opencl: fix incorrect local_size index in profiling log (#12868)

The grouped query attention optmization doesn't require a power of two ratio,
the only thing relying on it was the modulo operation written as bitwise &.

split_k need not depend on gqa_ratio - enable it any time there's only one
workgroup in the X dimension. The shader gets the split index from the x coord,
and multiple workgroups in the X dimension (pre-split) indicates a larger
FA operation that wouldn't need splitting.

.devops/cpu.Dockerfile

@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${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 ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(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 ${OPT_SYCL_F16} && \
+    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 ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -1,4 +1,4 @@
-ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
+ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10
 
 FROM ascendai/cann:$ASCEND_VERSION AS build
 
@@ -6,7 +6,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN yum install -y gcc g++ cmake make
+RUN yum install -y gcc g++ cmake make libcurl-devel
 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 ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(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=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
     && cmake --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 && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

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

@@ -0,0 +1,25 @@
+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

@@ -104,7 +104,6 @@ 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

@@ -0,0 +1,124 @@
+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', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.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']
@@ -54,6 +54,7 @@ jobs:
         continue-on-error: true
         run: |
           brew update
+          brew install curl
 
       - name: Build
         id: cmake_build
@@ -62,7 +63,6 @@ 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
@@ -92,7 +92,6 @@ 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
@@ -123,6 +122,7 @@ jobs:
         continue-on-error: true
         run: |
           brew update
+          brew install curl
 
       - name: Build
         id: cmake_build
@@ -133,7 +133,6 @@ 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)
@@ -162,7 +161,6 @@ 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
@@ -207,7 +205,6 @@ jobs:
         run: |
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_CURL=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -246,7 +243,6 @@ 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
@@ -281,7 +277,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -322,7 +318,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -360,7 +356,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -397,7 +393,7 @@ 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
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -431,7 +427,6 @@ 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
@@ -454,7 +449,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -530,7 +525,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -578,7 +573,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
 
       - name: install oneAPI MKL library
         shell: bash
@@ -606,6 +601,9 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
+  build-linux-cross:
+    uses: ./.github/workflows/build-linux-cross.yml
+
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -633,6 +631,7 @@ 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 \
@@ -668,6 +667,7 @@ 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 \
@@ -676,6 +676,36 @@ 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
 
@@ -707,6 +737,7 @@ 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 \
@@ -774,7 +805,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.3.261.1
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
@@ -867,10 +898,17 @@ 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 }}
+          cmake -S . -B build ${{ matrix.defines }} `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add libopenblas.dll
@@ -930,9 +968,10 @@ 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 LICENSE .\build\bin\Release\llama.cpp.txt
-          Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt
+          Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -958,7 +997,7 @@ jobs:
             DEBIAN_FRONTEND: noninteractive
           run: |
               apt update
-              apt install -y cmake build-essential ninja-build libgomp1 git
+              apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
 
         - name: ccache
           uses: hendrikmuhs/ccache-action@v1.2.16
@@ -1060,16 +1099,23 @@ 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
+            -DGGML_RPC=ON ^
+            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1090,7 +1136,10 @@ 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
@@ -1145,6 +1194,8 @@ 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
@@ -1230,8 +1281,14 @@ 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}"
@@ -1242,9 +1299,11 @@ jobs:
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_HIP=ON `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           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
@@ -1286,8 +1345,14 @@ 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}"
@@ -1299,7 +1364,8 @@ jobs:
             -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON `
+            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
@@ -1321,7 +1387,10 @@ 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
@@ -1346,6 +1415,7 @@ 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 \
@@ -1379,7 +1449,7 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          zip -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1696,16 +1766,17 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
+        shell: bash -el {0}
     strategy:
       matrix:
+        arch: [x86, aarch64]
         cann:
-          - '8.0.rc3.beta1-910b-openeuler22.03-py3.10'
+          - '8.1.RC1.alpha001-910b-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
@@ -1714,7 +1785,7 @@ jobs:
       - name: Dependencies
         run: |
           yum update -y
-          yum install -y git gcc gcc-c++ make cmake
+          yum install -y git gcc gcc-c++ make cmake libcurl-devel
 
       - 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, 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}
+          - { 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 }
           # 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, freediskspace: true }
+          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

.github/workflows/server.yml

@@ -129,7 +129,6 @@ 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 ;
@@ -142,7 +141,6 @@ 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 $(nproc) --target llama-server
@@ -154,7 +152,6 @@ jobs:
           cmake -B build \
               -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
-              -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
@@ -195,17 +192,14 @@ jobs:
 
       - name: libCURL
         id: get_libcurl
-        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
+        uses: ./.github/actions/windows-setup-curl
 
       - name: Build
         id: cmake_build
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
+          cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -221,8 +215,10 @@ jobs:
 
       - name: Copy Libcurl
         id: prepare_libcurl
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+          cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
 
       - name: Tests
         id: server_integration_tests

CMakeLists.txt

@@ -29,6 +29,8 @@ else()
     set(LLAMA_STANDALONE OFF)
 endif()
 
+option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
+
 if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
@@ -79,7 +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" OFF)
+option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 # Required for relocatable CMake package
@@ -145,7 +147,13 @@ endif()
 # 3rd-party
 #
 
-if (NOT TARGET ggml)
+if (LLAMA_USE_SYSTEM_GGML)
+    message(STATUS "Using system-provided libggml, skipping ggml build")
+    find_package(ggml REQUIRED)
+    add_library(ggml ALIAS ggml::ggml)
+endif()
+
+if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -160,6 +168,11 @@ 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()
@@ -234,3 +247,20 @@ 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

@@ -780,10 +780,6 @@ 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

@@ -9,13 +9,6 @@
 
 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)
@@ -104,6 +97,7 @@ 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)
@@ -112,6 +106,8 @@ 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
 
@@ -245,6 +241,7 @@ 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
 
@@ -263,7 +260,9 @@ 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 by using this CLI argument: `-hf <user>/<model>[:quant]`
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
+
+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/`.
 
 After downloading a model, use the CLI tools to run it locally - see below.
 
@@ -528,6 +527,35 @@ 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.
 

build-xcframework.sh

@@ -41,6 +41,11 @@ 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
@@ -325,21 +330,28 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if command -v vtool &>/dev/null; then
+        if command -v xcrun vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
-                    vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
+                    xcrun 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..."
-                    vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
+                    if [[ "$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} ${VISIONOS_MIN_OS_VERSION} -replace \
                         -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                     ;;
                 "tvos")
                     echo "Marking binary as a framework binary for tvOS..."
-                    vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
+                    xcrun vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
                         -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
                     ;;
             esac
@@ -399,6 +411,7 @@ cmake -B build-ios-sim -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-ios-sim --config Release -- -quiet
 
@@ -411,6 +424,7 @@ 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
 
@@ -421,6 +435,7 @@ 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
 
@@ -432,8 +447,9 @@ 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 -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}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
 
@@ -445,8 +461,9 @@ 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 -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}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
 
@@ -462,6 +479,7 @@ 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
 
@@ -476,6 +494,7 @@ 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

@@ -26,4 +26,43 @@ 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

@@ -16,6 +16,9 @@
 # # 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>"
@@ -36,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
@@ -52,13 +55,24 @@ 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
@@ -808,7 +822,7 @@ 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/models-mnt
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
     rm -rf ${SRC}/models-mnt
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
@@ -826,8 +840,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-
-test $ret -eq 0 && gg_run ctest_debug
+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_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -835,7 +851,9 @@ 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
-        test $ret -eq 0 && gg_run test_scripts_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run test_scripts_debug
+        fi
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -846,7 +864,9 @@ 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
-        test $ret -eq 0 && gg_run ctest_with_model_debug
+        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_release
     fi
 fi

cmake/common.cmake

@@ -1,3 +1,5 @@
+include("ggml/cmake/common.cmake")
+
 function(llama_add_compile_flags)
     if (LLAMA_FATAL_WARNINGS)
         if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")

common/CMakeLists.txt

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

common/arg.cpp

@@ -1,12 +1,24 @@
+#include "gguf.h" // for reading GGUF splits
 #include "arg.h"
 
+#include "common.h"
 #include "log.h"
 #include "sampling.h"
 #include "chat.h"
 
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 #include <algorithm>
 #include <climits>
 #include <cstdarg>
+#include <filesystem>
 #include <fstream>
 #include <regex>
 #include <set>
@@ -14,6 +26,14 @@
 #include <thread>
 #include <vector>
 
+//#define LLAMA_USE_CURL
+
+#if defined(LLAMA_USE_CURL)
+#include <curl/curl.h>
+#include <curl/easy.h>
+#include <future>
+#endif
+
 #include "json-schema-to-grammar.h"
 
 using json = nlohmann::ordered_json;
@@ -125,47 +145,554 @@ std::string common_arg::to_string() {
     return ss.str();
 }
 
+//
+// downloader
+//
+
+struct common_hf_file_res {
+    std::string repo; // repo name with ":tag" removed
+    std::string ggufFile;
+    std::string mmprojFile;
+};
+
+#ifdef LLAMA_USE_CURL
+
+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
+#elif defined(_AIX)
+#include <sys/limits.h>
+#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);
+        }
+    }
+};
+
+#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;
+}
+
+// download one single file from remote URL to local path
+static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) {
+    // 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);
+
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    // Check if hf-token or bearer-token was specified
+    if (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_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;
+}
+
+// download multiple files from remote URLs to local paths
+// the input is a vector of pairs <url, path>
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token) {
+    // Prepare download in parallel
+    std::vector<std::future<bool>> futures_download;
+    for (auto const & item : urls) {
+        futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair<std::string, std::string> & it) -> bool {
+            return common_download_file_single(it.first, it.second, bearer_token);
+        }, item));
+    }
+
+    // Wait for all downloads to complete
+    for (auto & f : futures_download) {
+        if (!f.get()) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool common_download_model(
+        const common_params_model & model,
+        const std::string & bearer_token) {
+    // Basic validation of the model.url
+    if (model.url.empty()) {
+        LOG_ERR("%s: invalid model url\n", __func__);
+        return false;
+    }
+
+    if (!common_download_file_single(model.url, model.path, bearer_token)) {
+        return false;
+    }
+
+    // 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(model.path.c_str(), gguf_params);
+        if (!ctx_gguf) {
+            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, model.path.c_str());
+            return false;
+        }
+
+        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), model.path.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split);
+                return false;
+            }
+
+            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 false;
+            }
+        }
+
+        std::vector<std::pair<std::string, std::string>> urls;
+        for (int idx = 1; idx < n_split; idx++) {
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
+
+            char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+            llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);
+
+            if (std::string(split_path) == model.path) {
+                continue; // skip the already downloaded file
+            }
+
+            urls.push_back({split_url, split_path});
+        }
+
+        // Download in parallel
+        common_download_file_multiple(urls, bearer_token);
+    }
+
+    return true;
+}
+
+/**
+ * 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.
+ */
+static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_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
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::string res_str;
+
+    std::string model_endpoint = get_model_endpoint();
+
+    std::string url = model_endpoint + "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 (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_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;
+    std::string ggufFile   = "";
+    std::string mmprojFile = "";
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+    if (res_code == 200) {
+        // extract ggufFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                ggufFile = match[1].str();
+            }
+        }
+        // extract mmprojFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                mmprojFile = match[1].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 (ggufFile.empty()) {
+        throw std::runtime_error("error: model does not have ggufFile");
+    }
+
+    return { hf_repo, ggufFile, mmprojFile };
+}
+
+#else
+
+static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from internet\n");
+    return false;
+}
+
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static bool common_download_model(
+        const common_params_model &,
+        const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return {};
+}
+
+#endif // LLAMA_USE_CURL
+
 //
 // utils
 //
 
-static void common_params_handle_model_default(
-        std::string & model,
-        const std::string & model_url,
-        std::string & hf_repo,
-        std::string & hf_file,
-        const std::string & hf_token,
-        const std::string & model_default) {
-    if (!hf_repo.empty()) {
-        // short-hand to avoid specifying --hf-file -> default it to --model
-        if (hf_file.empty()) {
-            if (model.empty()) {
-                auto auto_detected = common_get_hf_file(hf_repo, hf_token);
-                if (auto_detected.first.empty() || auto_detected.second.empty()) {
-                    exit(1); // built without CURL, error message already printed
+static void common_params_handle_model(
+        struct common_params_model & model,
+        const std::string & bearer_token,
+        const std::string & model_path_default,
+        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+    // handle pre-fill default model path and url based on hf_repo and hf_file
+    {
+        if (!model.hf_repo.empty()) {
+            // short-hand to avoid specifying --hf-file -> default it to --model
+            if (model.hf_file.empty()) {
+                if (model.path.empty()) {
+                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token);
+                    if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
+                        exit(1); // built without CURL, error message already printed
+                    }
+                    model.hf_repo = auto_detected.repo;
+                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                } else {
+                    model.hf_file = model.path;
                 }
-                hf_repo = auto_detected.first;
-                hf_file = auto_detected.second;
-            } else {
-                hf_file = model;
             }
+
+            std::string model_endpoint = get_model_endpoint();
+            model.url = model_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file;
+            // make sure model path is present (for caching purposes)
+            if (model.path.empty()) {
+                // this is to avoid different repo having same file name, or same file name in different subdirs
+                std::string filename = model.hf_repo + "_" + model.hf_file;
+                // to make sure we don't have any slashes in the filename
+                string_replace_all(filename, "/", "_");
+                model.path = fs_get_cache_file(filename);
+            }
+
+        } else if (!model.url.empty()) {
+            if (model.path.empty()) {
+                auto f = string_split<std::string>(model.url, '#').front();
+                f = string_split<std::string>(f, '?').front();
+                model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            }
+
+        } else if (model.path.empty()) {
+            model.path = model_path_default;
         }
-        // make sure model path is present (for caching purposes)
-        if (model.empty()) {
-            // this is to avoid different repo having same file name, or same file name in different subdirs
-            std::string filename = hf_repo + "_" + hf_file;
-            // to make sure we don't have any slashes in the filename
-            string_replace_all(filename, "/", "_");
-            model = fs_get_cache_file(filename);
+    }
+
+    // then, download it if needed
+    if (!model.url.empty()) {
+        bool ok = common_download_model(model, bearer_token);
+        if (!ok) {
+            LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
+            exit(1);
         }
-    } else if (!model_url.empty()) {
-        if (model.empty()) {
-            auto f = string_split<std::string>(model_url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            model = fs_get_cache_file(string_split<std::string>(f, '/').back());
-        }
-    } else if (model.empty()) {
-        model = model_default;
     }
 }
 
@@ -300,10 +827,16 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // TODO: refactor model params in a common struct
-    common_params_handle_model_default(params.model,             params.model_url,             params.hf_repo,             params.hf_file,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
-    common_params_handle_model_default(params.vocoder.model,     params.vocoder.model_url,     params.vocoder.hf_repo,     params.vocoder.hf_file,     params.hf_token, "");
+    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
+    common_params_handle_model(params.speculative.model, params.hf_token, "");
+    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
+
+    // allow --mmproj to be set from -hf
+    // assuming that mmproj is always in the same repo as text model
+    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
+        params.mmproj.hf_repo = params.model.hf_repo;
+    }
+    common_params_handle_model(params.mmproj,            params.hf_token, "", true);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -322,6 +855,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.kv_overrides.back().key[0] = 0;
     }
 
+    if (!params.tensor_buft_overrides.empty()) {
+        params.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
+
     if (params.reranking && params.embedding) {
         throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
     }
@@ -764,7 +1301,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CTX_SIZE"));
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
-        string_format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict),
+        string_format(
+            ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL
+                ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)"
+                : "number of tokens to predict (default: %d, -1 = infinity)",
+            params.n_predict),
         [](common_params & params, int value) {
             params.n_predict = value;
         }
@@ -849,6 +1390,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-sysf", "--system-prompt-file"}, "FNAME",
+        "a file containing the system prompt (default: none)",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
+                params.system_prompt.pop_back();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
         "an input file (repeat to specify multiple files)",
@@ -1543,7 +2098,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--mmproj"}, "FILE",
         "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
-            params.mmproj = value;
+            params.mmproj.path = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    add_opt(common_arg(
+        {"--mmproj-url"}, "URL",
+        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        [](common_params & params, const std::string & value) {
+            params.mmproj.url = value;
         }
     ).set_examples({LLAMA_EXAMPLE_LLAVA}));
     add_opt(common_arg(
@@ -1629,6 +2191,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
+    add_opt(common_arg(
+        {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
+        "override tensor buffer type", [](common_params & params, const std::string & value) {
+            /* 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;
+                    }
+                }
+            }
+
+            for (const auto & override : string_split<std::string>(value, ',')) {
+                std::string::size_type pos = override.find('=');
+                if (pos == std::string::npos) {
+                    throw std::invalid_argument("invalid value");
+                }
+                std::string tensor_name = override.substr(0, pos);
+                std::string buffer_type = override.substr(pos + 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));
+                    }
+                    throw std::invalid_argument("unknown buffer type");
+                }
+                // FIXME: this leaks memory
+                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+            }
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1772,14 +2369,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
             ),
         [](common_params & params, const std::string & value) {
-            params.model = value;
+            params.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.model_url = value;
+            params.model.url = value;
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
@@ -1788,35 +2385,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "example: unsloth/phi-4-GGUF:q4_k_m\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_repo = value;
+            params.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO"));
     add_opt(common_arg(
         {"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
         "Same as --hf-repo, but for the draft model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.hf_repo = value;
+            params.speculative.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HFD_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
         "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_file = value;
+            params.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE"));
     add_opt(common_arg(
         {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
         "Hugging Face model repository for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_repo = value;
+            params.vocoder.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO_V"));
     add_opt(common_arg(
         {"-hffv", "--hf-file-v"}, "FILE",
         "Hugging Face model file for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_file = value;
+            params.vocoder.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE_V"));
     add_opt(common_arg(
@@ -1871,7 +2468,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -1961,7 +2558,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
-        string_format("ip address to listen (default: %s)", params.hostname.c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
         [](common_params & params, const std::string & value) {
             params.hostname = value;
         }
@@ -2436,7 +3033,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.model = value;
+            params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
 
@@ -2444,7 +3041,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-mv", "--model-vocoder"}, "FNAME",
         "vocoder model for audio generation (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.model = value;
+            params.vocoder.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
      add_opt(common_arg(
@@ -2467,10 +3064,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf";
+            params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
+            params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
+            params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
+            params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
@@ -2478,8 +3075,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-bge-small-en-default"},
         string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
-            params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
+            params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2492,8 +3089,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-e5-small-en-default"},
         string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
-            params.hf_file = "e5-small-v2-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
+            params.model.hf_file = "e5-small-v2-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2506,8 +3103,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-gte-small-default"},
         string_format("use default gte-small model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
-            params.hf_file = "gte-small-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
+            params.model.hf_file = "gte-small-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2520,8 +3117,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-1.5b-default"},
         string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2536,8 +3133,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-3b-default"},
         string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2552,8 +3149,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-7b-default"},
         string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2568,10 +3165,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-7b-spec"},
         string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
-            params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.speculative.n_gpu_layers = 99;
             params.port = 8012;
             params.n_gpu_layers = 99;
@@ -2587,10 +3184,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-14b-spec"},
         string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
-            params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.speculative.n_gpu_layers = 99;
             params.port = 8012;
             params.n_gpu_layers = 99;

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()) {
+    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null() && params.tools.is_array() && params.json_schema.is_null()) {
         return common_chat_params_init_hermes_2_pro(tmpl, params);
     }
 

common/common.cpp

@@ -7,9 +7,6 @@
 
 #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>
@@ -51,47 +48,11 @@
 #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
 //
@@ -869,7 +830,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#ifdef __linux__
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -879,7 +840,9 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#endif // __linux__
+#else
+#  error Unknown architecture
+#endif
         cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
     }
@@ -900,22 +863,14 @@ 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 = 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);
-    }
-
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -950,13 +905,13 @@ 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.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
@@ -1033,6 +988,8 @@ struct common_init_result common_init_from_params(common_params & params) {
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
+        llama_set_warmup(lctx, true);
+
         std::vector<llama_token> tmp;
         llama_token bos = llama_vocab_bos(vocab);
         llama_token eos = llama_vocab_eos(vocab);
@@ -1060,9 +1017,10 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
+        llama_set_warmup(lctx, false);
     }
 
     iparams.model.reset(model);
@@ -1071,6 +1029,19 @@ 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) {
@@ -1086,15 +1057,18 @@ 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 {
@@ -1102,6 +1076,13 @@ 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;
 }
 
@@ -1161,451 +1142,6 @@ 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
 //
@@ -2029,26 +1565,3 @@ 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

@@ -121,10 +121,6 @@ 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
@@ -184,6 +180,13 @@ 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
 
@@ -197,19 +200,11 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    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_model model;
 };
 
 struct common_params_vocoder {
-    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
+    struct common_params_model model;
 
     std::string speaker_file = ""; // speaker file path                                      // NOLINT
 
@@ -267,12 +262,10 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    std::string model                = ""; // model path                                                    // NOLINT
+    struct common_params_model model;
+
     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
@@ -286,6 +279,7 @@ 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
@@ -347,7 +341,7 @@ struct common_params {
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    struct common_params_model mmproj;
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -546,26 +540,11 @@ 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
 //

common/llguidance.cpp

@@ -11,25 +11,24 @@ struct llama_sampler_llg {
     std::string         grammar_kind;
     std::string         grammar_data;
     LlgTokenizer *      tokenizer;
-    LlgConstraint *     grammar;
-    LlgMaskResult       llg_res;
-    bool                has_llg_res;
+    LlgMatcher *        grammar;
 };
 
-static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                             const char * grammar_data) {
+static LlgMatcher * 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_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);
+    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);
         return nullptr;
     }
+
     return c;
 }
 
@@ -40,39 +39,29 @@ 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) {
-        LlgCommitResult res;
-        llg_commit_token(ctx->grammar, token, &res);
-        ctx->has_llg_res = false;
+        llg_matcher_consume_token(ctx->grammar, token);
     }
 }
 
 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) {
-        if (!ctx->has_llg_res) {
-            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
-                ctx->has_llg_res = true;
+        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);
             } else {
-                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
-                llg_free_constraint(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
+                llg_free_matcher(ctx->grammar);
                 ctx->grammar = nullptr;
+                return;
             }
         }
-        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;
-                    }
-                }
+
+        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;
             }
         }
     }
@@ -80,14 +69,9 @@ 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) {
-        return;
+    if (ctx->grammar) {
+        llg_matcher_reset(ctx->grammar);
     }
-
-    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) {
@@ -102,7 +86,7 @@ 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_constraint(ctx->grammar);
+            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
             result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
         }
     }
@@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_llg *) smpl->ctx;
 
     if (ctx->grammar) {
-        llg_free_constraint(ctx->grammar);
+        llg_free_matcher(ctx->grammar);
         llg_free_tokenizer(ctx->tokenizer);
     }
 
@@ -239,9 +223,11 @@ 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,
@@ -249,15 +235,12 @@ 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

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

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

@@ -208,6 +208,9 @@ 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

@@ -173,7 +173,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         prompt.clear();
     } else {
@@ -192,14 +192,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -65,6 +65,7 @@ class Model:
     model_name: str | None
     metadata_override: Path | None
     dir_model_card: Path
+    remote_hf_model_id: str | None
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
@@ -73,7 +74,7 @@ class Model:
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
-                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -83,11 +84,24 @@ class Model:
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
         self.use_temp_file = use_temp_file
-        self.lazy = not eager
-        self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
-        self.is_safetensors = len(self.part_names) > 0
-        if not self.is_safetensors:
-            self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
+        self.lazy = not eager or (remote_hf_model_id is not None)
+        self.remote_hf_model_id = remote_hf_model_id
+        if remote_hf_model_id is not None:
+            self.is_safetensors = True
+
+            def get_remote_tensors() -> Iterator[tuple[str, Tensor]]:
+                logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
+                remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
+                self.tensor_names = set(name for name in remote_tensors.keys())
+                for name, remote_tensor in gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id).items():
+                    yield (name, LazyTorchTensor.from_remote_tensor(remote_tensor))
+
+            self.get_tensors = get_remote_tensors
+        else:
+            self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
+            self.is_safetensors = len(self.part_names) > 0
+            if not self.is_safetensors:
+                self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
@@ -180,7 +194,8 @@ class Model:
             extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
             missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
             if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                 f"Missing tensors: {missing}")
             else:
                 raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
                                  f"Missing tensors: {missing}\n"
@@ -392,6 +407,10 @@ class Model:
 
         self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params)
 
+        # If we are using HF model id, set the metadata name to the model id
+        if self.remote_hf_model_id:
+            self.metadata.name = self.remote_hf_model_id
+
         # Fallback to model directory name if metadata name is still missing
         if self.metadata.name is None:
             self.metadata.name = self.dir_model.name
@@ -528,6 +547,8 @@ class Model:
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
 
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
@@ -537,13 +558,13 @@ class Model:
                 if token in added_vocab:
                     # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
+                    if not added_tokens_decoder[i].normalized:
                         previous_token = token
                         token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
-                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
                         # NOTE: this was added for Gemma.
@@ -702,6 +723,21 @@ class Model:
         if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
             # ref: https://huggingface.co/Xenova/gpt-4o
             res = "gpt-4o"
+        if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f":
+            # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k
+            res = "superbpe"
+        if chkhsh == "1994ffd01900cfb37395608534236ecd63f2bd5995d6cb1004dda1af50240f15":
+            # ref: https://huggingface.co/trillionlabs/Trillion-7B-preview
+            res = "trillion"
+        if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224":
+            # ref: https://huggingface.co/inclusionAI/Ling-lite
+            res = "bailingmoe"
+        if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406":
+            # ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
+            res = "llama4"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
 
         if res is None:
             logger.warning("\n")
@@ -908,6 +944,40 @@ class Model:
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_rwkv_world(self):
+        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
+        vocab_size = self.hparams.get("vocab_size", 65536)
+
+        tokens: list[bytes] = ['<s>'.encode("utf-8")]
+        toktypes: list[int] = [gguf.TokenType.CONTROL]
+
+        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
+            lines = f.readlines()
+            for line in lines:
+                parts = line.split(' ')
+                assert len(parts) >= 3
+                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
+                token = token.encode("utf-8") if isinstance(token, str) else token
+                assert isinstance(token, bytes)
+                assert len(token) == token_len
+                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
+                tokens.append(token_text.encode("utf-8"))
+                toktypes.append(gguf.TokenType.NORMAL)
+        remainder = vocab_size - len(tokens)
+        assert remainder >= 0
+        for i in range(len(tokens), vocab_size):
+            tokens.append(f"[PAD{i}]".encode("utf-8"))
+            toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("rwkv")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int):
         tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf"
         logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
@@ -1065,13 +1135,6 @@ class BloomModel(Model):
 
         tensors.append((self.map_tensor_name(name), data_torch))
 
-        if name == "word_embeddings.weight":
-            assert self.tensor_names is not None
-
-            # TODO: tie them at runtime, don't duplicate in the model file
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -1569,6 +1632,7 @@ class StableLMModel(Model):
 @Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
 class LlamaModel(Model):
     model_arch = gguf.MODEL_ARCH.LLAMA
+    undo_permute = True
 
     def set_vocab(self):
         try:
@@ -1633,10 +1697,11 @@ class LlamaModel(Model):
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
 
-        if name.endswith(("q_proj.weight", "q_proj.bias")):
-            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
-        if name.endswith(("k_proj.weight", "k_proj.bias")):
-            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        if self.undo_permute:
+            if name.endswith(("q_proj.weight", "q_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+            if name.endswith(("k_proj.weight", "k_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
 
         # process the experts separately
         if name.find("block_sparse_moe.experts") != -1:
@@ -1688,7 +1753,7 @@ class LlamaModel(Model):
 
                 low_freq_wavelen = old_context_len / low_freq_factor
                 high_freq_wavelen = old_context_len / high_freq_factor
-                assert low_freq_wavelen != high_freq_wavelen
+                # assert low_freq_wavelen != high_freq_wavelen # Errors for Llama4
 
                 rope_factors = []
                 for freq in freqs:
@@ -1713,6 +1778,76 @@ class LlamaModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("Llama4ForConditionalGeneration")
+class Llama4Model(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA4
+    has_vision: bool = False
+    undo_permute = False
+
+    # TODO @ngxson : avoid duplicate this code everywhere by at least support "text_config"
+    # same with llama, but we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+        if "vision_config" in hparams:
+            logger.info("Has vision encoder, but it will be ignored")
+            self.has_vision = True
+        # IMPORTANT: the normal "intermediate_size" is renamed to "intermediate_size_mlp", we need to undo this
+        self.hparams["intermediate_size_moe"] = self.hparams["intermediate_size"]
+        self.hparams["intermediate_size"] = self.hparams["intermediate_size_mlp"]
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+        self.gguf_writer.add_add_bos_token(True)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_interleave_moe_layer_step(self.hparams["interleave_moe_layer_step"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # split the gate_up into gate and up
+        if "gate_up_proj" in name:
+            name_up = name.replace("gate_up_proj", "up_proj.weight")
+            name_gate = name.replace("gate_up_proj", "gate_proj.weight")
+            dim_half = data_torch.shape[-1] // 2
+            gate_proj_weight, up_proj_weight = data_torch.transpose(-1, -2).split(dim_half, dim=-2)
+            return [
+                (self.map_tensor_name(name_gate), gate_proj_weight),
+                (self.map_tensor_name(name_up), up_proj_weight)
+            ]
+
+        if name.endswith("down_proj"):
+            name += ".weight"
+            data_torch = data_torch.transpose(-1, -2)
+
+        if "multi_modal_projector" in name or "vision_model" in name:
+            return []
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@Model.register("Mistral3ForConditionalGeneration")
+class Mistral3Model(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        name = name.replace("language_model.", "")
+        if "multi_modal_projector" in name or "vision_tower" in name:
+            return []
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("DeciLMForCausalLM")
 class DeciModel(Model):
     model_arch = gguf.MODEL_ARCH.DECI
@@ -2217,7 +2352,7 @@ class Qwen2Model(Model):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
-@Model.register("Qwen2VLForConditionalGeneration")
+@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2VL
 
@@ -2341,6 +2476,16 @@ class Qwen2MoeModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("Qwen3ForCausalLM")
+class Qwen3Model(Qwen2Model):
+    model_arch = gguf.MODEL_ARCH.QWEN3
+
+
+@Model.register("Qwen3MoeForCausalLM")
+class Qwen3MoeModel(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3MOE
+
+
 @Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
     model_arch = gguf.MODEL_ARCH.GPT2
@@ -2370,10 +2515,6 @@ class GPT2Model(Model):
 
         tensors.append((new_name, data_torch))
 
-        # note: GPT2 output is tied to (same as) wte in original model
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -2703,21 +2844,26 @@ class CodeShellModel(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
+    _has_tok_embd = False
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
         new_name = self.map_tensor_name(name)
 
-        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+        # assuming token_embd.weight is seen before output.weight
+        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
+            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
+                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
+                self.tensor_names.remove("transformer.wte.weight")
+        elif new_name == tok_embd_name:
+            self._has_tok_embd = True
 
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            assert self.tensor_names is not None
-
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                # copy tok_embd.weight to output.weight
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
-        return tensors
+        return [(new_name, data_torch)]
 
 
 @Model.register("InternLM2ForCausalLM")
@@ -3332,7 +3478,7 @@ class Gemma3Model(Model):
 
     # we need to merge the text_config into the root level of hparams
     def __init__(self, *args, **kwargs):
-        hparams = Model.load_hparams(kwargs["dir_model"])
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
         if "text_config" in hparams:
             hparams = {**hparams, **hparams["text_config"]}
             kwargs["hparams"] = hparams
@@ -3412,38 +3558,7 @@ class Rwkv6Model(Model):
     model_arch = gguf.MODEL_ARCH.RWKV6
 
     def set_vocab(self):
-        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
-        vocab_size = self.hparams.get("vocab_size", 65536)
-
-        tokens: list[bytes] = ['<s>'.encode("utf-8")]
-        toktypes: list[int] = [gguf.TokenType.CONTROL]
-
-        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
-            lines = f.readlines()
-            for line in lines:
-                parts = line.split(' ')
-                assert len(parts) >= 3
-                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
-                token = token.encode("utf-8") if isinstance(token, str) else token
-                assert isinstance(token, bytes)
-                assert len(token) == token_len
-                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
-                tokens.append(token_text.encode("utf-8"))
-                toktypes.append(gguf.TokenType.NORMAL)
-        remainder = vocab_size - len(tokens)
-        assert remainder >= 0
-        for i in range(len(tokens), vocab_size):
-            tokens.append(f"[PAD{i}]".encode("utf-8"))
-            toktypes.append(gguf.TokenType.UNUSED)
-
-        self.gguf_writer.add_tokenizer_model("rwkv")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
-        # hack: Add '\n\n' as the EOT token to make it chat normally
-        special_vocab._set_special_token("eot", 261)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        self._set_vocab_rwkv_world()
 
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
@@ -3529,8 +3644,8 @@ class RWKV6Qwen2Model(Rwkv6Model):
         head_size = hidden_size // num_attention_heads
         rms_norm_eps = self.hparams["rms_norm_eps"]
         intermediate_size = self.hparams["intermediate_size"]
-        time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
-        time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
+        time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32)
+        time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64)
 
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
@@ -3565,6 +3680,168 @@ class RWKV6Qwen2Model(Rwkv6Model):
             yield (new_name, data)
 
 
+@Model.register("Rwkv7ForCausalLM", "RWKV7ForCausalLM")
+class Rwkv7Model(Model):
+    model_arch = gguf.MODEL_ARCH.RWKV7
+
+    def set_vocab(self):
+        self._set_vocab_rwkv_world()
+
+    def calc_lora_rank(self, hidden_size, exponent, multiplier):
+        return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        try:
+            head_size = self.hparams["head_size"]
+            layer_norm_eps = self.hparams["layer_norm_epsilon"]
+        except KeyError:
+            head_size = self.hparams["head_dim"]
+            layer_norm_eps = self.hparams["norm_eps"]
+        hidden_size = self.hparams["hidden_size"]
+        intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else (hidden_size * 4)
+
+        # ICLR: In-Context-Learning-Rate
+        try:
+            lora_rank_decay = self.hparams["lora_rank_decay"] if self.hparams["lora_rank_decay"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["lora_rank_iclr"] if self.hparams["lora_rank_iclr"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["lora_rank_value_residual_mix"] if self.hparams["lora_rank_value_residual_mix"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["lora_rank_gate"] if self.hparams["lora_rank_gate"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+        except KeyError:
+            lora_rank_decay = self.hparams["decay_low_rank_dim"] if self.hparams["decay_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["a_low_rank_dim"] if self.hparams["a_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["v_low_rank_dim"] if self.hparams["v_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["gate_low_rank_dim"] if self.hparams["gate_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    lerp_weights: dict[int, dict[str, Tensor]] = {}
+    lora_needs_transpose: bool = True
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # unify tensor names here to make life easier
+        name = name.replace("blocks", "layers").replace("ffn", "feed_forward")
+        name = name.replace("self_attn", "attention").replace("attn", "attention")
+        name = name.replace("time_mixer.", "")
+        # lora layer names in fla-hub's impl
+        if "_lora.lora" in name:
+            self.lora_needs_transpose = False
+        name = name.replace("_lora.lora.0.weight", "1.weight")
+        name = name.replace("_lora.lora.2.weight", "2.weight")
+        name = name.replace("_lora.lora.2.bias", "0.weight")
+
+        name = name.replace("feed_forward_norm", "ln2")
+        name = name.replace("g_norm", "ln_x")
+
+        if "attention.v" in name and "value" not in self.map_tensor_name(name) and bid == 0:
+            # some models have dummy v0/v1/v2 on first layer while others don't
+            # ignore them all since they are not used
+            return
+
+        wkv_has_gate = self.hparams.get("wkv_has_gate", True)
+        lerp_list = ["r", "w", "k", "v", "a", "g"] if wkv_has_gate else ["r", "w", "k", "v", "a"]
+
+        if bid is not None and "attention.x_" in name:
+            if "attention.x_x" in name:
+                # already concatenated
+                new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                data = data_torch.reshape(len(lerp_list), 1, 1, -1)
+                yield (new_name, data)
+            else:
+                try:
+                    self.lerp_weights[bid][name] = data_torch
+                except KeyError:
+                    self.lerp_weights[bid] = {name: data_torch}
+                if all(f"model.layers.{bid}.attention.x_{i}" in self.lerp_weights[bid].keys() for i in lerp_list):
+                    new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                    data = torch.stack([self.lerp_weights[bid][f"model.layers.{bid}.attention.x_{i}"] for i in lerp_list], dim=0)
+                    yield (new_name, data)
+            return
+        else:
+            data_torch = data_torch.squeeze()
+            new_name = self.map_tensor_name(name)
+
+            if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
+                new_name += ".weight"
+
+            if self.lora_needs_transpose and any(
+                new_name.endswith(t) for t in [
+                    "time_mix_w1.weight", "time_mix_w2.weight",
+                    "time_mix_a1.weight", "time_mix_a2.weight",
+                    "time_mix_v1.weight", "time_mix_v2.weight",
+                    "time_mix_g1.weight", "time_mix_g2.weight",
+                ]
+            ):
+                data_torch = data_torch.transpose(0, 1)
+
+            if 'r_k' in new_name:
+                data_torch = data_torch.flatten()
+
+            if bid == 0 and "time_mix_a" in new_name:
+                # dummy v0/v1/v2 on first layer
+                # easist way to make llama happy
+                yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch)
+
+            yield (new_name, data_torch)
+
+
+@Model.register("RwkvHybridForCausalLM")
+class ARwkv7Model(Rwkv7Model):
+    model_arch = gguf.MODEL_ARCH.ARWKV7
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        hidden_size = self.hparams["hidden_size"]
+        head_size = self.hparams["head_size"]
+        rms_norm_eps = self.hparams["rms_norm_eps"]
+        intermediate_size = self.hparams["intermediate_size"]
+        wkv_has_gate = self.hparams["wkv_has_gate"]
+        assert self.hparams["wkv_version"] == 7
+
+        # ICLR: In-Context-Learning-Rate
+        lora_rank_decay = 64
+        lora_rank_iclr = 64
+        lora_rank_value_residual_mix = 32
+        lora_rank_gate = 128 if wkv_has_gate else 0
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_token_shift_count(1)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+
 @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
@@ -3619,8 +3896,6 @@ class MambaModel(Model):
     _tok_embd = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
         tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
 
@@ -3630,6 +3905,10 @@ class MambaModel(Model):
             logger.debug("A_log --> A ==> " + new_name)
             data_torch = -torch.exp(data_torch)
 
+        # [4 1 8192 1] -> [4 8192 1 1]
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
         # assuming token_embd.weight is seen before output.weight
         if self._tok_embd is not None and new_name == output_name:
             if torch.equal(self._tok_embd, data_torch):
@@ -4143,6 +4422,10 @@ class DeepseekV2Model(Model):
         self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
+
+        # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
+        self.hparams["num_key_value_heads"] = 1
+
         super().set_gguf_parameters()
         hparams = self.hparams
 
@@ -4151,8 +4434,13 @@ class DeepseekV2Model(Model):
         if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
             self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
         self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
-        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
-        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+
+        # note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+        self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
+
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
@@ -4221,6 +4509,26 @@ class DeepseekV2Model(Model):
             else:
                 return []
 
+        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
+        if name.endswith("kv_b_proj.weight"):
+            name_kb = name.replace("kv_b_proj", "k_b_proj")
+            name_vb = name.replace("kv_b_proj", "v_b_proj")
+
+            n_head_kv = self.hparams["num_key_value_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+
+            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
+
+            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            return [
+                (self.map_tensor_name(name_kb), k_b),
+                (self.map_tensor_name(name_vb), v_b)
+            ]
+
         return [(self.map_tensor_name(name), data_torch)]
 
     def prepare_tensors(self):
@@ -4233,6 +4541,29 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("PLMForCausalLM")
+class PLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLM
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+
 @Model.register("T5WithLMHeadModel")
 @Model.register("T5ForConditionalGeneration")
 @Model.register("MT5ForConditionalGeneration")
@@ -4598,6 +4929,22 @@ class JaisModel(Model):
         self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
 
 
+@Model.register("Glm4ForCausalLM")
+class Glm4Model(Model):
+    model_arch = gguf.MODEL_ARCH.GLM4
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+
+
 @Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(Model):
     model_arch = gguf.MODEL_ARCH.CHATGLM
@@ -4921,6 +5268,105 @@ class GraniteMoeModel(GraniteModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@Model.register("BailingMoeForCausalLM")
+class BailingMoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
+
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_weights_scale(1.0)
+        self.gguf_writer.add_expert_count(hparams["num_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
+        if n_head_kv is not None and n_head != n_head_kv:
+            n_head = n_head_kv
+        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_embd = self.hparams["hidden_size"]
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
+
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+
+        if name.endswith("attention.dense.weight"):
+            return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)]
+        elif name.endswith("query_key_value.weight"):
+            q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2)
+
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+            ]
+        elif name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            tensors: list[tuple[str, Tensor]] = []
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+
+            return tensors
+
+        new_name = self.map_tensor_name(name)
+
+        if new_name == output_name and self.hparams.get("norm_head"):
+            data_torch = data_torch.float()
+            data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("ChameleonForConditionalGeneration")
 @Model.register("ChameleonForCausalLM")  # obsolete
 class ChameleonModel(Model):
@@ -5019,6 +5465,14 @@ class LazyTorchTensor(gguf.LazyBase):
         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 from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor):
+        dtype = cls._dtype_str_map[remote_tensor.dtype]
+        shape = remote_tensor.shape
+        meta = cls.meta_with_dtype_and_shape(dtype, shape)
+        lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=dtype).reshape(shape))
+        return cast(torch.Tensor, lazy)
+
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):
         del types  # unused
@@ -5096,6 +5550,10 @@ def parse_args() -> argparse.Namespace:
         "--print-supported-models", action="store_true",
         help="Print the supported models"
     )
+    parser.add_argument(
+        "--remote", action="store_true",
+        help="(Experimental) Read safetensors file remotely without downloading to disk. Config and tokenizer files will still be downloaded. To use this feature, you need to specify Hugging Face model repo name instead of a local directory. For example: 'HuggingFaceTB/SmolLM2-1.7B-Instruct'. Note: To access gated repo, set HF_TOKEN environment variable to your Hugging Face token.",
+    )
 
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
@@ -5136,6 +5594,14 @@ def main() -> None:
 
     dir_model = args.model
 
+    if args.remote:
+        from huggingface_hub import snapshot_download
+        local_dir = snapshot_download(
+            repo_id=str(dir_model),
+            allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
+        dir_model = Path(local_dir)
+        logger.info(f"Downloaded config and tokenizer to {local_dir}")
+
     if not dir_model.is_dir():
         logger.error(f'Error: {args.model} is not a directory')
         sys.exit(1)
@@ -5157,6 +5623,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
+    elif args.remote:
+        # if remote, use the model ID as the output file name
+        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -5167,20 +5636,20 @@ def main() -> None:
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
         model_architecture = hparams["architectures"][0]
-
         try:
             model_class = Model.from_model_architecture(model_architecture)
         except NotImplementedError:
             logger.error(f"Model {model_architecture} is not supported")
             sys.exit(1)
 
-        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+        model_instance = model_class(dir_model, output_type, fname_out,
                                      is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
                                      eager=args.no_lazy,
                                      metadata_override=args.metadata, model_name=args.model_name,
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
-                                     small_first_shard=args.no_tensor_first_split)
+                                     small_first_shard=args.no_tensor_first_split,
+                                     remote_hf_model_id=str(args.model) if args.remote else None)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")

convert_hf_to_gguf_update.py

@@ -110,6 +110,11 @@ 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", },
 ]
 
 

docs/backend/CUDA-FEDORA.md

@@ -14,9 +14,7 @@ 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-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 Nvidia Driver Libraries](#installing-nvidia-driver-libraries)
 - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
 - [Configuring the Environment](#configuring-the-environment)
 - [Verifying the Installation](#verifying-the-installation)
@@ -67,7 +65,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for
    sudo dnf distro-sync
    ```
 
-2. **Install the Default Text Editor (Optional):**
+2. **Install **Vim** the default text editor (Optional):**
 
    ```bash
    sudo dnf install vim-default-editor --allowerasing
@@ -97,36 +95,48 @@ After adding the repository, synchronize the package manager again:
 sudo dnf distro-sync
 ```
 
-## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+## Installing Nvidia Driver Libraries
 
-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).
+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):
 
 ```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:**
 
-- `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).
+#### Install the Nvidia Driver Libraries on Guest:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found).
+### If *`libcuda.so.1`* exists:
+```
+lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07
+```
 
-If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
+**Explanation:**
+The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:**
 
-#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies)
+#### 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)
 
 ```bash
-sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-#### 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/*
@@ -134,23 +144,26 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
 
 **Note:**
 
-- The `--justdb` option only updates the RPM database, without touching the filesystem.
+- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere.
 
-#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+##### 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.
 
-After manually installing the dependencies, run:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-You should receive a message indicating the package is already installed:
+*(this time it will not install anything, as the database things that these packages are already installed)*
 
 ```
 Updating and loading repositories:
 Repositories loaded.
-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.
+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.
 
 Nothing to do.
 ```
@@ -207,9 +220,9 @@ You should see output similar to:
 ```
 nvcc: NVIDIA (R) Cuda compiler driver
 Copyright (c) 2005-2025 NVIDIA Corporation
-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
+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
 ```
 
 This output confirms that the CUDA compiler is accessible and indicates the installed version.

docs/backend/OPENCL.md

@@ -145,8 +145,13 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
 * Clang 19
 * Ninja
 * Visual Studio 2022
+* Powershell 7
 
-Powershell is used for the following instructions.
+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.
 
 ### I. Setup Environment
 
@@ -196,10 +201,9 @@ ninja
 
 ## Known Issues
 
-- Qwen2.5 0.5B model produces gibberish output with Adreno kernels.
+- Currently OpenCL backend does not work on Adreno 6xx GPUs.
 
 ## 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. oneMKL and oneDNN)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
@@ -227,30 +227,19 @@ 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.
 
-
-**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.
+**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:
 
 ```sh
-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
+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
 ```
 
 - **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.
@@ -313,37 +302,39 @@ 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
 
-```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
+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
 # 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=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${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
 
 # 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
 
-```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
+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
 # Build LLAMA with rocBLAS acceleration through SYCL
 
 ## AMD
@@ -434,13 +425,13 @@ Examples:
 - Use device 0:
 
 ```sh
-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
+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
 ```
 
 - Use multiple devices:
 
 ```sh
-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
+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
 ```
 
 *Notes:*
@@ -484,6 +475,12 @@ 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:
@@ -514,13 +511,13 @@ 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:
 
@@ -549,13 +546,84 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-3. Visual Studio
+#### 3. Visual Studio
 
-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.
+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`
 
 *Notes:*
+- For a minimal experimental setup, you can build only the inference executable using:
 
-- 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`.
+    ```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.
 
 ### III. Run the inference
 
@@ -629,13 +697,13 @@ Examples:
 - Use device 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
+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
 ```
 
 - Use multiple devices:
 
 ```
-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
+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
 ```
 
 
@@ -660,8 +728,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
-| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)          | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
+| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
@@ -671,6 +740,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 

docs/build.md

@@ -132,12 +132,14 @@ 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](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
+We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
 
 **Recommended for:**
-
-- ***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).
+- ***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.
 - *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)
 
 
@@ -189,7 +191,7 @@ 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, 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, 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_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.                                                                         |
@@ -216,6 +218,7 @@ 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.
@@ -256,8 +259,6 @@ 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.
 
@@ -293,6 +294,10 @@ 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**
@@ -433,6 +438,116 @@ 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/install.md

@@ -9,6 +9,13 @@ 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

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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@@ -132,7 +132,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_clear(ctx);
+                llama_kv_self_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +141,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

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

examples/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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n" , __func__);

examples/cvector-generator/cvector-generator.cpp

@@ -342,7 +342,7 @@ 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_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;

examples/embedding/embedding.cpp

@@ -38,7 +38,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const struct llama_model * model = llama_get_model(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

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, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
+        lora_merge_ctx ctx(params.model.path, 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/gguf-split/gguf-split.cpp

@@ -408,8 +408,6 @@ 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();
 
@@ -453,7 +451,6 @@ 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);
             }
 
@@ -466,7 +463,6 @@ 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);
             }
 
@@ -479,7 +475,6 @@ 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);
             }
 
@@ -500,9 +495,11 @@ static void gguf_merge(const split_params & split_params) {
 
         fprintf(stderr, "\033[3Ddone\n");
     }
-
-    // placeholder for the meta data
-    {
+    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
         auto meta_size = gguf_get_meta_size(ctx_out);
         ::zeros(fout, meta_size);
     }
@@ -518,7 +515,9 @@ static void gguf_merge(const split_params & split_params) {
                 ggml_free(ctx_metas[i]);
             }
             gguf_free(ctx_out);
-            fout.close();
+            if (!split_params.dry_run) {
+                fout.close();
+            }
             exit(EXIT_FAILURE);
         }
         fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
@@ -540,10 +539,11 @@ 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);
-
-            // write tensor data + padding
-            fout.write((const char *)read_data.data(), n_bytes);
-            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - 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);
+            }
         }
 
         gguf_free(ctx_gguf);
@@ -552,16 +552,15 @@ 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

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
@@ -168,7 +168,7 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
 
     // create generation context
     llama_context * ctx = llama_init_from_model(model, cparams);

examples/imatrix/imatrix.cpp

@@ -495,7 +495,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

examples/infill/infill.cpp

@@ -332,8 +332,8 @@ int main(int argc, char ** argv) {
                 LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/llama-bench/llama-bench.cpp

@@ -1578,7 +1578,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // cool off before the test
         if (params.delay) {
@@ -1618,7 +1618,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_self_clear(ctx);
 
             uint64_t t_start = get_time_ns();
 

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

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

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

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

examples/llama.swiftui/README.md

@@ -16,7 +16,7 @@ Open `llama.swiftui.xcodeproj` project in Xcode and you should be able to build
 a simulator or a real device.
 
 To use the framework with a different project, the XCFramework can be added to the project by
-adding `build-ios/llama.xcframework` by dragging and dropping it into the project navigator, or
+adding `build-apple/llama.xcframework` by dragging and dropping it into the project navigator, or
 by manually selecting the framework in the "Frameworks, Libraries, and Embedded Content" section
 of the project settings.
 

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

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

examples/llava/CMakeLists.txt

@@ -1,3 +1,5 @@
+# llava (legacy)
+
 add_library(llava OBJECT
             llava.cpp
             llava.h
@@ -22,12 +24,41 @@ 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()
 
 set(TARGET llama-llava-cli)
@@ -55,7 +86,7 @@ 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 llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-llava-clip-quantize-cli)

examples/llava/README-gemma3.md

@@ -4,6 +4,26 @@
 >
 > 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-gemma3-cli
+
+# alternatively, install from brew (MacOS)
+brew install llama.cpp
+
+# run it
+llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# note: 1B model does not support vision
+```
+
 ## How to get mmproj.gguf?
 
 ```bash

examples/llava/clip-impl.h

@@ -0,0 +1,344 @@
+#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_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#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_PROJ_TYPE           "clip.projector_type"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+
+#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"
+
+
+//
+// tensor name constants
+//
+
+#define TN_TOKEN_EMBD      "%s.token_embd.weight"
+#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_UP          "%s.blk.%d.ffn_up.%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_TEXT_PROJ       "text_projection.weight"
+#define TN_VIS_PROJ        "visual_projection.weight"
+#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
+
+// 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"
+#define TN_GLM_BOI_W            "adapter.boi"
+#define TN_GLM_EOI_W            "adapter.eoi"
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_GLM_EDGE,
+    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_GEMMA3,
+    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_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
+    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+};
+
+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

@@ -1,6 +1,7 @@
 #ifndef CLIP_H
 #define CLIP_H
 
+#include "ggml.h"
 #include <stddef.h>
 #include <stdint.h>
 
@@ -29,19 +30,12 @@ struct clip_image_size {
     int height;
 };
 
-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_image_u8_batch;
+struct clip_image_f32_batch;
 
 struct clip_context_params {
     bool use_gpu;
-    int verbosity;
+    ggml_log_level verbosity;
 };
 
 // deprecated, use clip_init
@@ -54,9 +48,9 @@ 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_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);
+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);
 
 // TODO: should be enum, not string
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
@@ -72,15 +66,26 @@ 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_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
 
+// 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 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
@@ -105,6 +110,8 @@ 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);
 

examples/llava/gemma3-cli.cpp

@@ -2,15 +2,15 @@
 #include "log.h"
 #include "common.h"
 #include "sampling.h"
-#include "clip.h"
-#include "stb_image.h"
 #include "llama.h"
 #include "ggml.h"
 #include "console.h"
+#include "chat.h"
+#include "mtmd.h"
 
 #include <vector>
 #include <limits.h>
-#include <inttypes.h>
+#include <cinttypes>
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
 #include <signal.h>
@@ -57,13 +57,18 @@ static void sigint_handler(int signo) {
 #endif
 
 struct gemma3_context {
-    struct clip_ctx    * ctx_clip = NULL;
-    common_init_result   llama_init;
+    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;
 
     int n_threads    = 1;
     llama_pos n_past = 0;
@@ -74,16 +79,23 @@ struct gemma3_context {
         vocab = llama_model_get_vocab(model);
         n_threads = params.cpuparams.n_threads;
         batch = llama_batch_init(params.n_batch, 0, 1);
-        init_clip_model(params);
+        n_batch = params.n_batch;
+        tmpls = common_chat_templates_init(model, params.chat_template);
+        init_vision_context(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);
+    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 */   true,
+            /* timings */   true,
+            /* n_threads */ params.cpuparams.n_threads,
+            /* verbosity */ GGML_LOG_LEVEL_INFO,
+        }));
+        if (!ctx_vision.get()) {
+            LOG_ERR("Failed to load vision model from %s\n", clip_path);
+            exit(1);
+        }
     }
 };
 
@@ -120,77 +132,6 @@ struct decode_embd_batch {
     }
 };
 
-static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
-    llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
-    common_batch_clear(ctx.batch);
-    for (llama_token & t : tokens) {
-        common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false);
-    }
-    if (logits_last) {
-        ctx.batch.logits[ctx.batch.n_tokens - 1] = true;
-    }
-    // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
-    if (llama_decode(ctx.lctx, ctx.batch)) {
-        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);
-    decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0);
-    if (llama_decode(ctx.lctx, batch_img.batch)) {
-        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) {
@@ -220,6 +161,45 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
     return 0;
 }
 
+static int eval_message(gemma3_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_ptr chunks(mtmd_tokenize(ctx.ctx_vision.get(), text, bitmaps));
+    if (chunks == nullptr) {
+        LOG_ERR("Unable to tokenize prompt\n");
+        return 1;
+    }
+
+    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks.get(), 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.get());
+
+    return 0;
+}
+
 int main(int argc, char ** argv) {
     ggml_time_init();
 
@@ -232,13 +212,13 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty()) {
+    if (params.mmproj.path.empty()) {
         show_additional_info(argc, argv);
         return 1;
     }
 
     gemma3_context ctx(params);
-    printf("%s: %s\n", __func__, params.model.c_str());
+    printf("%s: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
 
@@ -261,21 +241,15 @@ int main(int argc, char ** argv) {
 #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;
+        if (params.prompt.find("<__image__>") == std::string::npos) {
+            params.prompt += " <__image__>";
         }
-        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)) {
+        common_chat_msg msg;
+        msg.role = "user";
+        msg.content = params.prompt;
+        if (eval_message(ctx, msg, params.image, true)) {
             return 1;
         }
         if (generate_response(ctx, smpl, n_predict)) {
@@ -289,9 +263,9 @@ int main(int argc, char ** argv) {
         LOG("\n   /quit or /exit   exit the program");
         LOG("\n");
 
-        if (eval_text(ctx, "<start_of_turn>user\n")) {
-            return 1;
-        }
+        bool is_first_msg = true;
+        std::vector<std::string> images_fname;
+        std::string content;
 
         while (true) {
             g_is_generating = false;
@@ -309,33 +283,40 @@ int main(int argc, char ** argv) {
             }
             if (line == "/clear") {
                 ctx.n_past = 0;
-                llama_kv_cache_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                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;
-                }
+                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 (ret == 2) {
+                // non-fatal error
+                images_fname.clear();
+                content.clear();
                 continue;
             }
-            if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
+            if (ret) {
                 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;
-            }
+            images_fname.clear();
+            content.clear();
+            is_first_msg = false;
         }
     }
-
+    llama_perf_context_print(ctx.lctx);
     return 0;
 }

examples/llava/llava-cli.cpp

@@ -225,7 +225,7 @@ 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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -234,14 +234,14 @@ 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.c_str();
+    const char * clip_path = params->mmproj.path.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);
+    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
 
     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
@@ -283,7 +283,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);
         return 1;
     }

examples/llava/llava.cpp

@@ -10,6 +10,7 @@
 #include <cstring>
 #include <limits>
 #include <vector>
+#include <memory>
 
 #if defined(LLAVA_LOG_OFF)
 #   define LOG_INF(...)
@@ -45,6 +46,17 @@ 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.
  *
@@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
         struct ggml_context * ctx;
     } model;
 
-    const int32_t image_size = clip_image_size(ctx_clip);
-    const int32_t patch_size = clip_patch_size(ctx_clip);
+    const int32_t image_size = clip_get_image_size(ctx_clip);
+    const int32_t patch_size = clip_get_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)
 
@@ -246,12 +258,9 @@ 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 img_res_v;
-    img_res_v.size = 0;
-    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
         LOG_ERR("%s: unable to preprocess image\n", __func__);
-        delete[] img_res_v.data;
         return false;
     }
 
@@ -259,66 +268,72 @@ 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(img_res_v.size);
-        struct clip_image_size * load_image_size = clip_image_size_init();
+        image_embd_v.resize(n_imgs);
+        clip_image_size load_image_size;
 
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        for (size_t i = 0; i < n_imgs; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            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);
+            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);
 
             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_v.data[i], image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
             }
             else {
-                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, 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) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 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)img_res_v.size, (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)n_imgs, (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)img_res_v.size, (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)n_imgs, (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, 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]);
+                clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
         }
         *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);
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
+        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);
     }
     else if (clip_is_glm(ctx_clip)){
         struct clip_image_size * load_image_size = clip_image_size_init();
-        load_image_size->width = img_res_v.data[0].nx;
-        load_image_size->height = img_res_v.data[0].ny;
+        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);
         clip_add_load_image_size(ctx_clip, load_image_size);
 
-        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);
+        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);
         *n_img_pos = (pos * pos + 2);
         if (!encoded){
             LOG_ERR("Unable to encode image \n");
@@ -328,8 +343,8 @@ 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);
-        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;
+        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
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
 
@@ -340,17 +355,18 @@ 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(img_res_v.size);
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        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[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_v.data[i], 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, 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) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 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)img_res_v.size, (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)n_imgs, (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);
@@ -360,12 +376,7 @@ 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]});
         }
 
-        // 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);
+        const int32_t image_size = clip_get_image_size(ctx_clip);
 
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 

examples/llava/minicpmv-cli.cpp

@@ -31,7 +31,7 @@ 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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -80,7 +80,7 @@ static void llava_free(struct llava_context * ctx_llava) {
 }
 
 static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.c_str();
+    const char * clip_path = params->mmproj.path.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
@@ -88,7 +88,7 @@ static struct clip_ctx * clip_init_context(common_params * params) {
     }
     struct clip_context_params clip_params = {
         /* use_gpu */   params->n_gpu_layers != 0,
-        /* verbosity */ params->verbosity,
+        /* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
     };
     auto * ctx_clip = clip_init(clip_path, clip_params);
     return ctx_clip;
@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty())) {
+    if (params.mmproj.path.empty() || (params.image.empty())) {
         show_additional_info(argc, argv);
         return 1;
     }

examples/llava/mtmd.cpp

@@ -0,0 +1,341 @@
+#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>
+
+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;
+
+    // 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;
+    }
+
+    ~mtmd_context() {
+        clip_free(ctx_clip);
+    }
+};
+
+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
+};
+
+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;
+}
+
+mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
+                                const mtmd_input_text & text,
+                                const std::vector<mtmd_bitmap> & bitmaps) {
+    mtmd_input_chunks * output = new mtmd_input_chunks;
+    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 (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        // <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);
+    }
+
+    std::vector<std::string> parts = string_split_str(text.text, ctx->image_marker);
+    output->clear();
+    output->reserve(parts.size());
+
+    size_t i_img = 0;
+
+    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 nullptr;
+            }
+
+            // shim layer
+            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);
+
+            // 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 nullptr;
+            }
+
+            mtmd_image_tokens * image_tokens = new mtmd_image_tokens;
+            image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image
+            image_tokens->ny = 1; // TODO
+            image_tokens->batch_f32 = std::move(batch_f32);
+
+            mtmd_input_chunk chunk{
+                MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                {},
+                image_tokens,
+            };
+            output->emplace_back(std::move(chunk));
+            i_img++;
+        }
+    }
+
+    return output;
+}
+
+void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
+    for (auto & chunk : *chunks) {
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE && chunk.tokens_image) {
+            delete chunk.tokens_image;
+        }
+    }
+    delete chunks;
+}
+
+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 = 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);
+
+    for (auto & chunk : *chunks) {
+        bool is_last = &chunk == &chunks->back();
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            // TODO @ngxson : may need to split into smaller batches
+            text_batch.n_tokens = chunk.tokens_text.size();
+            for (size_t i = 0; i < chunk.tokens_text.size(); 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...\n");
+            }
+            ret = mtmd_encode(ctx, chunk.tokens_image);
+            if (ret != 0) {
+                LOG_ERR("failed to encode image\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            if (ctx->print_timings) {
+                LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+            }
+
+            int32_t n_tokens = chunk.tokens_image->n_tokens();
+            float * embd = mtmd_get_output_embd(ctx);
+            decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
+            int64_t t1 = ggml_time_ms();
+            ret = llama_decode(lctx, batch_img.batch);
+            if (ret != 0) {
+                LOG_ERR("failed to decode image\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            if (ctx->print_timings) {
+                LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
+            }
+
+            n_past += n_tokens;
+
+        } 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;
+}

examples/llava/mtmd.h

@@ -0,0 +1,146 @@
+#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;
+};
+
+struct mtmd_input_chunk {
+    mtmd_input_chunk_type type;
+    std::vector<llama_token> tokens_text;
+    mtmd_image_tokens * tokens_image = nullptr;
+};
+
+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)
+MTMD_API mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
+                                const mtmd_input_text & text,
+                                const std::vector<mtmd_bitmap> & bitmaps);
+
+// free image chunk data
+MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
+
+// 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);
+
+//
+// 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>;
+
+struct mtmd_input_chunks_deleter {
+    void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
+};
+using mtmd_input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
+
+#else
+
+static_assert(false && "C header is not yet supported by this library");
+
+#endif
+
+#endif

examples/llava/qwen2vl-cli.cpp

@@ -314,7 +314,7 @@ 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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -323,14 +323,14 @@ 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.c_str();
+    const char * clip_path = params->mmproj.path.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);
+    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
 
     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
@@ -524,7 +524,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);
         return 1;
     }

examples/llava/tests.sh

@@ -0,0 +1,81 @@
+#!/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
+
+###############
+
+arr_bin=()
+arr_hf=()
+
+add_test() {
+    local bin=$1
+    local hf=$2
+    arr_bin+=("$bin")
+    arr_hf+=("$hf")
+}
+
+add_test "llama-gemma3-cli"   "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
+add_test "llama-llava-cli"    "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
+add_test "llama-llava-cli"    "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"
+add_test "llama-llava-cli"    "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
+add_test "llama-llava-cli"    "second-state/Llava-v1.5-7B-GGUF:Q2_K"
+add_test "llama-llava-cli"    "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"
+add_test "llama-llava-cli"    "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
+add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
+add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
+add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
+add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+
+###############
+
+cmake --build build -j --target "${arr_bin[@]}"
+
+arr_res=()
+
+for i in "${!arr_bin[@]}"; do
+    bin="${arr_bin[$i]}"
+    hf="${arr_hf[$i]}"
+
+    echo "Running test with binary: $bin and HF model: $hf"
+    echo ""
+    echo ""
+
+    output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty)
+
+    echo "$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

@@ -96,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -438,17 +438,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_cache_seq_keep(ctx, seq_id_best);
-            llama_kv_cache_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_cache_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_kv_self_seq_keep(ctx, seq_id_best);
+            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -192,7 +192,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/main/README.md

@@ -27,12 +27,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 ##### Input prompt (One-and-done)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 ##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -44,12 +56,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 
 ##### Input prompt (One-and-done)
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 #### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -77,6 +101,8 @@ The `llama-cli` program provides several ways to interact with the LLaMA models
 
 -   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
 -   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
+-   `--system-prompt PROMPT`: Provide a system prompt (will otherwise use the default one in the chat template (if provided)).
+-   `--system-prompt-file FNAME`: Provide a file containing a system prompt.
 -   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
 
 ## Interaction
@@ -89,7 +115,10 @@ In interactive mode, users can participate in text generation by injecting their
 
 -   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
 -   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
--   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false)
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default or provided chat template) (default: true if chat template found)
+-   `-no-cnv`:  Disable conversation mode (default: false)
+-   `-st, --single-turn`:  Only process a single conversation turn (user input) and then exit.
+-   `--jinja`:  Enable jinja chat template parser, will use the model's built-in template or a user-provided one (default: false)
 -   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
 
 By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
@@ -125,6 +154,8 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t
 
  Example usage: `--chat-template gemma`
 
+`--chat-template-file FNAME`:  Load a custom jinja chat template from an external file, useful if the model contains outdated or incompatible template, some examples can be found in models/templates. Up-to-date chat templates can be downloaded from Hugging Face using scripts/get_chat_template.py
+
 ## Context Management
 
 During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.

examples/main/main.cpp

@@ -354,7 +354,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
@@ -602,8 +602,8 @@ int main(int argc, char ** argv) {
                     LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_self_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -626,9 +626,9 @@ int main(int argc, char ** argv) {
                     LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_self_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 

examples/parallel/parallel.cpp

@@ -106,6 +106,8 @@ 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;
     }
@@ -202,7 +204,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -234,9 +236,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
+                llama_kv_self_seq_rm(ctx, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -372,8 +374,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
+                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 
@@ -405,7 +407,7 @@ 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.c_str());
+    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.path.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);

examples/passkey/passkey.cpp

@@ -64,7 +64,7 @@ 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.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
@@ -133,11 +133,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_update  (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -167,12 +167,12 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+      //llama_kv_self_defrag (ctx);
+        llama_kv_self_update (ctx);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -198,12 +198,12 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+          //llama_kv_self_defrag (ctx);
+            llama_kv_self_update (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -851,7 +851,7 @@ 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\n");
+    LOG("\ntask\tacc_norm\t95%% confidence interval\n");
 
     double acc = 0.0f;
 
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -985,8 +985,22 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
                 acc += 1.0;
             }
 
-            // Print the accumulated accuracy mean x 100
-            LOG("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0);
+            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);
         }
 
         i0 = i1 - 1;
@@ -1203,7 +1217,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1575,7 +1589,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1765,7 +1779,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

examples/quantize-stats/quantize-stats.cpp

@@ -1,6 +1,6 @@
 #include "ggml.h"
 #include "llama.h"
-#include "llama-context.h"
+#include "llama-model.h"
 #include "common.h"
 
 #include <algorithm>
@@ -328,7 +328,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const auto & tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(model);
 
     // check layer tensors
     int included_layers = 0;

examples/quantize/quantize.cpp

@@ -9,6 +9,7 @@
 #include <fstream>
 #include <cmath>
 #include <cctype>
+#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -16,7 +17,7 @@ struct quant_option {
     std::string desc;
 };
 
-static const std::vector<struct quant_option> QUANT_OPTIONS = {
+static const std::vector<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",  },
@@ -105,7 +106,8 @@ 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] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", 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("  --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");
@@ -114,6 +116,8 @@ 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");
@@ -244,6 +248,107 @@ 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]);
@@ -255,6 +360,7 @@ 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) {
@@ -277,6 +383,10 @@ 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]);
@@ -361,6 +471,9 @@ 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

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

examples/rpc/CMakeLists.txt

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

examples/rpc/README.md

@@ -72,3 +72,14 @@ $ 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

@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "ggml-cpu.h"
 
 #ifdef GGML_USE_CUDA
@@ -18,26 +22,144 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  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>
+
+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;
 };
 
 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, "  -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, "  -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, "  -c,      --cache          enable local file cache\n");
     fprintf(stderr, "\n");
 }
 
@@ -58,6 +180,8 @@ 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;
@@ -164,8 +288,20 @@ int main(int argc, char * argv[]) {
     } else {
         get_backend_memory(&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);
+    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\n");
+    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);
     ggml_backend_free(backend);
     return 0;
 }

examples/run/CMakeLists.txt

@@ -1,5 +1,16 @@
 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 ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_RUN_EXTRA_LIBS})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/run/run.cpp

@@ -38,24 +38,6 @@
 }
 #endif
 
-GGML_ATTRIBUTE_FORMAT(1, 2)
-static std::string fmt(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    const int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::string buf;
-    buf.resize(size);
-    const int size2 = vsnprintf(const_cast<char *>(buf.data()), buf.size() + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-
-    return buf;
-}
-
 GGML_ATTRIBUTE_FORMAT(1, 2)
 static int printe(const char * fmt, ...) {
     va_list args;
@@ -79,6 +61,7 @@ class Opt {
         ctx_params           = llama_context_default_params();
         model_params         = llama_model_default_params();
         context_size_default = ctx_params.n_batch;
+        n_threads_default    = ctx_params.n_threads;
         ngl_default          = model_params.n_gpu_layers;
         common_params_sampling sampling;
         temperature_default = sampling.temp;
@@ -104,6 +87,7 @@ class Opt {
 
         ctx_params.n_batch        = context_size >= 0 ? context_size : context_size_default;
         ctx_params.n_ctx          = ctx_params.n_batch;
+        ctx_params.n_threads = ctx_params.n_threads_batch = n_threads >= 0 ? n_threads : n_threads_default;
         model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default;
         temperature               = temperature >= 0 ? temperature : temperature_default;
 
@@ -116,12 +100,12 @@ class Opt {
     std::string chat_template_file;
     std::string          user;
     bool                 use_jinja   = false;
-    int                  context_size = -1, ngl = -1;
+    int                  context_size = -1, ngl = -1, n_threads = -1;
     float                temperature = -1;
     bool                 verbose     = false;
 
   private:
-    int   context_size_default = -1, ngl_default = -1;
+    int   context_size_default = -1, ngl_default = -1, n_threads_default = -1;
     float temperature_default = -1;
     bool  help                = false;
 
@@ -159,53 +143,94 @@ class Opt {
         return 0;
     }
 
+    int parse_options_with_value(int argc, const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
+            if (handle_option_with_value(argc, argv, i, context_size) == 1) {
+                return 1;
+            }
+        } else if (options_parsing &&
+                   (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
+            if (handle_option_with_value(argc, argv, i, ngl) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && (strcmp(argv[i], "-t") == 0 || strcmp(argv[i], "--threads") == 0)) {
+            if (handle_option_with_value(argc, argv, i, n_threads) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
+            if (handle_option_with_value(argc, argv, i, temperature) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0) {
+            if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
+                return 1;
+            }
+            use_jinja = true;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_options(const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
+            verbose = true;
+        } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
+            use_jinja = true;
+        } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
+            help = true;
+            return 0;
+        } else if (options_parsing && strcmp(argv[i], "--") == 0) {
+            options_parsing = false;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_positional_args(const char ** argv, int & i, int & positional_args_i) {
+        if (positional_args_i == 0) {
+            if (!argv[i][0] || argv[i][0] == '-') {
+                return 1;
+            }
+
+            ++positional_args_i;
+            model_ = argv[i];
+        } else if (positional_args_i == 1) {
+            ++positional_args_i;
+            user = argv[i];
+        } else {
+            user += " " + std::string(argv[i]);
+        }
+
+        return 0;
+    }
+
     int parse(int argc, const char ** argv) {
         bool options_parsing   = true;
         for (int i = 1, positional_args_i = 0; i < argc; ++i) {
-            if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
-                if (handle_option_with_value(argc, argv, i, context_size) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
-                if (handle_option_with_value(argc, argv, i, ngl) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
-                if (handle_option_with_value(argc, argv, i, temperature) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
-                verbose = true;
-            } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
-                use_jinja = true;
-            } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0){
-                if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
-                    return 1;
-                }
-                use_jinja = true;
-            } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
-                help = true;
-                return 0;
-            } else if (options_parsing && strcmp(argv[i], "--") == 0) {
-                options_parsing = false;
-            } else if (positional_args_i == 0) {
-                if (!argv[i][0] || argv[i][0] == '-') {
-                    return 1;
-                }
+            int ret = parse_options_with_value(argc, argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
 
-                ++positional_args_i;
-                model_ = argv[i];
-            } else if (positional_args_i == 1) {
-                ++positional_args_i;
-                user = argv[i];
-            } else {
-                user += " " + std::string(argv[i]);
+            ret = parse_options(argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
+
+            if (parse_positional_args(argv, i, positional_args_i)) {
+                return 1;
             }
         }
 
-        if (model_.empty()){
+        if (model_.empty()) {
             return 1;
         }
 
@@ -232,6 +257,8 @@ class Opt {
             "      Number of GPU layers (default: %d)\n"
             "  --temp <value>\n"
             "      Temperature (default: %.1f)\n"
+            "  -t, --threads <value>\n"
+            "      Number of threads to use during generation (default: %d)\n"
             "  -v, --verbose, --log-verbose\n"
             "      Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n"
             "  -h, --help\n"
@@ -260,7 +287,7 @@ class Opt {
             "  llama-run file://some-file3.gguf\n"
             "  llama-run --ngl 999 some-file4.gguf\n"
             "  llama-run --ngl 999 some-file5.gguf Hello World\n",
-            context_size_default, ngl_default, temperature_default);
+            context_size_default, ngl_default, temperature_default, n_threads_default);
     }
 };
 
@@ -480,11 +507,11 @@ class HttpClient {
         int secs = static_cast<int>(seconds) % 60;
 
         if (hrs > 0) {
-            return fmt("%dh %02dm %02ds", hrs, mins, secs);
+            return string_format("%dh %02dm %02ds", hrs, mins, secs);
         } else if (mins > 0) {
-            return fmt("%dm %02ds", mins, secs);
+            return string_format("%dm %02ds", mins, secs);
         } else {
-            return fmt("%ds", secs);
+            return string_format("%ds", secs);
         }
     }
 
@@ -499,7 +526,7 @@ class HttpClient {
             }
         }
 
-        return fmt("%.2f %s", dbl_size, suffix[i]);
+        return string_format("%.2f %s", dbl_size, suffix[i]);
     }
 
     static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
@@ -533,7 +560,9 @@ class HttpClient {
         return (now_downloaded_plus_file_size * 100) / total_to_download;
     }
 
-    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
+    static std::string generate_progress_prefix(curl_off_t percentage) {
+        return string_format("%3ld%% |", static_cast<long int>(percentage));
+    }
 
     static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
         const auto                          now             = std::chrono::steady_clock::now();
@@ -544,9 +573,9 @@ class HttpClient {
     static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download,
                                                 double speed, double estimated_time) {
         const int width = 10;
-        return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width,
-                   human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width,
-                   human_readable_time(estimated_time).c_str());
+        return string_format("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(),
+                             width, human_readable_size(total_to_download).c_str(), width,
+                             human_readable_size(speed).c_str(), width, human_readable_time(estimated_time).c_str());
     }
 
     static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) {
@@ -668,8 +697,10 @@ class LlamaData {
         std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
         std::string              url;
 
+        std::string model_endpoint = get_model_endpoint();
+
         if (pos == std::string::npos) {
-            auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/");
+            auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/");
             hfr                             = model_name;
 
             nlohmann::json manifest;
@@ -684,7 +715,7 @@ class LlamaData {
             hff = model.substr(pos + 1);
         }
 
-        url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
+        url = model_endpoint + hfr + "/resolve/main/" + hff;
 
         return download(url, bn, true, headers);
     }
@@ -891,7 +922,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;
 
     const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
     prompt_tokens.resize(n_prompt_tokens);
@@ -907,7 +938,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
     const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");

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

@@ -15,7 +15,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    print_build_info();
+    common_init();
 
     if (params.n_predict < 0) {
         params.n_predict = 16;
@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_cache_clear(ctx3);
+        llama_kv_self_clear(ctx3);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/server/server.cpp

@@ -133,7 +133,8 @@ struct slot_params {
 
         auto grammar_triggers = json::array();
         for (const auto & trigger : sampling.grammar_triggers) {
-            grammar_triggers.push_back(trigger.to_json<json>());
+            server_grammar_trigger ct(std::move(trigger));
+            grammar_triggers.push_back(ct.to_json());
         }
 
         return json {
@@ -372,9 +373,9 @@ struct server_task {
             const auto grammar_triggers = data.find("grammar_triggers");
             if (grammar_triggers != data.end()) {
                 for (const auto & t : *grammar_triggers) {
-                    auto ct = common_grammar_trigger::from_json(t);
-                    if (ct.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
-                        const auto & word = ct.value;
+                    server_grammar_trigger ct(t);
+                    if (ct.value.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
+                        const auto & word = ct.value.value;
                         auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true);
                         if (ids.size() == 1) {
                             auto token = ids[0];
@@ -392,7 +393,7 @@ struct server_task {
                             params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word});
                         }
                     } else {
-                        params.sampling.grammar_triggers.push_back(ct);
+                        params.sampling.grammar_triggers.push_back(std::move(ct.value));
                     }
                 }
             }
@@ -489,8 +490,12 @@ struct result_timings {
     double predicted_per_token_ms;
     double predicted_per_second;
 
+    // Optional speculative metrics - only included when > 0
+    int32_t draft_n = 0;
+    int32_t draft_n_accepted = 0;
+
     json to_json() const {
-        return {
+        json base = {
             {"prompt_n",               prompt_n},
             {"prompt_ms",              prompt_ms},
             {"prompt_per_token_ms",    prompt_per_token_ms},
@@ -501,6 +506,13 @@ struct result_timings {
             {"predicted_per_token_ms", predicted_per_token_ms},
             {"predicted_per_second",   predicted_per_second},
         };
+
+        if (draft_n > 0) {
+            base["draft_n"] = draft_n;
+            base["draft_n_accepted"] = draft_n_accepted;
+        }
+
+        return base;
     }
 };
 
@@ -830,6 +842,11 @@ struct server_task_result_cmpl_final : server_task_result {
             ret.push_back({"timings", timings.to_json()});
         }
 
+        // extra fields for debugging purposes
+        if (verbose) {
+            ret["__verbose"] = to_json_non_oaicompat();
+        }
+
         return ret;
     }
 };
@@ -1294,6 +1311,10 @@ struct server_slot {
 
     std::function<void(int)> callback_on_release;
 
+    // Speculative decoding stats
+    int32_t n_draft_total = 0;      // Total draft tokens generated
+    int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+
     void reset() {
         SLT_DBG(*this, "%s", "\n");
 
@@ -1310,6 +1331,10 @@ struct server_slot {
 
         generated_tokens.clear();
         generated_token_probs.clear();
+
+        // clear speculative decoding stats
+        n_draft_total = 0;
+        n_draft_accepted = 0;
     }
 
     bool is_non_causal() const {
@@ -1376,6 +1401,12 @@ struct server_slot {
         timings.predicted_per_token_ms = t_token_generation / n_decoded;
         timings.predicted_per_second = 1e3 / t_token_generation * n_decoded;
 
+        // Add speculative metrics
+        if (n_draft_total > 0) {
+            timings.draft_n = n_draft_total;
+            timings.draft_n_accepted = n_draft_accepted;
+        }
+
         return timings;
     }
 
@@ -1423,6 +1454,15 @@ struct server_slot {
                 t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                 t_token_generation, n_decoded, t_gen, n_gen_second,
                 t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
+
+        if (n_draft_total > 0) {
+            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            SLT_INF(*this,
+                    "\n"
+                    "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total
+            );
+        }
     }
 
     json to_json() const {
@@ -1665,6 +1705,8 @@ private:
 };
 
 struct server_response {
+    bool running = true;
+
     // for keeping track of all tasks waiting for the result
     std::unordered_set<int> waiting_task_ids;
 
@@ -1719,6 +1761,10 @@ struct server_response {
         while (true) {
             std::unique_lock<std::mutex> lock(mutex_results);
             condition_results.wait(lock, [&]{
+                if (!running) {
+                    SRV_DBG("%s : queue result stop\n", __func__);
+                    std::terminate(); // we cannot return here since the caller is HTTP code
+                }
                 return !queue_results.empty();
             });
 
@@ -1749,6 +1795,10 @@ struct server_response {
             }
 
             std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
+            if (!running) {
+                SRV_DBG("%s : queue result stop\n", __func__);
+                std::terminate(); // we cannot return here since the caller is HTTP code
+            }
             if (cr_res == std::cv_status::timeout) {
                 return nullptr;
             }
@@ -1778,6 +1828,12 @@ struct server_response {
             }
         }
     }
+
+    // terminate the waiting loop
+    void terminate() {
+        running = false;
+        condition_results.notify_all();
+    }
 };
 
 struct server_context {
@@ -1837,7 +1893,7 @@ struct server_context {
     }
 
     bool load_model(const common_params & params) {
-        SRV_INF("loading model '%s'\n", params.model.c_str());
+        SRV_INF("loading model '%s'\n", params.model.path.c_str());
 
         params_base = params;
 
@@ -1847,7 +1903,7 @@ struct server_context {
         ctx   = llama_init.context.get();
 
         if (model == nullptr) {
-            SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
+            SRV_ERR("failed to load model, '%s'\n", params_base.model.path.c_str());
             return false;
         }
 
@@ -1858,31 +1914,32 @@ struct server_context {
         add_bos_token = llama_vocab_get_add_bos(vocab);
         has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
 
-        if (!params_base.speculative.model.empty() || !params_base.speculative.hf_repo.empty()) {
-            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
+        if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
 
             auto params_dft = params_base;
 
             params_dft.devices      = params_base.speculative.devices;
-            params_dft.hf_file      = params_base.speculative.hf_file;
-            params_dft.hf_repo      = params_base.speculative.hf_repo;
             params_dft.model        = params_base.speculative.model;
-            params_dft.model_url    = params_base.speculative.model_url;
             params_dft.n_ctx        = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx;
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
 
+            // force F16 KV cache for the draft model for extra performance
+            params_dft.cache_type_k = GGML_TYPE_F16;
+            params_dft.cache_type_v = GGML_TYPE_F16;
+
             llama_init_dft = common_init_from_params(params_dft);
 
             model_dft = llama_init_dft.model.get();
 
             if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
+                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
                 return false;
             }
 
             if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
-                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
+                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
 
                 return false;
             }
@@ -1892,10 +1949,6 @@ struct server_context {
             cparams_dft = common_context_params_to_llama(params_dft);
             cparams_dft.n_batch = n_ctx_dft;
 
-            // force F16 KV cache for the draft model for extra performance
-            cparams_dft.type_k = GGML_TYPE_F16;
-            cparams_dft.type_v = GGML_TYPE_F16;
-
             // the context is not needed - we will create one for each slot
             llama_init_dft.context.reset();
         }
@@ -2040,6 +2093,18 @@ struct server_context {
         return ret;
     }
 
+    bool can_be_detokenized(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
+        const llama_model * model = llama_get_model(ctx);
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const int32_t n_vocab = llama_vocab_n_tokens(vocab);
+        for (const auto & token : tokens) {
+            if (token < 0 || token >= n_vocab) {
+                return false;
+            }
+        }
+        return true;
+    }
+
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot.reset();
         slot.id_task       = task.id;
@@ -2054,6 +2119,11 @@ struct server_context {
             slot.lora = task.params.lora;
         }
 
+        bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
+        if (!can_detokenize) {
+            send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST);
+            return false;
+        }
         SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@@ -2096,7 +2166,7 @@ struct server_context {
         SRV_DBG("%s", "clearing KV cache\n");
 
         // clear the entire KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         clean_kv_cache = false;
     }
 
@@ -2638,8 +2708,8 @@ struct server_context {
                     res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                     res->t_start             = metrics.t_start;
 
-                    res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx);
-                    res->kv_cache_used_cells   = llama_get_kv_cache_used_cells(ctx);
+                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
+                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);
 
                     res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                     res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
@@ -2755,7 +2825,7 @@ struct server_context {
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
+                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
                     slot->cache_tokens.clear();
 
                     auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2823,8 +2893,8 @@ struct server_context {
 
                 SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
+                llama_kv_self_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
+                llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 if (slot.params.cache_prompt) {
                     for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@@ -3015,8 +3085,8 @@ struct server_context {
 
                                             const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
 
-                                            llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
-                                            llama_kv_cache_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
+                                            llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
+                                            llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@@ -3054,9 +3124,9 @@ struct server_context {
                     }
 
                     // keep only the common part
-                    if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
                         // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
+                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
 
                         // there is no common part left
                         slot.n_past = 0;
@@ -3268,6 +3338,9 @@ struct server_context {
 
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
+                // keep track of total number of tokens generated in the draft
+                slot.n_draft_total += draft.size();
+
                 // ignore small drafts
                 if (slot.params.speculative.n_min > (int) draft.size()) {
                     SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
@@ -3293,10 +3366,13 @@ struct server_context {
                 slot.n_past    += ids.size();
                 slot.n_decoded += ids.size();
 
+                // update how many tokens out of draft was accepted
+                slot.n_draft_accepted += ids.size() - 1;
+
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
 
-                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
 
                 for (size_t i = 0; i < ids.size(); ++i) {
                     completion_token_output result;
@@ -3803,7 +3879,7 @@ int main(int argc, char ** argv) {
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
-            { "model_path",                  ctx_server.params_base.model },
+            { "model_path",                  ctx_server.params_base.model.path },
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -3831,6 +3907,21 @@ int main(int argc, char ** argv) {
         res_ok(res, {{ "success", true }});
     };
 
+    const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+        json data = {
+            {
+                "template", common_chat_templates_source(ctx_server.chat_templates.get()),
+            },
+            {
+                "model_info", {
+                    { "llama.context_length", ctx_server.slots.back().n_ctx, },
+                }
+            },
+        };
+
+        res_ok(res, data);
+    };
+
     // handle completion-like requests (completion, chat, infill)
     // we can optionally provide a custom format for partial results and final results
     const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
@@ -4069,7 +4160,7 @@ int main(int argc, char ** argv) {
             {"object", "list"},
             {"data", {
                 {
-                    {"id",       params.model_alias.empty() ? params.model : params.model_alias},
+                    {"id",       params.model_alias.empty() ? params.model.path : params.model_alias},
                     {"object",   "model"},
                     {"created",  std::time(0)},
                     {"owned_by", "llamacpp"},
@@ -4395,6 +4486,7 @@ int main(int argc, char ** argv) {
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Post("/props",               handle_props_change);
+    svr->Post("/api/show",            handle_api_show);
     svr->Get ("/models",              handle_models); // public endpoint (no API key check)
     svr->Get ("/v1/models",           handle_models); // public endpoint (no API key check)
     svr->Post("/completion",          handle_completions); // legacy
@@ -4431,21 +4523,31 @@ int main(int argc, char ** argv) {
     svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
 
     // clean up function, to be called before exit
-    auto clean_up = [&svr]() {
+    auto clean_up = [&svr, &ctx_server]() {
         SRV_INF("%s: cleaning up before exit...\n", __func__);
         svr->stop();
+        ctx_server.queue_results.terminate();
         llama_backend_free();
     };
 
-    // bind HTTP listen port
     bool was_bound = false;
-    if (params.port == 0) {
-        int bound_port = svr->bind_to_any_port(params.hostname);
-        if ((was_bound = (bound_port >= 0))) {
-            params.port = bound_port;
-        }
+    if (string_ends_with(std::string(params.hostname), ".sock")) {
+        LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
+        svr->set_address_family(AF_UNIX);
+        // bind_to_port requires a second arg, any value other than 0 should
+        // simply get ignored
+        was_bound = svr->bind_to_port(params.hostname, 8080);
     } else {
-        was_bound = svr->bind_to_port(params.hostname, params.port);
+        LOG_INF("%s: binding port with default address family\n", __func__);
+        // bind HTTP listen port
+        if (params.port == 0) {
+            int bound_port = svr->bind_to_any_port(params.hostname);
+            if ((was_bound = (bound_port >= 0))) {
+                params.port = bound_port;
+            }
+        } else {
+            was_bound = svr->bind_to_port(params.hostname, params.port);
+        }
     }
 
     if (!was_bound) {
@@ -4465,7 +4567,7 @@ int main(int argc, char ** argv) {
 
     if (!ctx_server.load_model(params)) {
         clean_up();
-        // t.join(); // FIXME: see below
+        t.join();
         LOG_ERR("%s: exiting due to model loading error\n", __func__);
         return 1;
     }
@@ -4513,7 +4615,7 @@ int main(int argc, char ** argv) {
     ctx_server.queue_tasks.start_loop();
 
     clean_up();
-    // t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
+    t.join();
 
     return 0;
 }

examples/server/tests/README.md

@@ -17,7 +17,7 @@ To mitigate it, you can increase values in `n_predict`, `kv_size`.
 
 ```shell
 cd ../../..
-cmake -B build -DLLAMA_CURL=ON
+cmake -B build
 cmake --build build --target llama-server
 ```
 

examples/server/tests/unit/test_embedding.py

@@ -49,6 +49,26 @@ def test_embedding_multiple():
         assert len(d['embedding']) > 1
 
 
+def test_embedding_multiple_with_fa():
+    server = ServerPreset.bert_bge_small_with_fa()
+    server.pooling = 'last'
+    server.start()
+    # one of these should trigger the FA branch (i.e. context size % 256 == 0)
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": [
+            "a "*253,
+            "b "*254,
+            "c "*255,
+            "d "*256,
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 4
+    for d in res.body['data']:
+        assert 'embedding' in d
+        assert len(d['embedding']) > 1
+
+
 @pytest.mark.parametrize(
     "input,is_multi_prompt",
     [

examples/server/tests/utils.py

@@ -302,7 +302,7 @@ class ServerPreset:
         server.model_hf_repo = "ggml-org/models"
         server.model_hf_file = "tinyllamas/stories260K.gguf"
         server.model_alias = "tinyllama-2"
-        server.n_ctx = 256
+        server.n_ctx = 512
         server.n_batch = 32
         server.n_slots = 2
         server.n_predict = 64
@@ -323,6 +323,21 @@ class ServerPreset:
         server.server_embeddings = True
         return server
 
+    @staticmethod
+    def bert_bge_small_with_fa() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
+        server.model_alias = "bert-bge-small"
+        server.n_ctx = 1024
+        server.n_batch = 300
+        server.n_ubatch = 300
+        server.n_slots = 2
+        server.fa = True
+        server.seed = 42
+        server.server_embeddings = True
+        return server
+
     @staticmethod
     def tinyllama_infill() -> ServerProcess:
         server = ServerProcess()

examples/server/utils.hpp

@@ -3,7 +3,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "common/base64.hpp"
+#include "base64.hpp"
 
 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
@@ -58,6 +58,32 @@ static T json_value(const json & body, const std::string & key, const T & defaul
 
 const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
 
+// thin wrapper around common_grammar_trigger with (de)serialization functions
+struct server_grammar_trigger {
+    common_grammar_trigger value;
+
+    server_grammar_trigger() = default;
+    server_grammar_trigger(const common_grammar_trigger & value) : value(value) {}
+    server_grammar_trigger(const json & in) {
+        value.type = (common_grammar_trigger_type) in.at("type").get<int>();
+        value.value = in.at("value").get<std::string>();
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            value.token = (llama_token) in.at("token").get<int>();
+        }
+    }
+
+    json to_json() const {
+        json out {
+            {"type", (int) value.type},
+            {"value", value.value},
+        };
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            out["token"] = (int) value.token;
+        }
+        return out;
+    }
+};
+
 //
 // tokenizer and input processing utils
 //
@@ -621,11 +647,14 @@ static json oaicompat_completion_params_parse(
 
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
     llama_params["prompt"]           = chat_params.prompt;
-    llama_params["grammar"]          = chat_params.grammar;
+    if (!chat_params.grammar.empty()) {
+        llama_params["grammar"] = chat_params.grammar;
+    }
     llama_params["grammar_lazy"]     = chat_params.grammar_lazy;
     auto grammar_triggers = json::array();
     for (const auto & trigger : chat_params.grammar_triggers) {
-        grammar_triggers.push_back(trigger.to_json<json>());
+        server_grammar_trigger ct(trigger);
+        grammar_triggers.push_back(ct.to_json());
     }
     llama_params["grammar_triggers"] = grammar_triggers;
     llama_params["preserved_tokens"] = chat_params.preserved_tokens;

examples/server/webui/package.json

@@ -13,9 +13,11 @@
   "dependencies": {
     "@heroicons/react": "^2.2.0",
     "@sec-ant/readable-stream": "^0.6.0",
+    "@tailwindcss/postcss": "^4.1.1",
+    "@tailwindcss/vite": "^4.1.1",
     "@vscode/markdown-it-katex": "^1.1.1",
     "autoprefixer": "^10.4.20",
-    "daisyui": "^4.12.14",
+    "daisyui": "^5.0.12",
     "dexie": "^4.0.11",
     "highlight.js": "^11.10.0",
     "katex": "^0.16.15",
@@ -29,7 +31,7 @@
     "remark-breaks": "^4.0.0",
     "remark-gfm": "^4.0.0",
     "remark-math": "^6.0.0",
-    "tailwindcss": "^3.4.15",
+    "tailwindcss": "^4.1.1",
     "textlinestream": "^1.1.1",
     "vite-plugin-singlefile": "^2.0.3"
   },

examples/server/webui/postcss.config.js

@@ -1,6 +1,5 @@
 export default {
   plugins: {
-    tailwindcss: {},
-    autoprefixer: {},
+    "@tailwindcss/postcss": {},
   },
 }

examples/server/webui/src/App.tsx

@@ -28,7 +28,7 @@ function AppLayout() {
     <>
       <Sidebar />
       <div
-        className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto"
+        className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto bg-base-100"
         id="main-scroll"
       >
         <Header />

examples/server/webui/src/Config.ts

@@ -1,4 +1,4 @@
-import daisyuiThemes from 'daisyui/src/theming/themes';
+import daisyuiThemes from 'daisyui/theme/object';
 import { isNumeric } from './utils/misc';
 
 export const isDev = import.meta.env.MODE === 'development';

examples/server/webui/src/components/ChatScreen.tsx

@@ -1,4 +1,4 @@
-import { useEffect, useMemo, useRef, useState } from 'react';
+import { useEffect, useMemo, useState } from 'react';
 import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
 import ChatMessage from './ChatMessage';
 import { CanvasType, Message, PendingMessage } from '../utils/types';
@@ -6,6 +6,7 @@ import { classNames, cleanCurrentUrl, throttle } from '../utils/misc';
 import CanvasPyInterpreter from './CanvasPyInterpreter';
 import StorageUtils from '../utils/storage';
 import { useVSCodeContext } from '../utils/llama-vscode';
+import { useChatTextarea, ChatTextareaApi } from './useChatTextarea.ts';
 
 /**
  * A message display is a message node with additional information for rendering.
@@ -99,13 +100,10 @@ export default function ChatScreen() {
     canvasData,
     replaceMessageAndGenerate,
   } = useAppContext();
-  const [inputMsg, setInputMsg] = useState(prefilledMsg.content());
-  const inputRef = useRef<HTMLTextAreaElement>(null);
 
-  const { extraContext, clearExtraContext } = useVSCodeContext(
-    inputRef,
-    setInputMsg
-  );
+  const textarea: ChatTextareaApi = useChatTextarea(prefilledMsg.content());
+
+  const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
   // TODO: improve this when we have "upload file" feature
   const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
 
@@ -135,9 +133,10 @@ export default function ChatScreen() {
   };
 
   const sendNewMessage = async () => {
-    if (inputMsg.trim().length === 0 || isGenerating(currConvId ?? '')) return;
-    const lastInpMsg = inputMsg;
-    setInputMsg('');
+    const lastInpMsg = textarea.value();
+    if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? ''))
+      return;
+    textarea.setValue('');
     scrollToBottom(false);
     setCurrNodeId(-1);
     // get the last message node
@@ -146,13 +145,13 @@ export default function ChatScreen() {
       !(await sendMessage(
         currConvId,
         lastMsgNodeId,
-        inputMsg,
+        lastInpMsg,
         currExtra,
         onChunk
       ))
     ) {
       // restore the input message if failed
-      setInputMsg(lastInpMsg);
+      textarea.setValue(lastInpMsg);
     }
     // OK
     clearExtraContext();
@@ -195,16 +194,13 @@ export default function ChatScreen() {
       // send the prefilled message if needed
       sendNewMessage();
     } else {
-      // otherwise, focus on the input and move the cursor to the end
-      if (inputRef.current) {
-        inputRef.current.focus();
-        inputRef.current.selectionStart = inputRef.current.value.length;
-      }
+      // otherwise, focus on the input
+      textarea.focus();
     }
     prefilledMsg.clear();
     // no need to keep track of sendNewMessage
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [inputRef]);
+  }, [textarea.ref]);
 
   // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg)
   const pendingMsgDisplay: MessageDisplay[] =
@@ -254,16 +250,16 @@ export default function ChatScreen() {
         </div>
 
         {/* chat input */}
-        <div className="flex flex-row items-center pt-8 pb-6 sticky bottom-0 bg-base-100">
+        <div className="flex flex-row items-end pt-8 pb-6 sticky bottom-0 bg-base-100">
           <textarea
-            className="textarea textarea-bordered w-full"
+            // Default (mobile): Enable vertical resize, overflow auto for scrolling if needed
+            // Large screens (lg:): Disable manual resize, apply max-height for autosize limit
+            className="textarea textarea-bordered w-full resize-vertical lg:resize-none lg:max-h-48 lg:overflow-y-auto" // Adjust lg:max-h-48 as needed (e.g., lg:max-h-60)
             placeholder="Type a message (Shift+Enter to add a new line)"
-            ref={inputRef}
-            value={inputMsg}
-            onChange={(e) => setInputMsg(e.target.value)}
+            ref={textarea.ref}
+            onInput={textarea.onInput} // Hook's input handler (will only resize height on lg+ screens)
             onKeyDown={(e) => {
               if (e.nativeEvent.isComposing || e.keyCode === 229) return;
-              if (e.key === 'Enter' && e.shiftKey) return;
               if (e.key === 'Enter' && !e.shiftKey) {
                 e.preventDefault();
                 sendNewMessage();
@@ -271,7 +267,11 @@ export default function ChatScreen() {
             }}
             id="msg-input"
             dir="auto"
+            // Set a base height of 2 rows for mobile views
+            // On lg+ screens, the hook will calculate and set the initial height anyway
+            rows={2}
           ></textarea>
+
           {isGenerating(currConvId ?? '') ? (
             <button
               className="btn btn-neutral ml-2"
@@ -280,11 +280,7 @@ export default function ChatScreen() {
               Stop
             </button>
           ) : (
-            <button
-              className="btn btn-primary ml-2"
-              onClick={sendNewMessage}
-              disabled={inputMsg.trim().length === 0}
-            >
+            <button className="btn btn-primary ml-2" onClick={sendNewMessage}>
               Send
             </button>
           )}

examples/server/webui/src/components/Header.tsx

@@ -2,7 +2,7 @@ import { useEffect, useState } from 'react';
 import StorageUtils from '../utils/storage';
 import { useAppContext } from '../utils/app.context';
 import { classNames } from '../utils/misc';
-import daisyuiThemes from 'daisyui/src/theming/themes';
+import daisyuiThemes from 'daisyui/theme/object';
 import { THEMES } from '../Config';
 import { useNavigate } from 'react-router';
 
@@ -20,7 +20,6 @@ export default function Header() {
     document.body.setAttribute('data-theme', selectedTheme);
     document.body.setAttribute(
       'data-color-scheme',
-      // @ts-expect-error daisyuiThemes complains about index type, but it should work
       daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto'
     );
   }, [selectedTheme]);

examples/server/webui/src/components/useChatTextarea.ts

@@ -0,0 +1,96 @@
+import { useEffect, useRef, useState, useCallback } from 'react';
+
+// Media Query for detecting "large" screens (matching Tailwind's lg: breakpoint)
+const LARGE_SCREEN_MQ = '(min-width: 1024px)';
+
+// Calculates and sets the textarea height based on its scrollHeight
+const adjustTextareaHeight = (textarea: HTMLTextAreaElement | null) => {
+  if (!textarea) return;
+
+  // Only perform auto-sizing on large screens
+  if (!window.matchMedia(LARGE_SCREEN_MQ).matches) {
+    // On small screens, reset inline height and max-height styles.
+    // This allows CSS (e.g., `rows` attribute or classes) to control the height,
+    // and enables manual resizing if `resize-vertical` is set.
+    textarea.style.height = ''; // Use 'auto' or '' to reset
+    textarea.style.maxHeight = '';
+    return; // Do not adjust height programmatically on small screens
+  }
+
+  const computedStyle = window.getComputedStyle(textarea);
+  // Get the max-height specified by CSS (e.g., from `lg:max-h-48`)
+  const currentMaxHeight = computedStyle.maxHeight;
+
+  // Temporarily remove max-height to allow scrollHeight to be calculated correctly
+  textarea.style.maxHeight = 'none';
+  // Reset height to 'auto' to measure the actual scrollHeight needed
+  textarea.style.height = 'auto';
+  // Set the height to the calculated scrollHeight
+  textarea.style.height = `${textarea.scrollHeight}px`;
+  // Re-apply the original max-height from CSS to enforce the limit
+  textarea.style.maxHeight = currentMaxHeight;
+};
+
+// Interface describing the API returned by the hook
+export interface ChatTextareaApi {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+  onInput: (event: React.FormEvent<HTMLTextAreaElement>) => void; // Input handler
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+// combined now with auto-sizing logic.
+export function useChatTextarea(initValue: string): ChatTextareaApi {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  // Effect to set initial value and height on mount or when initValue changes
+  useEffect(() => {
+    const textarea = textareaRef.current;
+    if (textarea) {
+      if (typeof savedInitValue === 'string' && savedInitValue.length > 0) {
+        textarea.value = savedInitValue;
+        // Call adjustTextareaHeight - it will check screen size internally
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+        setSavedInitValue(''); // Reset after applying
+      } else {
+        // Adjust height even if there's no initial value (for initial render)
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+      }
+    }
+  }, [textareaRef, savedInitValue]); // Depend on ref and savedInitValue
+
+  const handleInput = useCallback(
+    (event: React.FormEvent<HTMLTextAreaElement>) => {
+      // Call adjustTextareaHeight on every input - it will decide whether to act
+      adjustTextareaHeight(event.currentTarget);
+    },
+    []
+  );
+
+  return {
+    // Method to get the current value directly from the textarea
+    value: () => {
+      return textareaRef.current?.value ?? '';
+    },
+    // Method to programmatically set the value and trigger height adjustment
+    setValue: (value: string) => {
+      const textarea = textareaRef.current;
+      if (textarea) {
+        textarea.value = value;
+        // Call adjustTextareaHeight - it will check screen size internally
+        setTimeout(() => adjustTextareaHeight(textarea), 0);
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        textareaRef.current.focus();
+      }
+    },
+    ref: textareaRef,
+    onInput: handleInput,
+  };
+}

examples/server/webui/src/index.scss

@@ -1,8 +1,13 @@
 @use 'sass:meta';
+@use 'tailwindcss';
 
-@tailwind base;
-@tailwind components;
-@tailwind utilities;
+@plugin 'daisyui' {
+  themes: all;
+}
+
+html {
+  scrollbar-gutter: auto;
+}
 
 .markdown {
   h1,

examples/server/webui/src/utils/llama-vscode.ts

@@ -1,5 +1,6 @@
 import { useEffect, useState } from 'react';
 import { MessageExtraContext } from './types';
+import { ChatTextareaApi } from '../components/useChatTextarea.ts';
 
 // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
 // Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@@ -14,10 +15,7 @@ interface SetTextEvData {
  * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n  return 123' }, '*');
  */
 
-export const useVSCodeContext = (
-  inputRef: React.RefObject<HTMLTextAreaElement>,
-  setInputMsg: (text: string) => void
-) => {
+export const useVSCodeContext = (textarea: ChatTextareaApi) => {
   const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
     null
   );
@@ -27,20 +25,20 @@ export const useVSCodeContext = (
     const handleMessage = (event: MessageEvent) => {
       if (event.data?.command === 'setText') {
         const data: SetTextEvData = event.data;
-        setInputMsg(data?.text);
+        textarea.setValue(data?.text);
         if (data?.context && data.context.length > 0) {
           setExtraContext({
             type: 'context',
             content: data.context,
           });
         }
-        inputRef.current?.focus();
+        textarea.focus();
       }
     };
 
     window.addEventListener('message', handleMessage);
     return () => window.removeEventListener('message', handleMessage);
-  }, [inputRef, setInputMsg]);
+  }, [textarea]);
 
   // Add a keydown listener that sends the "escapePressed" message to the parent window
   useEffect(() => {

examples/server_embd.py

@@ -15,7 +15,7 @@ async def main():
     model_url = "http://127.0.0.1:6900"
     responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
         url= f"{model_url}/embedding",
-        json= {"content": str(0)*1024}
+        json= {"content": "a "*1022}
     ) for i in range(n)])
 
     for response in responses: