sync : ggml

ggml : fix ggml_gallocr_ptr type (ggml/1205)
cuda : fix unused variable compile warning (whisper/0)
2026-04-16 16:27:32 +03:00 · 2025-05-01 09:59:02 +03:00 · 2025-05-01 09:32:45 +03:00 · 2025-05-01 09:32:42 +03:00
413 changed files with 11477 additions and 20686 deletions
--- a/.devops/intel.Dockerfile
+++ b/.devops/intel.Dockerfile
@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04

 ## Build Image

--- a/.editorconfig
+++ b/.editorconfig
@@ -21,15 +21,15 @@ indent_style = tab
 [prompts/*.txt]
 insert_final_newline = unset

-[tools/server/public/*]
+[examples/server/public/*]
 indent_size = 2

-[tools/server/public/deps_*]
+[examples/server/public/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset

-[tools/server/deps_*]
+[examples/server/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset
@@ -37,7 +37,7 @@ indent_size = unset
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab

-[tools/cvector-generator/*.txt]
+[examples/cvector-generator/*.txt]
 trim_trailing_whitespace = unset
 insert_final_newline = unset

--- a/.flake8
+++ b/.flake8
@@ -2,9 +2,8 @@
 max-line-length = 125
 ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503
 exclude =
-    # Do not traverse examples and tools
+    # Do not traverse examples
    examples,
-    tools,
    # Do not include package initializers
    __init__.py,
    # No need to traverse our git directory
--- a/.github/actions/get-tag-name/action.yml
+++ b/.github/actions/get-tag-name/action.yml
@@ -1,22 +0,0 @@
-name: "Determine tag name"
-description: "Determine the tag name to use for a release"
-outputs:
-  name:
-    description: "The name of the tag"
-    value: ${{ steps.tag.outputs.name }}
-
-runs:
-  using: "composite"
-  steps:
-    - name: Determine tag name
-      id: tag
-      shell: bash
-      run: |
-        BUILD_NUMBER="$(git rev-list --count HEAD)"
-        SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-        if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-          echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-        else
-          SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-          echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-        fi
--- a/.github/actions/windows-setup-cuda/action.yml
+++ b/.github/actions/windows-setup-cuda/action.yml
@@ -1,67 +0,0 @@
-name: "Windows - Setup CUDA Toolkit"
-description: "Setup CUDA Toolkit for Windows"
-inputs:
-  cuda_version:
-    description: "CUDA toolkit version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install Cuda Toolkit 11.7
-      if: ${{ inputs.cuda_version == '11.7' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-    - name: Install Cuda Toolkit 12.4
-      if: ${{ inputs.cuda_version == '12.4' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
--- a/.github/actions/windows-setup-curl/action.yml
+++ b/.github/actions/windows-setup-curl/action.yml
@@ -5,10 +5,6 @@ inputs:
    description: 'CURL version'
    required: false
    default: '8.6.0_6'
-  architecture:
-    description: 'Architecture of the libcurl to download'
-    required: false
-    default: 'win64'
 outputs:
  curl_path:
    description: "Path to the downloaded libcurl"
@@ -22,9 +18,8 @@ runs:
      shell: powershell
      env:
        CURL_VERSION: ${{ inputs.curl_version }}
-        ARCHITECTURE: ${{ inputs.architecture }}
      run: |
-        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-${env:ARCHITECTURE}-mingw.zip"
+        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
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -45,9 +45,7 @@ build:
            - CMakePresets.json
 examples:
    - changed-files:
-        - any-glob-to-any-file:
-            - examples/**
-            - tools/**
+        - any-glob-to-any-file: examples/**
 devops:
    - changed-files:
        - any-glob-to-any-file:
@@ -72,7 +70,7 @@ android:
 server:
    - changed-files:
        - any-glob-to-any-file:
-            - tools/server/**
+            - examples/server/**
 ggml:
    - changed-files:
        - any-glob-to-any-file:
--- a/.github/workflows/bench.yml.disabled
+++ b/.github/workflows/bench.yml.disabled
@@ -27,10 +27,10 @@ on:
  push:
    branches:
      - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
  pull_request_target:
    types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
  schedule:
    -  cron: '04 2 * * *'

@@ -69,7 +69,7 @@ jobs:
      - name: Install python env
        id: pipenv
        run: |
-          cd tools/server/bench
+          cd examples/server/bench
          python3 -m venv venv
          source venv/bin/activate
          pip install -r requirements.txt
@@ -79,7 +79,7 @@ jobs:
        run: |
          wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
          tar xzf prometheus*.tar.gz --strip-components=1
-          ./prometheus --config.file=tools/server/bench/prometheus.yml &
+          ./prometheus --config.file=examples/server/bench/prometheus.yml &
          while ! nc -z localhost 9090; do
            sleep 0.1
          done
@@ -92,7 +92,7 @@ jobs:
      - name: Install k6 and xk6-sse
        id: k6_installation
        run: |
-          cd tools/server/bench
+          cd examples/server/bench
          go install go.k6.io/xk6/cmd/xk6@latest
          xk6 build master \
              --with github.com/phymbert/xk6-sse
@@ -116,7 +116,7 @@ jobs:
      - name: Download the dataset
        id: download_dataset
        run: |
-          cd tools/server/bench
+          cd examples/server/bench
          wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json

      - name: Server bench
@@ -126,7 +126,7 @@ jobs:
        run: |
          set -eux

-          cd tools/server/bench
+          cd examples/server/bench
          source venv/bin/activate
          python bench.py \
              --runner-label ${{ env.RUNNER_LABEL }} \
@@ -157,9 +157,9 @@ jobs:
          name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
          compression-level: 9
          path: |
-            tools/server/bench/*.jpg
-            tools/server/bench/*.json
-            tools/server/bench/*.log
+            examples/server/bench/*.jpg
+            examples/server/bench/*.json
+            examples/server/bench/*.log

      - name: Commit status
        uses: Sibz/github-status-action@v1
@@ -178,17 +178,17 @@ jobs:
        with:
          client_id: ${{secrets.IMGUR_CLIENT_ID}}
          path: |
-            tools/server/bench/prompt_tokens_seconds.jpg
-            tools/server/bench/predicted_tokens_seconds.jpg
-            tools/server/bench/kv_cache_usage_ratio.jpg
-            tools/server/bench/requests_processing.jpg
+            examples/server/bench/prompt_tokens_seconds.jpg
+            examples/server/bench/predicted_tokens_seconds.jpg
+            examples/server/bench/kv_cache_usage_ratio.jpg
+            examples/server/bench/requests_processing.jpg

      - name: Extract mermaid
        id: set_mermaid
        run: |
          set -eux

-          cd tools/server/bench
+          cd examples/server/bench
          PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
          echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
          echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV
--- a/.github/workflows/build-linux-cross.yml
+++ b/.github/workflows/build-linux-cross.yml
@@ -4,25 +4,18 @@ on:
  workflow_call:

 jobs:
-  ubuntu-24-riscv64-cpu-cross:
-    runs-on: ubuntu-24.04
+  ubuntu-latest-riscv64-cpu-cross:
+    runs-on: ubuntu-latest

    steps:
      - uses: actions/checkout@v4
      - name: Setup Riscv
        run: |
          sudo dpkg --add-architecture riscv64
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
+          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 \
@@ -34,7 +27,6 @@ jobs:
          cmake -B build -DCMAKE_BUILD_TYPE=Release \
                         -DGGML_OPENMP=OFF \
                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
                         -DLLAMA_BUILD_TESTS=OFF \
                         -DCMAKE_SYSTEM_NAME=Linux \
                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
@@ -48,25 +40,21 @@ jobs:

          cmake --build build --config Release -j $(nproc)

-  ubuntu-24-riscv64-vulkan-cross:
-    runs-on: ubuntu-24.04
+  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
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
+          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 \
@@ -81,7 +69,6 @@ jobs:
                         -DGGML_VULKAN=ON \
                         -DGGML_OPENMP=OFF \
                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
                         -DLLAMA_BUILD_TESTS=OFF \
                         -DCMAKE_SYSTEM_NAME=Linux \
                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
@@ -95,25 +82,21 @@ jobs:

          cmake --build build --config Release -j $(nproc)

-  ubuntu-24-arm64-vulkan-cross:
-    runs-on: ubuntu-24.04
+  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
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
+          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 \
@@ -127,7 +110,6 @@ jobs:
                         -DGGML_VULKAN=ON \
                         -DGGML_OPENMP=OFF \
                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
                         -DLLAMA_BUILD_TESTS=OFF \
                         -DCMAKE_SYSTEM_NAME=Linux \
                         -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
@@ -140,94 +122,3 @@ jobs:
                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH

          cmake --build build --config Release -j $(nproc)
-
-  ubuntu-24-ppc64el-cpu-cross:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup PowerPC64le
-        run: |
-          sudo dpkg --add-architecture ppc64el
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu \
-                  libcurl4-openssl-dev:ppc64el
-
-      - name: Build
-        run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-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-24-ppc64el-vulkan-cross:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup PowerPC64le
-        run: |
-          sudo dpkg --add-architecture ppc64el
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  glslc \
-                  gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu \
-                  libvulkan-dev:ppc64el \
-                  libcurl4-openssl-dev:ppc64el
-
-      - name: Build
-        run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_VULKAN=ON \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-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)
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -36,13 +36,10 @@ jobs:
      matrix:
        config:
          # Multi-stage build
-          # Note: the arm64 images are failing, which prevents the amd64 images from being built
-          # https://github.com/ggml-org/llama.cpp/issues/11888
-          #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, 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: 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, free_disk_space: true }
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -1,716 +0,0 @@
-name: Create Release
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      create_release:
-        description: 'Create new release'
-        required: true
-        type: boolean
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/release.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-  CMAKE_ARGS: "-DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON"
-
-jobs:
-  macOS-arm64:
-    runs-on: macos-14
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: macOS-latest-cmake-arm64
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DGGML_RPC=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
-          name: llama-bin-macos-arm64.zip
-
-  macOS-x64:
-    runs-on: macos-13
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: macOS-latest-cmake-x64
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
-          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
-          name: llama-bin-macos-x64.zip
-
-  ubuntu-22-cpu:
-    strategy:
-      matrix:
-        include:
-          - build: 'x64'
-            os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
-
-    runs-on: ${{ matrix.os }}
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: ubuntu-cpu-cmake
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j $(nproc)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
-          name: llama-bin-ubuntu-${{ matrix.build }}.zip
-
-  ubuntu-22-vulkan:
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: ubuntu-22-cmake-vulkan
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        run: |
-          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
-          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DGGML_VULKAN=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j $(nproc)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
-          name: llama-bin-ubuntu-vulkan-x64.zip
-
-  windows:
-    runs-on: windows-latest
-
-    env:
-      OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.309.0
-
-    strategy:
-      matrix:
-        include:
-          - build: 'cpu-x64'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
-          #- build: 'openblas-x64'
-          #  arch: 'x64'
-          #  defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'vulkan-x64'
-            arch: 'x64'
-            defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
-          - build: 'cpu-arm64'
-            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
-          - build: 'opencl-adreno-arm64'
-            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-${{ matrix.build }}
-          variant: ccache
-          evict-old-files: 1d
-
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
-          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/openblas
-          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
-          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
-
-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
-          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
-          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
-          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
-
-      - name: Install Ninja
-        id: install_ninja
-        run: |
-          choco install ninja
-
-      - name: Install OpenCL Headers and Libs
-        id: install_opencl
-        if: ${{ matrix.build == 'opencl-adreno-arm64' }}
-        run: |
-          git clone https://github.com/KhronosGroup/OpenCL-Headers
-          cd OpenCL-Headers
-          cmake -B build `
-            -DBUILD_TESTING=OFF `
-            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
-            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
-            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
-          cmake --build build --target install
-          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
-          cd OpenCL-ICD-Loader
-          cmake -B build-arm64-release `
-            -A arm64 `
-            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
-            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
-          cmake --build build-arm64-release --target install --config release
-
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
-
-      - name: Build
-        id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          cmake -S . -B build ${{ matrix.defines }} `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Add libopenblas.dll
-        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas-x64' }}
-        run: |
-          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
-          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
-
-  windows-cuda:
-    runs-on: windows-2019
-
-    strategy:
-      matrix:
-        cuda: ['12.4', '11.7']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
-
-      - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: windows-cuda-${{ matrix.cuda }}
-          variant: ccache
-          evict-old-files: 1d
-
-      - name: Install Cuda Toolkit
-        uses: ./.github/actions/windows-setup-cuda
-        with:
-          cuda_version: ${{ matrix.cuda }}
-
-      - name: Install Ninja
-        id: install_ninja
-        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" ^
-            -DGGML_NATIVE=OFF ^
-            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
-            -DGGML_CUDA=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
-            ${{ env.CMAKE_ARGS }}
-          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-
-      - name: Copy and pack Cuda runtime
-        run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
-          $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\*
-
-      - name: Upload Cuda runtime
-        uses: actions/upload-artifact@v4
-        with:
-          path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
-
-  windows-sycl:
-    runs-on: windows-latest
-
-    defaults:
-      run:
-        shell: bash
-
-    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-sycl
-          variant: ccache
-          evict-old-files: 1d
-
-      - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
-      # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
-      - name: Build
-        id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Build the release package
-        id: pack_artifacts
-        run: |
-          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
-
-          echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
-
-      - name: Upload the release package
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
-          name: llama-bin-win-sycl-x64.zip
-
-  windows-hip:
-    runs-on: windows-latest
-
-    strategy:
-      matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
-
-      - name: Clone rocWMMA repository
-        id: clone_rocwmma
-        run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
-
-      - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-hip-release
-          evict-old-files: 1d
-
-      - name: Install
-        id: depends
-        run: |
-          $ErrorActionPreference = "Stop"
-          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP SDK"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        id: verify
-        run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
-
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-
-      - name: Build
-        id: cmake_build
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DGGML_HIP=ON `
-            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" `
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
-          md "build\bin\rocblas\library\"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
-        run: |
-          cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-
-  ios-xcode-build:
-    runs-on: macos-latest
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_CURL=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=iOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          ./build-xcframework.sh
-
-      - name: Build Xcode project
-        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
-          name: llama-${{ steps.tag.outputs.name }}-xcframework
-
-  release:
-    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-
-    # Fine-grant permission
-    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-    permissions:
-        contents: write # for creating release
-
-    runs-on: ubuntu-latest
-
-    needs:
-      - ubuntu-22-cpu
-      - ubuntu-22-vulkan
-      - windows
-      - windows-cuda
-      - windows-sycl
-      - windows-hip
-      - macOS-arm64
-      - macOS-x64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Download artifacts
-        id: download-artifact
-        uses: actions/download-artifact@v4
-        with:
-          path: ./artifact
-
-      - name: Move artifacts
-        id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
-
-      - name: Create release
-        id: create_release
-        uses: ggml-org/action-create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          tag_name: ${{ steps.tag.outputs.name }}
-
-      - name: Upload release
-        id: upload_release
-        uses: actions/github-script@v3
-        with:
-          github-token: ${{secrets.GITHUB_TOKEN}}
-          script: |
-            const path = require('path');
-            const fs = require('fs');
-            const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
-              if (path.extname(file) === '.zip') {
-                console.log('uploadReleaseAsset', file);
-                await github.repos.uploadReleaseAsset({
-                  owner: context.repo.owner,
-                  repo: context.repo.repo,
-                  release_id: release_id,
-                  name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
-                });
-              }
-            }
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -15,10 +15,10 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']

 env:
  LLAMA_LOG_COLORS: 1
@@ -74,7 +74,7 @@ jobs:
      - name: Tests dependencies
        id: test_dependencies
        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip install -r examples/server/tests/requirements.txt

      # Setup nodejs (to be used for verifying bundled index.html)
      - uses: actions/setup-node@v4
@@ -84,14 +84,14 @@ jobs:
      - name: WebUI - Install dependencies
        id: webui_lint
        run: |
-          cd tools/server/webui
+          cd examples/server/webui
          npm ci

      - name: WebUI - Check code format
        id: webui_format
        run: |
          git config --global --add safe.directory $(realpath .)
-          cd tools/server/webui
+          cd examples/server/webui
          git status

          npm run format
@@ -108,7 +108,7 @@ jobs:
        id: verify_server_index_html
        run: |
          git config --global --add safe.directory $(realpath .)
-          cd tools/server/webui
+          cd examples/server/webui
          git status

          npm run build
@@ -161,21 +161,21 @@ jobs:
        env:
          GITHUB_ACTIONS: "true"
        run: |
-          cd tools/server/tests
+          cd examples/server/tests
          ./tests.sh

      - name: Tests (sanitizers)
        id: server_integration_tests_sanitizers
        if: ${{ matrix.sanitizer != '' }}
        run: |
-          cd tools/server/tests
+          cd examples/server/tests
          LLAMA_SANITIZE=1 ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
-          cd tools/server/tests
+          cd examples/server/tests
          SLOW_TESTS=1 ./tests.sh


@@ -211,7 +211,7 @@ jobs:
      - name: Tests dependencies
        id: test_dependencies
        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip install -r examples/server/tests/requirements.txt

      - name: Copy Libcurl
        id: prepare_libcurl
@@ -224,7 +224,7 @@ jobs:
        id: server_integration_tests
        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
-          cd tools/server/tests
+          cd examples/server/tests
          $env:PYTHONIOENCODING = ":replace"
          pytest -v -x -m "not slow"

@@ -232,6 +232,6 @@ jobs:
        id: server_integration_tests_slow
        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
-          cd tools/server/tests
+          cd examples/server/tests
          $env:SLOW_TESTS = "1"
          pytest -v -x
--- a/.gitignore
+++ b/.gitignore
@@ -96,11 +96,11 @@ perf-*.txt
 # Examples

 examples/jeopardy/results.txt
-tools/server/*.css.hpp
-tools/server/*.html.hpp
-tools/server/*.js.hpp
-tools/server/*.mjs.hpp
-tools/server/*.gz.hpp
+examples/server/*.css.hpp
+examples/server/*.html.hpp
+examples/server/*.js.hpp
+examples/server/*.mjs.hpp
+examples/server/*.gz.hpp
 !build_64.sh
 !examples/*.bat
 !examples/*/*.kts
@@ -110,7 +110,7 @@ tools/server/*.gz.hpp

 # Server Web UI temporary files
 node_modules
-tools/server/webui/dist
+examples/server/webui/dist

 # Python

--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -77,7 +77,6 @@ option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE

 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})

@@ -188,10 +187,6 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
    add_subdirectory(pocs)
 endif()

-if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
-    add_subdirectory(tools)
-endif()
-
 #
 # install
 #
@@ -252,3 +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()
+
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -38,6 +38,15 @@
        }
    },

+    {
+        "name": "arm64-windows-msvc", "hidden": true,
+        "architecture": { "value": "arm64",    "strategy": "external" },
+        "toolset":      { "value": "host=x64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
+        }
+    },
+
    {
        "name": "arm64-windows-llvm", "hidden": true,
        "architecture": { "value": "arm64",    "strategy": "external" },
@@ -64,6 +73,10 @@
    { "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
    { "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },

+    { "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
+
    { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] },
    { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] },
    { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] },
--- a/2
+++ b/2
@@ -2,7 +2,7 @@

 /ci/ @ggerganov
 /.devops/*.Dockerfile @ngxson
-/tools/server/ @ngxson
+/examples/server/ @ngxson
 /ggml/src/ggml-cuda/fattn* @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.* @JohannesGaessler
 /ggml/src/ggml-cuda/mmv.* @JohannesGaessler
--- a/97
+++ b/97
@@ -1156,10 +1156,10 @@ $(LIB_COMMON_S): $(OBJ_COMMON)

 # Clean generated server assets
 clean-server-assets:
-	find tools/server -type f -name "*.js.hpp"   -delete
-	find tools/server -type f -name "*.mjs.hpp"  -delete
-	find tools/server -type f -name "*.css.hpp"  -delete
-	find tools/server -type f -name "*.html.hpp" -delete
+	find examples/server -type f -name "*.js.hpp"   -delete
+	find examples/server -type f -name "*.mjs.hpp"  -delete
+	find examples/server -type f -name "*.css.hpp"  -delete
+	find examples/server -type f -name "*.html.hpp" -delete

 # Clean rule
 clean: clean-server-assets
@@ -1179,7 +1179,7 @@ clean: clean-server-assets
 # Helper function that replaces .c, .cpp, and .cu file endings with .o:
 GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))

-llama-cli: tools/main/main.cpp \
+llama-cli: examples/main/main.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1187,7 +1187,12 @@ llama-cli: tools/main/main.cpp \
 	@echo '====  Run ./llama-cli -h for help.  ===='
 	@echo

-llama-run: tools/run/run.cpp \
+llama-infill: examples/infill/infill.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+llama-run: examples/run/run.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1202,7 +1207,7 @@ llama-simple-chat: examples/simple-chat/simple-chat.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-tokenize: tools/tokenize/tokenize.cpp \
+llama-tokenize: examples/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1212,27 +1217,27 @@ llama-batched: examples/batched/batched.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-batched-bench: tools/batched-bench/batched-bench.cpp \
+llama-batched-bench: examples/batched-bench/batched-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-quantize: tools/quantize/quantize.cpp \
+llama-quantize: examples/quantize/quantize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-quantize-stats: tools/quantize-stats/quantize-stats.cpp \
+llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-perplexity: tools/perplexity/perplexity.cpp \
+llama-perplexity: examples/perplexity/perplexity.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-imatrix: tools/imatrix/imatrix.cpp \
+llama-imatrix: examples/imatrix/imatrix.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1274,7 +1279,7 @@ llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/s
 	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-gguf-split: tools/gguf-split/gguf-split.cpp \
+llama-gguf-split: examples/gguf-split/gguf-split.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1284,7 +1289,7 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-cvector-generator: tools/cvector-generator/cvector-generator.cpp \
+llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1294,12 +1299,12 @@ llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-bench: tools/llama-bench/llama-bench.cpp \
+llama-bench: examples/llama-bench/llama-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-llama-export-lora: tools/export-lora/export-lora.cpp \
+llama-export-lora: examples/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1355,17 +1360,17 @@ llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 ifdef GGML_RPC
-rpc-server: tools/rpc/rpc-server.cpp \
+rpc-server: examples/rpc/rpc-server.cpp \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 endif # GGML_RPC

 llama-server: \
-	tools/server/server.cpp \
-	tools/server/utils.hpp \
-	tools/server/httplib.h \
-	tools/server/index.html.hpp \
-	tools/server/loading.html.hpp \
+	examples/server/server.cpp \
+	examples/server/utils.hpp \
+	examples/server/httplib.h \
+	examples/server/index.html.hpp \
+	examples/server/loading.html.hpp \
 	common/chat.cpp \
 	common/chat.h \
 	common/chat-template.hpp \
@@ -1373,10 +1378,10 @@ llama-server: \
 	common/minja.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Itools/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)

-# Portable equivalent of `cd tools/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-tools/server/%.hpp: tools/server/public/% FORCE Makefile
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% FORCE Makefile
 	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
 		echo "unsigned char $${NAME}[] = {" && \
 		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
@@ -1389,36 +1394,36 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

-libllava.a: tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
+libllava.a: examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	common/stb_image.h \
 	common/base64.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual

-llama-llava-cli: tools/mtmd/llava-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
+llama-llava-cli: examples/llava/llava-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

-llama-minicpmv-cli: tools/mtmd/minicpmv-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
+llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

-llama-qwen2vl-cli: tools/mtmd/qwen2vl-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
+llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

@@ -1475,12 +1480,12 @@ tests/test-double-float: tests/test-double-float.cpp

 tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 tests/test-chat: tests/test-chat.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

 tests/test-opt: tests/test-opt.cpp \
--- a/README.md
+++ b/README.md
@@ -16,9 +16,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)

 ## Hot topics

- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
@@ -243,7 +242,7 @@ 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/tools/rpc) | All |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |

 ## Building the project

@@ -277,9 +276,9 @@ The Hugging Face platform provides a variety of online tools for converting, qua
 - Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
 - Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)

-To learn more about model quantization, [read this documentation](tools/quantize/README.md)
+To learn more about model quantization, [read this documentation](examples/quantize/README.md)

-## [`llama-cli`](tools/main)
+## [`llama-cli`](examples/main)

 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.

@@ -342,7 +341,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
    </details>


-## [`llama-server`](tools/server)
+## [`llama-server`](examples/server)

 #### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.

@@ -412,7 +411,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
    </details>


-## [`llama-perplexity`](tools/perplexity)
+## [`llama-perplexity`](examples/perplexity)

 #### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.

@@ -437,10 +436,10 @@ To learn more about model quantization, [read this documentation](tools/quantize

    </details>

-[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md)
+[^1]: [examples/perplexity/README.md](./examples/perplexity/README.md)
 [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)

-## [`llama-bench`](tools/llama-bench)
+## [`llama-bench`](examples/llama-bench)

 #### Benchmark the performance of the inference for various parameters.

@@ -461,7 +460,7 @@ To learn more about model quantization, [read this documentation](tools/quantize

    </details>

-## [`llama-run`](tools/run)
+## [`llama-run`](examples/run)

 #### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].

@@ -505,8 +504,8 @@ To learn more about model quantization, [read this documentation](tools/quantize

 ## Other documentation

- [main (cli)](tools/main/README.md)
- [server](tools/server/README.md)
+- [main (cli)](examples/main/README.md)
+- [server](examples/server/README.md)
 - [GBNF grammars](grammars/README.md)

 #### Development documentation
@@ -572,11 +571,4 @@ automatically. For example:
 $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 ```

-## Dependencies
-
- [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
- [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
- [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
- [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
- [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
+## References
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -40,7 +40,7 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks

 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
 * Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.

--- a/build-xcframework.sh
+++ b/build-xcframework.sh
@@ -8,7 +8,6 @@ TVOS_MIN_OS_VERSION=16.4

 BUILD_SHARED_LIBS=OFF
 LLAMA_BUILD_EXAMPLES=OFF
-LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 GGML_METAL=ON
@@ -32,7 +31,6 @@ COMMON_CMAKE_ARGS=(
    -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
    -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
    -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
-    -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
    -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
    -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
@@ -117,7 +115,6 @@ setup_framework_structure() {
    # Copy all required headers (common for all platforms)
    cp include/llama.h             ${header_path}
    cp ggml/include/ggml.h         ${header_path}
-    cp ggml/include/ggml-opt.h     ${header_path}
    cp ggml/include/ggml-alloc.h   ${header_path}
    cp ggml/include/ggml-backend.h ${header_path}
    cp ggml/include/ggml-metal.h   ${header_path}
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -187,8 +187,8 @@ function gg_run_test_scripts_debug {

    set -e

-    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log

    set +e
 }
@@ -211,8 +211,8 @@ function gg_run_test_scripts_release {

    set -e

-    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log

    set +e
 }
--- a/cmake/arm64-windows-msvc.cmake
+++ b/cmake/arm64-windows-msvc.cmake
@@ -0,0 +1,6 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
--- a/cmake/build-info.cmake
+++ b/cmake/build-info.cmake
@@ -41,20 +41,14 @@ endif()

 if(MSVC)
    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    if (CMAKE_VS_PLATFORM_NAME)
-        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
-    else()
-        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
-    endif()
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
 else()
    execute_process(
-        COMMAND ${CMAKE_C_COMPILER} --version
+        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
        OUTPUT_VARIABLE OUT
        OUTPUT_STRIP_TRAILING_WHITESPACE
    )
-    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
    set(BUILD_COMPILER ${OUT})
-
    execute_process(
        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
        OUTPUT_VARIABLE OUT
--- a/cmake/x64-windows-llvm.cmake
+++ b/cmake/x64-windows-llvm.cmake
@@ -3,3 +3,9 @@ set( CMAKE_SYSTEM_PROCESSOR x86_64 )

 set( CMAKE_C_COMPILER    clang )
 set( CMAKE_CXX_COMPILER  clang++ )
+
+set( arch_c_flags "-march=native" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" )
+
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -39,9 +39,7 @@ add_custom_command(
    COMMENT "Generating build details from Git"
    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
-            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
    VERBATIM
@@ -73,8 +71,6 @@ add_library(${TARGET} STATIC
    minja/minja.hpp
    ngram-cache.cpp
    ngram-cache.h
-    regex-partial.cpp
-    regex-partial.h
    sampling.cpp
    sampling.h
    speculative.cpp
@@ -121,8 +117,8 @@ if (LLAMA_LLGUIDANCE)

    ExternalProject_Add(llguidance_ext
        GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.20 (+ fix to build on GCC 15):
-        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
+        # v0.7.10:
+        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
        PREFIX ${CMAKE_BINARY_DIR}/llguidance
        SOURCE_DIR ${LLGUIDANCE_SRC}
        BUILD_IN_SOURCE TRUE
@@ -146,27 +142,3 @@ endif ()
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
-
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if (DEFINED ENV{GITHUB_ACTIONS} AND "$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)
-        add_custom_command(
-            POST_BUILD
-            TARGET ${TARGET}
-            COMMAND ${CMAKE_COMMAND} -E copy_if_different
-                "${LICENSE_FILE}"
-                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
-            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
-        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
-    endforeach()
-endif()
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -40,7 +40,7 @@ using json = nlohmann::ordered_json;

 std::initializer_list<enum llama_example> mmproj_examples = {
    LLAMA_EXAMPLE_LLAVA,
-    LLAMA_EXAMPLE_SERVER,
+    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
 };

 static std::string read_file(const std::string & fname) {
@@ -217,11 +217,13 @@ struct curl_slist_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, const char * method_name) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
    int remaining_attempts = max_attempts;
+    char * method = nullptr;
+    curl_easy_getinfo(curl, CURLINFO_EFFECTIVE_METHOD, &method);

    while (remaining_attempts > 0) {
-        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);

        CURLcode res = curl_easy_perform(curl);
        if (res == CURLE_OK) {
@@ -341,7 +343,7 @@ static bool common_download_file_single(const std::string & url, const std::stri

        // we only allow retrying once for HEAD requests
        // this is for the use case of using running offline (no internet), retrying can be annoying
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0);
        if (!was_perform_successful) {
            head_request_ok = false;
        }
@@ -423,7 +425,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
        // 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, "GET");
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
        if (!was_perform_successful) {
            return false;
        }
@@ -1283,7 +1285,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.use_color = true;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
    add_opt(common_arg(
        {"-t", "--threads"}, "N",
        string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
@@ -1416,7 +1418,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    add_opt(common_arg(
        {"-n", "--predict", "--n-predict"}, "N",
        string_format(
-            ex == LLAMA_EXAMPLE_MAIN
+            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),
@@ -1655,7 +1657,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.input_prefix = value;
            params.enable_chat_template = false;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
    add_opt(common_arg(
        {"--in-suffix"}, "STRING",
        "string to suffix after user inputs with (default: empty)",
@@ -1663,7 +1665,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.input_suffix = value;
            params.enable_chat_template = false;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
    add_opt(common_arg(
        {"--no-warmup"},
        "skip warming up the model with an empty run",
@@ -1680,7 +1682,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.spm_infill = true;
        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL}));
    add_opt(common_arg(
        {"--samplers"}, "SAMPLERS",
        string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
@@ -2097,6 +2099,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.cache_type_v = kv_cache_type_from_str(value);
        }
    ).set_env("LLAMA_ARG_CACHE_TYPE_V"));
+    add_opt(common_arg(
+        {"--perplexity", "--all-logits"},
+        string_format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"),
+        [](common_params & params) {
+            params.logits_all = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
    add_opt(common_arg(
        {"--hellaswag"},
        "compute HellaSwag score over random tasks from datafile supplied with -f",
@@ -2204,33 +2213,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
    add_opt(common_arg(
        {"--mmproj"}, "FILE",
-        "path to a multimodal projector file. see tools/mtmd/README.md\n"
-        "note: if -hf is used, this argument can be omitted",
+        "path to a multimodal projector file. see examples/llava/README.md",
        [](common_params & params, const std::string & value) {
            params.mmproj.path = value;
        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ"));
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--mmproj-url"}, "URL",
-        "URL to a multimodal projector file. see tools/mtmd/README.md",
+        "URL to a multimodal projector file. see examples/llava/README.md",
        [](common_params & params, const std::string & value) {
            params.mmproj.url = value;
        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_URL"));
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--no-mmproj"},
        "explicitly disable multimodal projector, useful when using -hf",
        [](common_params & params) {
            params.no_mmproj = true;
        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ"));
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--no-mmproj-offload"},
        "do not offload multimodal projector to GPU",
        [](common_params & params) {
            params.mmproj_use_gpu = false;
        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
+    ).set_examples(mmproj_examples));
    add_opt(common_arg(
        {"--image"}, "FILE",
        "path to an image file. use with multimodal models. Specify multiple times for batching",
@@ -2437,13 +2445,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            }
        }
    ));
-    add_opt(common_arg(
-        {"--no-op-offload"},
-        string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
-        [](common_params & params) {
-            params.no_op_offload = true;
-        }
-    ));
    add_opt(common_arg(
        {"--lora"}, "FNAME",
        "path to LoRA adapter (can be repeated to use multiple adapters)",
@@ -2585,7 +2586,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, int value) {
            params.n_junk = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_PASSKEY, LLAMA_EXAMPLE_PARALLEL}));
+    ).set_examples({LLAMA_EXAMPLE_PASSKEY}));
    add_opt(common_arg(
        {"--pos"}, "N",
        string_format("position of the passkey in the junk text (default: %d)", params.i_pos),
@@ -2635,20 +2636,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.i_chunk = value;
        }
    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
-    add_opt(common_arg(
-        {"--parse-special"},
-        string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
-        [](common_params & params) {
-            params.parse_special = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
    add_opt(common_arg(
        {"-pps"},
        string_format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"),
        [](common_params & params) {
            params.is_pp_shared = true;
        }
-    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
+    ).set_examples({LLAMA_EXAMPLE_BENCH}));
    add_opt(common_arg(
        {"-npp"}, "n0,n1,...",
        "number of prompt tokens",
@@ -2791,10 +2785,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
    add_opt(common_arg(
        {"--cache-reuse"}, "N",
-        string_format(
-            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
-            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
-        ),
+        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
        [](common_params & params, int value) {
            params.n_cache_reuse = value;
        }
@@ -2880,16 +2871,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.chat_template = read_file(value);
        }
    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
-    add_opt(common_arg(
-        {"--no-prefill-assistant"},
-        string_format(
-            "whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)\n"
-            "when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled\n"
-        ),
-        [](common_params & params) {
-            params.prefill_assistant = false;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_PREFILL_ASSISTANT"));
    add_opt(common_arg(
        {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
        string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),
@@ -2910,7 +2891,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.simple_io = true;
        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
    add_opt(common_arg(
        {"--positive-file"}, "FNAME",
        string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -6,15 +6,6 @@

 #include <optional>

-static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
-    auto time = std::chrono::system_clock::to_time_t(now);
-    auto local_time = *std::localtime(&time);
-    std::ostringstream ss;
-    ss << std::put_time(&local_time, format.c_str());
-    auto res = ss.str();
-    return res;
-}
-
 typedef minja::chat_template common_chat_template;

 struct common_chat_templates {
@@ -33,7 +24,6 @@ struct templates_params {
    std::string grammar;
    bool add_generation_prompt = true;
    bool extract_reasoning     = true;
-    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };

 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -135,9 +125,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
            msgs.push_back(msg);
        }
    } catch (const std::exception & e) {
-        // @ngxson : disable otherwise it's bloating the API response
-        // printf("%s\n", std::string("; messages = ") + messages.dump(2));
-        throw std::runtime_error("Failed to parse messages: " + std::string(e.what()));
+        throw std::runtime_error("Failed to parse messages: " + std::string(e.what()) + "; messages = " + messages.dump(2));
    }

    return msgs;
@@ -949,83 +937,78 @@ static void expect_tool_parameters(const std::string & name, const json & parame
    }
 }

-static common_chat_params common_chat_params_init_llama_3_x(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
+static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
    auto builtin_tools = json::array();
    common_chat_params data;
-    if (!inputs.tools.is_null()) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            std::vector<std::string> tool_rules;
+    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+        std::vector<std::string> tool_rules;

-            auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
-                if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
-                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
-                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
-                    expect_tool_parameters(name, parameters, {"query"});
-                } else if (name == "python" || name == "code_interpreter") {
-                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
-                    expect_tool_parameters(name, parameters, {"code"});
-                } else {
-                    return false;
-                }
-
-                std::vector<std::string> kvs;
-                for (const auto & [key, value] : parameters.at("properties").items()) {
-                    kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
-                }
-
-                tool_rules.push_back(
-                    builder.add_rule(
-                        name + "-call",
-                        "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
-                builtin_tools.push_back(name);
-
-                return true;
-            };
-
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string name = function.at("name");
-                auto parameters = function.at("parameters");
-                builder.resolve_refs(parameters);
-
-                // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
-                if (allow_python_tag_builtin_tools) {
-                    handle_builtin_tool(name, parameters);
-                }
-                tool_rules.push_back(
-                    builder.add_rule(
-                        name + "-call",
-                        "\"{\" space "
-                        "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
-                        "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
-                        "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
-                        "\"}\" space"));
-            });
-            // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
-            data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
-                "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
-            });
-            if (!builtin_tools.empty()) {
-                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-                data.preserved_tokens.push_back("<|python_tag|>");
+        auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
+            if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
+                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
+                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
+                expect_tool_parameters(name, parameters, {"query"});
+            } else if (name == "python" || name == "code_interpreter") {
+                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
+                expect_tool_parameters(name, parameters, {"code"});
+            } else {
+                return false;
            }
-            // Allow a few empty lines on top of the usual constrained json schema space rule.
-            builder.add_rule("root", string_join(tool_rules, " | "));
-            data.additional_stops.push_back("<|eom_id|>");
+
+            std::vector<std::string> kvs;
+            for (const auto & [key, value] : parameters.at("properties").items()) {
+                kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
+            }
+
+            tool_rules.push_back(
+                builder.add_rule(
+                    name + "-call",
+                    "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
+            builtin_tools.push_back(name);
+
+            return true;
+        };
+
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string name = function.at("name");
+            auto parameters = function.at("parameters");
+            builder.resolve_refs(parameters);
+
+            // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
+            if (allow_python_tag_builtin_tools) {
+                handle_builtin_tool(name, parameters);
+            }
+            tool_rules.push_back(
+                builder.add_rule(
+                    name + "-call",
+                    "\"{\" space "
+                    "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
+                    "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
+                    "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
+                    "\"}\" space"));
        });
-        data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
-            ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
-            : COMMON_CHAT_FORMAT_LLAMA_3_X;
-    } else {
-        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    }
+        // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
+        data.grammar_triggers.push_back({
+            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+            "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
+        });
+        if (!builtin_tools.empty()) {
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+            data.preserved_tokens.push_back("<|python_tag|>");
+        }
+        // Allow a few empty lines on top of the usual constrained json schema space rule.
+        builder.add_rule("root", string_join(tool_rules, " | "));
+    });
+    data.additional_stops.push_back("<|eom_id|>");
    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
-        {"date_string", format_time(inputs.now, "%d %b %Y")},
        {"tools_in_user_message", false},
        {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
    });
+    data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
+        ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
+        : COMMON_CHAT_FORMAT_LLAMA_3_X;
    return data;
 }
 static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {
@@ -1165,7 +1148,7 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
    LOG_DBG("%s\n", __func__);
    common_chat_params data;
    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
-        {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
+        {"datetime", "Jan 29 2025 13:00:00 GMT"},
        {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
    });
    if (inputs.tools.is_array() && !inputs.tools.empty()) {
@@ -1300,59 +1283,55 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in
 static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct templates_params & inputs) {
    // https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt
    common_chat_params data;
+    json tools = inputs.tools.is_null() ? inputs.tools : json::array();
+    std::string python_code_argument_name;
+    auto has_raw_python = false;

-    if (!inputs.tools.is_null()) {
-        std::string python_code_argument_name;
-        auto has_raw_python = false;
-
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            std::vector<std::string> tool_rules;
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                const auto & parameters = function.at("parameters");
-                std::string name = function.at("name");
-                if (name == "python" || name == "ipython") {
-                    if (!parameters.contains("type")) {
-                        throw std::runtime_error("Missing type in python tool");
-                    }
-                    has_raw_python = true;
-                    const auto & type = parameters.at("type");
-                    if (type == "object") {
-                        auto properties = parameters.at("properties");
-                        for (auto it = properties.begin(); it != properties.end(); ++it) {
-                            if (it.value().at("type") == "string") {
-                                if (!python_code_argument_name.empty()) {
-                                    throw std::runtime_error("Multiple string arguments found in python tool");
-                                }
-                                python_code_argument_name = it.key();
-                            }
-                        }
-                        if (python_code_argument_name.empty()) {
-                            throw std::runtime_error("No string argument found in python tool");
-                        }
-                    } else if (type != "string") {
-                        throw std::runtime_error("Invalid type in python tool: " + type.dump());
-                    }
+    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+        std::vector<std::string> tool_rules;
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            const auto & parameters = function.at("parameters");
+            std::string name = function.at("name");
+            if (name == "python" || name == "ipython") {
+                if (!parameters.contains("type")) {
+                    throw std::runtime_error("Missing type in python tool");
+                }
+                has_raw_python = true;
+                const auto & type = parameters.at("type");
+                if (type == "object") {
+                    auto properties = parameters.at("properties");
+                    for (auto it = properties.begin(); it != properties.end(); ++it) {
+                        if (it.value().at("type") == "string") {
+                            if (!python_code_argument_name.empty()) {
+                                throw std::runtime_error("Multiple string arguments found in python tool");
+                            }
+                            python_code_argument_name = it.key();
+                        }
+                    }
+                    if (python_code_argument_name.empty()) {
+                        throw std::runtime_error("No string argument found in python tool");
+                    }
+                } else if (type != "string") {
+                    throw std::runtime_error("Invalid type in python tool: " + type.dump());
                }
-                tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
-            });
-            if (has_raw_python) {
-                tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
-                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-                data.preserved_tokens.push_back("<|python_tag|>");
            }
-            auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
-            builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
-            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
+            tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
        });
-        data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
-    } else {
-        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    }
+        if (has_raw_python) {
+            tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+            data.preserved_tokens.push_back("<|python_tag|>");
+        }
+        auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
+        builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
+        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
+    });

    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
    // TODO: if (has_raw_python)
+    data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
    return data;
 }
 static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) {
@@ -1612,7 +1591,6 @@ static common_chat_params common_chat_templates_apply_jinja(
    params.extract_reasoning = inputs.extract_reasoning;
    params.tool_choice = inputs.tool_choice;
    params.grammar = inputs.grammar;
-    params.now = inputs.now;
    if (!inputs.json_schema.empty()) {
        params.json_schema = json::parse(inputs.json_schema);
    }
@@ -1664,21 +1642,21 @@ static common_chat_params common_chat_templates_apply_jinja(
        return common_chat_params_init_firefunction_v2(tmpl, params);
    }

+    // Plain handler (no tools)
+    if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+        return common_chat_params_init_without_tools(tmpl, params);
+    }
+
    // Functionary v3.1 (w/ tools)
    if (src.find("<|start_header_id|>") != std::string::npos
        && src.find("<function=") != std::string::npos) {
        return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, params);
    }

-    // Llama 3.1, 3.2, 3.3 (also requires date_string so using it even w/o tools)
+    // Llama 3.1, 3.2, 3.3 (w/ tools)
    if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
        auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
-        return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
-    }
-
-    // Plain handler (no tools)
-    if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-        return common_chat_params_init_without_tools(tmpl, params);
+        return common_chat_params_init_llama_3_1_tool_calls(tmpl, params, allow_python_tag_builtin_tools);
    }

    // Mistral Nemo (w/ tools)
--- a/common/chat.h
+++ b/common/chat.h
@@ -3,7 +3,6 @@
 #pragma once

 #include "common.h"
-#include <chrono>
 #include <string>
 #include <vector>

@@ -72,7 +71,6 @@ struct common_chat_templates_inputs {
    common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
    bool parallel_tool_calls = false;
    bool extract_reasoning     = true;
-    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };

 struct common_chat_params {
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -443,25 +443,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
    s = std::move(builder);
 }

-bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
-    if (!str.empty() && !stop.empty()) {
-        const char text_last_char = str.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
-            if (stop[char_index] == text_last_char) {
-                const auto current_partial = stop.substr(0, char_index + 1);
-                if (string_ends_with(str, current_partial)) {
-                    return str.size() - char_index - 1;
-                }
-            }
-        }
-    }
-
-    return std::string::npos;
-}
-
 std::string regex_escape(const std::string & s) {
    static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
    return std::regex_replace(s, special_chars, "\\$0");
@@ -1115,6 +1096,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.n_threads         = params.cpuparams.n_threads;
    cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+    cparams.logits_all        = params.logits_all;
    cparams.embeddings        = params.embedding;
    cparams.rope_scaling_type = params.rope_scaling_type;
    cparams.rope_freq_base    = params.rope_freq_base;
@@ -1132,7 +1114,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.offload_kqv       = !params.no_kv_offload;
    cparams.flash_attn        = params.flash_attn;
    cparams.no_perf           = params.no_perf;
-    cparams.op_offload        = !params.no_op_offload;

    if (params.reranking) {
        cparams.embeddings    = true;
@@ -1584,20 +1565,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c

    return result;
 }
-
-ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
-    const int64_t ne_datapoint = llama_n_ctx(ctx);
-    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
-    ggml_opt_dataset_t result = ggml_opt_dataset_init(
-        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
-
-    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
-    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
-
-    for (int64_t idata = 0; idata < ndata; ++idata) {
-        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
-        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
-    }
-
-    return result;
-}
--- a/common/common.h
+++ b/common/common.h
@@ -6,7 +6,6 @@

 #include <set>
 #include <string>
-#include <string_view>
 #include <vector>
 #include <sstream>

@@ -67,6 +66,7 @@ enum llama_example {
    LLAMA_EXAMPLE_COMMON,
    LLAMA_EXAMPLE_SPECULATIVE,
    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_INFILL,
    LLAMA_EXAMPLE_EMBEDDING,
    LLAMA_EXAMPLE_PERPLEXITY,
    LLAMA_EXAMPLE_RETRIEVAL,
@@ -96,7 +96,6 @@ enum common_sampler_type {
    COMMON_SAMPLER_TYPE_XTC         = 8,
    COMMON_SAMPLER_TYPE_INFILL      = 9,
    COMMON_SAMPLER_TYPE_PENALTIES   = 10,
-    COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
 };

 // dimensionality reduction methods, used by cvector-generator
@@ -162,7 +161,6 @@ struct common_params_sampling {
    std::vector<enum common_sampler_type> samplers = {
        COMMON_SAMPLER_TYPE_PENALTIES,
        COMMON_SAMPLER_TYPE_DRY,
-        COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
        COMMON_SAMPLER_TYPE_TOP_K,
        COMMON_SAMPLER_TYPE_TYPICAL_P,
        COMMON_SAMPLER_TYPE_TOP_P,
@@ -325,6 +323,7 @@ struct common_params {
    bool ctx_shift         = true;  // context shift on inifinite text generation

    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool logits_all        = false; // return logits for all tokens in the batch
    bool use_mmap          = true;  // use mmap for faster loads
    bool use_mlock         = false; // use mlock to keep model in memory
    bool verbose_prompt    = false; // print prompt tokens before generation
@@ -333,7 +332,6 @@ struct common_params {
    bool no_kv_offload     = false; // disable KV offloading
    bool warmup            = true;  // warmup run
    bool check_tensors     = false; // validate tensor data
-    bool no_op_offload     = false; // globally disable offload host tensor operations to device

    bool single_turn       = false; // single turn chat conversation

@@ -342,7 +340,7 @@ struct common_params {

    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;

-    // multimodal models (see tools/mtmd)
+    // multimodal models (see examples/llava)
    struct common_params_model mmproj;
    bool mmproj_use_gpu = true;     // use GPU for multimodal model
    bool no_mmproj = false;         // explicitly disable multimodal model
@@ -368,7 +366,6 @@ struct common_params {
    bool use_jinja = false;                                                                                 // NOLINT
    bool enable_chat_template = true;
    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
-    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response

    std::vector<std::string> api_keys;

@@ -412,14 +409,13 @@ struct common_params {

    bool process_output = false; // collect data for the output tensor
    bool compute_ppl    = true;  // whether to compute perplexity
-    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization

    // cvector-generator params
    int n_pca_batch = 100;
    int n_pca_iterations = 1000;
    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_positive_file = "tools/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";

    bool spm_infill = false; // suffix/prefix/middle pattern for infill

@@ -505,9 +501,10 @@ static bool string_starts_with(const std::string & str,
    return str.rfind(prefix, 0) == 0;
 }

-// While we wait for C++20's std::string::ends_with...
-bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
+static bool string_ends_with(const std::string & str,
+                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}

 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -667,9 +664,3 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

 }
-
-//
-// training utils
-//
-
-ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
--- a/common/llguidance.cpp
+++ b/common/llguidance.cpp
@@ -189,7 +189,6 @@ static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
        /* .tokenize_fn                        = */ llama_sampler_llg_tokenize_fn,
        /* .use_approximate_greedy_tokenize_fn = */ false,
        /* .tokenize_user_data                 = */ vocab,
-        /* .slices                             = */ nullptr,
    };

    char           error_buffer[1024];
--- a/common/minja/chat-template.hpp
+++ b/common/minja/chat-template.hpp
@@ -13,12 +13,10 @@
 #include <chrono>
 #include <cstddef>
 #include <cstdio>
-#include <ctime>
 #include <exception>
 #include <iomanip>
 #include <memory>
 #include <sstream>
-#include <stdexcept>
 #include <string>
 #include <vector>

@@ -395,8 +393,8 @@ class chat_template {

            for (const auto & message_ : adjusted_messages) {
                auto message = message_;
-                if (!message.contains("role") || (!message.contains("content") && !message.contains("tool_calls"))) {
-                    throw std::runtime_error("message must have 'role' and one of 'content' or 'tool_calls' fields: " + message.dump());
+                if (!message.contains("role") || !message.contains("content")) {
+                    throw std::runtime_error("message must have 'role' and 'content' fields: " + message.dump());
                }
                std::string role = message.at("role");

@@ -417,6 +415,7 @@ class chat_template {
                        }
                    }
                    if (polyfill_tool_calls) {
+                        auto content = message.at("content");
                        auto tool_calls = json::array();
                        for (const auto & tool_call : message.at("tool_calls")) {
                            if (tool_call.at("type") != "function") {
@@ -435,11 +434,8 @@ class chat_template {
                        auto obj = json {
                            {"tool_calls", tool_calls},
                        };
-                        if (message.contains("content")) {
-                            auto content = message.at("content");
-                            if (!content.is_null() && !content.empty()) {
-                                obj["content"] = content;
-                            }
+                        if (!content.is_null() && !content.empty()) {
+                            obj["content"] = content;
                        }
                        message["content"] = obj.dump(2);
                        message.erase("tool_calls");
--- a/common/minja/minja.hpp
+++ b/common/minja/minja.hpp
@@ -11,7 +11,6 @@
 #include <algorithm>
 #include <cctype>
 #include <cstddef>
-#include <cstdint>
 #include <cmath>
 #include <exception>
 #include <functional>
@@ -234,7 +233,7 @@ public:
      }
    } else if (is_object()) {
      if (!index.is_hashable())
-        throw std::runtime_error("Unhashable type: " + index.dump());
+        throw std::runtime_error("Unashable type: " + index.dump());
      auto it = object_->find(index.primitive_);
      if (it == object_->end())
        throw std::runtime_error("Key not found: " + index.dump());
@@ -253,7 +252,7 @@ public:
      auto index = key.get<int>();
      return array_->at(index < 0 ? array_->size() + index : index);
    } else if (object_) {
-      if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
+      if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
      auto it = object_->find(key.primitive_);
      if (it == object_->end()) return Value();
      return it->second;
@@ -262,7 +261,7 @@ public:
  }
  void set(const Value& key, const Value& value) {
    if (!object_) throw std::runtime_error("Value is not an object: " + dump());
-    if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
+    if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
    (*object_)[key.primitive_] = value;
  }
  Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
@@ -399,7 +398,7 @@ public:
      }
      return false;
    } else if (object_) {
-      if (!value.is_hashable()) throw std::runtime_error("Unhashable type: " + value.dump());
+      if (!value.is_hashable()) throw std::runtime_error("Unashable type: " + value.dump());
      return object_->find(value.primitive_) != object_->end();
    } else {
      throw std::runtime_error("contains can only be called on arrays and objects: " + dump());
@@ -417,7 +416,7 @@ public:
    return const_cast<Value*>(this)->at(index);
  }
  Value& at(const Value & index) {
-    if (!index.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
+    if (!index.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
    if (is_array()) return array_->at(index.get<int>());
    if (is_object()) return object_->at(index.primitive_);
    throw std::runtime_error("Value is not an array or object: " + dump());
@@ -677,8 +676,8 @@ public:
 class VariableExpr : public Expression {
    std::string name;
 public:
-    VariableExpr(const Location & loc, const std::string& n)
-      : Expression(loc), name(n) {}
+    VariableExpr(const Location & location, const std::string& n)
+      : Expression(location), name(n) {}
    std::string get_name() const { return name; }
    Value do_evaluate(const std::shared_ptr<Context> & context) const override {
        if (!context->contains(name)) {
@@ -1201,9 +1200,9 @@ public:

 class SliceExpr : public Expression {
 public:
-    std::shared_ptr<Expression> start, end, step;
-    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e, std::shared_ptr<Expression> && st = nullptr)
-      : Expression(loc), start(std::move(s)), end(std::move(e)), step(std::move(st)) {}
+    std::shared_ptr<Expression> start, end;
+    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
+      : Expression(loc), start(std::move(s)), end(std::move(e)) {}
    Value do_evaluate(const std::shared_ptr<Context> &) const override {
        throw std::runtime_error("SliceExpr not implemented");
    }
@@ -1220,35 +1219,18 @@ public:
        if (!index) throw std::runtime_error("SubscriptExpr.index is null");
        auto target_value = base->evaluate(context);
        if (auto slice = dynamic_cast<SliceExpr*>(index.get())) {
-          auto len = target_value.size();
-          auto wrap = [len](int64_t i) -> int64_t {
-            if (i < 0) {
-              return i + len;
-            }
-            return i;
-          };
-          int64_t step = slice->step ? slice->step->evaluate(context).get<int64_t>() : 1;
-          if (!step) {
-            throw std::runtime_error("slice step cannot be zero");
-          }
-          int64_t start = slice->start ? wrap(slice->start->evaluate(context).get<int64_t>()) : (step < 0 ? len - 1 : 0);
-          int64_t end = slice->end ? wrap(slice->end->evaluate(context).get<int64_t>()) : (step < 0 ? -1 : len);
+          auto start = slice->start ? slice->start->evaluate(context).get<int64_t>() : 0;
+          auto end = slice->end ? slice->end->evaluate(context).get<int64_t>() : (int64_t) target_value.size();
          if (target_value.is_string()) {
            std::string s = target_value.get<std::string>();
-
-            std::string result;
-            if (start < end && step == 1) {
-              result = s.substr(start, end - start);
-            } else {
-              for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
-                result += s[i];
-              }
-            }
-            return result;
-
+            if (start < 0) start = s.size() + start;
+            if (end < 0) end = s.size() + end;
+            return s.substr(start, end - start);
          } else if (target_value.is_array()) {
+            if (start < 0) start = target_value.size() + start;
+            if (end < 0) end = target_value.size() + end;
            auto result = Value::array();
-            for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
+            for (auto i = start; i < end; ++i) {
              result.push_back(target_value.at(i));
            }
            return result;
@@ -1323,8 +1305,6 @@ public:
              if (name == "iterable") return l.is_iterable();
              if (name == "sequence") return l.is_array();
              if (name == "defined") return !l.is_null();
-              if (name == "true") return l.to_bool();
-              if (name == "false") return !l.to_bool();
              throw std::runtime_error("Unknown type for 'is' operator: " + name);
            };
            auto value = eval();
@@ -1540,10 +1520,6 @@ public:
            vargs.expectArgs("endswith method", {1, 1}, {0, 0});
            auto suffix = vargs.args[0].get<std::string>();
            return suffix.length() <= str.length() && std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
-          } else if (method->get_name() == "startswith") {
-            vargs.expectArgs("startswith method", {1, 1}, {0, 0});
-            auto prefix = vargs.args[0].get<std::string>();
-            return prefix.length() <= str.length() && std::equal(prefix.begin(), prefix.end(), str.begin());
          } else if (method->get_name() == "title") {
            vargs.expectArgs("title method", {0, 0}, {0, 0});
            auto res = str;
@@ -2106,37 +2082,28 @@ private:

      while (it != end && consumeSpaces() && peekSymbols({ "[", "." })) {
        if (!consumeToken("[").empty()) {
-          std::shared_ptr<Expression> index;
-          auto slice_loc = get_location();
-          std::shared_ptr<Expression> start, end, step;
-          bool has_first_colon = false, has_second_colon = false;
-
-          if (!peekSymbols({ ":" })) {
-            start = parseExpression();
-          }
-
-          if (!consumeToken(":").empty()) {
-            has_first_colon = true;
-            if (!peekSymbols({ ":", "]" })) {
-              end = parseExpression();
-            }
+            std::shared_ptr<Expression> index;
            if (!consumeToken(":").empty()) {
-              has_second_colon = true;
-              if (!peekSymbols({ "]" })) {
-                step = parseExpression();
+              auto slice_end = parseExpression();
+              index = std::make_shared<SliceExpr>(slice_end->location, nullptr, std::move(slice_end));
+            } else {
+              auto slice_start = parseExpression();
+              if (!consumeToken(":").empty()) {
+                consumeSpaces();
+                if (peekSymbols({ "]" })) {
+                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), nullptr);
+                } else {
+                  auto slice_end = parseExpression();
+                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), std::move(slice_end));
+                }
+              } else {
+                index = std::move(slice_start);
              }
            }
-          }
+            if (!index) throw std::runtime_error("Empty index in subscript");
+            if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");

-          if ((has_first_colon || has_second_colon) && (start || end || step)) {
-            index = std::make_shared<SliceExpr>(slice_loc, std::move(start), std::move(end), std::move(step));
-          } else {
-            index = std::move(start);
-          }
-          if (!index) throw std::runtime_error("Empty index in subscript");
-          if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");
-
-          value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
+            value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
        } else if (!consumeToken(".").empty()) {
            auto identifier = parseIdentifier();
            if (!identifier) throw std::runtime_error("Expected identifier in subscript");
--- a/common/regex-partial.cpp
+++ b/common/regex-partial.cpp
@@ -1,204 +0,0 @@
-#include "regex-partial.h"
-#include "common.h"
-#include <functional>
-#include <optional>
-
-common_regex::common_regex(const std::string & pattern) :
-    pattern(pattern),
-    rx(pattern),
-    rx_reversed_partial(regex_to_reversed_partial_regex(pattern)) {}
-
-common_regex_match common_regex::search(const std::string & input, size_t pos, bool as_match) const {
-    std::smatch match;
-    if (pos > input.size()) {
-        throw std::runtime_error("Position out of bounds");
-    }
-    auto start = input.begin() + pos;
-    auto found = as_match
-        ? std::regex_match(start, input.end(), match, rx)
-        : std::regex_search(start, input.end(), match, rx);
-    if (found) {
-        common_regex_match res;
-        res.type = COMMON_REGEX_MATCH_TYPE_FULL;
-        for (size_t i = 0; i < match.size(); ++i) {
-            auto begin = pos + match.position(i);
-            res.groups.emplace_back(begin, begin + match.length(i));
-        }
-        return res;
-    }
-    std::match_results<std::string::const_reverse_iterator> srmatch;
-    if (std::regex_match(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial)) {
-        auto group = srmatch[1].str();
-        if (group.length() != 0) {
-            auto it = srmatch[1].second.base();
-            // auto position = static_cast<size_t>(std::distance(input.begin(), it));
-            if ((!as_match) || it == input.begin()) {
-                common_regex_match res;
-                res.type = COMMON_REGEX_MATCH_TYPE_PARTIAL;
-                const size_t begin = std::distance(input.begin(), it);
-                const size_t end = input.size();
-                if (begin == std::string::npos || end == std::string::npos || begin > end) {
-                    throw std::runtime_error("Invalid range");
-                }
-                res.groups.push_back({begin, end});
-                return res;
-            }
-        }
-    }
-    return {};
-}
-
-/*
-  Transforms a regex pattern to a partial match pattern that operates on a reversed input string to find partial final matches of the original pattern.
-
-  Ideally we'd like to use boost::match_partial (https://beta.boost.org/doc/libs/1_59_0/libs/regex/doc/html/boost_regex/partial_matches.html)
-  to see if a string ends with a partial regex match, but but it's not in std::regex yet.
-  Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input.
-
-  - /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:(?:d)?c)?b)?a).*
-  - /a|b/ -> (a|b).*
-  - /a*?/ -> error, could match ""
-  - /a*b/ -> ((?:b)?a*+).* (final repetitions become eager)
-  - /.*?ab/ -> ((?:b)?a).* (merge .*)
-  - /a.*?b/ -> ((?:b)?.*?a).* (keep reluctant matches)
-  - /a(bc)d/ -> ((?:(?:d)?(?:(?:c)?b))?a).*
-  - /a(bc|de)/ -> ((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a).*
-  - /ab{2,4}c/ -> abbb?b?c -> ((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a).*
-
-  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern
-  (i.e. just where the final .* starts in the inverted pattern; all other groups are turned into non-capturing groups, and reluctant quantifiers are ignored)
-*/
-std::string regex_to_reversed_partial_regex(const std::string & pattern) {
-    auto it = pattern.begin();
-    const auto end = pattern.end();
-
-    std::function<std::string()> process = [&]() {
-        std::vector<std::vector<std::string>> alternatives(1);
-        std::vector<std::string> * sequence = &alternatives.back();
-
-        while (it != end) {
-            if (*it == '[') {
-                auto start = it;
-                ++it;
-                while (it != end) {
-                    if ((*it == '\\') && (++it != end)) {
-                        ++it;
-                    } else if ((it != end) && (*it == ']')) {
-                        break;
-                    } else {
-                        ++it;
-                    }
-                }
-                if (it == end) {
-                    throw std::runtime_error("Unmatched '[' in pattern");
-                }
-                ++it;
-                sequence->push_back(std::string(start, it));
-            } else if (*it == '*' || *it == '?' || *it == '+') {
-                if (sequence->empty()) {
-                    throw std::runtime_error("Quantifier without preceding element");
-                }
-                sequence->back() += *it;
-                auto is_star = *it == '*';
-                ++it;
-                if (is_star) {
-                    if (*it == '?') {
-                        ++it;
-                    }
-                }
-            } else if (*it == '{') {
-                if (sequence->empty()) {
-                    throw std::runtime_error("Repetition without preceding element");
-                }
-                ++it;
-                auto start = it;
-                while (it != end && *it != '}') {
-                    ++it;
-                }
-                if (it == end) {
-                    throw std::runtime_error("Unmatched '{' in pattern");
-                }
-                auto parts = string_split(std::string(start, it), ",");
-                ++it;
-                if (parts.size() > 2) {
-                    throw std::runtime_error("Invalid repetition range in pattern");
-                }
-
-                auto parseOptInt = [&](const std::string & s, const std::optional<int> & def = std::nullopt) -> std::optional<int> {
-                    if (s.empty()) {
-                        return def;
-                    }
-                    return std::stoi(s);
-                };
-                auto min = parseOptInt(parts[0], 0);
-                auto max = parts.size() == 1 ? min : parseOptInt(parts[1]);
-                if (min && max && *max < *min) {
-                    throw std::runtime_error("Invalid repetition range in pattern");
-                }
-                // Brutal but... let's repeat at least min times, then ? for the delta between min & max (or * for unbounded)
-                auto part = sequence->back();
-                sequence->pop_back();
-                for (int i = 0; i < *min; i++) {
-                    sequence->push_back(part);
-                }
-                if (max) {
-                    for (int i = *min; i < *max; i++) {
-                        sequence->push_back(part + "?");
-                    }
-                } else {
-                    sequence->push_back(part + "*");
-                }
-            } else if (*it == '(') {
-                ++it;
-                if (it != end && *it == '?' && (it + 1 != end) && *(it + 1) == ':') {
-                    it += 2;
-                }
-                auto sub = process();
-                if (*it != ')') {
-                    throw std::runtime_error("Unmatched '(' in pattern");
-                }
-                ++it;
-                auto & part = sequence->emplace_back("(?:");
-                part += sub;
-                part += ")";
-            } else if (*it == ')') {
-                break;
-            } else if (*it == '|') {
-                ++it;
-                alternatives.emplace_back();
-                sequence = &alternatives.back();
-            } else if (*it == '\\' && (++it != end)) {
-                auto str = std::string("\\") + *it;
-                sequence->push_back(str);
-                ++it;
-            } else if (it != end) {
-                sequence->push_back(std::string(1, *it));
-                ++it;
-            }
-        }
-
-        // /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:d)?c)?b)?a).*
-        // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group
-        // We'll do the outermost capturing group and final .* in the enclosing function.
-        std::vector<std::string> res_alts;
-        for (const auto & parts : alternatives) {
-            auto & res = res_alts.emplace_back();
-            for (size_t i = 0; i < parts.size() - 1; i++) {
-                res += "(?:";
-            }
-            for (auto it = parts.rbegin(); it != parts.rend(); ++it) {
-                res += *it;
-                if (it != parts.rend() - 1) {
-                    res += ")?";
-                }
-            }
-        }
-        return string_join(res_alts, "|");
-    };
-    auto res = process();
-    if (it != end) {
-        throw std::runtime_error("Unmatched '(' in pattern");
-    }
-
-    return "(" + res + ")[\\s\\S]*";
-}
--- a/common/regex-partial.h
+++ b/common/regex-partial.h
@@ -1,56 +0,0 @@
-#pragma once
-
-#include <regex>
-#include <string>
-
-enum common_regex_match_type {
-    COMMON_REGEX_MATCH_TYPE_NONE,
-    COMMON_REGEX_MATCH_TYPE_PARTIAL,
-    COMMON_REGEX_MATCH_TYPE_FULL,
-};
-
-struct common_string_range {
-    size_t begin;
-    size_t end;
-    common_string_range(size_t begin, size_t end) : begin(begin), end(end) {
-        if (begin > end) {
-            throw std::runtime_error("Invalid range");
-        }
-    }
-    // prevent default ctor
-    common_string_range() = delete;
-    bool empty() const {
-        return begin == end;
-    }
-    bool operator==(const common_string_range & other) const {
-        return begin == other.begin && end == other.end;
-    }
-};
-
-struct common_regex_match {
-    common_regex_match_type type = COMMON_REGEX_MATCH_TYPE_NONE;
-    std::vector<common_string_range> groups;
-
-    bool operator==(const common_regex_match & other) const {
-        return type == other.type && groups == other.groups;
-    }
-    bool operator!=(const common_regex_match & other) const {
-        return !(*this == other);
-    }
-};
-
-class common_regex {
-    std::string pattern;
-    std::regex rx;
-    std::regex rx_reversed_partial;
-
-  public:
-    explicit common_regex(const std::string & pattern);
-
-    common_regex_match search(const std::string & input, size_t pos, bool as_match = false) const;
-
-    const std::string & str() const { return pattern; }
-};
-
-// For testing only (pretty print of failures).
-std::string regex_to_reversed_partial_regex(const std::string & pattern);
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -1,7 +1,6 @@
 #include "sampling.h"

 #include "common.h"
-#include "log.h"

 #include <cmath>
 #include <unordered_map>
@@ -230,48 +229,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                params.logit_bias.data()));

    if (params.mirostat == 0) {
-        for (const auto & cnstr : params.samplers) {
-            switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char *> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
+        if (params.top_n_sigma >= 0) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+        } else {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                        {
+                            std::vector<const char *> c_breakers;
+                            c_breakers.reserve(params.dry_sequence_breakers.size());
+                            for (const auto & str : params.dry_sequence_breakers) {
+                                c_breakers.push_back(str.c_str());
+                            }

-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
-                    }
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
-                    break;
-                case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                    break;
-                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                    break;
-                case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                    break;
-                default:
-                    GGML_ASSERT(false && "unknown sampler type");
+                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        }
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    case COMMON_SAMPLER_TYPE_INFILL:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                        break;
+                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
            }
        }
        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@@ -473,7 +475,6 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
        case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
-        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's';
        case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
        case COMMON_SAMPLER_TYPE_XTC:         return 'x';
@@ -489,7 +490,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
        case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
        case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
-        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma";
        case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
        case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
        case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
@@ -504,7 +504,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
        { "dry",         COMMON_SAMPLER_TYPE_DRY },
        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
        { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
@@ -518,7 +517,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
    std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
@@ -535,16 +533,14 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
        auto sampler = sampler_canonical_name_map.find(name);
        if (sampler != sampler_canonical_name_map.end()) {
            samplers.push_back(sampler->second);
-            continue;
-        }
-        if (allow_alt_names) {
-            sampler = sampler_alt_name_map.find(name);
-            if (sampler != sampler_alt_name_map.end()) {
-                samplers.push_back(sampler->second);
-                continue;
+        } else {
+            if (allow_alt_names) {
+                sampler = sampler_alt_name_map.find(name);
+                if (sampler != sampler_alt_name_map.end()) {
+                    samplers.push_back(sampler->second);
+                }
            }
        }
-        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
    }

    return samplers;
@@ -556,7 +552,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
@@ -571,8 +566,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
        const auto sampler = sampler_name_map.find(c);
        if (sampler != sampler_name_map.end()) {
            samplers.push_back(sampler->second);
-        } else {
-            LOG_WRN("%s: unable to match sampler by char '%c'\n", __func__, c);
        }
    }

--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -308,7 +308,6 @@ class ModelBase:
                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
                            gguf.MODEL_TENSOR.POSNET_NORM1,
                            gguf.MODEL_TENSOR.POSNET_NORM2,
-                            gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
                        )
                    )
                    or not new_name.endswith(".weight")
@@ -420,18 +419,12 @@ class ModelBase:
    @staticmethod
    def load_hparams(dir_model: Path):
        try:
-            # for security reason, we don't allow loading remote code by default
-            # if a model need remote code, we will fallback to config.json
-            return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
+            return AutoConfig.from_pretrained(dir_model).to_dict()
        except Exception as e:
            logger.warning(f"Failed to load model config from {dir_model}: {e}")
            logger.warning("Trying to load config.json instead")
            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-                config = json.load(f)
-                if "llm_config" in config:
-                    # rename for InternVL
-                    config["text_config"] = config["llm_config"]
-                return config
+                return json.load(f)

    @classmethod
    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -460,12 +453,8 @@ class ModelBase:


 class TextModel(ModelBase):
-    model_type = ModelType.TEXT
-    hf_arch: str
-
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
-        self.hf_arch = get_model_architecture(self.hparams, self.model_type)

        if "text_config" in self.hparams:
            # move the text_config to the root level
@@ -515,7 +504,7 @@ class TextModel(ModelBase):
    def set_gguf_parameters(self):
        self.gguf_writer.add_block_count(self.block_count)

-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
            self.gguf_writer.add_context_length(n_ctx)
            logger.info(f"gguf: context length = {n_ctx}")

@@ -799,9 +788,6 @@ class TextModel(ModelBase):
        if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
            # ref: https://huggingface.co/mistral-community/pixtral-12b
            res = "pixtral"
-        if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
-            # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
-            res = "seed-coder"

        if res is None:
            logger.warning("\n")
@@ -1087,36 +1073,10 @@ class TextModel(ModelBase):
        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None:
            self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0])

-    def _try_set_pooling_type(self) -> None:
-        # get pooling path
-        pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
-
-        # get pooling type
-        if pooling_path is not None:
-            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
-                pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
-                pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
-                pooling_type = gguf.PoolingType.CLS
-            elif pooling["pooling_mode_lasttoken"]:
-                pooling_type = gguf.PoolingType.LAST
-            else:
-                raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
-

 class VisionModel(ModelBase):
-    model_type = ModelType.VISION
    model_arch = gguf.MODEL_ARCH.CLIP_VISION
+    n_text_embd = 0
    preprocessor_config: dict[str, Any]
    global_config: dict[str, Any]

@@ -1127,8 +1087,6 @@ class VisionModel(ModelBase):
            raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")

        # get n_embd of the text model
-        if "text_config" not in self.hparams:
-            self.hparams["text_config"] = {}
        text_config = {**self.hparams, **self.hparams["text_config"]}
        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
        assert self.n_embd_text > 0, "n_embd not found in hparams"
@@ -1396,10 +1354,10 @@ class BaichuanModel(TextModel):
        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
        self.gguf_writer.add_file_type(self.ftype)

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        head_count = self.hparams["num_attention_heads"]
@@ -1520,10 +1478,10 @@ class XverseModel(TextModel):
        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
        self.gguf_writer.add_file_type(self.ftype)

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # unused
@@ -1786,12 +1744,6 @@ class LlamaModel(TextModel):
    model_arch = gguf.MODEL_ARCH.LLAMA
    undo_permute = True

-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # fix for SmolVLM2, missing `num_attention_heads` in config.json
-        if self.hf_arch == "VLlama3ForCausalLM":
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
-
    def set_vocab(self):
        try:
            self._set_vocab_sentencepiece()
@@ -1836,10 +1788,10 @@ class LlamaModel(TextModel):
            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
        self.gguf_writer.add_rope_dimension_count(rope_dim)

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    @staticmethod
    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -1947,10 +1899,7 @@ class LlamaModel(TextModel):
                raise ValueError(f"Unprocessed experts: {experts}")


-@ModelBase.register(
-    "LlavaForConditionalGeneration", # pixtral
-    "Mistral3ForConditionalGeneration", # mistral small 3.1
-)
+@ModelBase.register("LlavaForConditionalGeneration")
 class LlavaVisionModel(VisionModel):
    img_break_tok_id = -1

@@ -1959,38 +1908,17 @@ class LlavaVisionModel(VisionModel):
        if self.hparams["model_type"] == "pixtral":
            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
            self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
-            logger.info(f"Image break token id: {self.img_break_tok_id}")
+            self.img_break_tok_id = 12 # see tokenizer_config.json
        else:
            raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")

-    def get_token_id(self, token: str) -> int:
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-            added_tokens_decoder = json.load(f)['added_tokens_decoder']
-            for id_, token_data in added_tokens_decoder.items():
-                if token_data["content"] == token:
-                    return int(id_)
-        raise ValueError(f"Token '{token}' not found in tokenizer config.")
-
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hparams = self.hparams
        if hparams["model_type"] == "pixtral":
            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
-
-            # hidden_act
-            if hparams["hidden_act"] == "silu":
-                self.gguf_writer.add_vision_use_silu(True)
-            elif hparams["hidden_act"] == "gelu":
-                self.gguf_writer.add_vision_use_gelu(True)
-            else:
-                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
-
-            # spatial_merge_size
-            if "spatial_merge_size" in self.global_config:
-                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
+            self.gguf_writer.add_vision_use_silu(True)

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # unused
@@ -2070,9 +1998,6 @@ class Llama4Model(LlamaModel):
        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):
-        if name.startswith("language_model."):
-            name = name.replace("language_model.", "")
-
        # split the gate_up into gate and up
        if "gate_up_proj" in name:
            name_up = name.replace("gate_up_proj", "up_proj.weight")
@@ -2093,26 +2018,6 @@ class Llama4Model(LlamaModel):
        return super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Llama4ForConditionalGeneration")
-class Llama4VisionModel(VisionModel):
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.LLAMA4)
-        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams["norm_eps"])
-        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / self.hparams["pixel_shuffle_ratio"]))
-        assert self.hparams["hidden_act"] == "gelu"
-        self.gguf_writer.add_vision_use_gelu(True)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid # unused
-        if "multi_modal_projector" in name or "vision_model" in name:
-            # process vision tensors
-            if "positional_embedding_vlm" in name and ".weight" not in name:
-                name += ".weight"
-            return [(self.map_tensor_name(name), data_torch)]
-        return []
-
-
@ModelBase.register("Mistral3ForConditionalGeneration")
 class Mistral3Model(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA
@@ -2160,9 +2065,6 @@ class DeciModel(TextModel):
            # if n_heads_in_group is not None, then
            # _num_kv_heads[il] is num_attention_head // n_heads_in_group and
            # _num_heads[il] is num_attention_head
-            # ***dummy layer*** for nemotron 253B
-            # if n_heads_in_group is None and ffn_mult is None
-            # then _num_kv_heads[il] is 0 and _num_heads[il] is 0 and _ffn_dims is 0
            for il in range(len(_block_configs)):
                if _block_configs[il]["attention"]["n_heads_in_group"] is None:
                    if _block_configs[il]["attention"]["replace_with_linear"] is True:
@@ -2174,10 +2076,7 @@ class DeciModel(TextModel):
                else:
                    self._num_kv_heads.append(self.hparams["num_attention_heads"] // _block_configs[il]["attention"]["n_heads_in_group"])
                    self._num_heads.append(self.hparams["num_attention_heads"])
-                if _block_configs[il]["ffn"]["ffn_mult"] is None: # dummy layer
-                    _ffn_multipliers.append(0.0)
-                else:
-                    _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
+                _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
            assert self.block_count == len(self._num_kv_heads)
            assert self.block_count == len(self._num_heads)
            assert self.block_count == len(_ffn_multipliers)
@@ -2237,10 +2136,10 @@ class DeciModel(TextModel):
            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
        self.gguf_writer.add_rope_dimension_count(rope_dim)

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    @staticmethod
    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -2480,10 +2379,10 @@ class MiniCPMModel(TextModel):
        logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"]
        self.gguf_writer.add_logit_scale(logit_scale)
        logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}")
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "longrope":
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
-            logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
+        if self.hparams.get("rope_scaling") is not None:
+            if self.hparams["rope_scaling"].get("type") == "longrope":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
+                logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")

    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
        rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
@@ -2615,7 +2514,7 @@ class QwenModel(TextModel):
        self.gguf_writer.add_file_type(self.ftype)


-@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM")
+@ModelBase.register("Qwen2ForCausalLM")
 class Qwen2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.QWEN2

@@ -2627,22 +2526,11 @@ class Qwen2Model(TextModel):

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
-        self._try_set_pooling_type()
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if self.hf_arch == "Qwen2Model":
-            name = f"model.{name}"  # map to Qwen2ForCausalLM tensors
-        if "language_model." in name:
-            name = name.replace("language_model.", "") # for InternVL
-        if name.startswith("mlp") or name.startswith("vision_model"):
-            # skip visual tensors
-            return []
-        yield from super().modify_tensors(data_torch, name, bid)
+        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"])


@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
@@ -2669,138 +2557,6 @@ class Qwen2VLModel(TextModel):
        return [(self.map_tensor_name(name), data_torch)]


-@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
-class Qwen2VLVisionModel(VisionModel):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.hparams["image_size"] = self.hparams.get("image_size", 560)
-        # rename config.json values
-        self.hparams["num_attention_heads"] = self.hparams.get("num_heads")
-        self.hparams["num_hidden_layers"] = self.hparams.get("depth")
-        if "embed_dim" in self.hparams: # qwen2vl
-            self.hparams["intermediate_size"] = self.hparams.get("hidden_size")
-            self.hparams["hidden_size"] = self.hparams.get("embed_dim")
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hparams = self.hparams
-        if self.global_config['model_type'] == 'qwen2_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL)
-        elif self.global_config['model_type'] == 'qwen2_5_vl':
-            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
-            self.gguf_writer.add_vision_use_silu(True)
-            # find n_wa_pattern (window attention pattern)
-            fullatt_block_indexes = hparams.get("fullatt_block_indexes")
-            assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl"
-            n_wa_pattern = fullatt_block_indexes[0] + 1
-            # validate n_wa_pattern
-            for i in range(1, len(fullatt_block_indexes)):
-                if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern:
-                    raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}")
-            self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern)
-        else:
-            raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}")
-        # default values below are taken from HF tranformers code
-        self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
-
-    def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
-        if ".position_embd." in new_name:
-            return gguf.GGMLQuantizationType.F32
-        return False
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-        if name.startswith("visual."):
-            # process visual tensors
-            # split QKV tensors if needed
-            if ".qkv." in name:
-                if data_torch.ndim == 2: # weight
-                    c3, _ = data_torch.shape
-                else: # bias
-                    c3 = data_torch.shape[0]
-                assert c3 % 3 == 0
-                c = c3 // 3
-                wq = data_torch[:c]
-                wk = data_torch[c: c * 2]
-                wv = data_torch[c * 2:]
-                return [
-                    (self.map_tensor_name(name.replace("qkv", "q")), wq),
-                    (self.map_tensor_name(name.replace("qkv", "k")), wk),
-                    (self.map_tensor_name(name.replace("qkv", "v")), wv),
-                ]
-            elif 'patch_embed.proj.weight' in name:
-                # split Conv3D into Conv2Ds
-                c1, c2, kt, kh, kw = data_torch.shape
-                del c1, c2, kh, kw  # unused
-                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-                return [
-                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...]),
-                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
-                ]
-            else:
-                return [(self.map_tensor_name(name), data_torch)]
-        return [] # skip other tensors
-
-
-@ModelBase.register("InternVisionModel")
-class InternVisionModel(VisionModel):
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hparams = self.hparams
-        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.INTERNVL)
-        self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
-        # hidden_act
-        if hparams["hidden_act"] == "silu":
-            self.gguf_writer.add_vision_use_silu(True)
-        elif hparams["hidden_act"] == "gelu":
-            self.gguf_writer.add_vision_use_gelu(True)
-        else:
-            raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
-        # downsample_ratio
-        downsample_ratio = self.global_config.get("downsample_ratio")
-        assert downsample_ratio is not None
-        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
-
-    def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
-        if ".position_embd." in new_name:
-            return gguf.GGMLQuantizationType.F32
-        return False
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-        if name.startswith("vision_model") or name.startswith("mlp"):
-            # process visual tensors
-            # correct name
-            if name.startswith("vision_model"):
-                name = "vision_tower." + name
-            if (".ls" in name or "position_embedding" in name) and not name.endswith(".weight"):
-                name += ".weight"
-            # split QKV tensors if needed
-            if ".qkv." in name:
-                if data_torch.ndim == 2: # weight
-                    c3, _ = data_torch.shape
-                else: # bias
-                    c3 = data_torch.shape[0]
-                assert c3 % 3 == 0
-                c = c3 // 3
-                wq = data_torch[:c]
-                wk = data_torch[c: c * 2]
-                wv = data_torch[c * 2:]
-                return [
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.q_proj")), wq),
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.k_proj")), wk),
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.v_proj")), wv),
-                ]
-            return [(self.map_tensor_name(name), data_torch)]
-        return [] # skip other tensors
-
-
@ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
    model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -2853,13 +2609,6 @@ class Qwen2MoeModel(TextModel):
        if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
            self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
            logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
-        # YaRN is not enabled by default
-        # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])

    _experts: list[dict[str, Tensor]] | None = None

@@ -3127,7 +2876,7 @@ class Phi3MiniModel(TextModel):

        scale = max_pos_embds / orig_max_pos_embds

-        rope_scaling_type = rope_scaling.get('rope_type', rope_scaling.get('type', '')).lower()
+        rope_scaling_type = rope_scaling.get('type', '').lower()
        if len(rope_scaling_type) == 0:
            raise KeyError('Missing the required key rope_scaling.type')

@@ -3439,10 +3188,10 @@ class InternLM2Model(TextModel):
        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
        self.gguf_writer.add_file_type(self.ftype)
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        num_heads = self.hparams["num_attention_heads"]
@@ -3452,11 +3201,6 @@ class InternLM2Model(TextModel):
        head_dim = n_embd // num_heads
        num_groups = num_heads // q_per_kv

-        name = name.replace("language_model.", "") # InternVL
-        if name.startswith("mlp") or name.startswith("vision_model"):
-            # skip visual tensors
-            return []
-
        if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
            qkv = data_torch

@@ -3522,18 +3266,14 @@ class InternLM3Model(TextModel):
            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
        self.gguf_writer.add_rope_dimension_count(rope_dim)

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])

    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")
-        name = name.replace("language_model.", "") # InternVL
-        if name.startswith("mlp") or name.startswith("vision_model"):
-            # skip visual tensors
-            return []
        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")):
@@ -3552,7 +3292,29 @@ class BertModel(TextModel):
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        self.gguf_writer.add_causal_attention(False)
-        self._try_set_pooling_type()
+
+        # get pooling path
+        pooling_path = None
+        module_path = self.dir_model / "modules.json"
+        if module_path.is_file():
+            with open(module_path, encoding="utf-8") as f:
+                modules = json.load(f)
+            for mod in modules:
+                if mod["type"] == "sentence_transformers.models.Pooling":
+                    pooling_path = mod["path"]
+                    break
+
+        # get pooling type
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            else:
+                raise NotImplementedError("Only MEAN and CLS pooling types supported")
+            self.gguf_writer.add_pooling_type(pooling_type)

    def set_vocab(self):
        tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3761,13 +3523,8 @@ class NomicBertModel(BertModel):
        if self._tokenizer_is_xlmroberta:
            self._xlmroberta_tokenizer_init()

-        npos, mtp = self.hparams["n_positions"], self.hparams.get("max_trained_positions", 2048)
-        if npos == 8192 and mtp == 2048:
-            self.hparams["n_positions"] = 2048  # nomic-embed-text v1 and v1.5 are trained for 2048 tokens.
-        elif npos == 2048 and mtp == 2048:
-            self.hparams["n_positions"] = 512   # nomic-embed-text-v2-moe is trained for 512 tokens.
-        else:
-            raise ValueError(f"unrecognized parameters: n_positions={npos}, max_trained_positions={mtp}")
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048

        assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu"

@@ -4016,16 +3773,6 @@ class Gemma3VisionModel(VisionModel):
        # default values below are taken from HF tranformers code
        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
        self.gguf_writer.add_vision_use_gelu(True)
-        # calculate proj_scale_factor (used by tinygemma3 test model)
-        image_seq_length = self.preprocessor_config.get("image_seq_length", 256)
-        n_per_side = int(image_seq_length ** 0.5)
-        image_size = self.hparams["image_size"]
-        patch_size = self.hparams["patch_size"]
-        proj_scale_factor = (image_size // patch_size) // n_per_side
-        if proj_scale_factor > 0 and proj_scale_factor != 4:
-            # we only need to write this if it's not the default value
-            # in this case, we are converting a test model
-            self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)

    def tensor_force_quant(self, name, new_name, bid, n_dims):
        del bid, new_name, n_dims  # unused
@@ -4039,9 +3786,6 @@ class Gemma3VisionModel(VisionModel):
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # unused

-        if "vision_model.head." in name:
-            return [] # skip redundant tensors for tinygemma3
-
        if name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
                or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
            # process vision tensors
@@ -4967,12 +4711,12 @@ class DeepseekV2Model(TextModel):

        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])

-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
-            self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_scaling["mscale_all_dim"])
+        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"])
+                self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])

    _experts: list[dict[str, Tensor]] | None = None

@@ -5464,11 +5208,11 @@ class Glm4Model(TextModel):
        super().set_gguf_parameters()
        rope_dim = self.hparams["head_dim"]
        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+        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"])


@ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
@@ -5701,10 +5445,10 @@ class ExaoneModel(TextModel):
        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
        rotary_factor = rotary_factor if rotary_factor is not None else 1.0
        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
+            if hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])

    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
@@ -5770,20 +5514,11 @@ class GraniteModel(LlamaModel):
            logger.info("gguf: (granite) logits_scale = %s", logits_scale)


-@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
+@ModelBase.register("GraniteMoeForCausalLM")
 class GraniteMoeModel(GraniteModel):
    """Conversion for IBM's GraniteMoeForCausalLM"""
    model_arch = gguf.MODEL_ARCH.GRANITE_MOE

-    def set_gguf_parameters(self):
-        """GraniteMoeShared uses GraniteMoe parameters plus the following:
-        - shared_intermediate_size
-        """
-        super().set_gguf_parameters()
-        if shared_feed_forward_length := self.hparams.get("shared_intermediate_size"):
-            self.gguf_writer.add_expert_shared_feed_forward_length(shared_feed_forward_length)
-            logger.info("gguf: (granitemoeshared) shared_feed_forward_length = %s", shared_feed_forward_length)
-
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        """In modeling_granitemoe, the JetMoe implementation of parallel experts
        is used. This essentially merges w1 and w3 into a single tensor with 2x
@@ -5794,21 +5529,12 @@ class GraniteMoeModel(GraniteModel):
        if name.endswith("block_sparse_moe.input_linear.weight"):
            ffn_dim = self.hparams["intermediate_size"]
            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
-            gate, up = data_torch.split(ffn_dim, dim=-2)
+            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
            return [
                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
            ]

-        if name.endswith("shared_mlp.input_linear.weight"):
-            ffn_dim = self.hparams["shared_intermediate_size"]
-            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
-            gate, up = data_torch.split(ffn_dim, dim=-2)
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
-            ]
-
        return super().modify_tensors(data_torch, name, bid)


@@ -5825,13 +5551,7 @@ class BailingMoeModel(TextModel):
        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]

        self.gguf_writer.add_rope_dimension_count(rope_dim)
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
-        else:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        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"])
@@ -6133,7 +5853,8 @@ def split_str_to_n_bytes(split_str: str) -> int:
    return n


-def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str:
+def get_model_architecture(dir_model: Path, model_type: ModelType, hparams: Any = None) -> str:
+    hparams = ModelBase.load_hparams(dir_model) if hparams is None else hparams
    text_config = hparams.get("text_config", {})
    vision_config = hparams.get("vision_config", {})
    arch = hparams["architectures"][0]
@@ -6204,8 +5925,7 @@ def main() -> None:
    with torch.inference_mode():
        output_type = ftype_map[args.outtype]
        model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
-        hparams = ModelBase.load_hparams(dir_model)
-        model_architecture = get_model_architecture(hparams, model_type)
+        model_architecture = get_model_architecture(dir_model, model_type)
        logger.info(f"Model architecture: {model_architecture}")
        try:
            model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -116,7 +116,6 @@ models = [
    {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
    {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
-    {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]


--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -17,25 +17,25 @@

 **SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.

-**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to Intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:

 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.

 ### Llama.cpp + SYCL

-The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
-SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
+The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.

 ## Recommended Release

-The following releases are verified and recommended:
+The SYCL backend would be broken by some PRs due to no online CI.
+
+The following release is verified with good quality:

 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
-|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |ArcB580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||

@@ -106,14 +106,15 @@ SYCL backend supports Intel GPU Family:
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750, B580         |
-| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
-| Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake    |
+| Intel iGPU                    | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 |

 *Notes:*

 - **Memory**
  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
+
  - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.

 - **Execution Unit (EU)**
@@ -137,11 +138,9 @@ Note: AMD GPU support is highly experimental and is incompatible with F16.
 Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.

 ## Docker
-
-The docker build option is currently limited to *Intel GPU* targets.
+The docker build option is currently limited to *intel GPU* targets.

 ### Build image
-
 ```sh
 # Using FP16
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
@@ -149,10 +148,9 @@ docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f

 *Notes*:

-To build in default FP32 *(Slower than FP16 alternative)*, set `--build-arg="GGML_SYCL_F16=OFF"` in the previous command.
+To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.

 You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
-Check the [documentation for Docker](../docker.md) to see the available images.

 ### Run container

@@ -252,7 +250,7 @@ sycl-ls

 - **Intel GPU**

-When targeting an intel GPU, the user should expect one or more devices among the available SYCL devices. Please make sure that at least one GPU is present via `sycl-ls`, for instance `[level_zero:gpu]` in the sample output below:
+When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:

 ```
 [opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
@@ -284,7 +282,7 @@ For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:

 #### Intel GPU

-```sh
+```
 ./examples/sycl/build.sh
 ```

@@ -353,7 +351,7 @@ cmake --build build --config Release -j -v

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.

 ##### Check device

@@ -400,15 +398,11 @@ Choose one of following methods to run.

 ```sh
 ./examples/sycl/run-llama2.sh 0
-# OR
-./examples/sycl/run-llama3.sh 0
 ```
 - Use multiple devices:

 ```sh
 ./examples/sycl/run-llama2.sh
-# OR
-./examples/sycl/run-llama3.sh
 ```

 2. Command line
@@ -431,13 +425,13 @@ Examples:
 - Use device 0:

 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -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 -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer
+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:*
@@ -458,7 +452,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 1. Install GPU driver

-Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).

 2. Install Visual Studio

@@ -635,7 +629,7 @@ Once it is completed, final results will be in **build/Release/bin**

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.

 ##### Check device

@@ -654,7 +648,7 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow
 build\bin\llama-ls-sycl-device.exe
 ```

-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *Intel GPU* it would look like the following:
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -664,14 +658,13 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|

 ```
-
 #### Choose level-zero devices

 |Chosen Device ID|Setting|
 |-|-|
-|0|Default option. You may also want to `set ONEAPI_DEVICE_SELECTOR="level_zero:0"`|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|

 #### Execute

@@ -680,13 +673,7 @@ Choose one of following methods to run.
 1. Script

 ```
-examples\sycl\win-run-llama-2.bat
-```
-
-or
-
-```
-examples\sycl\win-run-llama-3.bat
+examples\sycl\win-run-llama2.bat
 ```

 2. Command line
@@ -710,13 +697,13 @@ Examples:
 - Use device 0:

 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -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 -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer
+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
 ```


@@ -727,9 +714,7 @@ Note:
 ```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
-
 Or
-
 ```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
@@ -741,17 +726,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
+| 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_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
+| 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). |
-| GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |

-1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
-
 #### Runtime

 | Name              | Value            | Function                                                                                                                  |
@@ -759,7 +741,6 @@ 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. |
-| GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |


@@ -769,7 +750,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512

 ## Q&A

- Error:  `error while loading shared libraries: libsycl.so: cannot open shared object file: No such file or directory`.
+- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.

  - Potential cause: Unavailable oneAPI installation or not set ENV variables.
  - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
@@ -798,18 +779,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512

  It's same for other projects including llama.cpp SYCL backend.

- `Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)`, `ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 3503030272 Bytes of memory on device`, or `failed to allocate SYCL0 buffer`
+- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`

-  You are running out of Device Memory.
+  Device Memory is not enough.

  |Reason|Solution|
  |-|-|
-  | The default context is too big. It leads to excessive memory usage.|Set `-c 8192` or a smaller value.|
-  | The model is too big and requires more memory than what is available.|Choose a smaller model or change to a smaller quantization, like Q5 -> Q4;<br>Alternatively, use more than one device to load model.|
+  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
+  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|

 ### **GitHub contribution**:
-Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
+Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.

 ## TODO

- Review ZES_ENABLE_SYSMAN: https://github.com/intel/compute-runtime/blob/master/programmers-guide/SYSMAN.md#support-and-limitations
+- NA
--- a/docs/development/HOWTO-add-model.md
+++ b/docs/development/HOWTO-add-model.md
@@ -9,10 +9,10 @@ Adding a model requires few steps:
 After following these steps, you can open PR.

 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
- [main](/tools/main/)
- [imatrix](/tools/imatrix/)
- [quantize](/tools/quantize/)
- [server](/tools/server/)
+- [main](/examples/main/)
+- [imatrix](/examples/imatrix/)
+- [quantize](/examples/quantize/)
+- [server](/examples/server/)

 ### 1. Convert the model to GGUF

--- a/docs/docker.md
+++ b/docs/docker.md
@@ -22,9 +22,6 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)

 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).

--- a/docs/multimodal.md
+++ b/docs/multimodal.md
@@ -1,80 +0,0 @@
-# Multimodal
-
-llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
- [llama-mtmd-cli](../tools/mtmd/README.md)
- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
-
-To enable it, can use use one of the 2 methods below:
-
- Use `-hf` option with a supported model (see a list of pre-quantized model below)
-    - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
-    - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
-
-By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload`
-
-For example:
-
-```sh
-# simple usage with CLI
-llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
-
-# simple usage with server
-llama-server -hf ggml-org/gemma-3-4b-it-GGUF
-
-# using local file
-llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf
-
-# no GPU offload
-llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
-```
-
-## Pre-quantized models
-
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/ggml-org
-
-Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
-
-NOTE: some models may require large context window, for example: `-c 8192`
-
-```sh
-# Gemma 3
-(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
-(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF
-(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF
-
-# SmolVLM
-(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF
-(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF
-(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF
-(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
-(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
-(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
-
-# Pixtral 12B
-(tool_name) -hf ggml-org/pixtral-12b-GGUF
-
-# Qwen 2 VL
-(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
-(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
-
-# Qwen 2.5 VL
-(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
-(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
-(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
-(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
-
-# Mistral Small 3.1 24B (IQ2_M quantization)
-(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
-
-# InternVL 2.5 and 3
-(tool_name) -hf ggml-org/InternVL2_5-1B-GGUF
-(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF
-(tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF
-(tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
-(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
-(tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
-
-# Llama 4 Scout
-(tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
-```
--- a/docs/multimodal/MobileVLM.md
+++ b/docs/multimodal/MobileVLM.md
@@ -33,13 +33,13 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:

 ```sh
-python ./tools/mtmd/llava_surgery.py -m path/to/MobileVLM-1.7B
+python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
 ```

 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:

 ```sh
-python ./tools/mtmd/convert_image_encoder_to_gguf.py \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
    -m path/to/clip-vit-large-patch14-336 \
    --llava-projector path/to/MobileVLM-1.7B/llava.projector \
    --output-dir path/to/MobileVLM-1.7B \
@@ -47,7 +47,7 @@ python ./tools/mtmd/convert_image_encoder_to_gguf.py \
 ```

 ```sh
-python ./tools/mtmd/convert_image_encoder_to_gguf.py \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
    -m path/to/clip-vit-large-patch14-336 \
    --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
    --output-dir path/to/MobileVLM-1.7B_V2 \
@@ -69,10 +69,10 @@ Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directo

 ## Android compile and run
 ### compile
-refer to `tools/mtmd/android/build_64.sh`
+refer to `examples/llava/android/build_64.sh`
 ```sh
-mkdir tools/mtmd/android/build_64
-cd tools/mtmd/android/build_64
+mkdir examples/llava/android/build_64
+cd examples/llava/android/build_64
 ../build_64.sh
 ```
 ### run on Android
--- a/docs/multimodal/glmedge.md
+++ b/docs/multimodal/glmedge.md
@@ -25,13 +25,13 @@ git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/T
 2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents:

 ```sh
-python ./tools/mtmd/glmedge-surgery.py -m ../model_path
+python ./examples/llava/glmedge-surgery.py -m ../model_path
 ```

 4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF:

 ```sh
-python ./tools/mtmd/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
+python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
 ```

 5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF:
--- a/docs/multimodal/llava.md
+++ b/docs/multimodal/llava.md
@@ -37,19 +37,19 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Install the required Python packages:

 ```sh
-pip install -r tools/mtmd/requirements.txt
+pip install -r examples/llava/requirements.txt
 ```

 3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:

 ```sh
-python ./tools/mtmd/llava_surgery.py -m ../llava-v1.5-7b
+python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
 ```

 4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:

 ```sh
-python ./tools/mtmd/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```

 5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
@@ -69,12 +69,12 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 2) Install the required Python packages:

 ```sh
-pip install -r tools/mtmd/requirements.txt
+pip install -r examples/llava/requirements.txt
 ```

 3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
-python tools/mtmd/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
+python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory

@@ -88,7 +88,7 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso

 5) Create the visual gguf model:
 ```console
-python ./tools/mtmd/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP

--- a/docs/multimodal/minicpmo2.6.md
+++ b/docs/multimodal/minicpmo2.6.md
@@ -29,8 +29,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)

 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model

 # quantize int4 version
--- a/docs/multimodal/minicpmv2.5.md
+++ b/docs/multimodal/minicpmv2.5.md
@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)

 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model

 # quantize int4 version
--- a/docs/multimodal/minicpmv2.6.md
+++ b/docs/multimodal/minicpmv2.6.md
@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)

 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model

 # quantize int4 version
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -12,30 +12,51 @@ llama_add_compile_flags()

 # examples

+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
 if (EMSCRIPTEN)
 else()
+    add_subdirectory(batched-bench)
    add_subdirectory(batched)
    add_subdirectory(embedding)
    add_subdirectory(eval-callback)

    add_subdirectory(gguf-hash)
+    add_subdirectory(gguf-split)
    add_subdirectory(gguf)
    add_subdirectory(gritlm)
+    add_subdirectory(imatrix)
+    add_subdirectory(infill)
+    add_subdirectory(llama-bench)
    add_subdirectory(lookahead)
    add_subdirectory(lookup)
+    add_subdirectory(main)
    add_subdirectory(parallel)
    add_subdirectory(passkey)
+    add_subdirectory(perplexity)
+    add_subdirectory(quantize)
    add_subdirectory(retrieval)
+    if (LLAMA_BUILD_SERVER)
+        add_subdirectory(server)
+    endif()
    add_subdirectory(save-load-state)
+    add_subdirectory(run)
    add_subdirectory(simple)
    add_subdirectory(simple-chat)
    add_subdirectory(speculative)
    add_subdirectory(speculative-simple)
+    add_subdirectory(tokenize)
+    add_subdirectory(tts)
    add_subdirectory(gen-docs)
-    add_subdirectory(training)
    if (NOT GGML_BACKEND_DL)
-        add_subdirectory(convert-llama2c-to-ggml)
        # these examples use the backends directly and cannot be built with dynamic loading
+        add_subdirectory(convert-llama2c-to-ggml)
+        add_subdirectory(cvector-generator)
+        add_subdirectory(export-lora)
+        add_subdirectory(llava)
+        if (GGML_RPC)
+            add_subdirectory(rpc)
+        endif()
        if (GGML_SYCL)
            add_subdirectory(sycl)
        endif()
--- a/examples/batched-bench/CMakeLists.txt
+++ b/examples/batched-bench/CMakeLists.txt
--- a/examples/batched-bench/README.md
+++ b/examples/batched-bench/README.md
--- a/examples/batched-bench/batched-bench.cpp
+++ b/examples/batched-bench/batched-bench.cpp
@@ -123,8 +123,8 @@ int main(int argc, char ** argv) {

                common_batch_clear(batch);

-                for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                    for (int i = 0; i < pp; ++i) {
+                for (int i = 0; i < pp; ++i) {
+                    for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
                        common_batch_add(batch, 0, i, { j }, false);
                    }
                }
--- a/examples/cvector-generator/CMakeLists.txt
+++ b/examples/cvector-generator/CMakeLists.txt
--- a/examples/cvector-generator/README.md
+++ b/examples/cvector-generator/README.md
--- a/examples/cvector-generator/completions.txt
+++ b/examples/cvector-generator/completions.txt
--- a/examples/cvector-generator/cvector-generator.cpp
+++ b/examples/cvector-generator/cvector-generator.cpp
--- a/examples/cvector-generator/mean.hpp
+++ b/examples/cvector-generator/mean.hpp
--- a/examples/cvector-generator/negative.txt
+++ b/examples/cvector-generator/negative.txt
--- a/examples/cvector-generator/pca.hpp
+++ b/examples/cvector-generator/pca.hpp
--- a/examples/cvector-generator/positive.txt
+++ b/examples/cvector-generator/positive.txt
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -35,14 +35,23 @@ 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, int embd_norm) {
    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);

    // clear previous kv_cache values (irrelevant for embeddings)
    llama_kv_self_clear(ctx);

    // run model
    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_encode(ctx, batch) < 0) {
-        LOG_ERR("%s : failed to encode\n", __func__);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            LOG_ERR("%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            LOG_ERR("%s : failed to decode\n", __func__);
+        }
    }

    for (int i = 0; i < batch.n_tokens; i++) {
--- a/examples/export-lora/CMakeLists.txt
+++ b/examples/export-lora/CMakeLists.txt
--- a/examples/export-lora/README.md
+++ b/examples/export-lora/README.md
--- a/examples/export-lora/export-lora.cpp
+++ b/examples/export-lora/export-lora.cpp
--- a/examples/gguf-split/CMakeLists.txt
+++ b/examples/gguf-split/CMakeLists.txt
--- a/examples/gguf-split/README.md
+++ b/examples/gguf-split/README.md
--- a/examples/gguf-split/gguf-split.cpp
+++ b/examples/gguf-split/gguf-split.cpp
--- a/examples/gguf-split/tests.sh
+++ b/examples/gguf-split/tests.sh
--- a/examples/imatrix/CMakeLists.txt
+++ b/examples/imatrix/CMakeLists.txt
--- a/examples/imatrix/README.md
+++ b/examples/imatrix/README.md
@@ -1,4 +1,4 @@
-# llama.cpp/tools/imatrix
+# llama.cpp/examples/imatrix

 Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
 More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861
--- a/examples/imatrix/imatrix.cpp
+++ b/examples/imatrix/imatrix.cpp
@@ -24,8 +24,7 @@ static void print_usage(int, char ** argv) {
    LOG("\n    %s \\\n"
            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n"
            "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
-            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...] \\\n"
-            "       [--parse-special]\n" , argv[0]);
+            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
    LOG("\n");
 }

@@ -47,7 +46,7 @@ private:
    common_params                          m_params;
    std::mutex                             m_mutex;
    int                                    m_last_call = 0;
-    std::vector<char>                      m_src1_data;
+    std::vector<float>                     m_src1_data;
    std::vector<char>                      m_ids; // the expert ids from ggml_mul_mat_id
 };

@@ -94,13 +93,11 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
    const bool is_host = ggml_backend_buffer_is_host(src1->buffer);

    if (!is_host) {
-        const size_t src1_nbytes = ggml_nbytes(src1);
-        m_src1_data.resize(src1_nbytes);
-        ggml_backend_tensor_get(src1, m_src1_data.data(), 0, src1_nbytes);
+        m_src1_data.resize(ggml_nelements(src1));
+        ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1));
    }

-    const char * data = is_host ? (const char *) src1->data : m_src1_data.data();
-    GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+    const float * data = is_host ? (const float *) src1->data : m_src1_data.data();

    // this has been adapted to the new format of storing merged experts in a single 3d tensor
    // ref: https://github.com/ggml-org/llama.cpp/pull/6387
@@ -147,7 +144,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *

                    const int64_t i11 = idx % src1->ne[1];
                    const int64_t i12 = row;
-                    const float * x = (const float *)(data + i11*src1->nb[1] + i12*src1->nb[2]);
+                    const float * x = (const float *)((const char *)data + i11*src1->nb[1] + i12*src1->nb[2]);

                    for (int j = 0; j < (int)src1->ne[0]; ++j) {
                        e.values[e_start + j] += x[j]*x[j];
@@ -183,7 +180,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
        ++e.ncall;
        LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
        for (int row = 0; row < (int)src1->ne[1]; ++row) {
-            const float * x = (const float *) (data + row * src1->nb[1]);
+            const float * x = data + row * src1->ne[0];
            for (int j = 0; j < (int)src1->ne[0]; ++j) {
                e.values[j] += x[j]*x[j];
                e.counts[j]++;
@@ -440,7 +437,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
    auto tim1 = std::chrono::high_resolution_clock::now();
    LOG_INF("%s: tokenizing the input ..\n", __func__);

-    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true, params.parse_special);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);

    auto tim2 = std::chrono::high_resolution_clock::now();
    LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
@@ -586,6 +583,7 @@ int main(int argc, char ** argv) {
    params.out_file = "imatrix.dat" ;

    params.n_ctx = 512;
+    params.logits_all = true;
    params.escape = false;

    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
--- a/examples/infill/CMakeLists.txt
+++ b/examples/infill/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET llama-infill)
+add_executable(${TARGET} infill.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
--- a/examples/infill/README.md
+++ b/examples/infill/README.md
@@ -0,0 +1,47 @@
+# llama.cpp/example/infill
+
+This example shows how to use the infill mode with Code Llama models supporting infill mode.
+Currently the 7B and 13B models support infill mode.
+
+Infill supports most of the options available in the main example.
+
+For further information have a look at the main README.md in llama.cpp/example/main/README.md
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference.
+-   `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
+
+## Input Prompts
+
+The `infill` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
+-   `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+### Example
+
+Download a model that supports infill, for example CodeLlama:
+```console
+scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
+```
+
+```bash
+./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+```
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -0,0 +1,590 @@
+#include "arg.h"
+#include "common.h"
+#include "console.h"
+#include "sampling.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static common_sampler          ** g_smpl;
+static common_params            * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+
+static bool is_interacting = false;
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting) {
+            is_interacting = true;
+        } else {
+            console::cleanup();
+            LOG("\n");
+            common_perf_print(*g_ctx, *g_smpl);
+
+            // make sure all logs are flushed
+            LOG("Interrupted by user\n");
+            common_log_pause(common_log_main());
+
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+    g_params = &params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
+        return 1;
+    }
+
+    common_init();
+
+    auto & sparams = params.sampling;
+
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.logits_all) {
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        LOG_ERR("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        LOG_ERR("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+
+    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        LOG_ERR("************\n\n");
+
+        return 0;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_INF("%s: llama backend init\n", __func__);
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    common_sampler * smpl = nullptr;
+
+    g_model = &model;
+    g_ctx = &ctx;
+    g_smpl = &smpl;
+
+    // load the model and apply lora adapter, if any
+    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
+    common_init_result llama_init = common_init_from_params(params);
+
+    model = llama_init.model.get();
+    ctx = llama_init.context.get();
+
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    const int n_ctx_train = llama_model_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG_DBG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_vocab_get_add_bos(vocab);
+    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
+
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> embd_end;
+    std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+    std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
+
+    GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
+    GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
+
+    inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
+    inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
+
+    embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+    embd_end = params.spm_infill ? inp_pfx : inp_sfx;
+    if (add_bos) {
+        embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
+    }
+    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+    const llama_token middle_token = llama_vocab_fim_mid(vocab);
+    if (middle_token >= 0) {
+        embd_inp.push_back(middle_token);
+    }
+
+    LOG_DBG("add_bos: %d\n", add_bos);
+    LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
+    LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
+    LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_vocab_bos(vocab));
+        LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
+    }
+
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    }
+
+    LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
+    LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_INF("\n");
+        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (params.n_keep > 0) {
+        LOG_INF("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_CNT("'\n");
+        }
+        LOG_INF("\n");
+    }
+
+    if (params.interactive) {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+        LOG_INF("%s: interactive mode on.\n", __func__);
+
+        if (params.input_prefix_bos) {
+            LOG_INF("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
+    }
+    smpl = common_sampler_init(model, sparams);
+
+    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
+    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
+    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
+
+    LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    LOG_INF("\n");
+    LOG_INF("\n#####  Infill mode  #####\n\n");
+    if (params.interactive) {
+        const char *control_message;
+        if (params.multiline_input) {
+            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n";
+        } else {
+            control_message = " - Press Return to return control to LLaMA.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n"
+                              " - If you want to submit another line, end your input with '\\'.\n";
+        }
+        LOG_INF("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_INF(       "%s\n", control_message);
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool input_echo = true;
+
+    int n_past     = 0;
+    int n_remain   = params.n_predict;
+    int n_consumed = 0;
+
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(console::prompt);
+
+    std::vector<llama_token> embd;
+
+    while (n_remain != 0 || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(console::error);
+                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(console::reset);
+            }
+
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() > n_ctx) {
+                if (params.n_predict == -2) {
+                    LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    break;
+                }
+
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
+
+                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_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;
+
+                LOG_DBG("after swap: n_past = %d\n", n_past);
+
+                LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                    LOG_ERR("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG_DBG("n_past = %d\n", n_past);
+            }
+
+        }
+
+        embd.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+            const llama_token id = common_sampler_sample(smpl, ctx, -1);
+
+            common_sampler_accept(smpl, id, true);
+
+            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
+
+            embd.push_back(id);
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG_DBG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                common_sampler_accept(smpl, embd_inp[n_consumed], false);
+
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo) {
+            for (auto id : embd) {
+                const std::string token_str = common_token_to_piece(ctx, id);
+                LOG("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+        }
+        // reset color to default if we there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(console::reset);
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+            // deal with eot token in infill mode
+            if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
+                if (is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
+                }
+                LOG("\n");
+                console::set_display(console::user_input);
+                std::string buffer;
+                std::string line;
+                bool another_line=true;
+                // set a new prefix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line, if so we use the old input
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_prefix = buffer;
+                }
+                buffer.clear();
+                // set a new suffix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_suffix = buffer;
+                }
+                buffer.clear();
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                if (params.escape) {
+                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
+                    string_process_escapes(params.input_prefix);
+                    string_process_escapes(params.input_suffix);
+                }
+
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+                std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
+
+                inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
+                inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
+
+                embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+                embd_end = params.spm_infill ? inp_pfx : inp_sfx;
+                if (add_bos) {
+                    embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
+                }
+                embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+                if (middle_token >= 0) {
+                    embd_inp.push_back(middle_token);
+                }
+
+                embd.clear();
+                n_remain = params.n_predict;
+                n_past = 0;
+                n_consumed = 0;
+                is_interacting = false;
+            }
+            // deal with end of generation tokens in interactive mode
+            else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
+                LOG_DBG("found EOS token\n");
+
+                if (params.interactive) {
+
+                    is_interacting = true;
+                    LOG("\n");
+                    console::set_display(console::user_input);
+               }
+            }
+
+            if (n_past > 0 && is_interacting && !params.interactive) {
+                LOG_DBG("waiting for user input\n");
+
+                if (params.input_prefix_bos) {
+                    LOG_DBG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_vocab_bos(vocab));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty()) {
+                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    LOG("%s", buffer.c_str());
+                }
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
+                        LOG("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    const auto line_inp = common_tokenize(ctx, buffer, false);
+                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
+
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << common_token_to_piece(ctx, token);
+                    }
+
+                    n_remain -= line_inp.size();
+                    LOG_DBG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG_DBG("empty line, passing control back\n");
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0) {
+                if (is_interacting) {
+                    common_sampler_reset(smpl);
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of generation
+        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) {
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+    if (!params.interactive && n_remain <= 0) {
+        LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
+    }
+
+    LOG("\n");
+    common_perf_print(ctx, smpl);
+
+    common_sampler_free(smpl);
+    llama_backend_free();
+
+    return 0;
+}
--- a/examples/llama-bench/CMakeLists.txt
+++ b/examples/llama-bench/CMakeLists.txt
--- a/examples/llama-bench/README.md
+++ b/examples/llama-bench/README.md
@@ -1,4 +1,4 @@
-# llama.cpp/tools/llama-bench
+# llama.cpp/examples/llama-bench

 Performance testing tool for llama.cpp.

@@ -20,20 +20,10 @@ Performance testing tool for llama.cpp.
 ## Syntax

 ```
-usage: llama-bench [options]
+usage: ./llama-bench [options]

 options:
  -h, --help
-  --numa <distribute|isolate|numactl>       numa mode (default: disabled)
-  -r, --repetitions <n>                     number of times to repeat each test (default: 5)
-  --prio <0|1|2|3>                          process/thread priority (default: 0)
-  --delay <0...N> (seconds)                 delay between each test (default: 0)
-  -o, --output <csv|json|jsonl|md|sql>      output format printed to stdout (default: md)
-  -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: none)
-  -v, --verbose                             verbose output
-  --progress                                print test progress indicators
-
-test parameters:
  -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
  -p, --n-prompt <n>                        (default: 512)
  -n, --n-gen <n>                           (default: 128)
@@ -43,27 +33,28 @@ test parameters:
  -ub, --ubatch-size <n>                    (default: 512)
  -ctk, --cache-type-k <t>                  (default: f16)
  -ctv, --cache-type-v <t>                  (default: f16)
-  -dt, --defrag-thold <f>                   (default: -1)
-  -t, --threads <n>                         (default: system dependent)
+  -t, --threads <n>                         (default: 8)
  -C, --cpu-mask <hex,hex>                  (default: 0x0)
  --cpu-strict <0|1>                        (default: 0)
  --poll <0...100>                          (default: 50)
  -ngl, --n-gpu-layers <n>                  (default: 99)
-  -rpc, --rpc <rpc_servers>                 (default: none)
+  -rpc, --rpc <rpc_servers>                 (default: )
  -sm, --split-mode <none|layer|row>        (default: layer)
  -mg, --main-gpu <i>                       (default: 0)
  -nkvo, --no-kv-offload <0|1>              (default: 0)
  -fa, --flash-attn <0|1>                   (default: 0)
  -mmp, --mmap <0|1>                        (default: 1)
+  --numa <distribute|isolate|numactl>       (default: disabled)
  -embd, --embeddings <0|1>                 (default: 0)
  -ts, --tensor-split <ts0/ts1/..>          (default: 0)
-  -ot --override-tensors <tensor name pattern>=<buffer type>;...
-                                            (default: disabled)
-  -nopo, --no-op-offload <0|1>              (default: 0)
+  -r, --repetitions <n>                     (default: 5)
+  --prio <0|1|2|3>                          (default: 0)
+  --delay <0...N> (seconds)                 (default: 0)
+  -o, --output <csv|json|jsonl|md|sql>      (default: md)
+  -oe, --output-err <csv|json|jsonl|md|sql> (default: none)
+  -v, --verbose                             (default: 0)

-Multiple values can be given for each parameter by separating them with ','
-or by specifying the parameter multiple times. Ranges can be given as
-'first-last' or 'first-last+step' or 'first-last*mult'.
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```

 llama-bench can perform three types of tests:
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
--- a/examples/llava/CMakeLists.txt
+++ b/examples/llava/CMakeLists.txt
@@ -1,8 +1,33 @@
+# llava (legacy)
+
+add_library(llava OBJECT
+            llava.cpp
+            llava.h
+            clip.cpp
+            clip.h
+            )
+
+target_link_libraries(llava PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(llava PUBLIC .)
+target_include_directories(llava PUBLIC ../..)
+target_include_directories(llava PUBLIC ../../common)
+
+target_compile_features(llava PRIVATE cxx_std_17)
+
+add_library(llava_static STATIC $<TARGET_OBJECTS:llava>)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(llava PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(llava PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(llava_shared SHARED $<TARGET_OBJECTS:llava>)
+    target_link_libraries(llava_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS llava_shared LIBRARY)
+endif()
+
 # mtmd

 add_library(mtmd OBJECT
            mtmd.cpp
-            mtmd-helper.cpp
            mtmd.h
            clip.cpp
            clip.h
@@ -27,10 +52,12 @@ if (BUILD_SHARED_LIBS)
 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()

@@ -45,3 +72,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+set(TARGET llama-llava-clip-quantize-cli)
+add_executable(${TARGET} clip-quantize-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-clip-quantize-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
--- a/examples/llava/README-quantize.md
+++ b/examples/llava/README-quantize.md
@@ -0,0 +1,44 @@
+# Quantizing CLIP Visual Projector
+
+This is the tool for quantizing the CLIP visual projector model. Quantization reduces the precision of the model's weights, which can significantly decrease the model size and improve inference speed, often with minimal impact on performance.
+
+## Usage
+
+To quantize a CLIP visual projector model, use the following command:
+
+```sh
+./bin/llama-llava-clip-quantize-cli /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf <type>
+```
+
+After the quantization, the visual projector can be used freely with the existing LLAVA cli (LLAVA, Qwen2VL, etc).
+
+### Arguments
+
+- `/path/to/ggml-model-f32.gguf`: The path to the input model file in FP32 or FP16 format.
+- `/path/to/ggml-model-quantized.gguf`: The path where the quantized model will be saved.
+- `<type>`: The quantization type to apply. This should be an integer corresponding to one of the quantization types defined in the `enum ggml_type`.
+
+### Quantization Types
+
+The following quantization types are supported, based on the `enum ggml_type` definition:
+
+- `2` - `q4_0`: 4-bit quantization with a single scale value.
+- `3` - `q4_1`: 4-bit quantization with a separate scale value for each block.
+- `6` - `q5_0`: 5-bit quantization with a single scale value.
+- `7` - `q5_1`: 5-bit quantization with a separate scale value for each block.
+- `8` - `q8_0`: 8-bit quantization with a single scale value.
+
+### Example
+
+To quantize a model using the `q4_0` quantization type, you would run:
+
+```sh
+./bin/llama-llava-clip-quantize-cli /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf 2
+```
+
+This command will generate a quantized model at `/path/to/ggml-model-quantized.gguf` using the `q4_0` quantization method.
+
+## Notes
+
+- Quantization can lead to a loss in model accuracy, depending on the chosen quantization type. It is recommended to evaluate the quantized model's performance on your specific task to ensure it meets your requirements.
+- The quantized model will typically be smaller in size and faster to run, making it more suitable for deployment in resource-constrained environments.
--- a/examples/llava/README.md
+++ b/examples/llava/README.md
@@ -16,7 +16,25 @@ The naming and structure related to multimodal support have evolved, which might

 ## Pre-quantized models

-See the list of pre-quantized model [here](../../docs/multimodal.md)
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
+
+```sh
+# Gemma 3
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+```

 ## How it works and what is `mmproj`?

@@ -39,20 +57,7 @@ Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advanta

 ## How to obtain `mmproj`

-Multimodal projector (`mmproj`) files are specific to each model architecture.
-
-For the following models, you can use `convert_hf_to_gguf.py` with `--mmproj` flag to get the `mmproj` file:
- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) ; See the guide [here](../../docs/multimodal/gemma3.md) - Note: 1B variant does not have vision support
- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
- Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
- InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported)
-
-For older models, please refer to the relevant guide for instructions on how to obtain or create them:
-
-NOTE: conversion scripts are located under `tools/mtmd/legacy-models`
+Multimodal projector (`mmproj`) files are specific to each model architecture. Please refer to the relevant guide for instructions on how to obtain or create them:

 - [LLaVA](../../docs/multimodal/llava.md)
 - [MobileVLM](../../docs/multimodal/MobileVLM.md)
@@ -61,3 +66,10 @@ NOTE: conversion scripts are located under `tools/mtmd/legacy-models`
 - [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
 - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
 - [IBM Granite Vision](../../docs/multimodal/granitevision.md)
+- [Google Gemma 3](../../docs/multimodal/gemma3.md)
+
+For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
--- a/examples/llava/android/adb_run.sh
+++ b/examples/llava/android/adb_run.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+model_dir="/Users/cxt/model/llm/mobileVLM/MobileVLM-1.7B_processed"
+projector_name="mmproj-model-f16.gguf"
+llama_name="ggml-model-q4_k.gguf"
+img_dir="/Users/cxt/model/llm"
+img_name="demo.jpg"
+prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:"
+# img_name="cat.jpeg"
+# prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
+
+program_dir="build_64/bin"
+binName="llama-mtmd-cli"
+n_threads=4
+
+
+deviceDir="/data/local/tmp"
+saveDir="output"
+if [ ! -d ${saveDir} ]; then
+    mkdir ${saveDir}
+fi
+
+
+function android_run() {
+    # # copy resource into device
+    # adb push ${model_dir}/${projector_name} ${deviceDir}/${projector_name}
+    # adb push ${model_dir}/${llama_name} ${deviceDir}/${llama_name}
+    adb push ${img_dir}/${img_name} ${deviceDir}/${img_name}
+    # copy program into device
+    adb push ${program_dir}/${binName} ${deviceDir}/${binName}
+    adb shell "chmod 0777 ${deviceDir}/${binName}"
+
+    # run
+    adb shell "echo cd ${deviceDir} ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 > ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt"
+    adb shell "cd ${deviceDir}; pwd; ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 >> ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt 2>&1"
+    adb pull ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt ${saveDir}
+}
+
+android_run
+
+echo "android_run is Done!"
--- a/examples/llava/android/build_64.sh
+++ b/examples/llava/android/build_64.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+cmake ../../../../ \
+-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+-DCMAKE_BUILD_TYPE=Release \
+-DANDROID_ABI="arm64-v8a" \
+-DANDROID_PLATFORM=android-23 $1
+
+make -j4
--- a/examples/llava/clip-impl.h
+++ b/examples/llava/clip-impl.h
@@ -4,7 +4,6 @@

 #include <climits>
 #include <cstdarg>
-#include <cinttypes>
 #include <string>
 #include <map>
 #include <sstream>
@@ -32,7 +31,9 @@
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
 #define KEY_PROJ_TYPE           "clip.projector_type"
-#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
+
+#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
+#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl

 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -45,7 +46,7 @@
 // tensor name constants
 //

-#define TN_POS_EMBD        "v.position_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"
@@ -54,16 +55,12 @@
 #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_ATTN_K_NORM     "%s.blk.%d.attn_k_norm.%s"
-#define TN_ATTN_Q_NORM     "%s.blk.%d.attn_q_norm.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
-#define TN_LN_1            "%s.blk.%d.ln1.%s" // layer norm
-#define TN_LN_2            "%s.blk.%d.ln2.%s" // layer norm
-#define TN_LS_1            "%s.blk.%d.ls1.%s" // layer scale
-#define TN_LS_2            "%s.blk.%d.ls2.%s" // layer scale
+#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
@@ -71,14 +68,10 @@
 #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_NORM     "mm.input_norm.weight"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
-#define TN_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
-#define TN_TOK_GLM_BOI     "adapter.boi"                // glm-edge (these embeddings are not in text model)
-#define TN_TOK_GLM_EOI     "adapter.eoi"                // glm-edge (these embeddings are not in text model)

 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -95,9 +88,6 @@
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"

-// align x to upper multiple of n
-#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
-
 enum projector_type {
    PROJECTOR_TYPE_MLP,
    PROJECTOR_TYPE_MLP_NORM,
@@ -110,8 +100,6 @@ enum projector_type {
    PROJECTOR_TYPE_IDEFICS3,
    PROJECTOR_TYPE_PIXTRAL,
    PROJECTOR_TYPE_QWEN25VL,
-    PROJECTOR_TYPE_INTERNVL,
-    PROJECTOR_TYPE_LLAMA4,
    PROJECTOR_TYPE_UNKNOWN,
 };

@@ -126,8 +114,6 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
-    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
-    { PROJECTOR_TYPE_LLAMA4,    "llama4"},
 };

 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -242,20 +228,6 @@ struct clip_image_u8_batch {

 struct clip_image_f32_batch {
    std::vector<clip_image_f32_ptr> entries;
-
-    // for llava-uhd style models, we need to know the grid size
-    // note: entries.size() == grid_x * grid_y + 1 (one overview image)
-    int grid_x = 0;
-    int grid_y = 0;
-
-    clip_image_f32_batch clone() const {
-        clip_image_f32_batch new_batch;
-        new_batch.entries.reserve(entries.size());
-        for (const auto & entry : entries) {
-            new_batch.entries.emplace_back(new clip_image_f32(*entry));
-        }
-        return new_batch;
-    }
 };

 //
@@ -366,70 +338,6 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
    }
 }

-//
-// debugging
-//
-
-static void print_tensor_shape(ggml_tensor * t) {
-    printf("%s.shape = [", t->name);
-    for (int i = 0; i < ggml_n_dims(t); ++i) {
-        printf("%" PRId64, t->ne[i]);
-        if (i < ggml_n_dims(t) - 1) {
-            printf(", ");
-        }
-    }
-    printf("]\n");
-}
-
-static void print_tensor_data(ggml_tensor * t, uint8_t * data, int64_t n) {
-    ggml_type type = t->type;
-    int64_t * ne = t->ne;
-    size_t * nb = t->nb;
-    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        printf("%s.data: [\n", t->name);
-        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
-            if (i2 == n && ne[2] > 2*n) {
-                printf("     ..., \n");
-                i2 = ne[2] - n;
-            }
-            printf("     [\n");
-            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
-                if (i1 == n && ne[1] > 2*n) {
-                    printf("      ..., \n");
-                    i1 = ne[1] - n;
-                }
-                printf("      [");
-                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
-                    if (i0 == n && ne[0] > 2*n) {
-                        printf("..., ");
-                        i0 = ne[0] - n;
-                    }
-                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
-                    float v;
-                    if (type == GGML_TYPE_F16) {
-                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
-                    } else if (type == GGML_TYPE_F32) {
-                        v = *(float *) &data[i];
-                    } else if (type == GGML_TYPE_I32) {
-                        v = (float) *(int32_t *) &data[i];
-                    } else if (type == GGML_TYPE_I16) {
-                        v = (float) *(int16_t *) &data[i];
-                    } else if (type == GGML_TYPE_I8) {
-                        v = (float) *(int8_t *) &data[i];
-                    } else {
-                        GGML_ABORT("fatal error");
-                    }
-                    printf("%8.4f", v);
-                    if (i0 < ne[0] - 1) printf(", ");
-                }
-                printf("],\n");
-            }
-            printf("     ],\n");
-        }
-        printf("    ]\n");
-    }
-}
-
 //
 // API used internally with mtmd
 //
--- a/examples/llava/clip-quantize-cli.cpp
+++ b/examples/llava/clip-quantize-cli.cpp
@@ -0,0 +1,59 @@
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+#include "ggml.h"
+
+static void print_usage(int argc, char ** argv) {
+    (void) argc;
+
+    fprintf(stderr, "usage: %s /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf type\n", argv[0]);
+    fprintf(stderr, "  type = 2 - q4_0\n");
+    fprintf(stderr, "  type = 3 - q4_1\n");
+    fprintf(stderr, "  type = 6 - q5_0\n");
+    fprintf(stderr, "  type = 7 - q5_1\n");
+    fprintf(stderr, "  type = 8 - q8_0\n");
+}
+
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        print_usage(argc, argv);
+        return 1;
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const int itype = atoi(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!clip_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us / 1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    return 0;
+}
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
--- a/examples/llava/clip.h
+++ b/examples/llava/clip.h
@@ -0,0 +1,135 @@
+#ifndef CLIP_H
+#define CLIP_H
+
+#include "ggml.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct clip_ctx;
+
+struct clip_image_size {
+    int width;
+    int height;
+};
+
+struct clip_image_f32;
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;
+
+struct clip_context_params {
+    bool use_gpu;
+    enum ggml_log_level verbosity;
+};
+
+// deprecated, use clip_init
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
+CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params);
+
+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_w, int img_h);
+
+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);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
+
+GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx),
+    "use clip_n_output_tokens instead");
+GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img),
+    "use clip_n_output_tokens instead");
+
+CLIP_API int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// for M-RoPE, this will be the number of token positions in X and Y directions
+// for other models, X will be the total number of tokens and Y will be 1
+CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// this should be equal to the embedding dimension of the text model
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
+CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
+CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
+
+CLIP_API struct clip_image_size      * clip_image_size_init();
+CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
+
+// nx, ny are the output image dimensions
+CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
+
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
+CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
+
+// use for accessing underlay data of clip_image_f32_batch
+CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
+CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
+CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
+CLIP_API struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+
+/**
+ * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
+ * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
+ */
+CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
+
+CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+
+/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
+/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
+
+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 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // CLIP_H
--- a/tools/mtmd/legacy-models/convert_image_encoder_to_gguf.py
+++ b/tools/mtmd/legacy-models/convert_image_encoder_to_gguf.py
--- a/examples/llava/deprecation-warning.cpp
+++ b/examples/llava/deprecation-warning.cpp
--- a/tools/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py
+++ b/tools/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py
--- a/tools/mtmd/legacy-models/glmedge-surgery.py
+++ b/tools/mtmd/legacy-models/glmedge-surgery.py
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@@ -0,0 +1,586 @@
+#include "clip.h"
+#include "llava.h"
+
+#include "llama.h"
+
+#include <algorithm>
+#include <cerrno>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
+#include <vector>
+#include <memory>
+
+#if defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...)
+#   define LOG_WRN(...)
+#   define LOG_ERR(...)
+#   define LOG_DBG(...)
+#else // defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#endif // defined(LLAVA_LOG_OFF)
+
+// 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;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    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.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) {
+    struct {
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_get_image_size(ctx_clip);
+    const int32_t patch_size = clip_get_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = ggml_graph_node(gf, -1);
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
+
+static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) {
+    int width = image->nx;
+    int height = image->ny;
+    int num_patches = (height / patch_size) * (width / patch_size);
+    clip_image_f32 * patch = clip_image_f32_init();
+    patch->nx = patch_size * num_patches;
+    patch->ny = patch_size;
+    patch->buf.resize(3 * patch->nx * patch->ny);
+
+    int patch_index = 0;
+
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            for (int pi = 0; pi < patch_size; ++pi) {
+                for (int pj = 0; pj < patch_size; ++pj) {
+                    int input_index = ((i + pi) * width + (j + pj)) * 3;
+                    int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
+                    patch->buf[output_index] = image->buf[input_index];
+                    patch->buf[output_index+1] = image->buf[input_index+1];
+                    patch->buf[output_index+2] = image->buf[input_index+2];
+                }
+            }
+            patch_index++;
+        }
+    }
+    return patch;
+}
+
+static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
+        LOG_ERR("%s: unable to preprocess image\n", __func__);
+        return false;
+    }
+
+    const int64_t t_img_enc_start_us = ggml_time_us();
+
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
+
+    if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(n_imgs);
+        clip_image_size load_image_size;
+
+        for (size_t i = 0; i < n_imgs; i++) {
+            const int64_t t_img_enc_step_start_us = ggml_time_us();
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            int patch_size = 14;
+            load_image_size.width = nx;
+            load_image_size.height = ny;
+            clip_add_load_image_size(ctx_clip, &load_image_size);
+
+            bool encoded = false;
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
+            if (clip_is_qwen2vl(ctx_clip)) {
+                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
+            }
+            else {
+                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) 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)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)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, nx, ny));
+            n_img_pos_out += clip_n_output_tokens(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);
+    }
+    else if (clip_is_glm(ctx_clip)){
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        load_image_size->width  = clip_image_f32_batch_nx(img_res_v.get(), 0);
+        load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
+        clip_add_load_image_size(ctx_clip, load_image_size);
+
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+        int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
+        *n_img_pos = (pos * pos + 2);
+        if (!encoded){
+            LOG_ERR("Unable to encode image \n");
+            return false;
+        }
+    }
+    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        *n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
+        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");
+
+            return false;
+        }
+    }
+    else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(n_imgs);
+        for (size_t i = 0; i < n_imgs; i++) {
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                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)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);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (size_t i = 0; i < num_gridpoints; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        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);
+
+        int n_img_pos_out;
+        clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0);
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input);
+        *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();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
+    }
+
+    LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
+    const int64_t t_img_enc_end_us = ggml_time_us();
+    float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
+
+    LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+
+    return true;
+}
+
+bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
+        // make sure that the correct mmproj was used, i.e., compare apples to apples
+    int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
+    auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
+    if (n_image_embd != n_llama_embd) {
+        LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        return false;
+    }
+    return true;
+}
+
+bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
+    if (clip_is_minicpmv(ctx_clip)) {
+        num_max_patches = 10;
+    }
+    if (clip_is_glm(ctx_clip)) {
+        num_max_patches = 1;
+    }
+    float * image_embd;
+    if (clip_is_qwen2vl(ctx_clip)) {
+        // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
+        image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
+    } else {
+        image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
+    }
+    if (!image_embd) {
+        LOG_ERR("Unable to allocate memory for image embeddings\n");
+        return false;
+    }
+
+    int n_img_pos;
+    if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
+        LOG_ERR("%s: cannot encode image, aborting\n", __func__);
+        free(image_embd);
+        return false;
+    }
+    *image_embd_out = image_embd;
+    *n_img_pos_out = n_img_pos;
+
+    return true;
+}
+
+struct llava_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
+bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
+    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
+
+    for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
+        int n_eval = image_embed->n_image_pos - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        float * embd = image_embed->embed+i*n_embd;
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
+        if (llama_decode(ctx_llama, llava_batch.batch)) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+    clip_image_u8 * img = clip_image_u8_init();
+    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
+        clip_image_u8_free(img);
+        LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
+        return NULL;
+    }
+
+    float* image_embed = NULL;
+    int n_image_pos = 0;
+    bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
+    if (!image_embed_result) {
+        clip_image_u8_free(img);
+        LOG_ERR("%s: couldn't embed the image\n", __func__);
+        return NULL;
+    }
+
+    clip_image_u8_free(img);
+    auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    result->embed = image_embed;
+    result->n_image_pos = n_image_pos;
+    return result;
+}
+
+static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
+    auto file = fopen(path, "rb");
+    if (file == NULL) {
+        LOG_ERR("%s: can't read file %s\n", __func__, path);
+        return false;
+    }
+
+    fseek(file, 0, SEEK_END);
+    auto fileSize = ftell(file);
+    fseek(file, 0, SEEK_SET);
+
+    auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
+    if (buffer == NULL) {
+        LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        perror("Memory allocation error");
+        fclose(file);
+        return false;
+    }
+    errno = 0;
+    size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
+    if (ferror(file)) {
+        LOG_ERR("read error: %s", strerror(errno));
+        free(buffer);
+        fclose(file);
+        return false;
+    }
+    if (ret != (size_t) fileSize) {
+        LOG_ERR("unexpectedly reached end of file");
+        free(buffer);
+        fclose(file);
+        return false;
+    }
+    fclose(file); // Close the file
+
+    *bytesOut = buffer;
+    *sizeOut = fileSize;
+    return true;
+}
+
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+    unsigned char* image_bytes;
+    long image_bytes_length;
+    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
+    if (!loaded) {
+        LOG_ERR("%s: failed to load %s\n", __func__, image_path);
+        return NULL;
+    }
+
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    free(image_bytes);
+
+    return embed;
+}
+
+void llava_image_embed_free(struct llava_image_embed * embed) {
+    free(embed->embed);
+    free(embed);
+}
--- a/examples/llava/llava.h
+++ b/examples/llava/llava.h
@@ -0,0 +1,49 @@
+#ifndef LLAVA_H
+#define LLAVA_H
+
+#include "ggml.h"
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAVA_API __declspec(dllexport)
+#        else
+#            define LLAVA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAVA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAVA_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct clip_ctx;
+struct llava_image_embed {
+    float * embed;
+    int n_image_pos;
+};
+
+/** sanity check for clip <-> llava embed size match */
+LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
+
+LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+
+/** build an image embed from image file bytes */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
+/** build an image embed from a path to an image filename */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
+/** free an embedding made with llava_image_embed_make_* */
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
+
+/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
+LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
--- a/tools/mtmd/legacy-models/llava_surgery.py
+++ b/tools/mtmd/legacy-models/llava_surgery.py
--- a/tools/mtmd/legacy-models/llava_surgery_v2.py
+++ b/tools/mtmd/legacy-models/llava_surgery_v2.py
--- a/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
+++ b/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
--- a/tools/mtmd/legacy-models/minicpmv-surgery.py
+++ b/tools/mtmd/legacy-models/minicpmv-surgery.py
--- a/examples/llava/mtmd-cli.cpp
+++ b/examples/llava/mtmd-cli.cpp
@@ -63,7 +63,7 @@ static void sigint_handler(int signo) {
 #endif

 struct mtmd_cli_context {
-    mtmd::context_ptr ctx_vision;
+    mtmd_context_ptr ctx_vision;
    common_init_result llama_init;

    llama_model       * model;
@@ -72,8 +72,6 @@ struct mtmd_cli_context {
    llama_batch         batch;
    int                 n_batch;

-    mtmd::bitmaps bitmaps;
-
    // 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;
@@ -92,15 +90,10 @@ struct mtmd_cli_context {
        batch = llama_batch_init(params.n_batch, 0, 1);
        n_batch = params.n_batch;

-        if (!model || !lctx) {
-            exit(1);
-        }
-
        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
            LOG_ERR("Model does not have chat template.\n");
            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
            LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
-            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
            exit(1);
        }

@@ -119,12 +112,12 @@ struct mtmd_cli_context {

    void init_vision_context(common_params & params) {
        const char * clip_path = params.mmproj.path.c_str();
-        mtmd_context_params mparams = mtmd_context_params_default();
-        mparams.use_gpu = params.mmproj_use_gpu;
-        mparams.print_timings = true;
-        mparams.n_threads = params.cpuparams.n_threads;
-        mparams.verbosity = params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO;
-        ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams));
+        ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
+            /* use_gpu */   params.mmproj_use_gpu,
+            /* timings */   true,
+            /* n_threads */ params.cpuparams.n_threads,
+            /* verbosity */ params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO,
+        }));
        if (!ctx_vision.get()) {
            LOG_ERR("Failed to load vision model from %s\n", clip_path);
            exit(1);
@@ -141,22 +134,13 @@ struct mtmd_cli_context {
            antiprompt_tokens.begin()
        );
    }
-
-    bool load_image(const std::string & fname) {
-        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(fname.c_str()));
-        if (!bmp.ptr) {
-            return false;
-        }
-        bitmaps.entries.push_back(std::move(bmp));
-        return true;
-    }
 };

 static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
    llama_tokens generated_tokens;
    for (int i = 0; i < n_predict; i++) {
        if (i > n_predict || !g_is_generating || g_is_interrupted) {
-            LOG("\n");
+            printf("\n");
            break;
        }

@@ -165,15 +149,15 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
        common_sampler_accept(smpl, token_id, true);

        if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
-            LOG("\n");
+            printf("\n");
            break; // end of generation
        }

-        LOG("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
        fflush(stdout);

        if (g_is_interrupted) {
-            LOG("\n");
+            printf("\n");
            break;
        }

@@ -188,7 +172,9 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
    return 0;
 }

-static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_bos = false) {
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
+    std::vector<mtmd_bitmap> bitmaps;
+
    common_chat_templates_inputs tmpl_inputs;
    tmpl_inputs.messages = {msg};
    tmpl_inputs.add_generation_prompt = true;
@@ -196,43 +182,35 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_
    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.c_str();
+    text.text          = formatted_chat.prompt;
    text.add_special   = add_bos;
    text.parse_special = true;
+    mtmd_input_chunks chunks;

    if (g_is_interrupted) return 0;

-    mtmd::input_chunks chunks(mtmd_input_chunks_init());
-    auto bitmaps_c_ptr = ctx.bitmaps.c_ptr();
-    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(),
-                        chunks.ptr.get(), // output
-                        &text, // text
-                        bitmaps_c_ptr.data(),
-                        bitmaps_c_ptr.size());
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
    if (res != 0) {
        LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
        return 1;
    }

-    ctx.bitmaps.entries.clear();
-
-    llama_pos new_n_past;
-    if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(),
-                ctx.lctx, // lctx
-                chunks.ptr.get(), // chunks
-                ctx.n_past, // n_past
-                0, // seq_id
-                ctx.n_batch, // n_batch
-                true, // logits_last
-                &new_n_past)) {
+    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
        LOG_ERR("Unable to eval prompt\n");
        return 1;
    }

-    ctx.n_past = new_n_past;
-
-    LOG("\n");
+    ctx.n_past += mtmd_helper_get_n_pos(chunks);

    return 0;
 }
@@ -256,14 +234,14 @@ int main(int argc, char ** argv) {
    }

    mtmd_cli_context ctx(params);
-    LOG("%s: loading model: %s\n", __func__, params.model.path.c_str());
+    printf("%s: %s\n", __func__, params.model.path.c_str());

    bool is_single_turn = !params.prompt.empty() && !params.image.empty();

    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;

-    // Ctrl+C handling
+    // ctrl+C handling
    {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
        struct sigaction sigint_action;
@@ -289,12 +267,7 @@ int main(int argc, char ** argv) {
        common_chat_msg msg;
        msg.role = "user";
        msg.content = params.prompt;
-        for (const auto & image : params.image) {
-            if (!ctx.load_image(image)) {
-                return 1; // error is already printed by libmtmd
-            }
-        }
-        if (eval_message(ctx, msg, true)) {
+        if (eval_message(ctx, msg, params.image, true)) {
            return 1;
        }
        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
@@ -309,6 +282,7 @@ int main(int argc, char ** argv) {
        LOG("\n");

        bool is_first_msg = true;
+        std::vector<std::string> images_fname;
        std::string content;

        while (!g_is_interrupted) {
@@ -333,17 +307,10 @@ int main(int argc, char ** argv) {
                continue;
            }
            g_is_generating = true;
-            if (line == "/image" || line.find("/image ") == 0) {
-                if (line.size() < 8) {
-                    LOG_ERR("ERR: Missing image filename\n");
-                    continue;
-                }
+            if (line.find("/image") == 0) {
                std::string image = line.substr(7);
-                if (ctx.load_image(image)) {
-                    LOG("Image %s loaded\n", image.c_str());
-                    content += "<__image__>";
-                }
-                // else, error is already printed by libmtmd
+                images_fname.push_back(string_strip(image));
+                content += "<__image__>";
                continue;
            } else {
                content += line;
@@ -351,14 +318,21 @@ int main(int argc, char ** argv) {
            common_chat_msg msg;
            msg.role = "user";
            msg.content = content;
-            int ret = eval_message(ctx, msg, is_first_msg);
+            int ret = eval_message(ctx, msg, images_fname, is_first_msg);
+            if (g_is_interrupted) break;
+            if (ret == 2) {
+                // non-fatal error
+                images_fname.clear();
+                content.clear();
+                continue;
+            }
            if (ret) {
                return 1;
            }
-            if (g_is_interrupted) break;
            if (generate_response(ctx, smpl, n_predict)) {
                return 1;
            }
+            images_fname.clear();
            content.clear();
            is_first_msg = false;
        }
--- a/examples/llava/mtmd.cpp
+++ b/examples/llava/mtmd.cpp
@@ -12,50 +12,15 @@
 #include <limits>
 #include <vector>

-// represents raw image data, layout is RGBRGBRGB...
-// length of data must be nx * ny * 3
-struct mtmd_bitmap {
-    uint32_t nx;
-    uint32_t ny;
-    std::vector<unsigned char> data;
-    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
-};
-
-struct mtmd_image_tokens_deleter {
-    void operator()(mtmd_image_tokens * val); // forward declaration
-};
-using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
-
-struct mtmd_input_chunk {
-    mtmd_input_chunk_type type;
-    std::vector<llama_token> tokens_text;
-    mtmd_image_tokens_ptr tokens_image;
-};
-
-struct mtmd_input_chunks {
-    std::vector<mtmd_input_chunk> entries;
-};
-
 // slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
 // models not having it (llava-1.6) will process embeddings without any special tokens in-between
 enum mtmd_slice_tmpl {
    MTMD_SLICE_TMPL_NONE,
    MTMD_SLICE_TMPL_MINICPMV_2_5,
    MTMD_SLICE_TMPL_MINICPMV_2_6,
-    MTMD_SLICE_TMPL_LLAMA4,
    // TODO @ngxson : add support for idefics (SmolVLM)
 };

-mtmd_context_params mtmd_context_params_default() {
-    mtmd_context_params params;
-    params.use_gpu = true;
-    params.print_timings = true;
-    params.n_threads = 4;
-    params.verbosity = GGML_LOG_LEVEL_INFO;
-    params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
-    return params;
-}
-
 struct mtmd_context {
    struct clip_ctx * ctx_clip;
    const struct llama_model * text_model;
@@ -65,19 +30,15 @@ struct mtmd_context {
    int n_threads;
    std::string image_marker;

-    // for llava-uhd style models, we need special tokens in-between slices
-    // minicpmv calls them "slices", llama 4 calls them "tiles"
+    // for minicpmv, we need special tokens in-between slices
    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
-    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice start
-    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice end
-    llama_token tok_sli_img_mid   = LLAMA_TOKEN_NULL; // between 2 slices
+    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
-    bool        tok_row_end_trail = false;
-    bool        ov_img_first      = false;

    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE

@@ -101,7 +62,6 @@ struct mtmd_context {

        use_mrope = clip_is_qwen2vl(ctx_clip);

-        projector_type proj = clip_get_projector_type(ctx_clip);
        int minicpmv_version = clip_is_minicpmv(ctx_clip);
        if (minicpmv_version == 2) {
            // minicpmv 2.5 format:
@@ -114,8 +74,6 @@ struct mtmd_context {
            tok_sli_img_start = tok_ov_img_start;
            tok_sli_img_end   = tok_ov_img_end;
            tok_row_end       = lookup_token("\n");
-            tok_row_end_trail = false; // no trailing end-of-row token
-            ov_img_first      = true;

        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
            // minicpmv 2.6 format:
@@ -126,25 +84,9 @@ struct mtmd_context {
            tok_sli_img_start = lookup_token("<slice>");
            tok_sli_img_end   = lookup_token("</slice>");
            tok_row_end       = lookup_token("\n");
-            tok_row_end_trail = false; // no trailing end-of-row token
-            ov_img_first      = true;

        } else if (minicpmv_version != 0) {
            GGML_ASSERT(false && "unsupported minicpmv version");
-        } else if (proj == PROJECTOR_TYPE_LLAMA4) {
-            // llama 4 format:
-            // <|image_start|>
-            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
-            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
-            //     ... <|tile_y_separator|>   <-- trailing end-of-row token
-            // <|image|> (overview)           <-- overview image is last
-            // <|image_end|>
-            slice_tmpl        = MTMD_SLICE_TMPL_LLAMA4;
-            tok_ov_img_start  = lookup_token("<|image|>");
-            tok_sli_img_mid   = lookup_token("<|tile_x_separator|>");
-            tok_row_end       = lookup_token("<|tile_y_separator|>");
-            tok_row_end_trail = true; // add trailing end-of-row token
-            ov_img_first      = false; // overview image is last
        }
    }

@@ -190,16 +132,6 @@ struct mtmd_image_tokens {
    uint32_t n_tokens() const { return nx * ny; }
    clip_image_f32_batch batch_f32; // preprocessed image patches
    std::string id; // optional user-defined ID, useful for KV cache tracking
-
-    mtmd_image_tokens clone() {
-        return mtmd_image_tokens{
-            nx,
-            ny,
-            use_mrope_pos,
-            batch_f32.clone(),
-            id
-        };
-    }
 };

 mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
@@ -240,13 +172,12 @@ static std::vector<llama_token> mtmd_tokenize_text_internal(
 }

 int32_t mtmd_tokenize(mtmd_context * ctx,
-            mtmd_input_chunks * output,
-            const mtmd_input_text * text,
-            const mtmd_bitmap ** bitmaps,
-            size_t n_bitmaps) {
+                        std::vector<mtmd_input_chunk> & output,
+                        const mtmd_input_text & text,
+                        const std::vector<mtmd_bitmap> & bitmaps) {
    auto vocab = llama_model_get_vocab(ctx->text_model);

-    std::string prompt_modified(text->text);
+    std::string prompt_modified(text.text);
    std::string marker_modified(ctx->image_marker);
    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);

@@ -258,6 +189,11 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
        marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);

+    } else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        // <|begin_of_image|> ... (image embeddings) ... <|end_of_image|>
+        marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
        // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
        marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
@@ -267,30 +203,20 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
        // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
        marker_modified = ctx->image_marker + "[IMG_END]";
        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+    }

-    } else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
+    else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
        // <|vision_start|> ... (image embeddings) ... <|vision_end|>
        marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>";
        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);

-    } else if (proj_type == PROJECTOR_TYPE_LLAMA4) {
-        // (more details in mtmd_context constructor)
-        marker_modified = "<|image_start|>" + ctx->image_marker + "<|image_end|>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-
-    } else if (proj_type == PROJECTOR_TYPE_INTERNVL) {
-        // <img> ... (image embeddings) ... </img>
-        marker_modified = "<img>" + ctx->image_marker + "</img>";
-        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
-
    }

    // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
-    // for glm-edge, BOI and EOI token's embeddings are not present in the text model

    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
-    output->entries.clear();
-    output->entries.reserve(parts.size());
+    output.clear();
+    output.reserve(parts.size());

    size_t i_img = 0;

@@ -301,7 +227,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
            std::move(tokens),
            {},
        };
-        output->entries.emplace_back(std::move(chunk));
+        output.emplace_back(std::move(chunk));
    };

    // utility for splitting batch of multiple images into chunks of batch having single images
@@ -329,7 +255,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
    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);
+        auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
        if (tokens.empty()) {
            continue;
        }
@@ -338,22 +264,23 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
            std::move(tokens),
            {},
        };
-        output->entries.emplace_back(std::move(chunk));
+        output.emplace_back(std::move(chunk));

        if (&parts.back() != &part) {
            // add image token to middle of 2 parts

-            if (i_img >= n_bitmaps) {
+            if (i_img >= bitmaps.size()) {
                LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
                return 1;
            }

            // convert mtmd_bitmap to clip_image_u8
            clip_image_u8_ptr img_u8(clip_image_u8_init());
-            img_u8->nx = bitmaps[i_img]->nx;
-            img_u8->ny = bitmaps[i_img]->ny;
-            img_u8->buf.resize(bitmaps[i_img]->data.size());
-            std::memcpy(img_u8->buf.data(), bitmaps[i_img]->data.data(), img_u8->nx * img_u8->ny * 3);
+            img_u8->nx = bitmaps[i_img].nx;
+            img_u8->ny = bitmaps[i_img].ny;
+            img_u8->buf.resize(bitmaps[i_img].data.size());
+            std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
+            clip_image_size img_u8_size{img_u8->nx, img_u8->ny};

            // preprocess image
            clip_image_f32_batch batch_f32;
@@ -363,52 +290,37 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                return 2;
            }

-            // handle llava-uhd style preprocessing
-            if (
-                ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5
-                || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6
-                || ctx->slice_tmpl == MTMD_SLICE_TMPL_LLAMA4
-            ) {
+            if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
                // split batch into chunks of single images
-                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img]->id);
+                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
                GGML_ASSERT(chunks.size() > 0);

-                auto ov_chunk = std::move(chunks.front());
+                // add overview image
+                add_text_chunk({ctx->tok_ov_img_start});
+                output.emplace_back(std::move(chunks.front()));
                chunks.erase(chunks.begin());
+                add_text_chunk({ctx->tok_ov_img_end});

-                // add overview image (first)
-                if (ctx->ov_img_first) {
-                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_ov_img_start});
-                    }
-                    output->entries.emplace_back(std::move(ov_chunk));
-                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_ov_img_end});
-                    }
-                }
-
-                // add slices (or tiles)
+                // add slices
                if (!chunks.empty()) {
-                    const int n_col = batch_f32.grid_x;
-                    const int n_row = batch_f32.grid_y;
+                    clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
+                    int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
+                    int n_row = (int)chunks.size() / n_col;
+                    GGML_ASSERT(n_row * n_col == (int)chunks.size());
                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
                        add_text_chunk({ctx->tok_slices_start});
                    }
                    for (int y = 0; y < n_row; y++) {
                        for (int x = 0; x < n_col; x++) {
-                            const bool is_last_in_row = (x == n_col - 1);
                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
                                add_text_chunk({ctx->tok_sli_img_start});
                            }
-                            output->entries.emplace_back(std::move(chunks[y * n_col + x]));
+                            output.emplace_back(std::move(chunks[y * n_col + x]));
                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
                                add_text_chunk({ctx->tok_sli_img_end});
                            }
-                            if (!is_last_in_row && ctx->tok_sli_img_mid != LLAMA_TOKEN_NULL) {
-                                add_text_chunk({ctx->tok_sli_img_mid});
-                            }
                        }
-                        if ((y != n_row - 1 || ctx->tok_row_end_trail) && ctx->tok_row_end != LLAMA_TOKEN_NULL) {
+                        if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
                            add_text_chunk({ctx->tok_row_end});
                        }
                    }
@@ -417,17 +329,6 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                    }
                }

-                // add overview image (last)
-                if (!ctx->ov_img_first) {
-                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_ov_img_start});
-                    }
-                    output->entries.emplace_back(std::move(ov_chunk));
-                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
-                        add_text_chunk({ctx->tok_ov_img_end});
-                    }
-                }
-
            } else {
                size_t n_tokens = 0;
                for (const auto & entry : batch_f32.entries) {
@@ -446,7 +347,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                    image_tokens->ny = 1;
                }
                image_tokens->batch_f32 = std::move(batch_f32);
-                image_tokens->id = bitmaps[i_img]->id; // optional
+                image_tokens->id = bitmaps[i_img].id; // optional

                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
@@ -457,7 +358,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                    {},
                    std::move(image_tokens),
                };
-                output->entries.emplace_back(std::move(chunk));
+                output.emplace_back(std::move(chunk));
            }

            i_img++; // move to next image
@@ -467,17 +368,48 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
    return 0;
 }

-static void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
+void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
    if (image_tokens) {
        delete image_tokens;
    }
 }

+size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->n_tokens();
+}
+
+size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->nx;
+}
+
+size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->ny;
+}
+
+std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->id;
+}
+
+llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
+    if (image_tokens->use_mrope_pos) {
+        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
+    }
+    return image_tokens->n_tokens();
+}
+
 int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
    bool ok = false;

+    // only effective for minicpmv and qwen2vl, other models will ignore load_image_size
+    {
+        clip_image_size slice_size{
+            image_tokens->batch_f32.entries[0]->nx,
+            image_tokens->batch_f32.entries[0]->ny};
+        clip_add_load_image_size(ctx->ctx_clip, &slice_size);
+    }
+
    if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
        const auto & entries = image_tokens->batch_f32.entries;
@@ -504,6 +436,261 @@ 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 += mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_tokens;
+}
+
+llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks) {
+    llama_pos n_pos = 0;
+    for (auto & chunk : chunks) {
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            n_pos += chunk.tokens_text.size();
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            n_pos += mtmd_image_tokens_get_n_pos(chunk.tokens_image.get());
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_pos;
+}
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
+    std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
+    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, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        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(),
+        };
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.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;
+        }
+    }
+
+    void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.resize(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                auto src = pos.begin() + i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(), src, src + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
+};
+
+int32_t mtmd_helper_eval(mtmd_context * ctx,
+        llama_context * lctx,
+        mtmd_input_chunks & chunks,
+        llama_pos pos0,
+        llama_seq_id seq_id,
+        int32_t n_batch) {
+    int32_t ret;
+    llama_pos n_past = pos0;
+    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
+
+    for (auto & chunk : chunks) {
+        bool is_last = &chunk == &chunks.back();
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            text_batch.n_tokens = chunk.tokens_text.size();
+            size_t i = 0;
+            while (i < chunk.tokens_text.size()) { // split into batches
+                for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
+                    text_batch.token   [i]    = chunk.tokens_text[i];
+                    text_batch.pos     [i]    = n_past++;
+                    text_batch.n_seq_id[i]    = 1;
+                    text_batch.seq_id  [i][0] = seq_id;
+                    text_batch.logits  [i]    = false;
+                }
+                if (is_last) {
+                    // always get logits for last input chunk
+                    text_batch.logits[text_batch.n_tokens - 1] = true;
+                }
+                ret = llama_decode(lctx, text_batch);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode text\n");
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
+            }
+
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
+            GGML_ASSERT(chunk.tokens_image != nullptr);
+            int64_t t0 = ggml_time_ms();
+            if (ctx->print_timings) {
+                LOG_INF("encoding image or slice...\n");
+            }
+            ret = mtmd_encode(ctx, chunk.tokens_image.get());
+            if (ret != 0) {
+                LOG_ERR("failed to encode image\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            if (ctx->print_timings) {
+                LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+            }
+
+            int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
+            int32_t i_batch = 0;
+            int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
+            float * embd = mtmd_get_output_embd(ctx);
+            decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+            const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
+            const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
+
+            if (mtmd_decode_use_mrope(ctx)) {
+                batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
+            } else {
+                batch_embd.set_position_normal(n_past, seq_id);
+            }
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, false);
+                // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+            }
+
+            while (i_batch < n_img_batches) { // split into batches
+                int pos_offset = i_batch*n_batch;
+                int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+                llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
+
+                LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+
+                int64_t t1 = ggml_time_ms();
+                ret = llama_decode(lctx, batch_embd_view);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode image\n");
+                    llama_set_causal_attn(lctx, true); // restore causal attn
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
+
+                if (ctx->print_timings) {
+                    LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
+                }
+
+                i_batch++;
+            }
+
+            // for mrope, one image is one single **temporal** position
+            n_past += mtmd_decode_use_mrope(ctx) ? 1 : n_tokens;
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, true);
+            }
+
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+
+    llama_batch_free(text_batch);
+    return 0;
+}
+
+int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) {
+    clip_image_u8_ptr img_u8(clip_image_u8_init());
+    bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
+    if (!ok) {
+        LOG_ERR("Unable to load image from buffer\n");
+        return 1;
+    }
+    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
+    output.data.resize(output.nx * output.ny * 3);
+    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
+    return 0;
+}
+
+int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) {
+    clip_image_u8_ptr img_u8(clip_image_u8_init());
+    bool ok = clip_image_load_from_file(fname, img_u8.get());
+    if (!ok) {
+        LOG_ERR("Unable to load image %s\n", fname);
+        return 1;
+    }
+    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
+    output.data.resize(output.nx * output.ny * 3);
+    std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
+    return 0;
+}
+
 bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
@@ -519,203 +706,3 @@ bool mtmd_decode_use_mrope(mtmd_context * ctx) {
 void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
    mtmd_image_tokens_free(val);
 }
-
-// these 2 helpers below use internal clip_image_u8_ptr,
-// so unfortunately they cannot moved to mtmd-helper.h
-// however, in theory, user can decode image file to bitmap using
-// whichever library they want, and then use mtmd_bitmap_init() to create bitmap
-
-mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len) {
-    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 nullptr;
-    }
-    uint32_t nx, ny;
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &nx, &ny);
-    return mtmd_bitmap_init(nx, ny, data);
-}
-
-mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname) {
-    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 nullptr;
-    }
-    uint32_t nx, ny;
-    unsigned char * data = clip_image_u8_get_data(img_u8.get(), &nx, &ny);
-    return mtmd_bitmap_init(nx, ny, data);
-}
-
-//
-// public API functions
-//
-
-// mtmd_bitmap
-
-mtmd_bitmap * mtmd_bitmap_init(uint32_t nx,
-                               uint32_t ny,
-                               const unsigned char * data) {
-    mtmd_bitmap * bitmap = new mtmd_bitmap;
-    bitmap->nx = nx;
-    bitmap->ny = ny;
-    size_t data_size = (size_t)nx * ny * 3;
-    bitmap->data.resize(data_size);
-    std::memcpy(bitmap->data.data(), data, data_size);
-    return bitmap;
-}
-
-uint32_t mtmd_bitmap_get_nx(const mtmd_bitmap * bitmap) {
-    return bitmap->nx;
-}
-
-uint32_t mtmd_bitmap_get_ny(const mtmd_bitmap * bitmap) {
-    return bitmap->ny;
-}
-
-const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap) {
-    return bitmap->data.data();
-}
-
-const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap) {
-    return bitmap->id.c_str();
-}
-
-void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id) {
-    if (id) {
-        bitmap->id = std::string(id);
-    } else {
-        bitmap->id.clear();
-    }
-}
-
-void mtmd_bitmap_free(mtmd_bitmap * bitmap) {
-    if (bitmap) {
-        delete bitmap;
-    }
-}
-
-// mtmd_input_chunks
-
-mtmd_input_chunks * mtmd_input_chunks_init() {
-    return new mtmd_input_chunks;
-}
-
-size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks) {
-    return chunks->entries.size();
-}
-
-const mtmd_input_chunk * mtmd_input_chunks_get(const mtmd_input_chunks * chunks, size_t idx) {
-    if (idx >= chunks->entries.size()) {
-        return nullptr;
-    }
-    return &chunks->entries[idx];
-}
-
-void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
-    if (chunks) {
-        delete chunks;
-    }
-}
-
-// mtmd_input_chunk
-
-enum mtmd_input_chunk_type mtmd_input_chunk_get_type(const mtmd_input_chunk * chunk) {
-    return chunk->type;
-}
-
-const llama_token * mtmd_input_chunk_get_tokens_text(const mtmd_input_chunk * chunk, size_t * n_tokens_output) {
-    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-        *n_tokens_output = chunk->tokens_text.size();
-        return chunk->tokens_text.data();
-    }
-    *n_tokens_output = 0;
-    return nullptr;
-}
-
-const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk) {
-    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-        return chunk->tokens_image.get();
-    }
-    return nullptr;
-}
-
-mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk) {
-    mtmd_input_chunk * copy = new mtmd_input_chunk{
-        chunk->type,
-        chunk->tokens_text,
-        mtmd_image_tokens_ptr(),
-    };
-    if (chunk->tokens_image) {
-        // copy the image tokens
-        copy->tokens_image = mtmd_image_tokens_ptr(new mtmd_image_tokens());
-        *copy->tokens_image = chunk->tokens_image->clone();
-    }
-    return copy;
-}
-
-void mtmd_input_chunk_free(mtmd_input_chunk * chunk) {
-    if (chunk) {
-        delete chunk;
-    }
-}
-
-// mtmd_image_tokens
-
-size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->n_tokens();
-}
-
-size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->nx;
-}
-
-size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->ny;
-}
-
-const char * mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
-    return image_tokens->id.c_str();
-}
-
-llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
-    if (image_tokens->use_mrope_pos) {
-        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
-    }
-    return image_tokens->n_tokens();
-}
-
-// test function
-
-mtmd_input_chunks * mtmd_test_create_input_chunks() {
-    mtmd_input_chunks * chunks = mtmd_input_chunks_init();
-    if (!chunks) {
-        return nullptr;
-    }
-
-    // create a text chunk
-    std::vector<llama_token> tokens_text = { 1, 2, 3, 4, 5 };
-    mtmd_input_chunk chunk_text{
-        MTMD_INPUT_CHUNK_TYPE_TEXT,
-        std::move(tokens_text),
-        {},
-    };
-    chunks->entries.emplace_back(std::move(chunk_text));
-
-    // create an image chunk
-    mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-    image_tokens->nx = 4;
-    image_tokens->ny = 4;
-    image_tokens->batch_f32.entries.resize(16);
-    image_tokens->id = "image_1";
-    mtmd_input_chunk chunk_image{
-        MTMD_INPUT_CHUNK_TYPE_IMAGE,
-        {},
-        std::move(image_tokens),
-    };
-    chunks->entries.emplace_back(std::move(chunk_image));
-
-    return chunks;
-}
--- a/examples/llava/mtmd.h
+++ b/examples/llava/mtmd.h
@@ -0,0 +1,168 @@
+#ifndef MTMD_H
+#define MTMD_H
+
+#include "ggml.h"
+#include "llama.h"
+#include "clip.h"
+
+#include <vector>
+#include <cinttypes>
+#include <memory>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define MTMD_API __declspec(dllexport)
+#        else
+#            define MTMD_API __declspec(dllimport)
+#        endif
+#    else
+#        define MTMD_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define MTMD_API
+#endif
+
+#ifdef __cplusplus
+
+enum mtmd_input_chunk_type {
+    MTMD_INPUT_CHUNK_TYPE_TEXT,
+    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+};
+
+struct mtmd_context;
+struct mtmd_image_tokens;
+
+// represents raw image data, layout is RGBRGBRGB...
+// length of data must be nx * ny * 3
+struct mtmd_bitmap {
+    uint32_t nx;
+    uint32_t ny;
+    std::vector<unsigned char> data;
+    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
+};
+
+struct mtmd_image_tokens_deleter {
+    void operator()(mtmd_image_tokens * val); // forward declaration
+};
+using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
+
+struct mtmd_input_chunk {
+    mtmd_input_chunk_type type;
+    std::vector<llama_token> tokens_text;
+    mtmd_image_tokens_ptr tokens_image;
+};
+
+using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
+
+struct mtmd_context_params {
+    bool use_gpu = true;
+    bool print_timings = true;
+    int n_threads = 4;
+    enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO;
+    const char * image_marker = "<__image__>";
+};
+
+struct mtmd_input_text {
+    std::string text;
+    bool add_special;
+    bool parse_special;
+};
+
+// initialize the mtmd context
+// return nullptr on failure
+MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+                                                const llama_model * text_model,
+                                                const mtmd_context_params ctx_params);
+
+MTMD_API void mtmd_free(mtmd_context * ctx);
+
+// tokenize an input text prompt and an image
+// the prompt must have the input image marker (default: "<__image__>") in it
+// the marker will be replaced with the image tokens
+// for example:
+//   "here is an image: <__image__>\ndescribe it in detail."
+//   this will gives 3 chunks:
+//   1. "here is an image: <start_of_image>"
+//   2. (image tokens)
+//   3. "<end_of_image>\ndescribe it in detail."
+// number of bitmaps must be equal to the number of image markers in the prompt
+// this function is thread-safe (shared ctx)
+// return values:
+//   0 on success
+//   1 on number of images not matching the number of markers
+//   2 on image preprocessing error
+MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
+                                std::vector<mtmd_input_chunk> & output,
+                                const mtmd_input_text & text,
+                                const std::vector<mtmd_bitmap> & bitmaps);
+
+// access mtmd_image_tokens
+MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
+MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
+MTMD_API llama_pos   mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens); // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
+
+// returns 0 on success
+MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
+                            const mtmd_image_tokens * image_tokens);
+
+// get output embeddings from the last encode pass
+MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
+
+// whether we need to set non-causal mask before llama_decode
+MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
+
+// whether the current model use M-RoPE for llama_decode
+MTMD_API bool mtmd_decode_use_mrope(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 KV cache
+MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
+
+// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
+MTMD_API llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks);
+
+// helper function that automatically:
+// 1. run llama_decode() on text chunks
+// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
+// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
+// otherwise, returns 0 on success
+MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
+                                llama_context * lctx,
+                                mtmd_input_chunks & chunks,
+                                llama_pos pos0,
+                                llama_seq_id seq_id,
+                                int32_t n_batch);
+
+// helper function to construct a mtmd_bitmap from a file
+// returns 0 on success
+// this function is thread-safe
+MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output);
+
+// helper function to construct a mtmd_bitmap from a buffer
+// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
+// returns 0 on success
+// this function is thread-safe
+MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output);
+
+// convenient unique_ptr wrappers
+struct mtmd_context_deleter {
+    void operator()(mtmd_context * val) { mtmd_free(val); }
+};
+using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
+
+#else
+
+static_assert(false && "C header is not yet supported by this library");
+
+#endif
+
+#endif
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Georgi Gerganov	65202d2985	sync : ggml	2025-05-01 09:59:02 +03:00
Diego Devesa	db1ff5b63a	ggml : fix ggml_gallocr_ptr type (ggml/1205)	2025-05-01 09:32:45 +03:00
Georgi Gerganov	610df4cc3b	cuda : fix unused variable compile warning (whisper/0) ggml-ci	2025-05-01 09:32:42 +03:00