor.from_safetensors_slice(data)
+                    yield name, data
+
+    def add_tensor(self, name: str, data_torch: Tensor):
+        is_1d = len(data_torch.shape) == 1
+        is_embd = ".embeddings." in name
+        old_dtype = data_torch.dtype
+        can_quantize = not is_1d and not is_embd
+        data_qtype = gguf.GGMLQuantizationType.F32
+
+        # this is to support old checkpoint
+        # TODO: remove this when we have the final model
+        name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.")
+        name = name.replace("multimodal_projector.", "multi_modal_projector.")
+
+        # filter only vision tensors
+        if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."):
+            return
+        # prefix
+        name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.")
+        name = name.replace("vision_tower.vision_model.", "v.")
+        # projector and input embd
+        name = name.replace(".embeddings.patch_embedding.", ".patch_embd.")
+        name = name.replace(".embeddings.position_embedding.", ".position_embd.")
+        name = name.replace(
+            "multi_modal_projector.mm_input_projection_weight",
+            "mm.input_projection.weight"
+        )
+        name = name.replace(
+            "multi_modal_projector.mm_soft_emb_norm.weight",
+            "mm.soft_emb_norm.weight"
+        )
+        name = name.replace("post_layernorm.", "post_ln.")
+        # each block
+        name = name.replace(".self_attn.k_proj.", ".attn_k.")
+        name = name.replace(".self_attn.v_proj.", ".attn_v.")
+        name = name.replace(".self_attn.q_proj.", ".attn_q.")
+        name = name.replace(".self_attn.out_proj.", ".attn_out.")
+        name = name.replace(".layer_norm1.", ".ln1.")
+        name = name.replace(".layer_norm2.", ".ln2.")
+        name = name.replace(".mlp.fc1.", ".ffn_down.")
+        name = name.replace(".mlp.fc2.", ".ffn_up.")
+
+        if can_quantize:
+            if self.ftype == gguf.LlamaFileType.ALL_F32:
+                data_qtype = gguf.GGMLQuantizationType.F32
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                data_qtype = gguf.GGMLQuantizationType.F16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                data_qtype = gguf.GGMLQuantizationType.BF16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                data_qtype = gguf.GGMLQuantizationType.Q8_0
+            else:
+                raise ValueError(f"Unsupported file type: {self.ftype}")
+
+        # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
+        # the other norm values are part of SigLIP model, and they are already correct
+        # ref code: Gemma3RMSNorm
+        if "soft_emb_norm.weight" in name:
+            logger.info(f"Correcting norm value for '{name}'")
+            data_torch = data_torch + 1
+
+        data = data_torch.numpy()
+
+        try:
+            data = gguf.quants.quantize(data, data_qtype)
+        except Exception as e:
+            logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16")
+            data_qtype = gguf.GGMLQuantizationType.F16
+            data = gguf.quants.quantize(data, data_qtype)
+
+        # reverse shape to make it similar to the internal ggml dimension order
+        shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}"
+        logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
+
+        self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
+
+    def write(self):
+        self.gguf_writer.write_header_to_file(path=self.fname_out)
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.write_tensors_to_file(progress=True)
+        self.gguf_writer.close()
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert Gemma 3 vision tower safetensors to GGUF format",)
+    parser.add_argument(
+        "--outfile", type=Path, default="mmproj.gguf",
+        help="path to write to",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
+        help="output format",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file",
+        nargs="?",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+
+    args = parser.parse_args()
+    if args.model is None:
+        parser.error("the following arguments are required: model")
+    return args
+
+
+def main() -> None:
+    args = parse_args()
+
+    if args.verbose:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    dir_model = args.model
+
+    if not dir_model.is_dir():
+        logger.error(f'Error: {args.model} is not a directory')
+        sys.exit(1)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+    }
+
+    logger.info(f"Loading model: {dir_model.name}")
+
+    with torch.inference_mode():
+        gemma3_vision_tower = Gemma3VisionTower(
+            dir_model=dir_model,
+            fname_out=args.outfile,
+            ftype=ftype_map[args.outtype],
+            is_big_endian=args.bigendian,
+        )
+        gemma3_vision_tower.write()
+
+
+if __name__ == '__main__':
+    main()
+

examples/llava/llava-cli.cpp

@@ -0,0 +1,333 @@
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <vector>
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        auto batch = llama_batch_ext_ptr::init_from_text(&tokens[i], n_eval, *n_past, 0, true);
+        if (llama_decode_ext(ctx_llama, batch.get())) {
+            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+    return true;
+}
+
+static const char * sample(struct common_sampler * smpl,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
+
+    const llama_model * model = llama_get_model(ctx_llama);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    static std::string ret;
+    if (llama_vocab_is_eog(vocab, id)) {
+        ret = "</s>";
+    } else {
+        ret = common_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void print_usage(int, char ** argv) {
+    LOG("\n example usage:\n");
+    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            LOG_INF("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
+        if (!embed) {
+            LOG_ERR("%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
+        if (!embed) {
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
+    int n_past = 0;
+
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+
+    std::string system_prompt, user_prompt;
+    size_t image_pos = prompt.find("<image>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
+        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
+        user_prompt = prompt + "\nASSISTANT:";
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
+    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+
+    // generate the response
+
+    LOG("\n");
+
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    if (!smpl) {
+        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
+    std::string response = "";
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
+        if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
+        LOG("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
+        fflush(stdout);
+    }
+
+    common_sampler_free(smpl);
+    LOG("\n");
+}
+
+static struct llama_model * llava_init(common_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = common_model_params_to_llama(*params);
+
+    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
+    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
+
+    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_model_free(ctx_llava->model);
+    llama_backend_free();
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        print_usage(argc, argv);
+        return 1;
+    }
+
+    auto * model = llava_init(&params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
+        return 1;
+    }
+
+    if (prompt_contains_image(params.prompt)) {
+        auto * ctx_llava = llava_init_context(&params, model);
+
+        auto * image_embed = load_image(ctx_llava, &params, "");
+
+        // process the prompt
+        process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+        llama_perf_context_print(ctx_llava->ctx_llama);
+        llava_image_embed_free(image_embed);
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+    } else {
+        for (auto & image : params.image) {
+            auto * ctx_llava = llava_init_context(&params, model);
+
+            auto * image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_perf_context_print(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
+    }
+
+    llama_model_free(model);
+
+    return 0;
+}

examples/llava/llava.cpp

@@ -2,6 +2,7 @@
 #include "llava.h"
 
 #include "llama.h"
+#include "llama-cpp.h"
 
 #include <algorithm>
 #include <cerrno>
@@ -10,7 +11,6 @@
 #include <cstring>
 #include <limits>
 #include <vector>
-#include <memory>
 
 #if defined(LLAVA_LOG_OFF)
 #   define LOG_INF(...)
@@ -46,17 +46,6 @@ struct clip_image_grid_shape {
     int second;
 };
 
-// convenience cpp wrapper
-struct clip_image_f32_batch_deleter {
-    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
-};
-typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
-
-struct clip_image_size_deleter {
-    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
-};
-typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
-
 /**
  * Selects the best resolution from a list of possible resolutions based on the original size.
  *
@@ -117,8 +106,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
         struct ggml_context * ctx;
     } model;
 
-    const int32_t image_size = clip_get_image_size(ctx_clip);
-    const int32_t patch_size = clip_get_patch_size(ctx_clip);
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
 
     int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
 
@@ -258,9 +247,12 @@ static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size)
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
-    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
-    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
         LOG_ERR("%s: unable to preprocess image\n", __func__);
+        delete[] img_res_v.data;
         return false;
     }
 
@@ -268,72 +260,66 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
-
     if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(n_imgs);
-        clip_image_size load_image_size;
+        image_embd_v.resize(img_res_v.size);
+        struct clip_image_size * load_image_size = clip_image_size_init();
 
-        for (size_t i = 0; i < n_imgs; i++) {
+        for (size_t i = 0; i < img_res_v.size; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
-            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
-            int patch_size = 14;
-            load_image_size.width = nx;
-            load_image_size.height = ny;
-            clip_add_load_image_size(ctx_clip, &load_image_size);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
+            int patch_size=14;
+            load_image_size->width = img_res_v.data[i].nx;
+            load_image_size->height = img_res_v.data[i].ny;
+            clip_add_load_image_size(ctx_clip, load_image_size);
 
             bool encoded = false;
-            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             if (clip_is_qwen2vl(ctx_clip)) {
-                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
             }
             else {
-                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
             }
 
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
             const int64_t t_img_enc_steop_batch_us = ggml_time_us();
-            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
-            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
-            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
-            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             std::memcpy(
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
-                clip_embd_nbytes_by_img(ctx_clip, nx, ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
+                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
             free(image_embd_v[i]);
         }
         image_embd_v.clear();
-        load_image_size.width = img->nx;
-        load_image_size.height = img->ny;
-        clip_add_load_image_size(ctx_clip, &load_image_size);
-        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height);
+        load_image_size->width = img->nx;
+        load_image_size->height = img->ny;
+        clip_add_load_image_size(ctx_clip, load_image_size);
+        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
     }
     else if (clip_is_glm(ctx_clip)){
         struct clip_image_size * load_image_size = clip_image_size_init();
-        load_image_size->width  = clip_image_f32_batch_nx(img_res_v.get(), 0);
-        load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
+        load_image_size->width = img_res_v.data[0].nx;
+        load_image_size->height = img_res_v.data[0].ny;
         clip_add_load_image_size(ctx_clip, load_image_size);
 
-        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
-        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
-        int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
+        int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
         *n_img_pos = (pos * pos + 2);
         if (!encoded){
             LOG_ERR("Unable to encode image \n");
@@ -343,8 +329,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
-        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
-        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
 
@@ -355,18 +341,17 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(n_imgs);
-        for (size_t i = 0; i < n_imgs; i++) {
-            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
         const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
@@ -376,7 +361,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
         }
 
-        const int32_t image_size = clip_get_image_size(ctx_clip);
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
 
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 
@@ -449,39 +439,6 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     return true;
 }
 
-struct llava_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
     int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
 
@@ -491,8 +448,8 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
             n_eval = n_batch;
         }
         float * embd = image_embed->embed+i*n_embd;
-        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
-        if (llama_decode(ctx_llama, llava_batch.batch)) {
+        auto batch = llama_batch_ext_ptr::init_from_embd(embd, n_eval, n_embd, 0, 0);
+        if (llama_decode_ext(ctx_llama, batch.get())) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }

examples/llava/minicpmv-cli.cpp

@@ -0,0 +1,355 @@
+#include "arg.h"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <algorithm>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <vector>
+#include <iostream> // TODO: remove me
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llama_model * llava_init(common_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = common_model_params_to_llama(*params);
+
+    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
+    if (params->n_ctx < 2048) {
+        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
+        LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
+        ctx_params.n_ctx = 2048;
+    } else {
+        ctx_params.n_ctx = params->n_ctx;
+    }
+
+    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_model_free(ctx_llava->model);
+    llama_backend_free();
+}
+
+static struct clip_ctx * clip_init_context(common_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+    struct clip_context_params clip_params = {
+        /* use_gpu */   params->n_gpu_layers != 0,
+        /* verbosity */ params->verbosity,
+    };
+    auto * ctx_clip = clip_init(clip_path, clip_params);
+    return ctx_clip;
+}
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        auto batch = llama_batch_ext_ptr::init_from_text(&tokens[i], n_eval, *n_past, 0, true);
+        if (llama_decode_ext(ctx_llama, batch.get())) {
+            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
+    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+}
+
+static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
+    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
+    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
+
+    auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    slice_embed->embed = image_embed;
+    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
+    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
+    llava_image_embed_free(slice_embed);
+}
+
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
+    std::string system_prompt;
+    int idx = 0;
+    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (has_minicpmv_projector == 2) {
+        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    }
+    else if (has_minicpmv_projector == 3) {
+        system_prompt = "<|im_start|>user\n";
+    }
+    else if (has_minicpmv_projector == 4) {
+        system_prompt = "<|im_start|>user\n";
+    }
+    LOG_INF("%s: image token past: %d\n", __func__, n_past);
+    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
+    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+    if (num_image_embeds > 1) {
+        if (has_minicpmv_projector == 2) {
+            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+            eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+                for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                    eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                    if (j == num_image_embeds_col - 1) {
+                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                    }
+                }
+            }
+            eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+        }
+        else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
+            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+                for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                    eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                    eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+                    if (j == num_image_embeds_col - 1) {
+                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                    }
+                }
+            }
+        }
+    }
+    LOG_INF("%s: image token past: %d\n", __func__, n_past);
+}
+
+static const char * sample(struct common_sampler * smpl,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
+
+    const llama_model * model = llama_get_model(ctx_llama);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    static std::string ret;
+    if (llama_vocab_is_eog(vocab, id)) {
+        ret = "</s>";
+    } else {
+        ret = common_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
+    auto * ctx_clip = clip_init_context(params);
+    auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
+    if (!embeds) {
+        LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
+        return NULL;
+    }
+
+    // process the prompt
+    if (params->prompt.empty() && params->interactive == false) {
+        LOG_ERR("prompt should be given or interactive mode should be on");
+        return NULL;
+    }
+
+    auto * model = llava_init(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
+        return NULL;
+    }
+    const int64_t t_llava_init_start_us = ggml_time_us();
+    auto * ctx_llava = llava_init_context(params, model);
+    ctx_llava->ctx_clip = ctx_clip;
+    const int64_t t_llava_init_end_us = ggml_time_us();
+    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
+    LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
+
+    const int64_t t_process_image_start_us = ggml_time_us();
+    process_image(ctx_llava, embeds, params, n_past);
+    const int64_t t_process_image_end_us = ggml_time_us();
+    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
+    LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
+
+    llava_image_embed_free(embeds);
+    return ctx_llava;
+}
+
+static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
+    std::string user_prompt = prompt;
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (!is_first) {
+        if (has_minicpmv_projector == 2) {
+            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+        }
+        else if (has_minicpmv_projector == 3) {
+            user_prompt = "<|im_start|>user\n" + prompt;
+        }
+        else if (has_minicpmv_projector == 4) {
+            user_prompt = "<|im_start|>user\n" + prompt;
+        }
+    }
+
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+    if (has_minicpmv_projector == 2) {
+        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    }
+    else if (has_minicpmv_projector == 3) {
+        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
+    }
+    else if (has_minicpmv_projector == 4) {
+        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
+    }
+
+    // generate the response
+
+    LOG_INF("\n");
+
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    return smpl;
+}
+
+static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
+
+    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
+    return tmp;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty() || (params.image.empty())) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    for (auto & image : params.image) {
+        int n_past = 0;
+        auto * ctx_llava = minicpmv_init(&params, image, n_past);
+
+        if (!params.prompt.empty()) {
+            LOG("<user>%s\n", params.prompt.c_str());
+            LOG("<assistant>");
+            auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
+            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+            std::string response;
+            bool have_tmp = false;
+            for (int i = 0; i < max_tgt_len; i++) {
+                const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
+                response += tmp;
+                if (strcmp(tmp, "</s>") == 0){
+                    if (!have_tmp) {
+                        continue;
+                    }
+                    break;
+                }
+                if (strstr(tmp, "###")) break; // Yi-VL behavior
+                have_tmp = true;
+                printf("%s", tmp);
+                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+
+                fflush(stdout);
+            }
+            common_sampler_free(smpl);
+        }else {
+            while (true) {
+                LOG("<user>");
+                std::string prompt;
+                std::getline(std::cin, prompt);
+                LOG("<assistant>");
+                auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
+                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+                std::string response;
+                for (int i = 0; i < max_tgt_len; i++) {
+                    const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
+                    response += tmp;
+                    if (strcmp(tmp, "</s>") == 0) break;
+                    printf("%s", tmp);// mistral llava-1.6
+                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+                    fflush(stdout);
+                }
+                common_sampler_free(smpl);
+            }
+        }
+        printf("\n");
+        llama_perf_context_print(ctx_llava->ctx_llama);
+
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+    }
+
+    return 0;
+}

examples/llava/mtmd-cli.cpp

@@ -1,376 +0,0 @@
-#include "arg.h"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "llama.h"
-#include "ggml.h"
-#include "console.h"
-#include "chat.h"
-#include "mtmd.h"
-
-#include <vector>
-#include <limits.h>
-#include <cinttypes>
-
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-#include <signal.h>
-#include <unistd.h>
-#elif defined (_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-#define NOMINMAX
-#endif
-#include <windows.h>
-#include <signal.h>
-#endif
-
-// volatile, because of signal being an interrupt
-static volatile bool g_is_generating = false;
-static volatile bool g_is_interrupted = false;
-
-/**
- * Please note that this is NOT a production-ready stuff.
- * It is a playground for trying multimodal support in llama.cpp.
- * For contributors: please keep this code simple and easy to understand.
- */
-
-static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG(
-        "Experimental CLI for multimodal\n\n"
-        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
-        "  -m and --mmproj are required\n"
-        "  -hf user/repo can replace both -m and --mmproj in most cases\n"
-        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n"
-        "  to disable using GPU for mmproj model, add --no-mmproj-offload\n",
-        argv[0]
-    );
-}
-
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-static void sigint_handler(int signo) {
-    if (signo == SIGINT) {
-        if (g_is_generating) {
-            g_is_generating = false;
-        } else {
-            console::cleanup();
-            if (g_is_interrupted) {
-                _exit(1);
-            }
-            g_is_interrupted = true;
-        }
-    }
-}
-#endif
-
-struct mtmd_cli_context {
-    mtmd_context_ptr ctx_vision;
-    common_init_result llama_init;
-
-    llama_model       * model;
-    llama_context     * lctx;
-    const llama_vocab * vocab;
-    llama_batch         batch;
-    int                 n_batch;
-
-    // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
-    // so here we don't need to keep track of chat history
-    common_chat_templates_ptr tmpls;
-
-    // support for legacy templates (models not having EOT token)
-    llama_tokens antiprompt_tokens;
-
-    int n_threads    = 1;
-    llama_pos n_past = 0;
-
-    mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
-        model = llama_init.model.get();
-        lctx = llama_init.context.get();
-        vocab = llama_model_get_vocab(model);
-        n_threads = params.cpuparams.n_threads;
-        batch = llama_batch_init(params.n_batch, 0, 1);
-        n_batch = params.n_batch;
-
-        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
-            LOG_ERR("Model does not have chat template.\n");
-            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
-            LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
-            exit(1);
-        }
-
-        tmpls = common_chat_templates_init(model, params.chat_template);
-        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja).c_str());
-
-        init_vision_context(params);
-
-        // load antiprompt tokens for legacy templates
-        if (params.chat_template == "vicuna") {
-            antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
-        } else if (params.chat_template == "deepseek") {
-            antiprompt_tokens = common_tokenize(lctx, "###", false, true);
-        }
-    }
-
-    void init_vision_context(common_params & params) {
-        const char * clip_path = params.mmproj.path.c_str();
-        ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
-            /* use_gpu */   params.mmproj_use_gpu,
-            /* timings */   true,
-            /* n_threads */ params.cpuparams.n_threads,
-            /* verbosity */ params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO,
-        }));
-        if (!ctx_vision.get()) {
-            LOG_ERR("Failed to load vision model from %s\n", clip_path);
-            exit(1);
-        }
-    }
-
-    bool check_antiprompt(const llama_tokens & generated_tokens) {
-        if (antiprompt_tokens.empty() || generated_tokens.size() < antiprompt_tokens.size()) {
-            return false;
-        }
-        return std::equal(
-            generated_tokens.end() - antiprompt_tokens.size(),
-            generated_tokens.end(),
-            antiprompt_tokens.begin()
-        );
-    }
-};
-
-struct decode_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
-static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
-    llama_tokens generated_tokens;
-    for (int i = 0; i < n_predict; i++) {
-        if (i > n_predict || !g_is_generating || g_is_interrupted) {
-            printf("\n");
-            break;
-        }
-
-        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
-        generated_tokens.push_back(token_id);
-        common_sampler_accept(smpl, token_id, true);
-
-        if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
-            printf("\n");
-            break; // end of generation
-        }
-
-        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
-        fflush(stdout);
-
-        if (g_is_interrupted) {
-            printf("\n");
-            break;
-        }
-
-        // eval the token
-        common_batch_clear(ctx.batch);
-        common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
-        if (llama_decode(ctx.lctx, ctx.batch)) {
-            LOG_ERR("failed to decode token\n");
-            return 1;
-        }
-    }
-    return 0;
-}
-
-static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
-    std::vector<mtmd_bitmap> bitmaps;
-
-    common_chat_templates_inputs tmpl_inputs;
-    tmpl_inputs.messages = {msg};
-    tmpl_inputs.add_generation_prompt = true;
-    tmpl_inputs.use_jinja = false; // jinja is buggy here
-    auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
-    LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
-
-    for (auto & fname : images_fname) {
-        mtmd_bitmap bitmap;
-        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
-            LOG_ERR("Unable to load image %s\n", fname.c_str());
-            return 2; // image not found
-        }
-        bitmaps.push_back(std::move(bitmap));
-    }
-
-    mtmd_input_text text;
-    text.text          = formatted_chat.prompt;
-    text.add_special   = add_bos;
-    text.parse_special = true;
-    mtmd_input_chunks chunks;
-
-    if (g_is_interrupted) return 0;
-
-    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
-    if (res != 0) {
-        LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
-        return 1;
-    }
-
-    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
-        LOG_ERR("Unable to eval prompt\n");
-        return 1;
-    }
-
-    ctx.n_past += mtmd_helper_get_n_tokens(chunks);
-
-    return 0;
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-    params.sampling.temp = 0.2; // lower temp by default for better quality
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty()) {
-        show_additional_info(argc, argv);
-        LOG_ERR("ERR: Missing --mmproj argument\n");
-        return 1;
-    }
-
-    mtmd_cli_context ctx(params);
-    printf("%s: %s\n", __func__, params.model.path.c_str());
-
-    bool is_single_turn = !params.prompt.empty() && !params.image.empty();
-
-    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
-    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
-
-    // ctrl+C handling
-    {
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-        struct sigaction sigint_action;
-        sigint_action.sa_handler = sigint_handler;
-        sigemptyset (&sigint_action.sa_mask);
-        sigint_action.sa_flags = 0;
-        sigaction(SIGINT, &sigint_action, NULL);
-#elif defined (_WIN32)
-        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
-        };
-        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-#endif
-    }
-
-    if (g_is_interrupted) return 130;
-
-    if (is_single_turn) {
-        g_is_generating = true;
-        if (params.prompt.find("<__image__>") == std::string::npos) {
-            params.prompt += " <__image__>";
-        }
-        common_chat_msg msg;
-        msg.role = "user";
-        msg.content = params.prompt;
-        if (eval_message(ctx, msg, params.image, true)) {
-            return 1;
-        }
-        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
-            return 1;
-        }
-
-    } else {
-        LOG("\n Running in chat mode, available commands:");
-        LOG("\n   /image <path>    load an image");
-        LOG("\n   /clear           clear the chat history");
-        LOG("\n   /quit or /exit   exit the program");
-        LOG("\n");
-
-        bool is_first_msg = true;
-        std::vector<std::string> images_fname;
-        std::string content;
-
-        while (!g_is_interrupted) {
-            g_is_generating = false;
-            LOG("\n> ");
-            console::set_display(console::user_input);
-            std::string line;
-            console::readline(line, false);
-            if (g_is_interrupted) break;
-            console::set_display(console::reset);
-            line = string_strip(line);
-            if (line.empty()) {
-                continue;
-            }
-            if (line == "/quit" || line == "/exit") {
-                break;
-            }
-            if (line == "/clear") {
-                ctx.n_past = 0;
-                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
-                LOG("Chat history cleared\n\n");
-                continue;
-            }
-            g_is_generating = true;
-            if (line.find("/image") == 0) {
-                std::string image = line.substr(7);
-                images_fname.push_back(string_strip(image));
-                content += "<__image__>";
-                continue;
-            } else {
-                content += line;
-            }
-            common_chat_msg msg;
-            msg.role = "user";
-            msg.content = content;
-            int ret = eval_message(ctx, msg, images_fname, is_first_msg);
-            if (g_is_interrupted) break;
-            if (ret == 2) {
-                // non-fatal error
-                images_fname.clear();
-                content.clear();
-                continue;
-            }
-            if (ret) {
-                return 1;
-            }
-            if (generate_response(ctx, smpl, n_predict)) {
-                return 1;
-            }
-            images_fname.clear();
-            content.clear();
-            is_first_msg = false;
-        }
-    }
-    if (g_is_interrupted) LOG("\nInterrupted by user\n");
-    llama_perf_context_print(ctx.lctx);
-    return g_is_interrupted ? 130 : 0;
-}

examples/llava/mtmd.cpp

@@ -1,600 +0,0 @@
-#include "clip.h"
-#include "clip-impl.h"
-#include "mtmd.h"
-
-#include "llama.h"
-
-#include <algorithm>
-#include <cerrno>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <limits>
-#include <vector>
-
-// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
-// models not having it (llava-1.6) will process embeddings without any special tokens in-between
-enum mtmd_slice_tmpl {
-    MTMD_SLICE_TMPL_NONE,
-    MTMD_SLICE_TMPL_MINICPMV_2_5,
-    MTMD_SLICE_TMPL_MINICPMV_2_6,
-    // TODO @ngxson : add support for idefics (SmolVLM)
-};
-
-struct mtmd_context {
-    struct clip_ctx * ctx_clip;
-    const struct llama_model * text_model;
-    std::vector<float> image_embd_v; // image embedding vector
-
-    bool print_timings;
-    int n_threads;
-    std::string image_marker;
-
-    // for minicpmv, we need special tokens in-between slices
-    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
-    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
-    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
-    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
-    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
-    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
-    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
-    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
-
-    // TODO @ngxson : add timings
-
-    mtmd_context(const char * mmproj_fname,
-                   const llama_model * text_model,
-                   const mtmd_context_params & ctx_params) :
-        print_timings(ctx_params.print_timings),
-        n_threads    (ctx_params.n_threads),
-        image_marker (ctx_params.image_marker)
-    {
-        clip_context_params ctx_clip_params;
-        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
-        ctx_clip_params.verbosity = ctx_params.verbosity;
-        ctx_clip = clip_init(mmproj_fname, ctx_clip_params);
-        if (!ctx_clip) {
-            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
-        }
-        this->text_model = text_model;
-
-        GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
-
-        int minicpmv_version = clip_is_minicpmv(ctx_clip);
-        if (minicpmv_version == 2) {
-            // minicpmv 2.5 format:
-            // <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
-            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_5;
-            tok_ov_img_start  = lookup_token("<image>");
-            tok_ov_img_end    = lookup_token("</image>");
-            tok_slices_start  = lookup_token("<slice>");
-            tok_slices_end    = lookup_token("</slice>");
-            tok_sli_img_start = tok_ov_img_start;
-            tok_sli_img_end   = tok_ov_img_end;
-            tok_row_end       = lookup_token("\n");
-
-        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
-            // minicpmv 2.6 format:
-            // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
-            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;
-            tok_ov_img_start  = lookup_token("<image>");
-            tok_ov_img_end    = lookup_token("</image>");
-            tok_sli_img_start = lookup_token("<slice>");
-            tok_sli_img_end   = lookup_token("</slice>");
-            tok_row_end       = lookup_token("\n");
-
-        } else if (minicpmv_version != 0) {
-            GGML_ASSERT(false && "unsupported minicpmv version");
-        }
-    }
-
-    ~mtmd_context() {
-        clip_free(ctx_clip);
-    }
-
-private:
-    llama_token lookup_token(const std::string & token_text) {
-        const llama_vocab * vocab = llama_model_get_vocab(text_model);
-        const int n_vocab = llama_vocab_n_tokens(vocab);
-        for (int i = 0; i < n_vocab; i++) {
-            if (token_to_piece(vocab, i, true) == token_text) {
-                return i;
-            }
-        }
-        return LLAMA_TOKEN_NULL;
-    }
-
-    std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
-        std::string piece;
-        piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
-        const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
-        if (n_chars < 0) {
-            piece.resize(-n_chars);
-            int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
-            GGML_ASSERT(check == -n_chars);
-        } else {
-            piece.resize(n_chars);
-        }
-        return piece;
-    }
-};
-
-struct mtmd_image_tokens_data {
-    clip_image_f32_batch batch_f32; // preprocessed image patches
-};
-
-struct mtmd_image_tokens {
-    uint32_t nx; // number of tokens in x direction
-    uint32_t ny; // number of tokens in y direction
-    uint32_t n_tokens() const { return nx * ny; }
-    clip_image_f32_batch batch_f32; // preprocessed image patches
-    std::string id; // optional user-defined ID, useful for KV cache tracking
-};
-
-mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
-        const struct llama_model * text_model,
-        const struct mtmd_context_params ctx_params) {
-    try {
-        return new mtmd_context(mmproj_fname, text_model, ctx_params);
-    } catch (const std::exception & e) {
-        LOG_ERR("%s: error: %s\n", __func__, e.what());
-        return nullptr;
-    }
-}
-
-void mtmd_free(mtmd_context * ctx) {
-    if (ctx) {
-        delete ctx;
-    }
-}
-
-// copied from common_tokenize
-static std::vector<llama_token> mtmd_tokenize_text_internal(
-    const struct llama_vocab * vocab,
-           const std::string & text,
-                        bool   add_special,
-                        bool   parse_special) {
-    // upper limit for the number of tokens
-    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 < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-    return result;
-}
-
-int32_t mtmd_tokenize(mtmd_context * ctx,
-                        std::vector<mtmd_input_chunk> & output,
-                        const mtmd_input_text & text,
-                        const std::vector<mtmd_bitmap> & bitmaps) {
-    auto vocab = llama_model_get_vocab(ctx->text_model);
-
-    std::string prompt_modified(text.text);
-    std::string marker_modified(ctx->image_marker);
-    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
-
-    // a bit hacky here, but works for now
-    // for some models, we need to add prefix and suffix to the image embeddings
-    if (clip_is_gemma3(ctx->ctx_clip)) {
-        // gemma 3
-        // <start_of_image> ... (image embeddings) ... <end_of_image>
-        marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-
-    } else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
-        // <|begin_of_image|> ... (image embeddings) ... <|end_of_image|>
-        marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-
-    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
-        // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
-        marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-
-    } else if (proj_type == PROJECTOR_TYPE_PIXTRAL) {
-        // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
-        marker_modified = ctx->image_marker + "[IMG_END]";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-    }
-
-    // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
-
-    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
-    output.clear();
-    output.reserve(parts.size());
-
-    size_t i_img = 0;
-
-    // utility for adding raw tokens
-    auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
-        mtmd_input_chunk chunk{
-            MTMD_INPUT_CHUNK_TYPE_TEXT,
-            std::move(tokens),
-            {},
-        };
-        output.emplace_back(std::move(chunk));
-    };
-
-    // utility for splitting batch of multiple images into chunks of batch having single images
-    auto split_batch_to_chunk = [&ctx](clip_image_f32_batch && batch_f32, const std::string & id) {
-        std::vector<mtmd_input_chunk> chunks;
-
-        for (auto & entry : batch_f32.entries) {
-            mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-            image_tokens->nx = clip_n_patches_by_img(ctx->ctx_clip, entry.get());
-            image_tokens->ny = 1;
-            image_tokens->batch_f32.entries.push_back(std::move(entry));
-            image_tokens->id = id;
-
-            mtmd_input_chunk chunk{
-                MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                {},
-                std::move(image_tokens),
-            };
-            chunks.emplace_back(std::move(chunk));
-        }
-
-        return chunks;
-    };
-
-    for (const auto & part : parts) {
-        // printf("tokenizing part: %s\n", part.c_str());
-        bool add_bos = &parts.front() == &part;
-        auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
-        if (tokens.empty()) {
-            continue;
-        }
-        mtmd_input_chunk chunk{
-            MTMD_INPUT_CHUNK_TYPE_TEXT,
-            std::move(tokens),
-            {},
-        };
-        output.emplace_back(std::move(chunk));
-
-        if (&parts.back() != &part) {
-            // add image token to middle of 2 parts
-
-            if (i_img >= bitmaps.size()) {
-                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
-                return 1;
-            }
-
-            // convert mtmd_bitmap to clip_image_u8
-            clip_image_u8_ptr img_u8(clip_image_u8_init());
-            img_u8->nx = bitmaps[i_img].nx;
-            img_u8->ny = bitmaps[i_img].ny;
-            img_u8->buf.resize(bitmaps[i_img].data.size());
-            std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
-            clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
-
-            // preprocess image
-            clip_image_f32_batch batch_f32;
-            bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32);
-            if (!ok) {
-                LOG_ERR("Unable to preprocess image\n");
-                return 2;
-            }
-
-            if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
-                // split batch into chunks of single images
-                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
-                GGML_ASSERT(chunks.size() > 0);
-
-                // add overview image
-                add_text_chunk({ctx->tok_ov_img_start});
-                output.emplace_back(std::move(chunks.front()));
-                chunks.erase(chunks.begin());
-                add_text_chunk({ctx->tok_ov_img_end});
-
-                // add slices
-                if (!chunks.empty()) {
-                    clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
-                    int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
-                    int n_row = (int)chunks.size() / n_col;
-                    GGML_ASSERT(n_row * n_col == (int)chunks.size());
-                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_slices_start});
-                    }
-                    for (int y = 0; y < n_row; y++) {
-                        for (int x = 0; x < n_col; x++) {
-                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
-                                add_text_chunk({ctx->tok_sli_img_start});
-                            }
-                            output.emplace_back(std::move(chunks[y * n_col + x]));
-                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
-                                add_text_chunk({ctx->tok_sli_img_end});
-                            }
-                        }
-                        if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
-                            add_text_chunk({ctx->tok_row_end});
-                        }
-                    }
-                    if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_slices_end});
-                    }
-                }
-
-            } else {
-                size_t n_tokens = 0;
-                for (const auto & entry : batch_f32.entries) {
-                    n_tokens += clip_n_patches_by_img(ctx->ctx_clip, entry.get());
-                }
-
-                mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-                image_tokens->nx = n_tokens;
-                image_tokens->ny = 1; // TODO
-                image_tokens->batch_f32 = std::move(batch_f32);
-                image_tokens->id = bitmaps[i_img].id; // optional
-
-                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
-                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
-                LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
-
-                mtmd_input_chunk chunk{
-                    MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                    {},
-                    std::move(image_tokens),
-                };
-                output.emplace_back(std::move(chunk));
-            }
-
-            i_img++; // move to next image
-        }
-    }
-
-    return 0;
-}
-
-void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
-    if (image_tokens) {
-        delete image_tokens;
-    }
-}
-
-size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->n_tokens();
-}
-
-size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->nx;
-}
-
-size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->ny;
-}
-
-std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->id;
-}
-
-int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
-    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
-    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
-    bool ok = false;
-
-    // only effective for minicpmv and qwen2vl, other models will ignore load_image_size
-    {
-        clip_image_size slice_size{
-            image_tokens->batch_f32.entries[0]->nx,
-            image_tokens->batch_f32.entries[0]->ny};
-        clip_add_load_image_size(ctx->ctx_clip, &slice_size);
-    }
-
-    if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
-        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
-        const auto & entries = image_tokens->batch_f32.entries;
-        for (size_t i = 0; i < entries.size(); i++) {
-            int n_tokens_per_image = clip_n_patches_by_img(ctx->ctx_clip, entries[i].get());
-            ok = clip_image_encode(
-                ctx->ctx_clip,
-                ctx->n_threads,
-                entries[i].get(),
-                ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
-        }
-    } else {
-        ok = clip_image_batch_encode(
-            ctx->ctx_clip,
-            ctx->n_threads,
-            &image_tokens->batch_f32,
-            ctx->image_embd_v.data());
-    }
-
-    return ok ? 0 : 1;
-}
-
-float * mtmd_get_output_embd(mtmd_context * ctx) {
-    return ctx->image_embd_v.data();
-}
-
-size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
-    size_t n_tokens = 0;
-    for (auto & chunk : chunks) {
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            n_tokens += chunk.tokens_text.size();
-        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            n_tokens += chunk.tokens_image->n_tokens();
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
-    }
-    return n_tokens;
-}
-
-// helper struct to make working with embd batch easier
-// note: this will be removed after llama_batch_ext refactoring
-struct decode_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
-int32_t mtmd_helper_eval(mtmd_context * ctx,
-        llama_context * lctx,
-        mtmd_input_chunks & chunks,
-        llama_pos pos0,
-        llama_seq_id seq_id,
-        int32_t n_batch) {
-    int32_t ret;
-    llama_pos n_past = pos0;
-    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
-    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
-
-    for (auto & chunk : chunks) {
-        bool is_last = &chunk == &chunks.back();
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            text_batch.n_tokens = chunk.tokens_text.size();
-            size_t i = 0;
-            while (i < chunk.tokens_text.size()) { // split into batches
-                for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
-                    text_batch.token   [i]    = chunk.tokens_text[i];
-                    text_batch.pos     [i]    = n_past++;
-                    text_batch.n_seq_id[i]    = 1;
-                    text_batch.seq_id  [i][0] = seq_id;
-                    text_batch.logits  [i]    = false;
-                }
-                if (is_last) {
-                    // always get logits for last input chunk
-                    text_batch.logits[text_batch.n_tokens - 1] = true;
-                }
-                ret = llama_decode(lctx, text_batch);
-                if (ret != 0) {
-                    LOG_ERR("failed to decode text\n");
-                    llama_batch_free(text_batch);
-                    return ret;
-                }
-            }
-
-        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
-            GGML_ASSERT(chunk.tokens_image != nullptr);
-            int64_t t0 = ggml_time_ms();
-            if (ctx->print_timings) {
-                LOG_INF("encoding image or slice...\n");
-            }
-            ret = mtmd_encode(ctx, chunk.tokens_image.get());
-            if (ret != 0) {
-                LOG_ERR("failed to encode image\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-            if (ctx->print_timings) {
-                LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
-            }
-
-            int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
-            int32_t i_batch = 0;
-            int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
-            float * embd = mtmd_get_output_embd(ctx);
-
-            if (mtmd_decode_use_non_causal(ctx)) {
-                llama_set_causal_attn(lctx, false);
-                // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
-            }
-
-            while (i_batch < n_img_batches) { // split into batches
-                int32_t pos_offset = i_batch*n_batch;
-                int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
-                float * embd_batch = embd + pos_offset*n_mmproj_embd;
-                decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
-
-                printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
-
-                int64_t t1 = ggml_time_ms();
-                ret = llama_decode(lctx, batch_img.batch);
-                if (ret != 0) {
-                    LOG_ERR("failed to decode image\n");
-                    llama_set_causal_attn(lctx, true); // restore causal attn
-                    llama_batch_free(text_batch);
-                    return ret;
-                }
-
-                if (ctx->print_timings) {
-                    LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
-                }
-
-                i_batch++;
-                n_past += n_tokens_batch;
-            }
-
-            if (mtmd_decode_use_non_causal(ctx)) {
-                llama_set_causal_attn(lctx, true);
-            }
-
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
-    }
-
-    llama_batch_free(text_batch);
-    return 0;
-}
-
-int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) {
-    clip_image_u8_ptr img_u8(clip_image_u8_init());
-    bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
-    if (!ok) {
-        LOG_ERR("Unable to load image from buffer\n");
-        return 1;
-    }
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
-    output.data.resize(output.nx * output.ny * 3);
-    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
-    return 0;
-}
-
-int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) {
-    clip_image_u8_ptr img_u8(clip_image_u8_init());
-    bool ok = clip_image_load_from_file(fname, img_u8.get());
-    if (!ok) {
-        LOG_ERR("Unable to load image %s\n", fname);
-        return 1;
-    }
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
-    output.data.resize(output.nx * output.ny * 3);
-    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
-    return 0;
-}
-
-bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
-    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
-    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
-        return true;
-    }
-    return false;
-}
-
-void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
-    mtmd_image_tokens_free(val);
-}

examples/llava/mtmd.h

@@ -1,161 +0,0 @@
-#ifndef MTMD_H
-#define MTMD_H
-
-#include "ggml.h"
-#include "llama.h"
-#include "clip.h"
-
-#include <vector>
-#include <cinttypes>
-#include <memory>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define MTMD_API __declspec(dllexport)
-#        else
-#            define MTMD_API __declspec(dllimport)
-#        endif
-#    else
-#        define MTMD_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define MTMD_API
-#endif
-
-#ifdef __cplusplus
-
-enum mtmd_input_chunk_type {
-    MTMD_INPUT_CHUNK_TYPE_TEXT,
-    MTMD_INPUT_CHUNK_TYPE_IMAGE,
-};
-
-struct mtmd_context;
-struct mtmd_image_tokens;
-
-// represents raw image data, layout is RGBRGBRGB...
-// length of data must be nx * ny * 3
-struct mtmd_bitmap {
-    uint32_t nx;
-    uint32_t ny;
-    std::vector<unsigned char> data;
-    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
-};
-
-struct mtmd_image_tokens_deleter {
-    void operator()(mtmd_image_tokens * val); // forward declaration
-};
-using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
-
-struct mtmd_input_chunk {
-    mtmd_input_chunk_type type;
-    std::vector<llama_token> tokens_text;
-    mtmd_image_tokens_ptr tokens_image;
-};
-
-using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
-
-struct mtmd_context_params {
-    bool use_gpu = true;
-    bool print_timings = true;
-    int n_threads = 4;
-    enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO;
-    const char * image_marker = "<__image__>";
-};
-
-struct mtmd_input_text {
-    std::string text;
-    bool add_special;
-    bool parse_special;
-};
-
-// initialize the mtmd context
-// return nullptr on failure
-MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
-                                                const llama_model * text_model,
-                                                const mtmd_context_params ctx_params);
-
-MTMD_API void mtmd_free(mtmd_context * ctx);
-
-// tokenize an input text prompt and an image
-// the prompt must have the input image marker (default: "<__image__>") in it
-// the marker will be replaced with the image tokens
-// for example:
-//   "here is an image: <__image__>\ndescribe it in detail."
-//   this will gives 3 chunks:
-//   1. "here is an image: <start_of_image>"
-//   2. (image tokens)
-//   3. "<end_of_image>\ndescribe it in detail."
-// number of bitmaps must be equal to the number of image markers in the prompt
-// this function is thread-safe (shared ctx)
-// return values:
-//   0 on success
-//   1 on number of images not matching the number of markers
-//   2 on image preprocessing error
-MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
-                                std::vector<mtmd_input_chunk> & output,
-                                const mtmd_input_text & text,
-                                const std::vector<mtmd_bitmap> & bitmaps);
-
-// access mtmd_image_tokens
-MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
-MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
-MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
-MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
-MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
-
-// returns 0 on success
-MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
-                            const mtmd_image_tokens * image_tokens);
-
-// get output embeddings from the last encode pass
-MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
-
-// whether we need to set non-causal mask before llama_decode
-MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
-
-
-
-//
-// helper functions (can be implemented based on other functions)
-//
-
-// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
-MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
-
-// helper function that automatically:
-// 1. run llama_decode() on text chunks
-// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
-// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
-// otherwise, returns 0 on success
-MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
-                                llama_context * lctx,
-                                mtmd_input_chunks & chunks,
-                                llama_pos pos0,
-                                llama_seq_id seq_id,
-                                int32_t n_batch);
-
-// helper function to construct a mtmd_bitmap from a file
-// returns 0 on success
-// this function is thread-safe
-MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output);
-
-// helper function to construct a mtmd_bitmap from a buffer
-// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
-// returns 0 on success
-// this function is thread-safe
-MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output);
-
-// convenient unique_ptr wrappers
-struct mtmd_context_deleter {
-    void operator()(mtmd_context * val) { mtmd_free(val); }
-};
-using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
-
-#else
-
-static_assert(false && "C header is not yet supported by this library");
-
-#endif
-
-#endif

examples/llava/qwen2_vl_surgery.py

@@ -1,16 +1,14 @@
 import argparse
-from typing import Dict, List, Optional
+from typing import Dict
 
 import torch
 import numpy as np
 from gguf import *
 from transformers import (
-    AutoProcessor,
-    Qwen2VLConfig,
-    Qwen2VLProcessor,
     Qwen2VLForConditionalGeneration,
-    Qwen2_5_VLConfig, # type: ignore[reportAttributeAccessIssue]
-    Qwen2_5_VLForConditionalGeneration, # type: ignore[reportAttributeAccessIssue]
+    Qwen2VLProcessor,
+    AutoProcessor,
+    Qwen2VLConfig
 )
 
 
@@ -21,93 +19,61 @@ def k(raw_key: str, arch: str) -> str:
     return raw_key.format(arch=arch)
 
 
-def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
-    if fullatt_block_indexes is None:
-        return 0
-    n_wa = fullatt_block_indexes[0]
-    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
-        if b - a - 1 != n_wa:
-            raise ValueError(
-                f"window/full attention layer should have fix pattern of "
-                f"for each full-attention layer followed by {n_wa} window-attention layers"
-            )
-    return n_wa + 1
+def to_gguf_name(name: str) -> str:
+    og = name
+    name = name.replace("text_model", "t").replace("vision_model", "v")
+    name = name.replace("blocks", "blk").replace("embeddings.", "")
+    name = name.replace("attn.", "attn_")
+    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
+    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
+    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+    name = name.replace("merger.mlp", 'mm')
+    print(f"[to_gguf_name] {og} --> {name}")
+    return name
 
 
-class VL2:
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-        # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[to_gguf_name] {og} --> {name}")
-        return name
-
-    @classmethod
-    def find_vision_tensors(cls, qwen2vl, dtype) -> Dict[str, np.ndarray]:
-        vision_model = qwen2vl.visual
-        tensor_map = {}
-        for name, ten in vision_model.state_dict().items():
-            ten = ten.numpy()
-            if 'qkv' in name:
-                if ten.ndim == 2: # weight
-                    c3, _ = ten.shape
-                else:             # bias
-                    c3 = ten.shape[0]
-                assert c3 % 3 == 0
-                c = c3 // 3
-                wq = ten[:c]
-                wk = ten[c: c * 2]
-                wv = ten[c * 2:]
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-            elif 'merger' in name:
-                if name.endswith("ln_q.weight"):
-                    tensor_map['v.post_ln.weight'] = ten
-                elif name.endswith("ln_q.bias"):
-                    tensor_map['v.post_ln.bias'] = ten
-                else:
-                    # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                    tensor_map[cls.to_gguf_name(name)] = ten
-            elif 'patch_embed.proj.weight' in name:
-                # NOTE: split Conv3D into Conv2Ds
-                c1, c2, kt, kh, kw = ten.shape
-                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-                tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-                tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
+def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
+    vision_model = qwen2vl.visual
+    tensor_map = {}
+    for name, ten in vision_model.state_dict().items():
+        ten = ten.numpy()
+        if 'qkv' in name:
+            if ten.ndim == 2: # weight
+                c3, _ = ten.shape
+            else:             # bias
+                c3 = ten.shape[0]
+            assert c3 % 3 == 0
+            c = c3 // 3
+            wq = ten[:c]
+            wk = ten[c: c * 2]
+            wv = ten[c * 2:]
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
+        elif 'merger' in name:
+            if name.endswith("ln_q.weight"):
+                tensor_map['v.post_ln.weight'] = ten
+            elif name.endswith("ln_q.bias"):
+                tensor_map['v.post_ln.bias'] = ten
             else:
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}")] = ten
+                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
+                tensor_map[to_gguf_name(name)] = ten
+        elif 'patch_embed.proj.weight' in name:
+            # NOTE: split Conv3D into Conv2Ds
+            c1, c2, kt, kh, kw = ten.shape
+            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
+            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
+        else:
+            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
 
-        for new_name, ten in tensor_map.items():
-            if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-                tensor_map[new_name] = ten.astype(np.float32)
-            else:
-                tensor_map[new_name] = ten.astype(dtype)
-        tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-        return tensor_map
-
-
-class VL25(VL2):
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.down_proj", "ffn_down").replace("mlp.up_proj", "ffn_up")
-        name = name.replace("mlp.gate_proj", "ffn_gate").replace("proj.", "out.")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[vl25][to_gguf_name] {og} --> {name}")
-        return name
+    for new_name, ten in tensor_map.items():
+        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
+            tensor_map[new_name] = ten.astype(np.float32)
+        else:
+            tensor_map[new_name] = ten.astype(dtype)
+    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
+    return tensor_map
 
 
 def main(args):
@@ -116,7 +82,7 @@ def main(args):
         np_dtype = np.float32
         ftype = 0
     elif args.data_type == 'fp16':
-        dtype = torch.float16
+        dtype = torch.float32
         np_dtype = np.float16
         ftype = 1
     else:
@@ -126,18 +92,11 @@ def main(args):
     model_path = ""
     model_name = args.model_name
     print("model_name: ", model_name)
-    if args.model_type == "qwen2vl":
-        qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
-    else:
-        qwen2vl = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2_5_VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
+    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
+        model_name, torch_dtype=dtype, device_map="cpu"
+    )
+    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+    vcfg = cfg.vision_config
 
     if os.path.isdir(model_name):
         local_model = True
@@ -154,6 +113,7 @@ def main(args):
     fout.add_bool("clip.has_text_encoder", False)
     fout.add_bool("clip.has_vision_encoder", True)
     fout.add_bool("clip.has_qwen2vl_merger", True)
+    fout.add_string("clip.projector_type", "qwen2vl_merger")
 
     print(cfg.vision_config)
     if 'silu' in cfg.vision_config.hidden_act.lower():
@@ -165,25 +125,14 @@ def main(args):
     else:
         raise ValueError()
 
-    if args.model_type == "qwen2.5vl":
-        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
-        fout.add_string("clip.projector_type", "qwen2.5vl_merger")
-    else:
-        fout.add_string("clip.projector_type", "qwen2vl_merger")
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
-
-    if args.model_type == "qwen2.5vl":
-        tensor_map = VL25.find_vision_tensors(qwen2vl, np_dtype)
-    else:
-        tensor_map = VL2.find_vision_tensors(qwen2vl, np_dtype)
+    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
     for name, data in tensor_map.items():
         fout.add_tensor(name, data)
 
     fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
     fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
+    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
     fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
@@ -211,7 +160,6 @@ def main(args):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
-    parser.add_argument("--model_type", nargs='?', choices=['qwen2vl', 'qwen2.5vl'], default="qwen2vl")
     parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
     args = parser.parse_args()
     main(args)

examples/llava/qwen2vl-cli.cpp

@@ -23,9 +23,6 @@
 #include <algorithm>
 #include <iostream>
 #include <fstream>
-#include <limits>
-#include <cassert>
-#include <cmath>
 
 
 static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
@@ -69,17 +66,11 @@ static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct lla
         memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos));
         memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos));
 
-        llama_batch batch = {
-            int32_t(n_eval),                // n_tokens
-            nullptr,                        // token
-            (image_embed->embed+i*n_embd),  // embed
-            batch_mrope_pos.data(),         // pos
-            nullptr,  // n_seq_id
-            nullptr,  // seq_id
-            nullptr,  // logits
-        };
+        float * batch_embd = image_embed->embed+i*n_embd;
+        auto batch = llama_batch_ext_ptr::init_from_embd(batch_embd, n_eval, n_embd, 0, 0);
+        llama_batch_ext_set_pos(batch.get(), batch_mrope_pos.data(), n_eval);
 
-        if (llama_decode(ctx_llama, batch)) {
+        if (llama_decode_ext(ctx_llama, batch.get())) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
@@ -92,14 +83,30 @@ static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct lla
 
 static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
     int N = (int) tokens.size();
+    std::vector<llama_pos> pos;
     for (int i = 0; i < N; i += n_batch) {
         int n_eval = (int) tokens.size() - i;
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        auto batch = llama_batch_get_one(&tokens[i], n_eval);
 
-        if (llama_decode(ctx_llama, batch)) {
+        // TODO: add mrope pos ids somewhere else
+        int n_tokens = n_eval;
+        pos.resize(n_tokens * 4);
+        std::fill(pos.begin(), pos.end(), 0);
+        for (int j = 0; j < n_tokens * 3; j ++) {
+            pos[j] = *st_pos_id + (j % n_tokens);
+        }
+
+        llama_batch_ext_ptr batch(llama_batch_ext_init(n_eval, 1));
+        for (int j = 0; j < n_eval; j++) {
+            llama_token token = tokens[i + j];
+            llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch.get(), token, pos[j], &seq_id, 1, false);
+        }
+        llama_batch_ext_set_output_last(batch.get());
+
+        if (llama_decode_ext(ctx_llama, batch.get())) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
@@ -309,7 +316,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -318,14 +325,14 @@ static struct llama_model * llava_init(common_params * params) {
 }
 
 static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.path.c_str();
+    const char * clip_path = params->mmproj.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
         prompt = "describe the image in detail.";
     }
 
-    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
     llama_context_params ctx_params = common_context_params_to_llama(*params);
     ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
@@ -362,14 +369,14 @@ static void debug_test_mrope_2d() {
     // 1. Initialize backend
     ggml_backend_t backend = NULL;
     std::string backend_name = "";
-// #ifdef GGML_USE_CUDA
-//     fprintf(stderr, "%s: using CUDA backend\n", __func__);
-//     backend = ggml_backend_cuda_init(0); // init device 0
-//     backend_name = "cuda";
-//     if (!backend) {
-//         fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
-//     }
-// #endif
+#ifdef GGML_USE_CUDA
+    fprintf(stderr, "%s: using CUDA backend\n", __func__);
+    backend = ggml_backend_cuda_init(0); // init device 0
+    backend_name = "cuda";
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+    }
+#endif
     // if there aren't GPU Backends fallback to CPU backend
     if (!backend) {
         backend = ggml_backend_cpu_init();
@@ -478,82 +485,28 @@ static void debug_test_mrope_2d() {
     ggml_backend_free(backend);
 }
 
-enum model_output_type {
-    conv3d,
-    patch_embed,
-    patch_win_attn_scatter,
-    first_attn_layer,
-    last_attn_layer,
-    attn_softmax,
-    final_layer,
-};
-
-static void debug_dump_img_embed(struct llava_context * ctx_llava, model_output_type output_type) {
-    constexpr int ih = 140;
-    constexpr int iw = 196;
-    // constexpr int ih = 56;
-    // constexpr int iw = 56;
-    // int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
-    int n_embd  = 1280;
-    int merge = 1;
-    if (output_type == model_output_type::final_layer) {
-        n_embd  = 2048;
-        merge = 2;
-    }
-    else if (output_type == model_output_type::attn_softmax) {
-        merge = 1;
-        n_embd = (ih/14/merge) * (iw/14/merge) * 16;
-    }
-
-    int ne = (ih/14/merge) * (iw/14/merge) * n_embd;
-    float vals[iw * ih * 3];
+static void debug_dump_img_embed(struct llava_context * ctx_llava) {
+    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
+    int ne = n_embd * 4;
+    float vals[56 * 56 * 3];
     // float embd[ne];
     std::vector<float> embd;
     embd.resize(ne);
 
-    for (int i = 0; i < iw*ih; i++)
+    for (int i = 0; i < 56*56; i++)
     {
         for (int c = 0; c < 3; c++)
-            vals[i * 3 + c] = (float)i / (iw*ih);
+            vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
     }
 
-    clip_encode_float_image(ctx_llava->ctx_clip, 8, vals, ih, iw, embd.data());
+    clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
 
-    std::string file_postfix = "";
-    switch (output_type)
-    {
-    case model_output_type::conv3d:
-        file_postfix = "conv3d";
-        break;
-    case model_output_type::patch_embed:
-        file_postfix = "patch_embed";
-        break;
-    case model_output_type::patch_win_attn_scatter:
-        file_postfix = "scatter";
-        break;
-    case model_output_type::first_attn_layer:
-        file_postfix = "first_attn";
-        break;
-    case model_output_type::last_attn_layer:
-        file_postfix = "last_attn";
-        break;
-    case model_output_type::attn_softmax:
-        file_postfix = "attn_softmax";
-        break;
-    case model_output_type::final_layer:
-        file_postfix = "final";
-        break;
-    default:
-        break;
-    }
-    auto output_path = "img_embed_" + file_postfix + ".bin";
-
-    std::ofstream outFile(output_path, std::ios::binary);
+    std::ofstream outFile("img_embed.bin", std::ios::binary);
     if (outFile.is_open()) {
         outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
 
         outFile.close();
-        std::cout << "Data successfully written to ::[ " << output_path << std::endl;
+        std::cout << "Data successfully written to mrope.bin" << std::endl;
     } else {
         std::cerr << "Error opening file!" << std::endl;
     }
@@ -573,7 +526,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);
         return 1;
     }
@@ -600,9 +553,8 @@ int main(int argc, char ** argv) {
     } else if (params.image[0].empty()) {
         auto ctx_llava = llava_init_context(&params, model);
 
-        // debug_test_mrope_2d();
-        debug_dump_img_embed(ctx_llava, model_output_type::final_layer);
-        // debug_dump_img_embed(ctx_llava, model_output_type::last_attn_layer);
+        debug_test_mrope_2d();
+        debug_dump_img_embed(ctx_llava);
 
         llama_perf_context_print(ctx_llava->ctx_llama);
         ctx_llava->model = NULL;

examples/llava/tests.sh

@@ -1,120 +0,0 @@
-#!/bin/bash
-
-# make sure we are in the right directory
-SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
-cd $SCRIPT_DIR
-
-#export LLAMA_CACHE="$SCRIPT_DIR/tmp"
-
-set -eux
-
-mkdir -p $SCRIPT_DIR/output
-
-PROJ_ROOT="$SCRIPT_DIR/../.."
-cd $PROJ_ROOT
-
-# Check if the first argument is "big", then run test with big models
-# This is useful if we're running the script on a larger machine, so we can test the big models
-RUN_BIG_TESTS=false
-if [ "${1:-}" = "big" ]; then
-    RUN_BIG_TESTS=true
-    echo "Include BIG models..."
-fi
-
-###############
-
-arr_bin=()
-arr_hf=()
-arr_tmpl=() # chat template
-
-add_test() {
-    local bin=$1
-    local hf=$2
-    local tmpl=${3:-""} # default to empty string if not provided
-    arr_bin+=("$bin")
-    arr_hf+=("$hf")
-    arr_tmpl+=("$tmpl")
-}
-
-add_test_big() {
-    if [ "$RUN_BIG_TESTS" = true ]; then
-        add_test "$@"
-    fi
-}
-
-add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
-add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
-add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
-add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
-add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
-add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
-add_test "llama-mtmd-cli"  "second-state/Llava-v1.5-7B-GGUF:Q2_K"            "vicuna"
-add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"           "vicuna"
-add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
-add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
-add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
-add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
-add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
-add_test "llama-qwen2vl-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
-
-# to test the big models, run: ./tests.sh big
-add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
-
-# these models always give the wrong answer, not sure why
-# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"
-# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0"
-# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF:Q8_0"
-
-# this model has broken chat template, not usable
-# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
-
-###############
-
-cmake --build build -j --target "${arr_bin[@]}"
-
-arr_res=()
-
-for i in "${!arr_bin[@]}"; do
-    bin="${arr_bin[$i]}"
-    hf="${arr_hf[$i]}"
-    tmpl="${arr_tmpl[$i]}"
-
-    echo "Running test with binary: $bin and HF model: $hf"
-    echo ""
-    echo ""
-
-    output=$(\
-        "$PROJ_ROOT/build/bin/$bin" \
-        -hf "$hf" \
-        --image $SCRIPT_DIR/test-1.jpeg \
-        -p "what is the publisher name of the newspaper?" \
-        --temp 0 -n 128 \
-        ${tmpl:+--chat-template "$tmpl"} \
-        2>&1 | tee /dev/tty)
-
-    echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
-
-    if echo "$output" | grep -iq "new york"; then
-        result="\033[32mOK\033[0m:   $bin $hf"
-    else
-        result="\033[31mFAIL\033[0m: $bin $hf"
-    fi
-    echo -e "$result"
-    arr_res+=("$result")
-
-    echo ""
-    echo ""
-    echo ""
-    echo "#################################################"
-    echo "#################################################"
-    echo ""
-    echo ""
-done
-
-set +x
-
-for i in "${!arr_res[@]}"; do
-    echo -e "${arr_res[$i]}"
-done
-echo ""
-echo "Output logs are saved in $SCRIPT_DIR/output"

examples/lookahead/lookahead.cpp

@@ -92,8 +92,10 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
+    llama_batch_ext_ptr batch0(llama_batch_ext_init_from_text( inp.data(), n_input - 1, 0,           0, true));
+    llama_batch_ext_ptr batch1(llama_batch_ext_init_from_text(&inp.back(),           1, n_input - 1, 0, true));
+    llama_decode_ext(ctx, batch0.get());
+    llama_decode_ext(ctx, batch1.get());
 
     for (int s = 1; s < W + G + 1; ++s) {
         llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
@@ -115,7 +117,7 @@ int main(int argc, char ** argv) {
     // seq_id == 0           : the current input token
     // seq_id [1, W]         : tokens from the past N - 1 Jacobi iterations
     // seq_id [W + 1, W + G] : verification n-grams
-    llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
+    llama_batch_ext * batch = llama_batch_ext_init(params.n_ctx, W + G + 1);
 
     // target model sampling context
     struct common_sampler * smpl = common_sampler_init(model, params.sampling);
@@ -204,10 +206,10 @@ int main(int argc, char ** argv) {
         //                                                      V  V  V  V  V  V
         //                                                             id
         {
-            common_batch_clear(batch);
+            llama_batch_ext_clear(batch);
 
             // current token - first token of the first level
-            common_batch_add(batch, id, n_past, seq_id_all, true);
+            llama_batch_ext_add_text(batch, id, n_past, seq_id_all.data(), seq_id_all.size(), true);
 
             // verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation
             {
@@ -230,9 +232,10 @@ int main(int argc, char ** argv) {
                         const llama_token t = ngrams_observed.tokens[idx + j];
 
                         ngrams_cur[g].tokens [j + 1] = t;
-                        ngrams_cur[g].i_batch[j + 1] = batch.n_tokens;
+                        ngrams_cur[g].i_batch[j + 1] = llama_batch_ext_get_n_tokens(batch);
 
-                        common_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
+                        llama_seq_id seq_id = W + 1 + g;
+                        llama_batch_ext_add_text(batch, t, n_past + j + 1, &seq_id, 1, true);
                     }
                 }
             }
@@ -244,18 +247,20 @@ int main(int argc, char ** argv) {
                     seq_id_look[j] = i + j + 1;
                 }
 
-                common_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
+                llama_batch_ext_add_text(batch, tokens_j[0][i], n_past + i,
+                    seq_id_look.data(), seq_id_look.size(), false);
             }
 
             // fill the rest of the levels
             for (int j = 1; j < N - 1; j++) {
                 for (int i = 0; i < W; i++) {
-                    common_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
+                    llama_seq_id seq_id = i + 1;
+                    llama_batch_ext_add_text(batch, tokens_j[j][i], n_past + j + i, &seq_id, 1, j == N - 2);
                 }
             }
         }
 
-        if (llama_decode(ctx, batch) != 0) {
+        if (llama_decode_ext(ctx, batch) != 0) {
             LOG_ERR("\n\n%s: llama_decode failed - increase KV cache size\n", __func__);
             return 1;
         }
@@ -475,7 +480,7 @@ int main(int argc, char ** argv) {
 
     llama_kv_cache_view_free(&kvc_view);
 
-    llama_batch_free(batch);
+    llama_batch_ext_free(batch);
 
     llama_backend_free();
 

examples/lookup/lookup.cpp

@@ -91,8 +91,10 @@ int main(int argc, char ** argv){
 
     const auto t_enc_start = ggml_time_us();
 
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
+    llama_batch_ext_ptr batch0(llama_batch_ext_init_from_text( inp.data(), n_input - 1, 0,           0, true));
+    llama_batch_ext_ptr batch1(llama_batch_ext_init_from_text(&inp.back(),           1, n_input - 1, 0, true));
+    llama_decode_ext(ctx, batch0.get());
+    llama_decode_ext(ctx, batch1.get());
 
     const auto t_enc_end = ggml_time_us();
 
@@ -108,7 +110,7 @@ int main(int argc, char ** argv){
 
     std::vector<llama_token> draft;
 
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+    llama_batch_ext * batch_tgt = llama_batch_ext_init(params.n_ctx, 1);
 
     // debug
     struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
@@ -194,8 +196,9 @@ int main(int argc, char ** argv){
         // clean the cache of draft tokens that weren't accepted
         llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
-        common_batch_clear(batch_tgt);
-        common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+        const llama_seq_id seq_id = 0;
+        llama_batch_ext_clear(batch_tgt);
+        llama_batch_ext_add_text(batch_tgt, draft[0], n_past, &seq_id, 1, true);
 
         // Draft already contains a single token sampled from the model:
         GGML_ASSERT(draft.size() == 1);
@@ -205,13 +208,13 @@ int main(int argc, char ** argv){
         common_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
 
         for (size_t i = 1; i < draft.size(); ++i) {
-            common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+            llama_batch_ext_add_text(batch_tgt, draft[i], n_past + i, &seq_id, 1, true);
         }
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
         n_drafted += draft.size() - 1;
 
-        llama_decode(ctx, batch_tgt);
+        llama_decode_ext(ctx, batch_tgt);
         ++n_past;
 
         draft.erase(draft.begin());
@@ -243,7 +246,7 @@ int main(int argc, char ** argv){
 
     common_sampler_free(smpl);
 
-    llama_batch_free(batch_tgt);
+    llama_batch_ext_free(batch_tgt);
 
     llama_backend_free();
 

examples/main/main.cpp

@@ -548,7 +548,8 @@ int main(int argc, char ** argv) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
 
-        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
+        auto batch = llama_batch_ext_ptr::init_from_text(enc_input_buf, enc_input_size, 0, 0, true);
+        if (llama_decode_ext(ctx, batch.get())) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
@@ -668,7 +669,8 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                auto batch = llama_batch_ext_ptr::init_from_text(&embd[i], n_eval, n_past, 0, true);
+                if (llama_decode_ext(ctx, batch.get())) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -865,22 +867,9 @@ int main(int argc, char ** argv) {
                 console::set_display(console::reset);
                 display = true;
 
-                if (buffer.empty()) { // Ctrl+D on empty line exits
-                    LOG("EOF by user\n");
-                    break;
-                }
-
-                if (buffer.back() == '\n') {
-                    // Implement #587:
-                    // If the user wants the text to end in a newline,
-                    // this should be accomplished by explicitly adding a newline by using \ followed by return,
-                    // then returning control by pressing return again.
-                    buffer.pop_back();
-                }
-
-                if (buffer.empty()) { // Enter key on empty line lets the user pass control back
-                    LOG_DBG("empty line, passing control back\n");
-                } else { // Add tokens to embd only if the input buffer is non-empty
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty() && !params.conversation_mode) {
                         LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
@@ -928,6 +917,8 @@ int main(int argc, char ** argv) {
 
                     n_remain -= line_inp.size();
                     LOG_DBG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG_DBG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again

examples/parallel/parallel.cpp

@@ -106,8 +106,6 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
-    params.n_predict = 128;
-
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
@@ -176,7 +174,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(n_ctx, 0, 1);
+    llama_batch_ext * batch = llama_batch_ext_init(n_ctx, 1);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -194,10 +192,11 @@ int main(int argc, char ** argv) {
         LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
 
         for (int32_t i = 0; i < n_tokens_system; ++i) {
-            common_batch_add(batch, tokens_system[i], i, { 0 }, false);
+            llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch, tokens_system[i], i, &seq_id, 1, false);
         }
 
-        if (llama_decode(ctx, batch) != 0) {
+        if (llama_decode_ext(ctx, batch) != 0) {
             LOG_ERR("%s: llama_decode() failed\n", __func__);
             return 1;
         }
@@ -218,7 +217,7 @@ int main(int argc, char ** argv) {
             common_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch);
 
         // decode any currently ongoing sequences
         for (auto & client : clients) {
@@ -226,14 +225,15 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            client.i_batch = batch.n_tokens;
+            client.i_batch = llama_batch_ext_get_n_tokens(batch);
 
-            common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            llama_seq_id seq_id = client.id + 1;
+            llama_batch_ext_add_text(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, &seq_id, 1, true);
 
             client.n_decoded += 1;
         }
 
-        if (batch.n_tokens == 0) {
+        if (llama_batch_ext_get_n_tokens(batch) == 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);
@@ -245,7 +245,7 @@ int main(int argc, char ** argv) {
         }
 
         // insert new sequences for decoding
-        if (cont_batching || batch.n_tokens == 0) {
+        if (cont_batching || llama_batch_ext_get_n_tokens(batch) == 0) {
             for (auto & client : clients) {
                 if (client.seq_id == -1 && g_seq_id < n_seq) {
                     client.seq_id = g_seq_id;
@@ -264,17 +264,18 @@ int main(int argc, char ** argv) {
                     tokens_prompt = common_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        llama_seq_id seq_id = client.id + 1;
+                        llama_batch_ext_add_text(batch, tokens_prompt[i], i + n_tokens_system, &seq_id, 1, false);
                     }
 
                     // extract the logits only for the last token
-                    if (batch.n_tokens > 0) {
-                        batch.logits[batch.n_tokens - 1] = true;
+                    if (llama_batch_ext_get_n_tokens(batch) > 0) {
+                        llama_batch_ext_set_output_last(batch);
                     }
 
                     client.n_prompt  = tokens_prompt.size();
                     client.n_decoded = 0;
-                    client.i_batch   = batch.n_tokens - 1;
+                    client.i_batch   = llama_batch_ext_get_n_tokens(batch) - 1;
 
                     LOG_INF("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
 
@@ -288,14 +289,15 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (batch.n_tokens == 0) {
+        if (llama_batch_ext_get_n_tokens(batch) == 0) {
             break;
         }
 
         // 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 n_tokens_in_batch = llama_batch_ext_get_n_tokens(batch);
+        for (int32_t i = 0; i < (int32_t) n_tokens_in_batch; i += n_batch) {
             // experiment: process in powers of 2
             //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
             //    n_batch /= 2;
@@ -303,19 +305,11 @@ 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, (int32_t) (n_tokens_in_batch - i));
 
-            llama_batch batch_view = {
-                n_tokens,
-                batch.token    + i,
-                nullptr,
-                batch.pos      + i,
-                batch.n_seq_id + i,
-                batch.seq_id   + i,
-                batch.logits   + i,
-            };
-
-            const int ret = llama_decode(ctx, batch_view);
+            llama_batch_ext * batch_view = llama_batch_ext_get_view(batch, i, n_tokens);
+            const int ret = llama_decode_ext(ctx, batch_view);
+            llama_batch_ext_free(batch_view);
             if (ret != 0) {
                 if (n_batch == 1 || ret < 0) {
                     // if you get here, it means the KV cache is full - try increasing it via the context size
@@ -407,7 +401,7 @@ int main(int argc, char ** argv) {
         params.prompt_file = "used built-in defaults";
     }
     LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
-    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.path.c_str());
+    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
 
     LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
     LOG_INF("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
@@ -419,7 +413,7 @@ int main(int argc, char ** argv) {
     // TODO: print sampling/grammar timings for all clients
     llama_perf_context_print(ctx);
 
-    llama_batch_free(batch);
+    llama_batch_ext_free(batch);
 
     llama_backend_free();
 

examples/passkey/passkey.cpp

@@ -2,6 +2,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "llama-cpp.h"
 
 #include <cmath>
 #include <cstdio>
@@ -64,7 +65,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
@@ -122,7 +123,7 @@ int main(int argc, char ** argv) {
     LOG_INF("prompt tokens: %d\n", n_tokens_all);
     //LOG_INF("prompt: %s\n", params.prompt.c_str());
 
-    llama_batch batch = llama_batch_init(params.n_batch, 0, 1);
+    llama_batch_ext_ptr batch(llama_batch_ext_init(params.n_batch, 1));
 
     int n_past = 0;
 
@@ -140,17 +141,18 @@ int main(int argc, char ** argv) {
             n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
 
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch.get());
 
         for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
-            common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+            llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch.get(), tokens_list[i + j], n_past++, &seq_id, 1, false);
         }
 
         if (i + n_batch >= n_tokens_all) {
-            batch.logits[batch.n_tokens - 1] = true;
+            llama_batch_ext_set_output_last(batch.get());
         }
 
-        if (llama_decode(ctx, batch) != 0) {
+        if (llama_decode_ext(ctx, batch.get()) != 0) {
             LOG_INF("%s: llama_decode() failed\n", __func__);
             return 1;
         }
@@ -174,17 +176,18 @@ int main(int argc, char ** argv) {
 
         n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
 
-        common_batch_clear(batch);
+        llama_batch_ext_clear(batch.get());
 
         for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
-            common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+            llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch.get(), tokens_list[i + j], n_past++, &seq_id, 1, false);
         }
 
         if (i + n_batch >= n_tokens_all) {
-            batch.logits[batch.n_tokens - 1] = true;
+            llama_batch_ext_set_output_last(batch.get());
         }
 
-        if (llama_decode(ctx, batch) != 0) {
+        if (llama_decode_ext(ctx, batch.get()) != 0) {
             LOG_ERR("%s: llama_decode() failed\n", __func__);
             return 1;
         }
@@ -223,7 +226,7 @@ int main(int argc, char ** argv) {
     while (n_cur <= n_len) {
         // sample the next token
         {
-            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, llama_batch_ext_get_n_tokens(batch.get()) - 1);
 
             // is it an end of generation?
             if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
@@ -237,16 +240,17 @@ int main(int argc, char ** argv) {
             n_decode += 1;
 
             // prepare the next batch
-            common_batch_clear(batch);
+            llama_batch_ext_clear(batch.get());
 
             // push this new token for next evaluation
-            common_batch_add(batch, new_token_id, n_past++, { 0 }, true);
+            llama_seq_id seq_id = 0;
+            llama_batch_ext_add_text(batch.get(), new_token_id, n_past++, &seq_id, 1, true);
         }
 
         n_cur += 1;
 
         // evaluate the current batch with the transformer model
-        if (llama_decode(ctx, batch)) {
+        if (llama_decode_ext(ctx, batch.get())) {
             LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
@@ -266,8 +270,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler_free(smpl);
 
-    llama_batch_free(batch);
-
     llama_free(ctx);
     llama_model_free(model);
 

examples/perplexity/perplexity.cpp

@@ -363,21 +363,20 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         // clear the KV cache
         llama_kv_self_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+        common_batch batch(n_batch, 1);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
-            common_batch_clear(batch);
+            batch.clear();
             for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+                batch.add_text(tokens[batch_start + i], j*n_batch + i, 0, true);
             }
 
             //LOG_DBG("    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode_ext(ctx, batch.get())) {
                 //LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
                 return {tokens, -1, logit_history, prob_history};
             }
 
@@ -397,8 +396,6 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
             }
         }
 
-        llama_batch_free(batch);
-
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -504,7 +501,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
     GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);
     GGML_ASSERT(params.n_ctx == n_seq * n_ctx);
 
-    llama_batch batch = llama_batch_init(std::min(n_batch, n_ctx*n_seq), 0, 1);
+    common_batch batch(std::min(n_batch, n_ctx*n_seq), 1);
 
     std::vector<float> logits;
     if (num_batches > 1) {
@@ -555,7 +552,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
 
             int n_outputs = 0;
 
-            batch.n_tokens = 0;
+            batch.clear();
             for (int seq = 0; seq < n_seq_batch; seq++) {
                 int seq_start = batch_start + seq*n_ctx;
 
@@ -568,22 +565,18 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
                 }
 
                 for (int k = 0; k < batch_size; ++k) {
-                    const int idx = seq*n_ctx + k;
-                    batch.token   [idx]    = tokens[seq_start + k];
-                    batch.pos     [idx]    = j*n_batch + k;
-                    batch.n_seq_id[idx]    = 1;
-                    batch.seq_id  [idx][0] = seq;
-                    batch.logits  [idx]    = batch.pos[idx] >= first ? 1 : 0;
+                    const llama_pos pos = j*n_batch + k;
+                    bool output = pos >= first;
+                    batch.add_text(tokens[seq_start + k], pos, seq, output);
 
-                    n_outputs += batch.logits[idx] != 0;
+                    n_outputs += output ? 1 : 0;
                 }
-                batch.n_tokens += batch_size;
 
                 // restore the original token in case it was set to BOS
                 tokens[seq_start] = token_org;
             }
 
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode_ext(ctx, batch.get())) {
                 LOG_INF("%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
@@ -653,36 +646,23 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
         LOG_ERR("Unexpected negative standard deviation of log(prob)\n");
     }
 
-    llama_batch_free(batch);
-
     return {tokens, ppl, logit_history, prob_history};
 }
 
-static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int n_batch, int n_vocab) {
+static bool decode_helper(llama_context * ctx, common_batch & batch, std::vector<float> & batch_logits, int n_batch, int n_vocab) {
     int prev_outputs = 0;
-    for (int i = 0; i < (int) batch.n_tokens; i += n_batch) {
-        const int n_tokens = std::min<int>(n_batch, batch.n_tokens - i);
+    for (int i = 0; i < (int) batch.get_n_tokens(); i += n_batch) {
+        const int n_tokens = std::min<int>(n_batch, batch.get_n_tokens() - i);
 
-        llama_batch batch_view = {
-            n_tokens,
-            batch.token    + i,
-            nullptr,
-            batch.pos      + i,
-            batch.n_seq_id + i,
-            batch.seq_id   + i,
-            batch.logits   + i,
-        };
+        common_batch batch_view = batch.get_view(i, n_tokens);
 
-        const int ret = llama_decode(ctx, batch_view);
+        const int ret = llama_decode_ext(ctx, batch_view.get());
         if (ret != 0) {
             LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
             return false;
         }
 
-        int n_outputs = 0;
-        for (int i = 0; i < n_tokens; ++i) {
-            n_outputs += batch_view.logits[i] != 0;
-        }
+        int n_outputs = batch_view.n_outputs;
 
         memcpy(batch_logits.data() + size_t(prev_outputs)*n_vocab, llama_get_logits(ctx), size_t(n_outputs)*n_vocab*sizeof(float));
 
@@ -851,7 +831,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
 
     LOG_INF("%s : calculating hellaswag score over selected tasks.\n", __func__);
 
-    LOG("\ntask\tacc_norm\t95%% confidence interval\n");
+    LOG("\ntask\tacc_norm\n");
 
     double acc = 0.0f;
 
@@ -863,7 +843,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, 4);
+    common_batch batch(n_ctx, 4);
 
     std::vector<float> tok_logits(n_vocab);
     // TODO: this could be made smaller; it's currently the worst-case size
@@ -879,7 +859,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
         size_t i1 = i0;
         size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
-        common_batch_clear(batch);
+        batch.clear();
 
         // batch as much tasks as possible into the available context
         // each task has 4 unique sequence ids - one for each ending
@@ -895,9 +875,9 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             }
 
             for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
-                common_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
+                batch.add_text_multi_seq(hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
             }
-            batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            llama_batch_ext_set_output_last(batch.get());
             n_logits += 1;
 
             for (int s = 0; s < 4; ++s) {
@@ -905,7 +885,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
                 // TODO: don't evaluate the last token of each sequence
                 for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
                     const bool needs_logits = i < seq_tokens_size - 1;
-                    common_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    batch.add_text_multi_seq(hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
                     n_logits += needs_logits;
                 }
             }
@@ -985,29 +965,13 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
                 acc += 1.0;
             }
 
-            double freq = acc / double(i + 1);
-
-            const double za = 1.95996398454;
-
-            // // Wald normal approx
-            // double conf =za*sqrt(freq*(1-freq)/double(i + 1));
-            // LOG("%zu\t%.8lf +/- %.8lf\n", i + 1, freq*100.0, conf*100.0);
-
-            // Wilson score interval, more accurate
-            double z   = za * za / double(i + 1);
-            double cnf = z * sqrt(double(i + 1) * (4.0 * freq * (1 - freq) + z)) / (za + za);
-            double a   = (freq + z * 0.5 - cnf) / (1.0 + z);
-            double b   = (freq + z * 0.5 + cnf) / (1.0 + z);
-
-            // Print the accumulated accuracy mean x 100 and confidence interval
-            LOG("%zu\t%3.8lf%%\t[%3.4lf%%, %3.4lf%%]\n", i + 1, freq * 100.0, a * 100.0, b * 100.0);
+            // Print the accumulated accuracy mean x 100
+            LOG("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0);
         }
 
         i0 = i1 - 1;
     }
 
-    llama_batch_free(batch);
-
     LOG("\n");
 }
 
@@ -1161,7 +1125,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
     const int max_tasks_per_batch = 128;
     const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, 2);
+    common_batch batch(n_ctx, 2);
 
     std::vector<float> tok_logits(n_vocab);
     // TODO: this could be made smaller; it's currently the worst-case size
@@ -1180,7 +1144,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
         size_t i1 = i0;
         size_t i_logits = 0;
 
-        common_batch_clear(batch);
+        batch.clear();
 
         while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
             int n_logits = 0;
@@ -1190,15 +1154,15 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             }
 
             for (size_t i = 0; i < data[i1].common_prefix; ++i) {
-                common_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
+                batch.add_text_multi_seq(data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
             }
-            batch.logits[batch.n_tokens - 1] = true;
+            llama_batch_ext_set_output_last(batch.get());
             n_logits += 1;
 
             for (int s = 0; s < 2; ++s) {
                 // TODO: end before the last token, no need to predict past the end of the sequences
                 for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
-                    common_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
+                    batch.add_text_multi_seq(data[i1].seq_tokens[s][i], i, { s0 + s }, true);
                     n_logits += 1;
                 }
             }
@@ -1515,7 +1479,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    common_batch batch(n_ctx, max_seq);
 
     std::vector<float> tok_logits(n_vocab);
     std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);
@@ -1535,7 +1499,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
         size_t i1 = i0;
         size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
-        common_batch_clear(batch);
+        batch.clear();
 
         // batch as much tasks as possible into the available context
         // each task has 4 unique sequence ids - one for each ending
@@ -1558,9 +1522,9 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
 
             for (size_t i = 0; i < cur_task.common_prefix; ++i) {
                 //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
-                common_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
+                batch.add_text_multi_seq(cur_task.seq_tokens[0][i], i, batch_indeces, false);
             }
-            batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            llama_batch_ext_set_output_last(batch.get()); // we need logits for the last token of the common prefix
             n_logits += 1;
 
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
@@ -1568,7 +1532,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
                 // TODO: don't evaluate the last token of each sequence
                 for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
                     const bool needs_logits = i < seq_tokens_size - 1;
-                    common_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    batch.add_text_multi_seq(cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
                     n_logits += needs_logits;
                 }
             }
@@ -1667,8 +1631,6 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
         i0 = i1 - 1;
     }
 
-    llama_batch_free(batch);
-
     if (n_done < 100 && (params.multiple_choice_tasks != 0 && params.multiple_choice_tasks < (size_t)n_task)) return;
 
     float p = 1.f*n_correct/n_done;
@@ -1781,7 +1743,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_self_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+        common_batch batch(n_batch, 1);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -1795,14 +1757,13 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_vocab_bos(vocab);
             }
 
-            common_batch_clear(batch);
+            batch.clear();
             for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+                batch.add_text_multi_seq(tokens[batch_start + i], j*n_batch + i, {0}, true);
             }
 
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode_ext(ctx, batch.get())) {
                 LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
                 return;
             }
 
@@ -1815,8 +1776,6 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
             }
         }
 
-        llama_batch_free(batch);
-
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {

examples/quantize-stats/CMakeLists.txt

@@ -0,0 +1,6 @@
+set(TARGET llama-quantize-stats)
+add_executable(${TARGET} quantize-stats.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/quantize-stats/quantize-stats.cpp

@@ -1,9 +1,8 @@
 #include "ggml.h"
 #include "llama.h"
+#include "llama-model.h"
 #include "common.h"
 
-#include "../src/llama-model.h"
-
 #include <algorithm>
 #include <cassert>
 #include <cinttypes>

examples/quantize/quantize.cpp

@@ -9,7 +9,6 @@
 #include <fstream>
 #include <cmath>
 #include <cctype>
-#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -17,7 +16,7 @@ struct quant_option {
     std::string desc;
 };
 
-static const std::vector<quant_option> QUANT_OPTIONS = {
+static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
     { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
     { "Q5_0",     LLAMA_FTYPE_MOSTLY_Q5_0,     " 5.21G, +0.1316 ppl @ Llama-3-8B",  },
@@ -106,8 +105,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
-    printf("       [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@@ -116,8 +114,6 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
-    printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
-    printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@@ -248,107 +244,6 @@ static ggml_type parse_ggml_type(const char * arg) {
     return GGML_TYPE_COUNT;
 }
 
-// Allowed tensors for arbitrary quantization with --tensor-type option
-static const std::vector<std::string> ALLOWED_TENSOR_TYPE = {
-    "attn_k",
-    "attn_kv_a_mqa",
-    "attn_kv_b",
-    "attn_o",
-    "attn_output",
-    "attn_q",
-    "attn_q_a",
-    "attn_q_b",
-    "attn_qkv",
-    "attn_v",
-    "channel_mix_key",
-    "channel_mix_receptance",
-    "channel_mix_value",
-    "cls",
-    "cls.output",
-    "cross_attn_k",
-    "cross_attn_o",
-    "cross_attn_q",
-    "cross_attn_v",
-    "ffn_act",
-    "ffn_down",
-    "ffn_down_exps",
-    "ffn_down_shexp",
-    "ffn_gate",
-    "ffn_gate_exps",
-    "ffn_gate_shexp",
-    "ffn_up",
-    "ffn_up_exps",
-    "ffn_up_shexp",
-    "ssm_in",
-    "ssm_out",
-    "time_mix_gate",
-    "time_mix_key",
-    "time_mix_output",
-    "time_mix_receptance",
-    "time_mix_value",
-};
-
-// changes to this struct must be replicated in llama-quant.cpp
-struct tensor_quantization {
-    std::string name;
-    ggml_type quant = GGML_TYPE_COUNT;
-};
-
-static bool parse_tensor_type(const char * data, std::vector<tensor_quantization> & tensor_type) {
-    const char * sep = strchr(data, '=');
-    if (sep == nullptr) {
-        printf("\n%s: malformed tensor type '%s'\n\n", __func__, data);
-        return false;
-    }
-
-    const size_t tn_len = sep - data;
-    if (tn_len == 0) {
-        printf("\n%s: missing tensor name\n\n", __func__);
-        return false;
-    }
-
-    if (const size_t qt_len = strlen(sep); qt_len == 1) {
-        printf("\n%s: missing quantization type\n\n", __func__);
-        return false;
-    }
-
-    std::string tn(data, tn_len);
-    std::transform(tn.begin(), tn.end(), tn.begin(), tolower);
-    sep++;
-    const std::string qt(sep);
-
-    bool found = false;
-    for (const auto & allowed : ALLOWED_TENSOR_TYPE) {
-        std::string tensor;
-        tensor = tn.rfind('.') != std::string::npos ? tn.substr(tn.rfind('.') + 1) : tn;
-        // handle special case of cls.output
-        std::string cls_output = "cls.output";
-        if (tn.find(cls_output) != std::string::npos) {
-            tensor = "cls.output";
-        }
-        // check if an allowed tensor exists and it's at the end of the kv string
-        if (tensor == allowed) {
-            found = true;
-            break;
-        }
-    }
-    if (!found) {
-        printf("\n%s: invalid tensor name '%s'\n\n", __func__, tn.c_str());
-        return false;
-    }
-
-    if (parse_ggml_type(qt.c_str()) == GGML_TYPE_COUNT) {
-        printf("\n%s: invalid quantization type '%s'\n\n", __func__, qt.c_str());
-        return false;
-    }
-
-    tensor_quantization tqz;
-    tqz.name = tn;
-    tqz.quant = parse_ggml_type(qt.c_str());
-    tensor_type.emplace_back(std::move(tqz));
-    return true;
-}
-
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -360,7 +255,6 @@ int main(int argc, char ** argv) {
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
     std::vector<llama_model_kv_override> kv_overrides;
-    std::vector<tensor_quantization> tensor_types;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -383,10 +277,6 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
-        } else if (strcmp(argv[arg_idx], "--tensor-type") == 0) {
-            if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
-                usage(argv[0]);
-            }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
@@ -471,9 +361,6 @@ int main(int argc, char ** argv) {
         kv_overrides.back().key[0] = 0;
         params.kv_overrides = &kv_overrides;
     }
-    if (!tensor_types.empty()) {
-        params.tensor_types = &tensor_types;
-    }
 
     llama_backend_init();
 

examples/retrieval/retrieval.cpp

@@ -74,40 +74,56 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
     return chunks;
 }
 
-static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
+static void batch_add_seq(common_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
     size_t n_tokens = tokens.size();
     for (size_t i = 0; i < n_tokens; i++) {
-        common_batch_add(batch, tokens[i], i, { seq_id }, true);
+        batch.add_text(tokens[i], i, seq_id, true);
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+static void batch_decode(llama_context * ctx, common_batch & batch, float * output, int n_seq, int n_embd, int embd_norm = 2) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_self_clear(ctx);
 
     // 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_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, llama_batch_ext_get_n_tokens(batch.get()), n_seq);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode_ext(ctx, batch.get()) < 0) {
+            LOG_ERR("%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode_ext(ctx, batch.get()) < 0) {
+            LOG_ERR("%s : failed to decode\n", __func__);
+        }
     }
 
-    for (int i = 0; i < batch.n_tokens; i++) {
-        if (!batch.logits[i]) {
+    for (int i = 0; i < llama_batch_ext_get_n_tokens(batch.get()); i++) {
+        if (!batch.tokens[i].logits) {
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        if (embd == NULL) {
+        const float * embd = nullptr;
+        int embd_pos = 0;
+
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
             embd = llama_get_embeddings_ith(ctx, i);
-            if (embd == NULL) {
-                LOG_ERR("%s: failed to get embeddings for token %d\n", __func__, i);
-                continue;
-            }
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.tokens[i].seq_id);
+            embd_pos = batch.tokens[i].seq_id;
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
         }
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
-        common_embd_normalize(embd, out, n_embd, 2);
+        float * out = output + embd_pos * n_embd;
+        common_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
 
@@ -214,7 +230,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_chunks = chunks.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    struct common_batch batch = common_batch(n_batch, 1);
 
     // allocate output
     const int n_embd = llama_model_n_embd(model);
@@ -231,10 +247,10 @@ int main(int argc, char ** argv) {
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity
-        if (batch.n_tokens + n_toks > n_batch) {
+        if (llama_batch_ext_get_n_tokens(batch.get()) + n_toks > n_batch) {
             float * out = emb + p * n_embd;
             batch_decode(ctx, batch, out, s, n_embd);
-            common_batch_clear(batch);
+            batch.clear();
             p += s;
             s = 0;
         }
@@ -255,7 +271,7 @@ int main(int argc, char ** argv) {
         chunks[i].tokens.clear();
     }
 
-    struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
+    struct common_batch query_batch = common_batch(n_batch, 1);
 
     // start loop, receive query and return top k similar chunks based on cosine similarity
     std::string query;
@@ -269,7 +285,7 @@ int main(int argc, char ** argv) {
         std::vector<float> query_emb(n_embd, 0);
         batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
 
-        common_batch_clear(query_batch);
+        query_batch.clear();
 
         // compute cosine similarities
         {
@@ -299,6 +315,5 @@ int main(int argc, char ** argv) {
     llama_perf_context_print(ctx);
 
     // clean up
-    llama_batch_free(query_batch);
     llama_backend_free();
 }

examples/rpc/CMakeLists.txt

@@ -1,4 +1,2 @@
-set(TARGET rpc-server)
-add_executable(${TARGET} rpc-server.cpp)
-target_link_libraries(${TARGET} PRIVATE ggml)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+add_executable(rpc-server rpc-server.cpp)
+target_link_libraries(rpc-server PRIVATE ggml llama)

examples/rpc/README.md

@@ -72,14 +72,3 @@ $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name
 
 This way you can offload model layers to both local and remote devices.
 
-### Local cache
-
-The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
-This can speed up model loading significantly, especially when using large models.
-To enable the cache, use the `-c` option:
-
-```bash
-$ bin/rpc-server -c
-```
-
-By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.

examples/rpc/rpc-server.cpp

@@ -1,7 +1,3 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "ggml-cpu.h"
 
 #ifdef GGML_USE_CUDA
@@ -22,149 +18,26 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
-#  define NOMINMAX
-#  define DIRECTORY_SEPARATOR '\\'
-#  include <locale>
 #  include <windows.h>
-#  include <fcntl.h>
-#  include <io.h>
 #else
-#  define DIRECTORY_SEPARATOR '/'
 #  include <unistd.h>
-#  include <sys/stat.h>
 #endif
-#include <codecvt>
 #include <string>
 #include <stdio.h>
-#include <vector>
-#include <filesystem>
-#include <algorithm>
-#include <thread>
-
-namespace fs = std::filesystem;
-
-// NOTE: this is copied from common.cpp to avoid linking with libcommon
-// returns true if successful, false otherwise
-static bool fs_create_directory_with_parents(const std::string & path) {
-#ifdef _WIN32
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
-    std::wstring wpath = converter.from_bytes(path);
-
-    // if the path already exists, check whether it's a directory
-    const DWORD attributes = GetFileAttributesW(wpath.c_str());
-    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
-        return true;
-    }
-
-    size_t pos_slash = 0;
-
-    // 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);
-        if (!success) {
-            const DWORD error = GetLastError();
-
-            // if the path already exists, ensure that it's a directory
-            if (error == ERROR_ALREADY_EXISTS) {
-                const DWORD attributes = GetFileAttributesW(subpath.c_str());
-                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
-                    return false;
-                }
-            } else {
-                return false;
-            }
-        }
-
-        pos_slash += 1;
-    }
-
-    return true;
-#else
-    // if the path already exists, check whether it's a directory
-    struct stat info;
-    if (stat(path.c_str(), &info) == 0) {
-        return S_ISDIR(info.st_mode);
-    }
-
-    size_t pos_slash = 1; // skip leading slashes for directory creation
-
-    // process path from front to back, procedurally creating directories
-    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
-        const std::string subpath = path.substr(0, pos_slash);
-        struct stat info;
-
-        // if the path already exists, ensure that it's a directory
-        if (stat(subpath.c_str(), &info) == 0) {
-            if (!S_ISDIR(info.st_mode)) {
-                return false;
-            }
-        } else {
-            // create parent directories
-            const int ret = mkdir(subpath.c_str(), 0755);
-            if (ret != 0) {
-                return false;
-            }
-        }
-
-        pos_slash += 1;
-    }
-
-    return true;
-#endif // _WIN32
-}
-
-// NOTE: this is copied from common.cpp to avoid linking with libcommon
-static std::string fs_get_cache_directory() {
-    std::string cache_directory = "";
-    auto ensure_trailing_slash = [](std::string p) {
-        // Make sure to add trailing slash
-        if (p.back() != DIRECTORY_SEPARATOR) {
-            p += DIRECTORY_SEPARATOR;
-        }
-        return p;
-    };
-    if (getenv("LLAMA_CACHE")) {
-        cache_directory = std::getenv("LLAMA_CACHE");
-    } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
-        if (std::getenv("XDG_CACHE_HOME")) {
-            cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else {
-            cache_directory = std::getenv("HOME") + std::string("/.cache/");
-        }
-#elif defined(__APPLE__)
-        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
-#elif defined(_WIN32)
-        cache_directory = std::getenv("LOCALAPPDATA");
-#else
-#  error Unknown architecture
-#endif
-        cache_directory = ensure_trailing_slash(cache_directory);
-        cache_directory += "llama.cpp";
-    }
-    return ensure_trailing_slash(cache_directory);
-}
 
 struct rpc_server_params {
     std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
-    bool        use_cache   = false;
-    int         n_threads   = std::max(1U, std::thread::hardware_concurrency()/2);
 };
 
 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
     fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help                show this help message and exit\n");
-    fprintf(stderr, "  -t,      --threads        number of threads for the CPU backend (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
-    fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
-    fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
-    fprintf(stderr, "  -c,      --cache          enable local file cache\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
     fprintf(stderr, "\n");
 }
 
@@ -177,15 +50,6 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
                 return false;
             }
             params.host = argv[i];
-        } else if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                return false;
-            }
-            params.n_threads = std::stoi(argv[i]);
-            if (params.n_threads <= 0) {
-                fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
-                return false;
-            }
         } else if (arg == "-p" || arg == "--port") {
             if (++i >= argc) {
                 return false;
@@ -194,8 +58,6 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
             if (params.port <= 0 || params.port > 65535) {
                 return false;
             }
-        } else if (arg == "-c" || arg == "--cache") {
-            params.use_cache = true;
         } else if (arg == "-m" || arg == "--mem") {
             if (++i >= argc) {
                 return false;
@@ -213,7 +75,7 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
     return true;
 }
 
-static ggml_backend_t create_backend(const rpc_server_params & params) {
+static ggml_backend_t create_backend() {
     ggml_backend_t backend = NULL;
 #ifdef GGML_USE_CUDA
     fprintf(stderr, "%s: using CUDA backend\n", __func__);
@@ -245,7 +107,6 @@ static ggml_backend_t create_backend(const rpc_server_params & params) {
     if (!backend) {
         fprintf(stderr, "%s: using CPU backend\n", __func__);
         backend = ggml_backend_cpu_init();
-        ggml_backend_cpu_set_n_threads(backend, params.n_threads);
     }
     return backend;
 }
@@ -290,7 +151,7 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "\n");
     }
 
-    ggml_backend_t backend = create_backend(params);
+    ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");
         return 1;
@@ -303,23 +164,8 @@ int main(int argc, char * argv[]) {
     } else {
         get_backend_memory(&free_mem, &total_mem);
     }
-    const char * cache_dir = nullptr;
-    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
-    if (params.use_cache) {
-        if (!fs_create_directory_with_parents(cache_dir_str)) {
-            fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
-            return 1;
-        }
-        cache_dir = cache_dir_str.c_str();
-    }
-    printf("Starting RPC server v%d.%d.%d\n",
-           RPC_PROTO_MAJOR_VERSION,
-           RPC_PROTO_MINOR_VERSION,
-           RPC_PROTO_PATCH_VERSION);
-    printf("  endpoint       : %s\n", endpoint.c_str());
-    printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
-    printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));
-    ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
+    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
     ggml_backend_free(backend);
     return 0;
 }

examples/run/CMakeLists.txt

@@ -1,16 +1,5 @@
 set(TARGET llama-run)
 add_executable(${TARGET} run.cpp linenoise.cpp/linenoise.cpp)
-
-# TODO: avoid copying this code block from common/CMakeLists.txt
-set(LLAMA_RUN_EXTRA_LIBS "")
-if (LLAMA_CURL)
-    find_package(CURL REQUIRED)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
-    include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_RUN_EXTRA_LIBS ${LLAMA_RUN_EXTRA_LIBS} ${CURL_LIBRARY})
-endif ()
-
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_RUN_EXTRA_LIBS})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)