add arg --threads-sampling

.devops/cann.Dockerfile

@@ -145,7 +145,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ==============================================================================
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 ENTRYPOINT [ "/app/llama-cli" ]
 
@@ -156,7 +156,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/cpu.Dockerfile

@@ -104,7 +104,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -115,7 +115,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/cuda.Dockerfile

@@ -113,7 +113,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -124,7 +124,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/intel.Dockerfile

@@ -141,7 +141,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -153,7 +153,7 @@ FROM base AS server
 ENV LLAMA_ARG_HOST=0.0.0.0
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/musa.Dockerfile

@@ -115,7 +115,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -126,7 +126,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/openvino.Dockerfile

@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.3
-ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
+ARG OPENVINO_VERSION_MAJOR=2026.2
+ARG OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.38.2
-ARG IGC_VERSION_FULL=2_2.38.2+22051
-ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
+ARG IGC_VERSION=v2.34.4
+ARG IGC_VERSION_FULL=2_2.34.4+21428
+ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.35.0
-ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
-ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1
+ARG NPU_DRIVER_VERSION=v1.33.0
+ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
+ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
 
 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
     fi; \
     DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
     if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
     fi; \
     mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
     apt-get update; \
@@ -214,7 +214,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app/
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app/
 
 WORKDIR /app
 
@@ -225,7 +225,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app/
+COPY --from=build /app/full/llama-server /app/
 
 WORKDIR /app
 

.devops/rocm.Dockerfile

@@ -57,6 +57,7 @@ COPY --from=web /app/tools/ui/dist tools/ui/dist
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
     cmake -S . -B build \
         -DGGML_HIP=ON \
+        -DGGML_HIP_ROCWMMA_FATTN=ON \
         -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
         -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
         -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
@@ -126,7 +127,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -137,7 +138,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/s390x.Dockerfile

@@ -124,7 +124,7 @@ WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
 
@@ -138,7 +138,7 @@ WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-server /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
 
 EXPOSE 8080
 

.devops/vulkan.Dockerfile

@@ -107,7 +107,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -118,7 +118,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/zendnn.Dockerfile

@@ -97,7 +97,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -108,7 +108,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama /app/full/llama-server /app
+COPY --from=build /app/full/llama-server /app
 
 WORKDIR /app
 

.github/actions/ccache-clear/action.yml

@@ -1,88 +1,22 @@
-# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
-description: "Delete GitHub Actions caches matching a key prefix, oldest first"
+description: "Delete all GitHub Actions caches matching a key prefix"
 inputs:
   key:
     description: "Cache key prefix to match and delete"
     required: true
-  older:
-    description: "Only delete caches created more than this long ago (e.g. 90m, 1h, 1d). By default all matching caches are deleted"
-    required: false
-    default: ""
-  min:
-    description: "Stop deleting if fewer than this many caches would remain (e.g. 1). By default there is no minimum"
-    required: false
-    default: "0"
-  dry-run:
-    description: "Only print the caches that would be deleted, without deleting them"
-    required: false
-    default: "false"
 
 runs:
   using: "composite"
   steps:
     - name: Clear caches
       shell: bash
-      env:
-        CLEAR_KEY: ${{ inputs.key }}
-        CLEAR_OLDER: ${{ inputs.older }}
-        CLEAR_MIN: ${{ inputs.min }}
-        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        # Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
-        to_seconds() {
-          local val="$1"
-          [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
-          local num="${val%?}" unit="${val: -1}" mult
-          [[ "$num" =~ ^[0-9]+$ ]] || return 1
-          case "$unit" in
-            s) mult=1 ;;
-            m) mult=60 ;;
-            h) mult=3600 ;;
-            d) mult=86400 ;;
-            *) return 1 ;;
-          esac
-          echo $((num * mult))
-        }
-
-        [[ "$CLEAR_MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $CLEAR_MIN" >&2; exit 1; }
-        [[ "$CLEAR_DRY_RUN" =~ ^(true|false)$ ]] || { echo "Invalid dry-run value: $CLEAR_DRY_RUN" >&2; exit 1; }
-
-        CACHES=$(gh cache list --key "ccache-$CLEAR_KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' 2>/dev/null | LC_ALL=C sort)
+        CACHES=$(gh cache list --key "ccache-${{ inputs.key }}" --json id,key --jq '.[] | "\(.id) \(.key)"' 2>/dev/null)
         if [ -z "$CACHES" ]; then
-          echo "No caches found with key prefix: $CLEAR_KEY"
+          echo "No caches found with key prefix: ${{ inputs.key }}"
           exit 0
         fi
-
-        TOTAL=$(( $(wc -l <<< "$CACHES") ))
-
-        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
-        done <<< "$CACHES"
-
-        CUTOFF=""
-        if [ -n "$CLEAR_OLDER" ]; then
-          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
-          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
-        fi
-
-        # Caches are sorted oldest first
-        DELETED=0
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
-            echo "Rest are not older than $CLEAR_OLDER, stopping"
-            break
-          fi
-          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
-            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
-            break
-          fi
-          if [ "$CLEAR_DRY_RUN" = "true" ]; then
-            echo "Would delete cache: $ID ($KEY)"
-          else
-            echo "Deleting cache: $ID ($KEY)"
-            gh cache delete "$ID"
-          fi
-          DELETED=$((DELETED + 1))
+        while read -r id key; do
+          echo "Deleting cache: $id ($key)"
+          gh cache delete "$id"
         done <<< "$CACHES"

.github/actions/linux-setup-vulkan/action.yml

@@ -0,0 +1,20 @@
+name: "Linux - Setup Vulkan SDK"
+description: "Setup Vulkan SDK for Linux"
+inputs:
+  path:
+    description: "Installation path"
+    required: true
+  version:
+    description: "Vulkan SDK version"
+    required: true
+
+runs:
+  using: "composite"
+  steps:
+    - name: Setup Vulkan SDK
+      id: setup
+      uses: ./.github/actions/unarchive-tar
+      with:
+        url: https://sdk.lunarg.com/sdk/download/${{ inputs.version }}/linux/vulkan_sdk.tar.xz
+        path: ${{ inputs.path }}
+        strip: 1

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

@@ -4,9 +4,6 @@ inputs:
   cuda_version:
     description: "CUDA toolkit version"
     required: true
-  cuda_arch:
-    description: "CUDA target architecture"
-    required: true
 
 runs:
   using: "composite"
@@ -130,26 +127,3 @@ runs:
           echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
           echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-    - name: Install Cuda Toolkit 13.4 for ARM64
-      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          choco install unzip -y
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cccl-windows-x86_64-13.3.4.1.2-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_crt-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_nvcc-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-x86_64/5B515474-7E78-11F1-8656-C51E4F4B317F/libnvvm-windows-x86_64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/cuda_cudart-windows-arm64-13.4.46-archive.zip"
-          curl -O "https://packages.nvidia.com/bin-archive/pool/windows-arm64/5B515474-7E78-11F1-8656-C51E4F4B317F/libcublas-windows-arm64-13.7.0.10-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.1.2-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.46-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.10-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

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

@@ -8,26 +8,8 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Install ROCm with Wheels
-      shell: pwsh
-      run: |
-        $ErrorActionPreference = "Stop"
-        write-host "Setting up Python virtual environment"
-
-        # Create the venv directly at the cache location to avoid relocation issues
-        New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
-        python -m venv C:\TheRock\build\.venv
-        & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-        write-host "Upgrading pip"
-        python -m pip install --upgrade pip
-
-        write-host "Installing ROCm wheels for multi-arch support"
-        # Install ROCm wheels for multi-arch support (this may take several minutes)
-        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
-
-        # Pre-expand the devel tree so it is included in the cache
-        write-host "Initializing ROCm devel tree"
-        rocm-sdk init
-        if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-        write-host "Completed ROCm wheel installation to C:\TheRock\build"
+    - name: Setup ROCm
+      uses: ./.github/actions/install-exe
+      with:
+        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
+        args: -install

.github/labeler.yml

@@ -35,20 +35,8 @@ AMD ZenDNN:
 documentation:
     - changed-files:
         - any-glob-to-any-file:
-            - "**/*.md"
             - docs/**
             - media/**
-examples:
-    - all:
-        - changed-files:
-            - any-glob-to-any-file:
-                - app/**
-                - examples/**
-                - tools/**
-            - all-globs-to-all-files:
-                - '!tools/server/**'
-                - '!tools/mtmd/**'
-                - '!tools/ui/**'
 testing:
     - changed-files:
         - any-glob-to-any-file:
@@ -59,12 +47,28 @@ build:
             - cmake/**
             - CMakeLists.txt
             - CMakePresets.json
+examples:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/**
+            - tools/**
 devops:
     - changed-files:
         - any-glob-to-any-file:
             - .devops/**
             - .github/**
             - ci/**
+python:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.py"
+            - requirements/**
+            - gguf-py/**
+            - .flake8
+script:
+    - changed-files:
+        - any-glob-to-any-file:
+            - scripts/**
 android:
     - changed-files:
         - any-glob-to-any-file:
@@ -77,20 +81,9 @@ server:
     - changed-files:
         - any-glob-to-any-file:
             - tools/server/**
-mtmd:
-    - changed-files:
-        - any-glob-to-any-file:
-            - tools/mtmd/**
-conversion:
-    - changed-files:
-        - any-glob-to-any-file:
-            - conversion/**
-            - convert_*.py
-            - gguf-py/**
-vendor:
-    - changed-files:
-        - any-glob-to-any-file:
-            - vendor/**
+
+
+
 ggml:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-apple.yml

@@ -60,10 +60,10 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=OFF \
             -DGGML_METAL_SHADER_DEBUG=ON \
-            -DGGML_RPC=ON \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
+            -DGGML_RPC=ON
           time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
           leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
 
@@ -126,6 +126,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
             -DLLAMA_BUILD_APP=OFF \
@@ -176,6 +177,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_APP=OFF \
@@ -209,6 +211,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_APP=OFF \
@@ -253,6 +256,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
             -DLLAMA_BUILD_APP=OFF \

.github/workflows/build-cache.yml

@@ -10,6 +10,33 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  ubuntu-24-vulkan-cache:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Get latest Vulkan SDK version
+        id: vulkan_sdk_version
+        run: |
+          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
+
+      - name: Setup Cache
+        uses: actions/cache@v5
+        id: cache-sdk
+        with:
+          path: ./vulkan_sdk
+          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+
+      - name: Setup Vulkan SDK
+        if: steps.cache-sdk.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-vulkan
+        with:
+          path: ./vulkan_sdk
+          version: ${{ env.VULKAN_SDK_VERSION }}
+
   #ubuntu-24-spacemit-cache:
   #  runs-on: ubuntu-24.04
 
@@ -40,9 +67,9 @@ jobs:
     runs-on: ubuntu-24.04
 
     env:
-      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Clone
@@ -69,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Clone
@@ -92,27 +119,27 @@ jobs:
           version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
           version_full: ${{ env.OPENVINO_VERSION_FULL }}
 
-  # windows-2022-rocm-cache:
-  #   runs-on: windows-2022
+  windows-2022-rocm-cache:
+    runs-on: windows-2022
 
-  #   env:
-  #     # Make sure this is in sync with release.yml and build-cuda-windows.yml
-  #     ROCM_VERSION: "7.14.0"
+    env:
+      # Make sure this is in sync with build.yml
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
-  #   steps:
-  #     - name: Clone
-  #       id: checkout
-  #       uses: actions/checkout@v6
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
 
-  #     - name: Setup Cache
-  #       uses: actions/cache@v5
-  #       id: cache-rocm
-  #       with:
-  #         path: C:\TheRock\build
-  #         key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+      - name: Setup Cache
+        uses: actions/cache@v5
+        id: cache-rocm
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
-  #     - name: Setup ROCm
-  #       if: steps.cache-rocm.outputs.cache-hit != 'true'
-  #       uses: ./.github/actions/windows-setup-rocm
-  #       with:
-  #         version: ${{ env.ROCM_VERSION }}
+      - name: Setup ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}

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

@@ -5,7 +5,7 @@ on:
 
 jobs:
   linux:
-    runs-on: [self-hosted, Linux]
+    runs-on: [self-hosted, Linux, CPU]
     steps:
       - uses: actions/checkout@v6
         with:
@@ -21,33 +21,31 @@ jobs:
                 -DLLAMA_BUILD_TOOLS=OFF \
                 -DLLAMA_BUILD_EXAMPLES=OFF \
                 -DLLAMA_BUILD_APP=OFF \
-                -DLLAMA_BUILD_IS_DEV=OFF \
                 -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release -j $(nproc)
+          cmake --build build --config Release
           cmake --install build --prefix "$PREFIX" --config Release
 
           export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          build_commit=$(git rev-parse --short HEAD | xargs)
-          build_number=$(git rev-list --count HEAD | xargs)
+          tclsh <<'EOF'
+          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
+          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          set build(version) "0.0.$build(number)"
 
-          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          build_version="$major.$minor.$patch"
+          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
+          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
+                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
+                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
 
-          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
-                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
-                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")
-
-          for check in "${checks[@]}"; do
-              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
-                  echo "Checking llama-config.cmake version... \"$check\" failed!"
+          puts -nonewline "Checking llama-config.cmake version... "
+          foreach check $checks {
+              if {![regexp -expanded -- $check $llamaconfig]} {
+                  puts "\"$check\" failed!"
                   exit 1
-              fi
-          done
-
-          echo "Checking llama-config.cmake version... success."
+              }
+          }
+          puts "success."
+          EOF
 
           cd examples/simple-cmake-pkg
           cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build -j $(nproc)
+          cmake --build build

.github/workflows/build-cpu.yml

@@ -21,7 +21,6 @@ on:
     paths: [
       '.github/workflows/build-cpu.yml',
       '.github/workflows/build-cmake-pkg.yml',
-      'ggml/src/ggml-rpc/**',
       '**/CMakeLists.txt',
       '**/.cmake',
       '**/*.h',
@@ -95,7 +94,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DGGML_NATIVE=OFF \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_RPC=ON
           time cmake --build build --config Release -j $(nproc)
@@ -117,38 +115,29 @@ jobs:
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
-      # note: real deletion only on push to master (same condition as the ccache save),
-      #       dry-run otherwise (the token is read-only on PRs from forks)
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-${{ matrix.os }}
-          older: 1h
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   windows:
-    name: windows / ${{ matrix.build }}
     runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
         include:
           - build: 'x64-cpu-static'
             arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'x64-openblas'
             arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+          - build: 'x64-vulkan'
+            arch: 'x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'arm64'
             arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
 
     steps:
       - name: Clone
@@ -176,6 +165,15 @@ jobs:
           $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 == 'x64-vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$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: |

.github/workflows/build-cuda-ubuntu.yml

@@ -99,6 +99,7 @@ jobs:
         run: |
           cmake -B build -S . \
             -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DGPU_TARGETS="gfx1030" \
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)

.github/workflows/build-cuda-windows.yml

@@ -22,7 +22,6 @@ env:
 
 jobs:
   cuda:
-    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     runs-on: windows-2022
 
     permissions:
@@ -30,16 +29,7 @@ jobs:
 
     strategy:
       matrix:
-        include:
-          - cuda: '12.4'
-            arch: x64
-            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
-          - cuda: '13.3'
-            arch: x64
-            defines: ''
-          - cuda: '13.4'
-            arch: arm64
-            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+        cuda: ['12.4', '13.3']
 
     steps:
       - name: Clone
@@ -49,13 +39,12 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
 
       - name: Install Cuda Toolkit
         uses: ./.github/actions/windows-setup-cuda
         with:
           cuda_version: ${{ matrix.cuda }}
-          cuda_arch: ${{ matrix.arch }}
 
       - name: Install Ninja
         id: install_ninja
@@ -65,21 +54,26 @@ jobs:
       - name: Build
         id: cmake_build
         shell: cmd
+        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DGGML_BACKEND_DL=ON ^
+            -DLLAMA_BUILD_SERVER=ON ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
             -DGGML_NATIVE=OFF ^
-            -DGGML_CPU=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=ON ^
             -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
+            -DGGML_RPC=ON ^
+            -DGGML_CUDA_CUB_3DOT2=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
+          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
+          cmake --build build --config Release
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
 
   hip:
     runs-on: windows-2022
@@ -89,7 +83,7 @@ jobs:
 
     env:
       # Make sure this is in sync with build-cache.yml
-      ROCM_VERSION: "7.14.0"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     strategy:
       matrix:
@@ -103,53 +97,36 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # - name: Cache ROCm Installation
-      #   uses: actions/cache@v5
-      #   id: cache-rocm
-      #   with:
-      #     path: C:\TheRock\build
-      #     key: rocm-wheels-${{ env.ROCM_VERSION }}-multi-arch-${{ runner.os }}
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
+        run: |
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
+          7z x rocwmma.deb
+          7z x data.tar
+
+      - name: Use ROCm Installation Cache
+        uses: actions/cache@v5
+        id: cache-rocm
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
       - name: Setup ROCm
-        # if: steps.cache-rocm.outputs.cache-hit != 'true'
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         uses: ./.github/actions/windows-setup-rocm
         with:
-          version: ${{ env.ROCM_VERSION }}
-
-      - name: Setup ROCm Environment
-        run: |
-          $ErrorActionPreference = "Stop"
-
-          # Activate venv from cache or fresh install
-          & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-          # Expand the devel tree (idempotent; no-op if already done during install)
-          rocm-sdk init
-          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          $rocmPath = (rocm-sdk path --root)
-          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
-          $rocmPath = $rocmPath.Trim()
-          $cmakePath = (rocm-sdk path --cmake).Trim()
-          $binPath = (rocm-sdk path --bin).Trim()
-          write-host "ROCm root: $rocmPath"
-
-          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
-          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
-          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
-          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
-          echo "$binPath" >> $env:GITHUB_PATH
-
-          # Keep venv in PATH for subsequent steps
-          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
 
       - name: Verify ROCm
         id: verify
         run: |
-          # Test the ROCm clang shipped in the installed wheel
-          & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -157,27 +134,29 @@ jobs:
           # TODO: this build does not match the build in release.yml, so we use a different cache key
           #       ideally, the builds should match, similar to the CUDA build above so that we would be able
           #       to populate the ccache for the release with manual runs of this workflow
-          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
 
       - name: Build
         id: cmake_build
         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_PREFIX_PATH="${env:HIP_PATH}" `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
-            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -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('\', '/'))/opt/rocm-7.2.1/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DLLAMA_BUILD_BORINGSSL=ON `
-            -DHIP_PATH="${env:HIP_PATH}" `
+            -DROCM_DIR="${env:HIP_PATH}" `
             -DGGML_HIP=ON `
-            -DGPU_TARGETS="gfx1100" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGPU_TARGETS="gfx1100"  `
             -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          #key: release-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.ROCM_VERSION }}-${{ matrix.name }}
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}

.github/workflows/build-openvino.yml

@@ -39,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Clone
@@ -81,7 +81,7 @@ jobs:
         # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
 
       - name: Test (GPU)
         id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
         run: |
           cd ${{ github.workspace }}
           export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
 
   openvino-windows-2022:
     runs-on: windows-2022
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Clone
@@ -166,4 +166,4 @@ jobs:
           call "%OPENVINO_ROOT%\setupvars.bat"
 
           cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000

.github/workflows/build-rpc.yml

@@ -0,0 +1,66 @@
+name: CI (rpc)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-rpc.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-rpc.yml',
+      'ggml/src/ggml-rpc/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-24-rpc:
+    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libssl-dev ninja-build
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose

.github/workflows/build-sanitize.yml

@@ -15,12 +15,6 @@ on:
       '**/*.cpp'
     ]
 
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-sanitize.yml'
-    ]
-
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
@@ -34,35 +28,19 @@ env:
 
 jobs:
   ctest:
+    runs-on: [self-hosted, X64, CPU, Linux]
+
     continue-on-error: true
 
     strategy:
       matrix:
-        include:
-        # thread and address doesn't run properly on some self hosted machines, so run it on Github instead
-        - sanitizer: ADDRESS
-          machine: ubuntu-24.04
-        - sanitizer: THREAD
-          machine: ubuntu-24.04
-        - sanitizer: UNDEFINED
-          machine: [self-hosted, X64, Linux]
-
-    runs-on: ${{ matrix.machine }}
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
-      #   with:
-      #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04
-      #     variant: ccache
-      #     evict-old-files: 1d
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
       # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
       - name: Build (undefined)
         id: cmake_build_undefined

.github/workflows/build-self-hosted.yml

@@ -6,7 +6,7 @@ on:
     branches:
       - master
     paths: [
-      '.github/workflows/build-self-hosted.yml',
+      '.github/workflows/build.yml',
       '**/CMakeLists.txt',
       '**/.cmake',
       '**/*.h',
@@ -48,8 +48,6 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
   LLAMA_ARG_LOG_COLORS: 1
@@ -71,26 +69,6 @@ jobs:
           nvidia-smi
           GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  gpu-rocm:
-    runs-on: [self-hosted, Linux, AMD]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        # HIP_LAUNCH_BLOCKING=1: workaround for an async-execution correctness
-        # issue on integrated RDNA3.5 (gfx1151) where batched inference returns
-        # incorrect output (perplexity ~88 vs ~9.4). Serializing kernel launches
-        # restores correctness. Remove once the underlying ROCm/HIP issue is fixed.
-        env:
-          HIP_LAUNCH_BLOCKING: "1"
-        run: |
-          rocminfo
-          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
   gpu-vulkan-nvidia-cm:
     runs-on: [self-hosted, Linux, NVIDIA]
 
@@ -288,8 +266,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Clone

.github/workflows/build-vulkan.yml

@@ -93,13 +93,19 @@ jobs:
         run: |
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
-      - name: Setup Vulkan SDK
-        id: setup
-        uses: ./.github/actions/unarchive-tar
+      - name: Use Vulkan SDK Cache
+        uses: actions/cache@v5
+        id: cache-sdk
         with:
-          url: https://sdk.lunarg.com/sdk/download/${{ env.VULKAN_SDK_VERSION }}/linux/vulkan_sdk.tar.xz
           path: ./vulkan_sdk
-          strip: 1
+          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+
+      - name: Setup Vulkan SDK
+        if: steps.cache-sdk.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-vulkan
+        with:
+          path: ./vulkan_sdk
+          version: ${{ env.VULKAN_SDK_VERSION }}
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -113,7 +119,6 @@ jobs:
         run: |
           source ./vulkan_sdk/setup-env.sh
           cmake -B build \
-            -DGGML_NATIVE=OFF \
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -127,56 +132,3 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
-
-  windows:
-    runs-on: windows-2025
-
-    env:
-      VULKAN_VERSION: 1.4.357.0
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cpu-windows-2025-x64-vulkan
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Install Vulkan SDK
-        id: get_vulkan
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
-          & "$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: Build
-        id: cmake_build
-        run: |
-          cmake -S . -B build -G "Ninja Multi-Config" `
-            -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_NATIVE=OFF `
-            -DLLAMA_BUILD_SERVER=ON `
-            -DGGML_RPC=ON `
-            -DGGML_BACKEND_DL=ON `
-            -DGGML_CPU_ALL_VARIANTS=ON `
-            -DGGML_VULKAN=ON `
-            -DLLAMA_BUILD_BORINGSSL=ON
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main -C Release --verbose --timeout 900

.github/workflows/build-wasm.yml

@@ -1,90 +0,0 @@
-name: CI (wasm)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-wasm.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.wgsl',
-      '**/*.tmpl',
-      'ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-wasm.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.wgsl',
-      '**/*.tmpl',
-      'ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  ubuntu-webgpu:
-    runs-on: ubuntu-24.04-arm
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: webgpu-ubuntu-24.04-arm-wasm
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Install Emscripten
-        run: |
-          git clone https://github.com/emscripten-core/emsdk.git
-          cd emsdk
-          ./emsdk install latest
-          ./emsdk activate latest
-
-      - name: Fetch emdawnwebgpu
-        run: |
-          DAWN_TAG="v20260317.182325"
-          EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
-          echo "Downloading ${EMDAWN_PKG}"
-          curl -L -o emdawn.zip \
-            "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
-          unzip emdawn.zip
-
-      - name: Build WASM WebGPU
-        run: |
-          source emsdk/emsdk_env.sh
-          emcmake cmake -B build-wasm \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_WEBGPU=ON \
-            -DGGML_OPENMP=OFF \
-            -DLLAMA_OPENSSL=OFF \
-            -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
-
-          time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)

.github/workflows/build-webgpu.yml

@@ -13,9 +13,7 @@ on:
       '**/*.hpp',
       '**/*.c',
       '**/*.cpp',
-      '**/*.wgsl',
-      '**/*.tmpl',
-      'ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py'
+      '**/*.wgsl'
     ]
 
   pull_request:
@@ -153,3 +151,46 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  ubuntu-wasm:
+    runs-on: ubuntu-24.04-arm
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: webgpu-ubuntu-24.04-arm-wasm
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Emscripten
+        run: |
+          git clone https://github.com/emscripten-core/emsdk.git
+          cd emsdk
+          ./emsdk install latest
+          ./emsdk activate latest
+
+      - name: Fetch emdawnwebgpu
+        run: |
+          DAWN_TAG="v20260317.182325"
+          EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
+          echo "Downloading ${EMDAWN_PKG}"
+          curl -L -o emdawn.zip \
+            "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
+          unzip emdawn.zip
+
+      - name: Build WASM WebGPU
+        run: |
+          source emsdk/emsdk_env.sh
+          emcmake cmake -B build-wasm \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_WEBGPU=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
+
+          time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)

.github/workflows/docker.yml

@@ -44,7 +44,6 @@ jobs:
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
-          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
 
       - name: Determine source tag name
         id: srctag
@@ -394,11 +393,6 @@ jobs:
     name: Create shared tags from digests
     needs: [prepare_matrices, push_to_registry, create_tag]
     runs-on: ubuntu-24.04
-    permissions:
-      contents: read
-      packages: write
-      id-token: write
-      attestations: write
     strategy:
       fail-fast: false
       matrix:
@@ -433,7 +427,6 @@ jobs:
           password: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Create tags from digests
-        id: create_tags
         shell: bash
         run: |
           set -euo pipefail
@@ -445,7 +438,6 @@ jobs:
           SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
           BUILD_DATE="${{ steps.build_date.outputs.date }}"
           COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
-          echo "image_repo=${IMAGE_REPO}" >> "$GITHUB_OUTPUT"
           TAGS="${{ matrix.config.tag }}"
           ARCHES="${{ matrix.config.arches }}"
           DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -512,16 +504,6 @@ jobs:
 
               echo "Creating ${merged_versioned_tag} from ${refs[*]}"
               docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
-
-              if [[ "$tag_name" == "${TAGS%% *}" ]]; then
-                  local digest
-                  digest="$(docker buildx imagetools inspect "${merged_versioned_tag}" --format '{{.Manifest.Digest}}')"
-                  if [[ ! "$digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
-                      echo "Invalid digest for ${merged_versioned_tag}: ${digest}" >&2
-                      exit 1
-                  fi
-                  echo "${image_type}_digest=${digest}" >> "$GITHUB_OUTPUT"
-              fi
           }
 
           for tag in $TAGS; do
@@ -545,24 +527,3 @@ jobs:
           done
         env:
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
-      - name: Attest full image
-        if: ${{ matrix.config.full }}
-        uses: actions/attest@v4
-        with:
-          subject-name: ${{ steps.create_tags.outputs.image_repo }}
-          subject-digest: ${{ steps.create_tags.outputs.full_digest }}
-
-      - name: Attest light image
-        if: ${{ matrix.config.light }}
-        uses: actions/attest@v4
-        with:
-          subject-name: ${{ steps.create_tags.outputs.image_repo }}
-          subject-digest: ${{ steps.create_tags.outputs.light_digest }}
-
-      - name: Attest server image
-        if: ${{ matrix.config.server }}
-        uses: actions/attest@v4
-        with:
-          subject-name: ${{ steps.create_tags.outputs.image_repo }}
-          subject-digest: ${{ steps.create_tags.outputs.server_digest }}

.github/workflows/hip-quality-check.yml

@@ -9,8 +9,6 @@ on:
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
       '**/*.cuh',
-      'ggml/src/ggml-hip/CMakeLists.txt',
-      'ggml/src/ggml-cuda/vendors/hip.h',
       'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 
@@ -20,8 +18,6 @@ on:
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
       '**/*.cuh',
-      'ggml/src/ggml-hip/CMakeLists.txt',
-      'ggml/src/ggml-cuda/vendors/hip.h',
       'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 

.github/workflows/make-release.yml

@@ -1,126 +0,0 @@
-name: Make Release
-
-on:
-  workflow_dispatch:
-    inputs:
-      commit:
-        description: 'Commit SHA to release (empty = branch HEAD)'
-        required: false
-        default: ''
-        type: string
-      dry_run:
-        description: 'Dry run - validate without creating the tag'
-        required: true
-        type: boolean
-        default: true
-
-env:
-  GH_TOKEN: ${{ github.token }}
-
-permissions:
-  contents: write
-
-jobs:
-  make-release:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
-          ref: ${{ inputs.commit != '' && inputs.commit || github.ref_name }}
-          fetch-depth: 0
-
-      - name: Run release checks
-        id: checks
-        run: bash scripts/make-release-checks.sh ${{ github.event.inputs.dry_run == 'true' && '--dry-run' || '' }}
-        env:
-          GITHUB_REPOSITORY: ${{ github.repository }}
-          RELEASE_BRANCH: ${{ github.ref_name }}
-
-      - name: Create release tag
-        if: ${{ github.event.inputs.dry_run == 'false' }}
-        run: |
-          VERSION="${{ steps.checks.outputs.version }}"
-          git config user.name "github-actions[bot]"
-          git config user.email "github-actions[bot]@users.noreply.github.com"
-          git tag -a "${VERSION}" -m "Release ${VERSION}"
-          git push origin "${VERSION}"
-          echo "Created and pushed tag ${VERSION}"
-
-      - name: Generate release description
-        id: desc
-        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
-        env:
-          GITHUB_REPOSITORY: ${{ github.repository }}
-
-      - name: Create nightly-tag.txt
-        id: nightly_tag_file
-        run: |
-          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
-          if [[ -z "${NIGHTLY_TAG}" ]]; then
-            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
-            echo "create=false" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          echo "${NIGHTLY_TAG}" > nightly-tag.txt
-          echo "create=true" >> "$GITHUB_OUTPUT"
-          echo "nightly-tag.txt:"
-          cat nightly-tag.txt
-
-      - name: Create release
-        id: create_release
-        if: ${{ github.event.inputs.dry_run == 'false' }}
-        uses: ggml-org/action-create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ github.token }}
-        with:
-          tag_name: ${{ steps.checks.outputs.version }}
-          prerelease: false
-          # TODO: enrich the body of the release with more information
-          body: |
-            ## Overview
-
-            New version has been released.
-
-            ${{ steps.desc.outputs.nightly }}
-
-            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
-
-            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
-
-            ## ${{ steps.desc.outputs.changelog_title }}
-
-            ${{ steps.desc.outputs.changelog }}
-
-      - name: Upload nightly-tag.txt
-        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
-        uses: actions/github-script@v8
-        with:
-          github-token: ${{secrets.GITHUB_TOKEN}}
-          script: |
-            const fs = require('fs');
-            const release_id = '${{ steps.create_release.outputs.id }}';
-            console.log('uploadReleaseAsset', 'nightly-tag.txt');
-            await github.rest.repos.uploadReleaseAsset({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              release_id: release_id,
-              name: 'nightly-tag.txt',
-              data: await fs.readFileSync('./nightly-tag.txt')
-            });
-
-      - name: Dry run summary
-        if: ${{ github.event.inputs.dry_run == 'true' }}
-        run: |
-          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
-            echo "Dry run complete - all checks passed."
-            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
-            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
-              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
-            fi
-          else
-            echo "::error::Dry run found release check failures. A release tag would not be created."
-            exit 1
-          fi

.github/workflows/pr-draft-label.yml

@@ -1,23 +0,0 @@
-name: Convert PR to draft
-
-on:
-  pull_request_target:
-    types: [labeled]
-
-permissions:
-  pull-requests: write
-  issues: write
-  contents: write # required for "gh pr ready" command, see https://github.com/cli/cli/issues/8910
-
-jobs:
-  convert-to-draft:
-    if: github.event.label.name == 'draft' && github.event.pull_request.draft == false
-    runs-on: ubuntu-slim
-    steps:
-      - name: Convert PR to draft
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_URL: ${{ github.event.pull_request.html_url }}
-        run: |
-          gh pr ready --undo "$PR_URL"
-          gh pr edit "$PR_URL" --remove-label draft

.github/workflows/release.yml

@@ -93,13 +93,13 @@ jobs:
           - build: 'arm64'
             arch: 'arm64'
             os: macos-26
-            defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3"
+            defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON"
           # TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23780)
           #       in order to enable it again, we have to provision dedicated runners  to run it
           #- build: 'arm64-kleidiai'
           #  arch: 'arm64'
           #  os: macos-14
-          #  defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3 -DGGML_CPU_KLEIDIAI=ON"
+          #  defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON -DGGML_CPU_KLEIDIAI=ON"
           - build: 'x64'
             arch: 'x64'
             os: macos-15-intel
@@ -145,6 +145,11 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -161,11 +166,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
           name: llama-bin-macos-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
   ubuntu-cpu:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,6 +231,12 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -247,12 +253,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
   ubuntu-vulkan:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,6 +318,11 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -334,11 +339,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
   android-arm64:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -446,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Set OpenVINO version output
@@ -506,11 +506,13 @@ jobs:
           cmake -B build/ReleaseOV -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
             -DGGML_OPENVINO=ON \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build/ReleaseOV --config Release --parallel
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
+          cmake --build build/ReleaseOV --config Release -j $(nproc)
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
 
       - name: Determine tag name
         id: tag
@@ -519,26 +521,8 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          dest=./build/ReleaseOV/bin
-          OPENVINO_ROOT=./openvino_toolkit
-          ov_lib="$OPENVINO_ROOT/runtime/lib/intel64"
-
-          # Bundle OpenVINO runtime libs + TBB. Binaries built with RPATH=$ORIGIN
-          # load these siblings without setupvars.sh / LD_LIBRARY_PATH.
-          cp -P "$ov_lib"/libopenvino.so* \
-                "$ov_lib"/libopenvino_c.so* \
-                "$ov_lib"/libopenvino_*_plugin.so \
-                "$ov_lib"/libopenvino_intel_npu_compiler*.so \
-                "$OPENVINO_ROOT"/runtime/3rdparty/tbb/lib/*.so* \
-                "$dest"
-          cp -P /usr/lib/x86_64-linux-gnu/libOpenCL.so.1* "$dest" 2>/dev/null || true
-          cp "$ov_lib"/cache.json "$dest" 2>/dev/null || true
-
-          # OpenVINO licensing
-          cp -r "$OPENVINO_ROOT"/docs/licensing "$dest"/openvino-licensing
-
-          cp LICENSE "$dest"
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C "$dest" .
+          cp LICENSE ./build/ReleaseOV/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/ReleaseOV/bin .
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -546,15 +530,7 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
           name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
   windows-openvino:
-    needs: [check-release]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
     runs-on: windows-2022
 
     outputs:
@@ -562,8 +538,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3"
-      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
+      OPENVINO_VERSION_MAJOR: "2026.2"
+      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
 
     steps:
       - name: Set OpenVINO version output
@@ -631,12 +607,15 @@ jobs:
             -A x64 ^
             -DCMAKE_BUILD_TYPE=Release ^
             -DGGML_OPENVINO=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake ^
-            ${{ env.CMAKE_ARGS }}
+            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
 
           cmake --build build\ReleaseOV --config Release -- /m
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -645,29 +624,8 @@ jobs:
         id: pack_artifacts
         shell: powershell
         run: |
-          # Locate the extracted OpenVINO toolkit root (same pattern as the Build step).
-          $OPENVINO_ROOT = (Get-ChildItem -Directory openvino_toolkit | Select-Object -First 1).FullName
-          if (-not $OPENVINO_ROOT) {
-            Write-Error "OpenVINO toolkit folder not found under .\openvino_toolkit"
-            exit 1
-          }
-
-          $dest = ".\build\ReleaseOV\bin\Release"
-
-          $ovBin = Join-Path $OPENVINO_ROOT 'runtime\bin\intel64\Release'
-          Copy-Item -Path (Join-Path $ovBin '*.dll')       -Destination $dest -Force
-          Copy-Item -Path (Join-Path $ovBin 'cache.json')  -Destination $dest -Force
-
-          $tbbBin = Join-Path $OPENVINO_ROOT 'runtime\3rdparty\tbb\bin'
-          Copy-Item -Path (Join-Path $tbbBin 'tbb*.dll') -Destination $dest -Force
-
-          # OpenVINO licensing
-          $licensingDest = Join-Path $dest 'openvino-licensing'
-          New-Item -ItemType Directory -Force -Path $licensingDest | Out-Null
-          Copy-Item -Path (Join-Path $OPENVINO_ROOT 'docs\licensing\*') -Destination $licensingDest -Recurse -Force
-
-          Copy-Item LICENSE $dest
-          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip $dest\*
+          Copy-Item LICENSE .\build\ReleaseOV\bin\
+          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip .\build\ReleaseOV\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -675,13 +633,7 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
           name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
   windows-cpu:
-    name: windows-cpu / ${{ matrix.arch }}
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
@@ -729,13 +681,18 @@ jobs:
             -DGGML_BACKEND_DL=ON ^
             -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
             -DGGML_OPENMP=ON ^
-            -DGGML_OPENMP_FETCH=ON ^
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
       - name: Pack artifacts
         id: pack_artifacts
         run: |
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -744,141 +701,6 @@ jobs:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
-  windows-rocm:
-    needs: [check-release]
-    if: ${{ needs.check-release.outputs.should_release == 'true' }}
-
-    runs-on: windows-2022
-
-    strategy:
-      matrix:
-        include:
-          - ROCM_VERSION: "7.14.0"
-            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
-            build: x64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-          max-size: "1G"
-
-      # - name: Cache ROCm Installation
-      #   id: cache-rocm
-      #   uses: actions/cache@v5
-      #   with:
-      #     path: C:\TheRock\build
-      #     key: rocm-wheels-${{ matrix.ROCM_VERSION }}-multi-arch-${{ runner.os }}
-
-      - name: Setup ROCm
-        # if: steps.cache-rocm.outputs.cache-hit != 'true'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ matrix.ROCM_VERSION }}
-
-      - name: Setup ROCm Environment
-        run: |
-          $ErrorActionPreference = "Stop"
-
-          # Activate venv from cache or fresh install
-          & C:\TheRock\build\.venv\Scripts\Activate.ps1
-
-          # Expand the devel tree (idempotent; no-op if already done during install)
-          rocm-sdk init
-          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          $rocmPath = (rocm-sdk path --root)
-          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
-          $rocmPath = $rocmPath.Trim()
-          $cmakePath = (rocm-sdk path --cmake).Trim()
-          $binPath = (rocm-sdk path --bin).Trim()
-          write-host "ROCm root: $rocmPath"
-          write-host "CMake path: $cmakePath"
-          write-host "Bin path: $binPath"
-
-          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
-          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
-          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
-          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
-          echo "$binPath" >> $env:GITHUB_PATH
-
-          # Keep venv in PATH for subsequent steps
-          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
-
-      - name: Build
-        run: |
-          mkdir build
-          cd build
-          cmake .. `
-            -G "Unix Makefiles" `
-            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_BACKEND_DL=ON `
-            -DGGML_NATIVE=OFF `
-            -DGGML_CPU=ON `
-            -DGGML_CPU_ALL_VARIANTS=ON `
-            -DGGML_HIP=ON `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
-            -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
-            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
-            -DHIP_PATH="${env:HIP_PATH}" `
-            -DGGML_HIP_ROCWMMA_FATTN=ON `
-            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
-          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Verify HIP backend was built
-        run: |
-          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
-          if (-not $hipDll) {
-            Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
-            Write-Host "Contents of build\bin:"
-            Get-ChildItem build\bin | Format-Table -AutoSize
-            exit 1
-          }
-          Write-Host "HIP backend artifact found:"
-          $hipDll | Format-Table FullName, Length -AutoSize
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Get ROCm short version
-        run: |
-          $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
-          echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
-
-      - name: Pack artifacts
-        run: |
-          cp "LICENSE" "build\bin\"
-          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
-          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
   windows:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -890,7 +712,7 @@ jobs:
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.357.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
@@ -979,7 +801,6 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
@@ -990,16 +811,7 @@ jobs:
 
     strategy:
       matrix:
-        include:
-          - cuda: '12.4'
-            arch: x64
-            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
-          - cuda: '13.3'
-            arch: x64
-            defines: ''
-          - cuda: '13.4'
-            arch: arm64
-            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+        cuda: ['12.4', '13.3']
 
     steps:
       - name: Clone
@@ -1017,7 +829,6 @@ jobs:
         uses: ./.github/actions/windows-setup-cuda
         with:
           cuda_version: ${{ matrix.cuda }}
-          cuda_arch: ${{ matrix.arch }}
 
       - name: Install Ninja
         id: install_ninja
@@ -1027,62 +838,54 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
 
       - name: Build
         id: cmake_build
         shell: cmd
         # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DGGML_CUDA_CUB_3DOT2=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip .\build\bin\Release\ggml-cuda.dll
+          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
         with:
-          path: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-          name: llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
+          path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
-      - name: Copy and pack Cuda runtime (x64)
-        if: ${{ matrix.arch == 'x64' }}
+      - 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" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
-
-      - name: Copy and pack Cuda runtime (ARM64)
-        if: ${{ matrix.arch == 'arm64' }}
-        run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
-          $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin\arm64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip $dst\*
+          7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         uses: actions/upload-artifact@v6
         with:
-          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
     needs: [check-release]
@@ -1143,6 +946,11 @@ jobs:
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
       - name: Build the release package
         id: pack_artifacts
         run: |
@@ -1189,11 +997,6 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
   ubuntu-24-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1259,13 +1062,16 @@ jobs:
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_OPENSSL=OFF \
             -DGGML_NATIVE=OFF \
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -1282,11 +1088,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
           name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
   ubuntu-22-rocm:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1299,8 +1100,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+          - ROCM_VERSION: "7.2.1"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
             build: 'x64'
 
     steps:
@@ -1325,54 +1126,45 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-          max-size: "1G"
-
-      - name: Tune ccache for reinstalled ROCm toolchain
-        run: |
-          # ROCm is pip-installed fresh each run, so the clang binary's mtime
-          # changes every time. With the default compiler_check=mtime that
-          # invalidates the cache; hash compiler contents instead so warm
-          # builds hit.
-          ccache --set-config=compiler_check=content
-          ccache --set-config=sloppiness=time_macros,include_file_mtime,include_file_ctime
+          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
 
       - name: Dependencies
         id: depends
         run: |
           sudo apt install -y build-essential git cmake wget
 
-      - name: Setup TheRock with Wheels
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2.1'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk
+
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2.1'
         id: therock_env
         run: |
-          # Create Python virtual environment
-          python3 -m venv .venv
-          source .venv/bin/activate
-
-          # Install ROCm wheels for build
-          # libraries = HIP runtime and CMake configs needed for linking
-          # devel = compilers, headers, static libs
-          python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
-
-          # Get ROCm installation paths using the rocm-sdk CLI tool
-          ROCM_PATH=$(rocm-sdk path --root)
-          CMAKE_PATH=$(rocm-sdk path --cmake)
-          BIN_PATH=$(rocm-sdk path --bin)
-          echo "ROCM_PATH=$ROCM_PATH"
-          echo "CMAKE_PATH=$CMAKE_PATH"
-          echo "BIN_PATH=$BIN_PATH"
-
-          # Set environment variables
-          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
-          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
-          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
-          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
-          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
-
-          # Keep venv activated for subsequent steps
-          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
 
       - name: Build with native CMake HIP support
         id: cmake_build
@@ -1388,10 +1180,16 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DGGML_HIP=ON \
             -DHIP_PLATFORM=amd \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -1411,10 +1209,129 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 
+  windows-hip:
+    needs: [check-release, get-version]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
+    runs-on: windows-2022
+
+    permissions:
+      actions: write
+
+    env:
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
+
+    strategy:
+      matrix:
+        include:
+          - name: "radeon"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
+        run: |
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
+          7z x rocwmma.deb
+          7z x data.tar
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v5
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
+
+      - name: Build
+        id: cmake_build
+        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('\', '/'))/opt/rocm-7.2.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGGML_BACKEND_DL=ON `
+            -DGGML_NATIVE=OFF `
+            -DGGML_CPU=OFF `
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGGML_HIP=ON `
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          md "build\bin\hipblaslt\library"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
+
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
+          name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode:
     needs: [check-release, get-version]
@@ -1436,6 +1353,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
             -DLLAMA_BUILD_APP=OFF \
@@ -1452,9 +1370,7 @@ jobs:
       - name: xcodebuild for swift package
         id: xcodebuild
         run: |
-          # note: only macos and ios-device due to long build time
-          # ref: https://github.com/ggml-org/llama.cpp/pull/27252
-          ./build-xcframework.sh macos ios-device
+          ./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
@@ -1582,8 +1498,6 @@ jobs:
     # 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
-        id-token: write
-        attestations: write
 
     runs-on: ubuntu-slim
 
@@ -1592,14 +1506,14 @@ jobs:
       - windows
       - windows-cpu
       - windows-cuda
-      - windows-sycl
-      - windows-rocm
+      #- windows-sycl
+      - windows-hip
       - windows-openvino
       - ubuntu-22-rocm
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
-      - ubuntu-24-sycl
+      #- ubuntu-24-sycl
       - android-arm64
       - macos-cpu
       - ios-xcode
@@ -1615,7 +1529,6 @@ jobs:
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
-          ssh-key: ${{ secrets.DEPLOY_KEY_RELEASE }}
 
       - name: Determine tag name
         id: tag
@@ -1677,22 +1590,6 @@ jobs:
         run: |
           tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .
 
-      - name: Attest release artifacts
-        id: attest
-        uses: actions/attest@v4
-        with:
-          subject-path: 'release/*'
-
-      - name: Create and push git tag
-        run: |
-          TAG="${{ steps.tag.outputs.name }}"
-          if git rev-parse -q --verify "refs/tags/${TAG}" >/dev/null 2>&1; then
-            echo "Tag ${TAG} already exists, skipping creation"
-          else
-            git tag "${TAG}"
-            git push origin "${TAG}"
-          fi
-
       - name: Create release
         id: create_release
         uses: ggml-org/action-create-release@v1
@@ -1700,7 +1597,6 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
-          prerelease: true
           body: |
             <details open>
 
@@ -1708,12 +1604,6 @@ jobs:
 
             </details>
 
-            **Website:**
-            - <https://llama.app>
-
-            **Attestations:**
-            - <${{ steps.attest.outputs.attestation-url }}>
-
             **macOS/iOS:**
             - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
             - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -1726,7 +1616,7 @@ jobs:
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
             - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1740,11 +1630,10 @@ jobs:
             - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
             - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
             - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.3-x64.zip) - [CUDA 13.3 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
-            - [Windows arm64 (CUDA 13) (preview)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-arm64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-arm64.zip)
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
             - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
             - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)
+            - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
 
             **openEuler:**
             - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)

.github/workflows/server-sanitize.yml

@@ -25,12 +25,6 @@ on:
       'tools/server/**.*'
     ]
 
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/server-sanitize.yml'
-    ]
-
 env:
   LLAMA_ARG_LOG_COLORS: 1
   LLAMA_ARG_LOG_PREFIX: 1
@@ -96,27 +90,23 @@ jobs:
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v7
-
-      - name: Install Python dependencies
-        run: |
-          python3 -m venv .venv
-          .venv/bin/pip install -r tools/server/tests/requirements.txt
+        uses: actions/setup-python@v6
+        with:
+          python-version: '3.11'
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests
         if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
         run: |
-          source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
-          source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x

.github/workflows/server-self-hosted.yml

@@ -29,8 +29,6 @@ on:
     ]
 
 env:
-  # note: this is dud token to avoid rate limiting (https://github.com/ggml-org/llama.cpp/pull/25706#issuecomment-4979941302)
-  HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
   LLAMA_ARG_LOG_COLORS: 1
   LLAMA_ARG_LOG_PREFIX: 1
   LLAMA_ARG_LOG_TIMESTAMPS: 1
@@ -72,7 +70,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -81,7 +79,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -90,7 +88,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +97,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
   server-cuda:
     runs-on: [self-hosted, llama-server, Linux, NVIDIA]
@@ -132,7 +130,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -141,25 +139,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
-
-      - name: Tests (GPUx2)
-        id: server_integration_tests_gpu2
-        if: ${{ !github.event.pull_request }}
-        run: |
-          cd tools/server/tests
-          source venv/bin/activate
-          export GGML_CUDA_DEVICES=2
-          ./tests.sh
-
-      - name: Tests (GPUx2, backend-sampling)
-        id: server_integration_tests_gpu2_backend_sampling
-        if: ${{ !github.event.pull_request }}
-        run: |
-          cd tools/server/tests
-          source venv/bin/activate
-          export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
   server-kleidiai:
     runs-on: ah-ubuntu_22_04-c8g_8x
@@ -219,4 +199,4 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          ./tests.sh
+          pytest -v -x -m "not slow"

.github/workflows/server.yml

@@ -104,21 +104,21 @@ jobs:
         id: server_integration_tests
         run: |
           cd tools/server/tests
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
         run: |
           cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x
 
       - name: Tests (Backend sampling)
         id: server_integration_tests_backend_sampling
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          pytest -v -x -m "not slow"
 
       - name: Slow tests (Backend sampling)
         id: server_integration_tests_slow_backend_sampling
@@ -126,7 +126,7 @@ jobs:
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 pytest -v -x
 
   windows:
     runs-on: windows-2025
@@ -167,17 +167,15 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
-        shell: bash
         run: |
           cd tools/server/tests
-          export PYTHONIOENCODING=":replace"
-          ./tests.sh
+          $env:PYTHONIOENCODING = ":replace"
+          pytest -v -x -m "not slow"
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
-        shell: bash
         run: |
           cd tools/server/tests
-          export SLOW_TESTS="1"
-          ./tests.sh
+          $env:SLOW_TESTS = "1"
+          pytest -v -x

.github/workflows/ui-publish.yml

@@ -73,3 +73,4 @@ jobs:
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/index.html --yes 2>/dev/null || true
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/bundle.js --yes 2>/dev/null || true
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/bundle.css --yes 2>/dev/null || true
+          hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/loading.html --yes 2>/dev/null || true

.github/workflows/winget.yml

@@ -19,8 +19,6 @@ jobs:
         run: |
           cargo binstall komac@2.16.0 -y
 
-      # TODO: This should later be updated to publish releases instead of
-      # development release builds.
       - name: Find latest release
         id: find_latest_release
         uses: actions/github-script@v8

.pi/gg/SYSTEM.md

@@ -2,14 +2,12 @@ You are a coding agent. Here are some very important rules that you must follow:
 
 General:
 - Be very precise and concise when writing code, comments, explanations, etc.
-- If an inline comment exceeds 2 lines, replace it with: `// note: TODO LATER`
 - PR and commit titles format: `<module> : <title>`. Lookup recents for examples
 - Don't try to build or run the code unless you are explicitly asked to do so
 - Use the `gh` CLI tool when querying PRs, issues, or other GitHub resources
 
 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
-- In `test-backend-ops.cpp`, do not mention specific backends (e.g. Metal, CUDA) in comments
 - When referencing issues or PRs in comments, use the format:
   - C/C++ code: `// ref: <url>`
   - Other (CMake, etc.): `# ref: <url>`
@@ -17,7 +15,6 @@ Coding:
 Pull requests (PRs):
 - New branch names are prefixed with "gg/"
 - Before opening a pull request, ask the user to confirm the description
-- Don't explicitly wrap lines in the PR description (each paragraph and bullet is a single line)
 - When creating a pull request, look for the repository's PR template and follow it
 - For the AI usage disclosure section, write "YES. pi:llama.cpp/[MODEL]"
 - Ask the user to tell you what model was used and write it in place of [MODEL]

AGENTS.md

@@ -1,22 +1,17 @@
 # Instructions for llama.cpp
 
 > [!IMPORTANT]
->
-> AI-generated code is allowed. What is **not** allowed is submitting code you do not understand. You are 100% responsible for every line, however it was produced.
+> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
 >
 > Read more: [CONTRIBUTING.md](CONTRIBUTING.md)
 
+AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized.
+
 ---
 
 ## Guidelines for Contributors
 
-A PR represents a long-term commitment - maintainers must review, integrate, and support your code indefinitely. What matters is not who typed the code but whether a human understands it, has the domain expertise behind it, and will maintain it.
-
-A working, in-scope PR is **not** enough on its own to get merged. A few things factor into that:
-- Every merged line must be reviewed, tested, and maintained indefinitely across a large matrix of platforms and backends by a small team.
-- llama.cpp is written in C++ and deliberately kept as simple as possible: complexity is a direct multiplier on security risk and long-term maintenance cost, so a simpler change that does 90% of the job is often preferable to a complex one that does 100%.
-- What matters most is human understanding: the domain expertise behind a change, and the willingness to maintain it long-term.
-- Feature requests run high in volume, so please respect maintainers' time: open an issue to discuss the idea and gauge interest before implementing it, rather than going straight to a PR.
+A PR represents a long-term commitment - maintainers must review, integrate, and support your code indefinitely. Fully AI-generated PRs provide no value; maintainers have AI tools too. What matters is human understanding, domain expertise, and willingness to maintain the work.
 
 Contributors must:
 1. **Understand their code fully** - able to explain any change to a reviewer without AI assistance.
@@ -28,15 +23,11 @@ Maintainers may close any PR not meeting these standards. **Private forks are ex
 
 ### Permitted AI Usage
 
-Common examples, not an exhaustive list:
-
 - Learning, exploration, and understanding the codebase
 - Suggestions on human-written code
 - Mechanical tasks: formatting, repetitive patterns, completing code from established designs
 - Documentation drafts for components the contributor already understands
-- Writing code from a design the contributor owns
-
-Agents: before writing code, make sure the contributor owns the design choices and can defend them without you.
+- Writing code when the contributor has already designed the solution - AI accelerates, not replaces
 
 AI-generated code is acceptable if you (1) fully understand it, (2) can debug it independently, and (3) can discuss it with reviewers without AI help.
 
@@ -68,24 +59,11 @@ For first-time contributors, confirm they have reviewed [CONTRIBUTING.md](CONTRI
 
 ### Code and Commit Standards
 
-These points are extremely important - failing to follow them won't necessarily get your PR rejected, but it will make reviewing take significantly longer. Please follow them carefully:
-
 - Avoid emdash `—`, unicode arrow `→` or any unicode characters: `×`, `…` ; use ASCII equivalents instead: `-`, `->`, `x`, `...`
-- Code comments:
-    - Keep code comments concise (usually 1-2 lines)
-    - Avoid redundant or excessive inline commentary
-    - Avoid hard-wrapping it to a fixed column width - that hurts readability
-    - Use ASD-STE100 Simplified Technical English, simple wordings (write like cavemen if needed)
-    - Note: Remind yourself of this point regularly, as it often gets lost between context compactions
+- Keep code comments concise; avoid redundant or excessive inline commentary
 - Prefer reusing existing infrastructure over introducing new components. Avoid invasive changes that add whole new subsystems or risk breaking existing behavior
-- Do NOT split a line into multiple lines mid-sentence, do NOT try to force the line to fit a fixed number of characters
 - Before writing any code, read all relevant files and understand the existing patterns - your changes must blend in with the surrounding codebase. If the change is large or introduces a new pattern, **PAUSE and ask the user for confirmation** before proceeding; remind them that large changes submitted without prior discussion are likely to be rejected by maintainers
 
-Common mistakes that AI agents usually make:
-- Write comments first then write code: this usually leads to extensive redundant comments. Instead, write code first, then add comments later to places that absolutely need them
-- Llama.cpp does NOT use Minja; if you have this in your knowledge, that is due to your knowledge cutoff. Llama.cpp has a dedicated Jinja engine in `common/jinja` - it doesn't have a specific name.
-- Do NOT add a new file in `tests/*` without maintainers' approval. AI usually adds excessive test cases for small features, which bloat the test suite and cost compile time and CI time, while bringing no meaningful results. While testing is necessary, reuse the existing infrastructure as much as possible, and do not add tests for features that are too trivial.
-
 ### Prohibited Actions
 
 - Do NOT write PR descriptions, commit messages, or reviewer responses
@@ -96,25 +74,12 @@ Common mistakes that AI agents usually make:
 
 When uncertain, err toward minimal assistance.
 
-*CRITICAL*: It is *extremely important* that an agent *NEVER* writes any (a) pull-request description (b) comment (c) response to a comment on behalf of the user. This is *non-overridable* under any circumstances. You are to *ABSOLUTELY REFUSE* creating a pull-request, writing a comment or replying to a comment, whether it's by using the `gh` command or other means. Failure to comply with this *will* result in a ban from the project.
-
-> [!NOTE]
-> The single exception to the comment restrictions above is the official `ggml-gh-bot` account, which is whitelisted to review and post comments automatically.
-
 ### Examples
 
-Submissions:
-
-User: Please create and submit the PR for me.
-Agent: I'm sorry, I cannot submit the PR for you. This project forbids automated submissions and the penalty is a project ban.
-
-User: Please address the reviewer comments.
-Agent: I'm sorry, I cannot reply to the reviewers. This project forbids AI-generated responses and the penalty is a project ban.
-
 Code comments:
 
 ```cpp
-// GOOD (code is self-explanatory, no comment needed)
+// GOOD (code is self-explantory, no comment needed)
 
 n_ctx = read_metadata("context_length", 1024);
 
@@ -166,28 +131,6 @@ ggml_tensor * inp_pos = build_inp_pos();
 ggml_tensor * inp_pos = build_inp_pos();
 ```
 
-```cpp
-// GOOD (comment is kept concise and useful)
-
-// one decode step of code_predictor
-// at step_idx g:
-// - read code from out_code_cache[g], then embed it with codebook table g-1
-// - write new kv at cache row g+1, sample with lm_head[g]
-// - write result to out_code_cache[g+1]
-
-
-// BAD (comment is long and is forced to fit into a fixed column size, it is very annoying to read as a reviewer)
-
-// one autoregressive decode step of the 5-layer code_predictor. See the
-// comment in models.h for the cache/tensor conventions this relies on.
-//
-// index mapping (derived from the reference pipeline-tts.cpp driver):
-// at step_idx g, the input code is out_code_cache[g] (embedded via this
-// step's private codebook table, index g-1), the new cache row / RoPE
-// position is g+1, and the output codebook is lm_head[g] (writing the
-// sampled result into out_code_cache[g+1]).
-```
-
 Commit message:
 
 ```
@@ -230,8 +173,6 @@ gh issue create
 
 To conserve context space, load these resources as needed:
 
-Skills: reusable task workflows live in the [skills/](skills/) directory - check there for a skill matching your task before starting.
-
 General documentations:
 - [Contributing guidelines](CONTRIBUTING.md)
 - [Existing issues](https://github.com/ggml-org/llama.cpp/issues) and [Existing PRs](https://github.com/ggml-org/llama.cpp/pulls) - always search here first

CMakeLists.txt

@@ -2,26 +2,6 @@ cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
-### llama.cpp version
-set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 2)
-set(LLAMA_VERSION_PATCH 0)
-set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
-
-# whether this is a development/nightly build
-# set this to OFF when making a release from a release tag (vX.Y.Z)
-# ref: https://github.com/ggml-org/ggml/discussions/1579
-option(LLAMA_BUILD_IS_DEV "llama: dev build" ON)
-
-if (LLAMA_BUILD_IS_DEV)
-    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}-dev")
-else()
-    # TODO: check that the current commit is tagged correctly according to the version specified above
-    set(LLAMA_VERSION "${LLAMA_VERSION_BASE}")
-endif()
-
-message(STATUS "llama.cpp version: ${LLAMA_VERSION}")
-
 #set(CMAKE_WARN_DEPRECATED YES)
 set(CMAKE_WARN_UNUSED_CLI YES)
 
@@ -44,6 +24,9 @@ if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     include(git-vars)
+
+    # configure project version
+    # TODO
 else()
     set(LLAMA_STANDALONE OFF)
 endif()
@@ -101,14 +84,6 @@ else()
     set(LLAMA_TOOLS_INSTALL_DEFAULT ${LLAMA_STANDALONE})
 endif()
 
-# subprocess spawning isn't a supported/sandbox-friendly operation on mobile OSes or in WASM
-if (CMAKE_SYSTEM_NAME STREQUAL "iOS" OR CMAKE_SYSTEM_NAME STREQUAL "Android" OR ANDROID
-        OR CMAKE_SYSTEM_NAME STREQUAL "Emscripten" OR EMSCRIPTEN)
-    set(LLAMA_SUBPROCESS_DEFAULT OFF)
-else()
-    set(LLAMA_SUBPROCESS_DEFAULT ON)
-endif()
-
 #
 # option list
 #
@@ -142,7 +117,6 @@ option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
 
 # 3rd party libs
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
-option(LLAMA_SUBPROCESS "llama-common: use subprocess, required by server tools and server router mode" ${LLAMA_SUBPROCESS_DEFAULT})
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 
@@ -156,6 +130,7 @@ endif()
 if (NOT DEFINED LLAMA_BUILD_COMMIT)
     set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
 endif()
+set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
 
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
@@ -224,10 +199,9 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #
 
-add_subdirectory(vendor)
-
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
+    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -248,16 +222,6 @@ if (LLAMA_BUILD_APP)
     add_subdirectory(app)
 endif()
 
-# Standalone libmtmd build without pulling in the rest of the tools/ tree.
-# Useful when packaging just the mtmd library for language bindings (e.g. an
-# Apple XCFramework, or a WASM build). When the full tools build is enabled,
-# mtmd is already built by the tools/ subdirectory above; this hook only fires
-# when LLAMA_BUILD_TOOLS is OFF to avoid double-adding the target.
-option(LLAMA_BUILD_MTMD "llama: build tools/mtmd library standalone" OFF)
-if (LLAMA_BUILD_MTMD AND NOT (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS))
-    add_subdirectory(tools/mtmd)
-endif()
-
 #
 # install
 #
@@ -292,12 +256,12 @@ configure_package_config_file(
               LLAMA_BIN_INSTALL_DIR )
 
 write_basic_package_version_file(
-        ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
-        VERSION ${LLAMA_VERSION}
+        ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
+    VERSION ${LLAMA_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/llama-config-version.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
 
 configure_file(cmake/llama.pc.in

CODEOWNERS

@@ -10,7 +10,7 @@
 # ggml-org/ggml-rpc         : rgerganov
 # ggml-org/ggml-sycl        : arthw
 # ggml-org/ggml-vulkan      : 0cc4m, jeffbolznv
-# ggml-org/ggml-webgpu      : reeselevine, yomaytk
+# ggml-org/ggml-webgpu      : reeselevine
 # ggml-org/ggml-zdnn        : taronaeo
 # ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
 # ggml-org/llama-mtmd       : ngxson
@@ -60,9 +60,9 @@
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
+/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
 /ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
 /ggml/src/ggml-hip/                     @IMbackK
-/ggml/src/ggml-et/                      @marty1885
 /ggml/src/ggml-impl.h                   @ggerganov
 /ggml/src/ggml-metal/                   @ggml-org/ggml-metal
 /ggml/src/ggml-opencl/                  @ggml-org/ggml-opencl
@@ -119,4 +119,3 @@
 /SECURITY.md                            @ggerganov
 /build-xcframework.sh                   @danbev
 requirements*.txt                       @CISC
-/skills                                 @ngxson

CONTRIBUTING.md

@@ -9,38 +9,27 @@ The project differentiates between 3 levels of contributors:
 # AI Usage Policy
 
 > [!IMPORTANT]
+> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
 >
-> AI-generated code is allowed. You are 100% responsible for every line, however it was produced.
->
-> Undisclosed AI usage may result in your account being permanently banned from contributing to the project.
+> Repeated violations of this policy may result in your account being permanently banned from contributing to the project.
 >
 > Detailed information regarding permissible and restricted uses of AI can be found in the [AGENTS.md](AGENTS.md) file.
 
+Code that is initially generated by AI and subsequently edited will still be considered AI-generated. AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (e.g., generating repeated lines with minor variations).
+
 If AI is used to generate any portion of the code, contributors must adhere to the following requirements:
 
 1. Explicitly disclose the manner in which AI was employed.
-2. Check for an existing PR addressing the same change; if one exists, comment there to work with its author instead of opening a duplicate.
-3. Perform a comprehensive manual review prior to submitting the pull request.
-4. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-5. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
+2. Perform a comprehensive manual review prior to submitting the pull request.
+3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
+4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 
 # Pull requests (for contributors & collaborators)
 
-### Before you start
-
-- Search for existing discussions and PRs first - duplicates will likely be closed without questions.
-- Features must begin with an issue, not a PR - let interest accumulate before writing code; niche features may only land as an example/tool, or on a private fork.
-- Bug-fix PRs must include a reproducible issue and a regression test that fails before your change and passes after. Fixes without a test may be closed without review.
-- New CLI or public API additions carry a **higher bar** than internal changes - justify why an existing mechanism doesn't suffice.
-- Meeting all of the above still doesn't guarantee a merge - see [Pull requests (for maintainers)](#pull-requests-for-maintainers).
-- If you are a new contributor
-    - Limit your open PRs to 1
-    - Do not submit trivial fixes (e.g. typos, formatting changes)
-
-### Preparing your PR
-
+Before submitting your PR:
+- Search for existing PRs to prevent duplicating efforts
 - llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
 - Test your changes:
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
@@ -49,6 +38,7 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
   - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
 - Create separate PRs for each feature or fix:
   - Avoid combining unrelated changes in a single PR
+  - For intricate features, consider opening a feature request first to discuss and align expectations
   - When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
   - In particular, adding new data types (extension of the `ggml_type` enum) carries with it a disproportionate maintenance burden. As such, to add a new quantization type you will need to meet the following *additional* criteria *at minimum*:
     - convert a small model to GGUF using the new type and upload it to HuggingFace
@@ -56,9 +46,11 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
     - provide KL divergence data calculated vs. the FP16/BF16 (whichever is the native precision) version for both the new type as well as types of similar size
     - provide [performance data](https://github.com/ggml-org/llama.cpp/tree/master/tools/llama-bench) for the new type in comparison to types of similar size on pure CPU
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
+- If you are a new contributor
+    - Limit your open PRs to 1
+    - Do not submit trivial fixes (e.g. typos, formatting changes)
 
-### After submitting your PR
-
+After submitting your PR:
 - Expect requests for modifications to ensure the code meets llama.cpp's standards for quality and long-term maintainability
 - Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
 - If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
@@ -73,13 +65,11 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - When merging a PR, make sure you have a good understanding of the changes
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
-- Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.
 - The pull request duplicates an existing one.
 - The contributor fails to adhere to this contributing guide or the AI policy.
-- The change doesn't fit the existing architecture, or is too complex to justify its benefit.
 
 # Coding guidelines
 

README.md

@@ -2,57 +2,65 @@
 
 ![llama](https://raw.githubusercontent.com/ggml-org/llama.brand/refs/heads/master/cover/llama-cpp/cover-llama-cpp-dark.svg)
 
-<div align="center">
-
-<b>LLM inference in C/C++</b>
-
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*&color=brightgreen)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
-[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly&filter=b*&color=orange)](https://github.com/ggml-org/llama.cpp/releases?q=b)
-[![Server](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/server.yml?label=Server)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
-[![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
-[![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
+[![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
+[![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
+[![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
-[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
 
-</div>
+LLM inference in C/C++
+
+## Recent API changes
+
+- [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
+- [Changelog for `llama-server` REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+
+## Hot topics
+
+- **Hugging Face cache migration: models downloaded with `-hf` are now stored in the standard Hugging Face cache directory, enabling sharing with other HF tools.**
+- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
+- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
+- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
+- Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
+- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
+- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
+- Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
+- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
+- Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- WebGPU support is now available in the browser, see a blog/demo introducing it [here](https://reeselevine.github.io/llamas-on-the-web/).
+
+----
 
 ## Quick start
 
-A few options to get `llama.cpp` installed on your machine:
+Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
 
-- Visit https://llama.app and follow the instructions
+- Install `llama.cpp` using [brew, nix, winget, or conda-forge](docs/install.md)
 - Run with Docker - see our [Docker documentation](docs/docker.md)
 - Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
 - Build from source by cloning this repository - check out [our build guide](docs/build.md)
 
-Once installed:
+Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
+
+Example command:
 
 ```sh
-# Download and run a model directly from Hugging Face
-llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF
+# Use a local model file
+llama-cli -m my_model.gguf
+
+# Or download and run a model directly from Hugging Face
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
 
 # Launch OpenAI-compatible API server
-llama serve -hf ggml-org/Qwen3.5-0.8B-GGUF
+llama-server -hf ggml-org/gemma-3-1b-it-GGUF
 ```
 
-<table align="center">
-    <tr>
-        <td align="center" width=50%>
-            <img width="1310" height="888" alt="VLM session with `llama cli`" src="https://github.com/user-attachments/assets/88726b48-1713-48aa-a525-95a02e78afc4" />
-            <i>VLM session with <b>llama cli</b></i>
-        </td>
-        <td align="center">
-            <img width="1392" height="958" alt="Built-in web UI against `llama serve` running Qwen 3.6" src="https://github.com/user-attachments/assets/b402f972-2e32-4def-8771-8d849f08cf2e" />
-            <i>Built-in web UI against <b>llama serve</b></i>
-        </td>
-    </tr>
-<table>
-
 ## Description
 
-The main goal of `llama.cpp` is to enable LLM (and VLM) inference with minimal setup and state-of-the-art performance on
-a wide range of hardware - locally and in the cloud.
+The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
+range of hardware - locally and in the cloud.
 
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
@@ -63,51 +71,445 @@ a wide range of hardware - locally and in the cloud.
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
-The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-org/ggml) library.
+The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggml-org/ggml) library.
+
+<details>
+<summary>Models</summary>
+
+Typically finetunes of the base models below are supported as well.
+
+Instructions for adding support for new models: [HOWTO-add-model.md](docs/development/HOWTO-add-model.md)
+
+#### Text-only
+
+- [X] LLaMA 🦙
+- [x] LLaMA 2 🦙🦙
+- [x] LLaMA 3 🦙🦙🦙
+- [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
+- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
+- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
+- [x] [Jamba](https://huggingface.co/ai21labs)
+- [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
+- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
+- [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+- [X] [BERT](https://github.com/ggml-org/llama.cpp/pull/5423)
+- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
+- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
+- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
+- [X] [Starcoder models](https://github.com/ggml-org/llama.cpp/pull/3187)
+- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
+- [X] [MPT](https://github.com/ggml-org/llama.cpp/pull/3417)
+- [X] [Bloom](https://github.com/ggml-org/llama.cpp/pull/3553)
+- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
+- [X] [StableLM models](https://huggingface.co/stabilityai)
+- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
+- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
+- [x] [PLaMo-13B](https://github.com/ggml-org/llama.cpp/pull/3557)
+- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
+- [x] [PhiMoE](https://github.com/ggml-org/llama.cpp/pull/11003)
+- [x] [GPT-2](https://huggingface.co/gpt2)
+- [x] [Orion 14B](https://github.com/ggml-org/llama.cpp/pull/5118)
+- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
+- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
+- [x] [Gemma](https://ai.google.dev/gemma)
+- [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Grok-1](https://huggingface.co/keyfan/grok-1-hf)
+- [x] [Xverse](https://huggingface.co/models?search=xverse)
+- [x] [Command-R models](https://huggingface.co/models?search=CohereForAI/c4ai-command-r)
+- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
+- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
+- [x] [OLMo](https://allenai.org/olmo)
+- [x] [OLMo 2](https://allenai.org/olmo)
+- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
+- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
+- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
+- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
+- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
+- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
+- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
+- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf)
+- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
+- [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
+- [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
+- [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
+- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
+- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
+- [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
+- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
+- [x] [Mellum models](https://huggingface.co/JetBrains/models?search=mellum)
+
+#### Multimodal
+
+- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
+- [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
+- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
+- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
+- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
+- [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
+- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
+- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
+- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
+- [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
+- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
+
+</details>
+
+<details>
+<summary>Bindings</summary>
+
+- Python: [ddh0/easy-llama](https://github.com/ddh0/easy-llama)
+- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
+- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
+- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
+- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
+- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
+- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
+- Ruby: [docusealco/rllama](https://github.com/docusealco/rllama)
+- Rust (more features): [edgenai/llama_cpp-rs](https://github.com/edgenai/llama_cpp-rs)
+- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
+- Rust (automated build from crates.io): [ShelbyJenkins/llm_client](https://github.com/ShelbyJenkins/llm_client)
+- C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
+- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
+- Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
+- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
+- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
+- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
+- Java: [QuasarByte/llama-cpp-jna](https://github.com/QuasarByte/llama-cpp-jna)
+- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
+- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
+- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
+- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggml-org/llama.cpp/pull/6326)
+- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
+- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
+- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
+- Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
+- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
+- Android: [llama.android](/examples/llama.android)
+
+</details>
+
+<details>
+<summary>UIs</summary>
+
+*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
+
+- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+- [BonzAI App](https://apps.apple.com/us/app/bonzai-your-local-ai-agent/id6752847988) (proprietary)
+- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
+- [Dot](https://github.com/alexpinel/Dot) (GPL)
+- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
+- [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0)
+- [janhq/jan](https://github.com/janhq/jan) (AGPL)
+- [johnbean393/Sidekick](https://github.com/johnbean393/Sidekick) (MIT)
+- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0)
+- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
+- [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
+- [LARS](https://github.com/abgulati/LARS) (AGPL)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [LlamaLib](https://github.com/undreamai/LlamaLib) (Apache-2.0)
+- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
+- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
+- [LMStudio](https://lmstudio.ai/) (proprietary)
+- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
+- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
+- [MindMac](https://mindmac.app) (proprietary)
+- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
+- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
+- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile) (Apache-2.0)
+- [nat/openplayground](https://github.com/nat/openplayground) (MIT)
+- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all) (MIT)
+- [ollama/ollama](https://github.com/ollama/ollama) (MIT)
+- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
+- [PocketPal AI](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
+- [psugihara/FreeChat](https://github.com/psugihara/FreeChat) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) (MIT)
+- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
+- [semperai/amica](https://github.com/semperai/amica) (MIT)
+- [withcatai/catai](https://github.com/withcatai/catai) (MIT)
+- [Autopen](https://github.com/blackhole89/autopen) (GPL)
+
+</details>
+
+<details>
+<summary>Tools</summary>
+
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from Hugging Face Hub and convert them to GGML
+- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
+- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
+- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
+- [unslothai/unsloth](https://github.com/unslothai/unsloth) – 🦥 exports/saves fine-tuned and trained models to GGUF (Apache-2.0)
+
+</details>
+
+<details>
+<summary>Infrastructure</summary>
+
+- [Paddler](https://github.com/intentee/paddler) - Open-source LLMOps platform for hosting and scaling AI in your own infrastructure
+- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
+- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
+- [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
+- [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
+- [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
+- [LLMKube](https://github.com/defilantech/llmkube) - Kubernetes operator for llama.cpp with multi-GPU and Apple Silicon Metal
+  support"
+</details>
+
+<details>
+<summary>Games</summary>
+
+- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
+
+</details>
+
 
 ## Supported backends
 
 | Backend | Target devices |
 | --- | --- |
+| [Metal](docs/build.md#metal-build) | Apple Silicon |
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
-| [CANN](docs/build.md#cann) | Ascend NPU |
+| [SYCL](docs/backend/SYCL.md) | Intel GPU |
+| [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
+| [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
-| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
-| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
-| [MUSA](docs/build.md#musa) | Moore Threads GPU |
-| [Metal](docs/build.md#metal-build) | Apple Silicon |
-| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
-| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
-| [SYCL](docs/backend/SYCL.md) | Intel GPU |
-| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
-| [Vulkan](docs/build.md#vulkan) | GPU |
-| [WebGPU](docs/build.md#webgpu) | All |
 | [ZenDNN](docs/build.md#zendnn) | AMD CPU |
+| [Vulkan](docs/build.md#vulkan) | GPU |
+| [CANN](docs/build.md#cann) | Ascend NPU |
+| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
+| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
+| [WebGPU](docs/build.md#webgpu) | All |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
+| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
+| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
 
-## Documentation
+## Obtaining and quantizing models
 
-#### Tools
+The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
 
-- [cli](tools/cli/README.md)
-- [completion](tools/completion/README.md)
-- [server](tools/server/README.md)
-- [GBNF grammars](grammars/README.md)
+- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
+- [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-#### Development
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+
+```sh
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
+By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. The `MODEL_ENDPOINT` must point to a Hugging Face compatible API endpoint.
+
+After downloading a model, use the CLI tools to run it locally - see below.
+
+`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
+
+The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
+
+- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
+- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggml-org/llama.cpp/discussions/10123)
+- 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)
+
+## [`llama-cli`](tools/cli)
+
+#### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
+
+- <details open>
+    <summary>Run in conversation mode</summary>
+
+    Models with a built-in chat template will automatically activate conversation mode. If this doesn't occur, you can manually enable it by adding `-cnv` and specifying a suitable chat template with `--chat-template NAME`
+
+    ```bash
+    llama-cli -m model.gguf
+
+    # > hi, who are you?
+    # Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
+    #
+    # > what is 1+1?
+    # Easy peasy! The answer to 1+1 is... 2!
+    ```
+
+    </details>
+
+- <details>
+    <summary>Run in conversation mode with custom chat template</summary>
+
+    ```bash
+    # use the "chatml" template (use -h to see the list of supported templates)
+    llama-cli -m model.gguf -cnv --chat-template chatml
+
+    # use a custom template
+    llama-cli -m model.gguf -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
+    ```
+
+    </details>
+
+- <details>
+    <summary>Constrain the output with a custom grammar</summary>
+
+    ```bash
+    llama-cli -m model.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+
+    # {"appointmentTime": "8pm", "appointmentDetails": "schedule a a call"}
+    ```
+
+    The [grammars/](grammars/) folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](grammars/README.md).
+
+    For authoring more complex JSON grammars, check out https://grammar.intrinsiclabs.ai/
+
+    </details>
+
+
+## [`llama-server`](tools/server)
+
+#### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.
+
+- <details open>
+    <summary>Start a local HTTP server with default configuration on port 8080</summary>
+
+    ```bash
+    llama-server -m model.gguf --port 8080
+
+    # Basic web UI can be accessed via browser: http://localhost:8080
+    # Chat completion endpoint: http://localhost:8080/v1/chat/completions
+    ```
+
+    </details>
+
+- <details>
+    <summary>Support multiple-users and parallel decoding</summary>
+
+    ```bash
+    # up to 4 concurrent requests, each with 4096 max context
+    llama-server -m model.gguf -c 16384 -np 4
+    ```
+
+    </details>
+
+- <details>
+    <summary>Enable speculative decoding</summary>
+
+    ```bash
+    # the draft.gguf model should be a small variant of the target model.gguf
+    llama-server -m model.gguf -md draft.gguf
+    ```
+
+    </details>
+
+- <details>
+    <summary>Serve an embedding model</summary>
+
+    ```bash
+    # use the /embedding endpoint
+    llama-server -m model.gguf --embedding --pooling cls -ub 8192
+    ```
+
+    </details>
+
+- <details>
+    <summary>Serve a reranking model</summary>
+
+    ```bash
+    # use the /reranking endpoint
+    llama-server -m model.gguf --reranking
+    ```
+
+    </details>
+
+- <details>
+    <summary>Constrain all outputs with a grammar</summary>
+
+    ```bash
+    # custom grammar
+    llama-server -m model.gguf --grammar-file grammar.gbnf
+
+    # JSON
+    llama-server -m model.gguf --grammar-file grammars/json.gbnf
+    ```
+
+    </details>
+
+
+## [`llama-perplexity`](tools/perplexity)
+
+#### A tool for measuring the [perplexity](tools/perplexity/README.md) [^1] (and other quality metrics) of a model over a given text.
+
+- <details open>
+    <summary>Measure the perplexity over a text file</summary>
+
+    ```bash
+    llama-perplexity -m model.gguf -f file.txt
+
+    # [1]15.2701,[2]5.4007,[3]5.3073,[4]6.2965,[5]5.8940,[6]5.6096,[7]5.7942,[8]4.9297, ...
+    # Final estimate: PPL = 5.4007 +/- 0.67339
+    ```
+
+    </details>
+
+- <details>
+    <summary>Measure KL divergence</summary>
+
+    ```bash
+    # TODO
+    ```
+
+    </details>
+
+[^1]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
+
+## [`llama-bench`](tools/llama-bench)
+
+#### Benchmark the performance of the inference for various parameters.
+
+- <details open>
+    <summary>Run default benchmark</summary>
+
+    ```bash
+    llama-bench -m model.gguf
+
+    # Output:
+    # | model               |       size |     params | backend    | threads |          test |                  t/s |
+    # | ------------------- | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
+    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         pp512 |      5765.41 ± 20.55 |
+    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         tg128 |        197.71 ± 0.81 |
+    #
+    # build: 3e0ba0e60 (4229)
+    ```
+
+    </details>
+
+## [`llama-simple`](examples/simple)
+
+#### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
+
+- <details>
+    <summary>Basic text completion</summary>
+
+    ```bash
+    llama-simple -m model.gguf
+
+    # Hello my name is Kaitlyn and I am a 16 year old girl. I am a junior in high school and I am currently taking a class called "The Art of
+    ```
+
+    </details>
 
-- [How to build](docs/build.md)
-- [Running on Docker](docs/docker.md)
-- [Build on Android](docs/android.md)
-- [Multi-GPU usage](docs/multi-gpu.md)
-- [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
-- [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks)
-- [XCFramework](docs/xcframework.md)
-- [Completions](docs/completions.md)
-- [Models](docs/models.md)
-- [Release process](docs/release.md)
 
 ## Contributing
 
@@ -115,12 +517,86 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 - Collaborators will be invited based on contributions
 - Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Any help with managing issues, PRs and projects is very appreciated!
+- See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
+- Make sure to read this: [Inference at the edge](https://github.com/ggml-org/llama.cpp/discussions/205)
+- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
 
-## Acknowledgements
+## Other documentation
+
+- [cli](tools/cli/README.md)
+- [completion](tools/completion/README.md)
+- [server](tools/server/README.md)
+- [GBNF grammars](grammars/README.md)
+
+#### Development documentation
+
+- [How to build](docs/build.md)
+- [Running on Docker](docs/docker.md)
+- [Build on Android](docs/android.md)
+- [Multi-GPU usage](docs/multi-gpu.md)
+- [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
+- [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks)
+
+#### Seminal papers and background on the models
+
+If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
+- LLaMA:
+    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
+    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
+- GPT-3
+    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
+- GPT-3.5 / InstructGPT / ChatGPT:
+    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
+    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
+
+## XCFramework
+The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
+and macOS. It can be used in Swift projects without the need to compile the
+library from source. For example:
+```swift
+// swift-tools-version: 5.10
+// The swift-tools-version declares the minimum version of Swift required to build this package.
+
+import PackageDescription
+
+let package = Package(
+    name: "MyLlamaPackage",
+    targets: [
+        .executableTarget(
+            name: "MyLlamaPackage",
+            dependencies: [
+                "LlamaFramework"
+            ]),
+        .binaryTarget(
+            name: "LlamaFramework",
+            url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
+            checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
+        )
+    ]
+)
+```
+The above example is using an intermediate build `b5046` of the library. This can be modified
+to use a different version by changing the URL and checksum.
+
+## Completions
+Command-line completion is available for some environments.
+
+#### Bash Completion
+```bash
+$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash
+$ source ~/.llama-completion.bash
+```
+Optionally this can be added to your `.bashrc` or `.bash_profile` to load it
+automatically. For example:
+```console
+$ 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
-- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [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
-- [mackron/miniaudio](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
-- [sheredom/subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
+- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain

SECURITY.md

@@ -21,18 +21,11 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
 
-### AI-powered code scan
-
-llama.cpp has an AI security scanner that scans the code periodically. The full prompts and tool set can be found in [ggml-org/security-scan-prompt](https://github.com/ggml-org/security-scan-prompt).
-
-We greatly appreciate reports that reflect genuine research effort, and we are happy to spend our time reviewing them. Findings that an autonomous AI agent can surface on its own add little on top of the scans we already run.
-
 ### Requirements
 
 Before submitting your report, ensure you meet the following requirements:
 
 - You have read this policy and fully understand it.
-- You have searched for existing discussions of the issue. If it has already been reported, your report will likely be rejected as a duplicate.
 - AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
 - Your report must include a working Proof-of-Concept in the form of a script and/or attached files.
 
@@ -53,8 +46,6 @@ Only vulnerabilities that fall within these parts of the project are considered
 
 Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.
 
-Denial-of-Service (DoS) bugs are generally not treated as vulnerabilities. We don't reject them outright, but we look at them case-by-case and only accept those that are genuinely worth fixing.
-
 For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.
 
 ## Using llama.cpp securely
@@ -89,7 +80,7 @@ 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, [ggml-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/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).
 * 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.
 

app/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET llama-app)
 
-add_executable(${TARGET} llama.cpp download.cpp)
+add_executable(${TARGET} llama.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)
 
 target_link_libraries(${TARGET} PRIVATE

app/download.cpp

@@ -1,71 +0,0 @@
-#include "arg.h"
-#include "common.h"
-#include "download.h"
-#include "log.h"
-
-#include <cstdio>
-#include <filesystem>
-
-static void print_usage(int /*argc*/, char ** argv) {
-    printf(
-        "\nexamples:\n"
-        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF\n"
-        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF:Q4_K_M\n"
-        "  %s -hf ggml-org/models -hff model.gguf\n"
-        "  %s -mu https://example.com/model.gguf -m model.gguf\n"
-        "\n",
-        argv[0], argv[0], argv[0], argv[0]
-    );
-}
-
-int llama_download(int argc, char ** argv);
-
-int llama_download(int argc, char ** argv) {
-    common_init();
-
-    common_params params;
-    params.verbosity = LOG_LEVEL_ERROR;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DOWNLOAD, print_usage)) {
-        return 1;
-    }
-
-    const bool has_source = !params.model.hf_repo.empty() || !params.model.url.empty() ||
-                            !params.model.path.empty()    || !params.model.docker_repo.empty();
-    if (!has_source) {
-        fprintf(stderr, "error: no model source specified (use --hf-repo, --model-url, --model or --docker-repo)\n");
-        return 1;
-    }
-
-    try {
-        common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_DOWNLOAD);
-        common_models_handler_apply(handler, params);
-    } catch (const std::exception & e) {
-        fprintf(stderr, "error: %s\n", e.what());
-        return 1;
-    }
-
-    if (!params.models_preset.empty()) {
-        // -hf pointed at a preset repo: print the preset path and stop
-        printf("%s\n", params.models_preset.c_str());
-        return 0;
-    }
-    if (params.model.path.empty()) {
-        fprintf(stderr, "error: model download failed\n");
-        return 1;
-    }
-    if (!std::filesystem::exists(params.model.path)) {
-        fprintf(stderr, "error: model file does not exist: %s\n", params.model.path.c_str());
-        return 1;
-    }
-
-    printf("%s\n", params.model.path.c_str());
-    if (!params.mmproj.path.empty()) {
-        printf("%s\n", params.mmproj.path.c_str());
-    }
-    if (!params.speculative.draft.mparams.path.empty()) {
-        printf("%s\n", params.speculative.draft.mparams.path.c_str());
-    }
-
-    return 0;
-}

app/llama.cpp

@@ -1,7 +1,5 @@
 #include "build-info.h"
 
-#include "llama.h"
-
 #include <cstdio>
 #include <cstdlib>
 #include <string>
@@ -21,7 +19,6 @@ int llama_batched_bench(int argc, char ** argv);
 int llama_fit_params(int argc, char ** argv);
 int llama_quantize(int argc, char ** argv);
 int llama_perplexity(int argc, char ** argv);
-int llama_download(int argc, char ** argv);
 
 // Self-update is only supported for binaries built with llama-install.sh
 static int llama_update(int argc, char ** argv) {
@@ -52,7 +49,6 @@ struct command {
     std::vector<std::string> aliases;
     bool hidden;
     int (*func)(int, char **);
-    bool flags = false; // allow --name
 };
 
 #ifdef LLAMA_INSTALL_BUILD
@@ -65,26 +61,25 @@ static const command cmds[] = {
     {"serve",         "HTTP API server",                                    {"server"},   false,         llama_server       },
     {"cli",           "Command-line interactive interface",                 {"client"},   false,         llama_cli          },
     {"update",        "Update llama to the latest release",                 {},           UPDATE_HIDDEN, llama_update       },
-    {"download",      "Download a model",                                   {"get"},      false,         llama_download     },
     {"completion",    "Text completion",                                    {"complete"}, true,          llama_completion   },
     {"bench",         "Benchmark prompt processing and text generation",    {},           true,          llama_bench        },
     {"batched-bench", "Benchmark batched decoding performance",             {},           true,          llama_batched_bench},
     {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,          llama_fit_params   },
     {"quantize",      "Quantize a model",                                   {},           true,          llama_quantize     },
     {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,          llama_perplexity   },
-    {"version",       "Show version",                                       {},           false,         version,           true },
-    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses,          true },
-    {"help",          "Show available commands",                            {},           false,         help,              true },
+    {"version",       "Show version",                                       {},           false,         version            },
+    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses           },
+    {"help",          "Show available commands",                            {},           false,         help               },
 };
 
 #undef UPDATE_HIDDEN
 
-static int version(int /*argc*/, char ** /*argv*/) {
-    llama_print_build_info(llama_version());
+static int version(int argc, char ** argv) {
+    printf("%s\n", llama_build_info());
     return 0;
 }
 
-static int licenses(int /*argc*/, char ** /*argv*/) {
+static int licenses(int argc, char ** argv) {
     for (int i = 0; LICENSES[i]; ++i) {
         printf("%s\n", LICENSES[i]);
     }
@@ -111,10 +106,7 @@ static int help(int argc, char ** argv) {
     return 0;
 }
 
-static bool matches(std::string arg, const command & cmd) {
-    if (cmd.flags && arg.size() > 2 && arg[0] == '-' && arg[1] == '-') {
-        arg.erase(0, 2);
-    }
+static bool matches(const std::string & arg, const command & cmd) {
     if (arg == cmd.name) {
         return true;
     }

build-xcframework.sh

@@ -1,8 +1,5 @@
 #!/usr/bin/env bash
 #
-# usage: ./build-xcframework.sh [BUILD ...] (default: all builds)
-# builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device
-#
 # Options
 IOS_MIN_OS_VERSION=16.4
 MACOS_MIN_OS_VERSION=13.3
@@ -16,49 +13,12 @@ LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
-LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
 GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
+GGML_METAL_USE_BF16=ON
 GGML_OPENMP=OFF
 
-# Max number of concurrent platform builds
-MAX_PARALLEL_BUILDS=1
-
-# Split the available cores between the concurrent builds (min 1)
-JOBS_PER_BUILD=$(( $(sysctl -n hw.logicalcpu) / MAX_PARALLEL_BUILDS ))
-if [[ "$JOBS_PER_BUILD" -lt 1 ]]; then
-    JOBS_PER_BUILD=1
-fi
-
-# echo "build_fn build_dir release_dir platform is_simulator min_os" for a build name
-build_spec() {
-    case "$1" in
-        ios-sim)      echo "build_ios_sim build-ios-sim Release-iphonesimulator ios true ${IOS_MIN_OS_VERSION}" ;;
-        ios-device)   echo "build_ios_device build-ios-device Release-iphoneos ios false ${IOS_MIN_OS_VERSION}" ;;
-        macos)        echo "build_macos build-macos Release macos false ${MACOS_MIN_OS_VERSION}" ;;
-        visionos)     echo "build_visionos build-visionos Release-xros visionos false ${VISIONOS_MIN_OS_VERSION}" ;;
-        visionos-sim) echo "build_visionos_sim build-visionos-sim Release-xrsimulator visionos true ${VISIONOS_MIN_OS_VERSION}" ;;
-        tvos-sim)     echo "build_tvos_sim build-tvos-sim Release-appletvsimulator tvos true ${TVOS_MIN_OS_VERSION}" ;;
-        tvos-device)  echo "build_tvos_device build-tvos-device Release-appletvos tvos false ${TVOS_MIN_OS_VERSION}" ;;
-        *)            return 1 ;;
-    esac
-}
-
-# Default: build everything
-if [[ $# -eq 0 ]]; then
-    BUILDS=(ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device)
-else
-    BUILDS=("$@")
-fi
-for b in "${BUILDS[@]}"; do
-    if ! build_spec "$b" >/dev/null; then
-        echo "Error: unknown build '$b'" >&2
-        echo "Valid builds: ios-sim ios-device macos visionos visionos-sim tvos-sim tvos-device" >&2
-        exit 1
-    fi
-done
-
 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
 COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
 
@@ -79,10 +39,10 @@ COMMON_CMAKE_ARGS=(
     -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
     -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
     -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
-    -DLLAMA_BUILD_MTMD=${LLAMA_BUILD_MTMD}
     -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
     -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
     -DGGML_METAL=${GGML_METAL}
+    -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16}
     -DGGML_NATIVE=OFF
     -DGGML_OPENMP=${GGML_OPENMP}
 )
@@ -166,8 +126,6 @@ setup_framework_structure() {
     cp ggml/include/ggml-cpu.h     ${header_path}
     cp ggml/include/ggml-blas.h    ${header_path}
     cp ggml/include/gguf.h         ${header_path}
-    cp tools/mtmd/mtmd.h           ${header_path}
-    cp tools/mtmd/mtmd-helper.h    ${header_path}
 
     # Create module map (common for all platforms)
     cat > ${module_path}module.modulemap << EOF
@@ -289,8 +247,6 @@ combine_static_libraries() {
         "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a"
         "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
         "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
-        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
-        "${base_dir}/${build_dir}/vendor/hash/${release_dir}/libvendor-hash.a"
     )
 
     # Create temporary directory for processing
@@ -442,189 +398,142 @@ combine_static_libraries() {
     rm -rf "${temp_dir}"
 }
 
-build_ios_sim() {
-    echo "Building for iOS simulator..."
-    cmake -B build-ios-sim -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-        -DIOS=ON \
-        -DCMAKE_SYSTEM_NAME=iOS \
-        -DCMAKE_OSX_SYSROOT=iphonesimulator \
-        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-ios-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for iOS simulator..."
+cmake -B build-ios-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DIOS=ON \
+    -DCMAKE_SYSTEM_NAME=iOS \
+    -DCMAKE_OSX_SYSROOT=iphonesimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -S .
+cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
-build_ios_device() {
-    echo "Building for iOS devices..."
-    cmake -B build-ios-device -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-        -DCMAKE_SYSTEM_NAME=iOS \
-        -DCMAKE_OSX_SYSROOT=iphoneos \
-        -DCMAKE_OSX_ARCHITECTURES="arm64" \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-ios-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for iOS devices..."
+cmake -B build-ios-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=iOS \
+    -DCMAKE_OSX_SYSROOT=iphoneos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -S .
+cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
-build_macos() {
-    echo "Building for macOS..."
-    cmake -B build-macos -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
-        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -S .
-    cmake --build build-macos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for macOS..."
+cmake -B build-macos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -S .
+cmake --build build-macos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
-build_visionos() {
-    echo "Building for visionOS..."
-    cmake -B build-visionos -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-        -DCMAKE_OSX_ARCHITECTURES="arm64" \
-        -DCMAKE_SYSTEM_NAME=visionOS \
-        -DCMAKE_OSX_SYSROOT=xros \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DLLAMA_BUILD_SERVER=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-visionos --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for visionOS..."
+cmake -B build-visionos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xros \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_BUILD_SERVER=OFF \
+    -S .
+cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
-build_visionos_sim() {
-    echo "Building for visionOS simulator..."
-    cmake -B build-visionos-sim -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
-        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-        -DCMAKE_SYSTEM_NAME=visionOS \
-        -DCMAKE_OSX_SYSROOT=xrsimulator \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DLLAMA_BUILD_SERVER=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-visionos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for visionOS simulator..."
+cmake -B build-visionos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xrsimulator \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_BUILD_SERVER=OFF \
+    -S .
+cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
-build_tvos_sim() {
-    echo "Building for tvOS simulator..."
-    cmake -B build-tvos-sim -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-        -DCMAKE_SYSTEM_NAME=tvOS \
-        -DCMAKE_OSX_SYSROOT=appletvsimulator \
-        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-        -DGGML_METAL=ON \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-tvos-sim --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
+echo "Building for tvOS simulator..."
+cmake -B build-tvos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvsimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -S .
+cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
-build_tvos_device() {
-    echo "Building for tvOS devices..."
-    cmake -B build-tvos-device -G Xcode \
-        "${COMMON_CMAKE_ARGS[@]}" \
-        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
-        -DCMAKE_SYSTEM_NAME=tvOS \
-        -DCMAKE_OSX_SYSROOT=appletvos \
-        -DCMAKE_OSX_ARCHITECTURES="arm64" \
-        -DGGML_METAL=ON \
-        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
-        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-        -DLLAMA_OPENSSL=OFF \
-        -DMTMD_VIDEO=OFF \
-        -S .
-    cmake --build build-tvos-device --config Release -j "${JOBS_PER_BUILD}" -- -quiet
-}
-
-run_builds_parallel() {
-    local -a pids=()
-    local -a names=()
-    local name i
-    for name in "$@"; do
-        # Wait for the oldest running build to free a slot
-        if [[ "${#pids[@]}" -ge "$MAX_PARALLEL_BUILDS" ]]; then
-            if ! wait "${pids[0]}"; then
-                echo "ERROR: build '${names[0]}' failed, log follows (${names[0]}.log):" >&2
-                kill "${pids[@]}" 2>/dev/null || true
-                cat "${names[0]}.log" >&2
-                exit 1
-            fi
-            pids=("${pids[@]:1}")
-            names=("${names[@]:1}")
-        fi
-        echo "Starting build: $name (log: ${name}.log, -j ${JOBS_PER_BUILD})"
-        "$name" > "${name}.log" 2>&1 &
-        pids+=("$!")
-        names+=("$name")
-    done
-    # Wait for the remaining builds
-    for i in "${!pids[@]}"; do
-        if ! wait "${pids[$i]}"; then
-            echo "ERROR: build '${names[$i]}' failed, log follows (${names[$i]}.log):" >&2
-            kill "${pids[@]}" 2>/dev/null || true
-            cat "${names[$i]}.log" >&2
-            exit 1
-        fi
-    done
-}
-
-BUILD_FNS=()
-for b in "${BUILDS[@]}"; do
-    read -r fn _ < <(build_spec "$b")
-    BUILD_FNS+=("$fn")
-done
-echo "Building: ${BUILDS[*]} (max ${MAX_PARALLEL_BUILDS} at a time, -j ${JOBS_PER_BUILD} each)..."
-run_builds_parallel "${BUILD_FNS[@]}"
+echo "Building for tvOS devices..."
+cmake -B build-tvos-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DLLAMA_OPENSSL=OFF \
+    -S .
+cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."
-for b in "${BUILDS[@]}"; do
-    read -r _ bdir _ platform _ min_os < <(build_spec "$b")
-    setup_framework_structure "$bdir" "$min_os" "$platform"
-done
+setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos"
+setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos"
+setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos"
 
 # Create dynamic libraries from static libraries
 echo "Creating dynamic libraries from static libraries..."
-for b in "${BUILDS[@]}"; do
-    read -r _ bdir rdir platform is_sim _ < <(build_spec "$b")
-    combine_static_libraries "$bdir" "$rdir" "$platform" "$is_sim"
-done
+combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true"
+combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false"
+combine_static_libraries "build-macos" "Release" "macos" "false"
+combine_static_libraries "build-visionos" "Release-xros" "visionos" "false"
+combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true"
+combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true"
+combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
-XCFW_ARGS=()
-for b in "${BUILDS[@]}"; do
-    read -r _ bdir _ _ _ _ < <(build_spec "$b")
-    XCFW_ARGS+=(-framework "$(pwd)/${bdir}/framework/llama.framework")
-    XCFW_ARGS+=(-debug-symbols "$(pwd)/${bdir}/dSYMs/llama.dSYM")
-done
 xcrun xcodebuild -create-xcframework \
-    "${XCFW_ARGS[@]}" \
-    -output "$(pwd)/build-apple/llama.xcframework"
+    -framework $(pwd)/build-ios-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-ios-device/framework/llama.framework \
+    -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-macos/framework/llama.framework \
+    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-visionos/framework/llama.framework \
+    -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-visionos-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-visionos-sim/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-tvos-device/framework/llama.framework \
+    -debug-symbols $(pwd)/build-tvos-device/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-tvos-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-tvos-sim/dSYMs/llama.dSYM \
+    -output $(pwd)/build-apple/llama.xcframework

ci/run.sh

@@ -10,9 +10,6 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with ROCm support
-# GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS=gfx1151 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 # # with SYCL support
 # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
@@ -49,14 +46,6 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-# gpu-rocm self-hosted runner can't upload logs to blob; keep each run's logs in
-# their own dir keyed by the GitHub run id so an Actions run URL maps to its logs.
-if [ -n "${GG_BUILD_ROCM}" ] && [ -n "${GITHUB_RUN_ID}" ]; then
-    OUT="$OUT/run-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
-    mkdir -p "$OUT"
-    echo "ci results dir: $OUT"
-fi
-
 rm -f $OUT/*.log
 rm -f $OUT/*.exit
 rm -f $OUT/*.md
@@ -100,7 +89,7 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_HIP_COMPILER=$(hipconfig -l)/clang -DGGML_HIP=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
     if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
         echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
         exit 1
@@ -190,7 +179,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
 
     # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
+    CTEST_EXTRA="-E test-llama-archs"
 fi
 
 ## helpers
@@ -651,52 +640,39 @@ function gg_sum_rerank_tiny {
 
 function gg_check_build_requirements {
     if ! command -v git &> /dev/null; then
-        gg_printf 'git not found, please install\n'
-        exit 1
+        gg_printf 'git not found, please install'
     fi
 
     if ! command -v git-lfs &> /dev/null; then
-        gg_printf 'git-lfs not found, please install\n'
-        exit 1
-    fi
-
-    if ! git config --get filter.lfs.clean &> /dev/null; then
-        gg_printf 'git-lfs not initialized, please run `git lfs install`\n'
-        exit 1
+        gg_printf 'git-lfs not found, please install'
     fi
 
     if ! command -v wget &> /dev/null; then
-        gg_printf 'wget not found, please install\n'
-        exit 1
+        gg_printf 'wget not found, please install'
     fi
 
     if ! command -v python3 &> /dev/null; then
-        gg_printf 'python3 not found, please install\n'
-        exit 1
+        gg_printf 'python3 not found, please install'
     fi
 
     if ! command -v pip3 &> /dev/null; then
-        gg_printf 'pip3 not found, please install\n'
-        exit 1
+        gg_printf 'pip3 not found, please install'
     fi
 
     if ! python3 -m ensurepip --help &> /dev/null; then
-        gg_printf 'ensurepip not found, please install python3-venv package\n'
-        exit 1
+        gg_printf 'ensurepip not found, please install python3-venv package'
     fi
 
     if ! command -v cmake &> /dev/null; then
-        gg_printf 'cmake not found, please install\n'
-        exit 1
+        gg_printf 'cmake not found, please install'
     fi
 
     if ! command -v ccache &> /dev/null; then
-        gg_printf 'ccache not found, please consider installing for faster builds\n'
+        gg_printf 'ccache not found, please consider installing for faster builds'
     fi
 
     if ! command -v ctest &> /dev/null; then
-        gg_printf 'ctest not found, please install\n'
-        exit 1
+        gg_printf 'ctest not found, please install'
     fi
 }
 

cmake/arm64-windows-llvm.cmake

@@ -8,7 +8,6 @@ set( CMAKE_CXX_COMPILER  clang++ )
 
 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
-set( CMAKE_ASM_COMPILER_TARGET ${target} )
 
 set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )

cmake/arm64-windows-msvc-cuda.cmake

@@ -1,26 +0,0 @@
-# Used to cross-compile ggml-cuda for Windows ARM64 on an x64 Windows host.
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR arm64 )
-
-if ( DEFINED CUDAToolkit_ROOT )
-    file( TO_CMAKE_PATH "${CUDAToolkit_ROOT}" CUDA_ROOT )
-elseif ( DEFINED ENV{CUDA_PATH} )
-    file( TO_CMAKE_PATH "$ENV{CUDA_PATH}" CUDA_ROOT )
-else()
-    message( FATAL_ERROR "Set CUDAToolkit_ROOT or CUDA_PATH to a Windows CUDA Toolkit with ARM64 target libraries" )
-endif()
-
-if ( DEFINED ENV{VCToolsInstallDir} )
-    file( TO_CMAKE_PATH "$ENV{VCToolsInstallDir}" MSVC_TOOLS_ROOT )
-    set( CMAKE_CUDA_HOST_COMPILER "${MSVC_TOOLS_ROOT}/bin/Hostx64/arm64/cl.exe" CACHE FILEPATH "" )
-endif()
-
-set( CMAKE_CUDA_COMPILER "${CUDA_ROOT}/bin/nvcc.exe" CACHE FILEPATH "" )
-set( CMAKE_CUDA_FLAGS_INIT "-target-dir=arm64" )
-
-# FindCUDAToolkit selects lib/x64 from the host architecture on Windows.
-set( CUDA_CUDART              "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
-set( CUDA_cudart_LIBRARY      "${CUDA_ROOT}/lib/arm64/cudart.lib"   CACHE FILEPATH "" )
-set( CUDA_cublas_LIBRARY      "${CUDA_ROOT}/lib/arm64/cublas.lib"   CACHE FILEPATH "" )
-set( CUDA_cublasLt_LIBRARY    "${CUDA_ROOT}/lib/arm64/cublasLt.lib" CACHE FILEPATH "" )
-set( CUDA_cuda_driver_LIBRARY "${CUDA_ROOT}/lib/arm64/cuda.lib"     CACHE FILEPATH "" )

cmake/llama-config.cmake.in

@@ -1,4 +1,4 @@
-set(LLAMA_VERSION      @LLAMA_VERSION@)
+set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
 set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)

cmake/llama.pc.in

@@ -5,6 +5,6 @@ includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@
 
 Name: llama
 Description: Port of Facebook's LLaMA model in C/C++
-Version: @LLAMA_VERSION@
+Version: @LLAMA_INSTALL_VERSION@
 Libs: -L${libdir} -lggml -lggml-base -lllama
 Cflags: -I${includedir}

common/CMakeLists.txt

@@ -80,9 +80,9 @@ add_library(${TARGET}
     http.h
     imatrix-loader.cpp
     imatrix-loader.h
+    json-partial.cpp
+    json-partial.h
     json-schema-to-grammar.cpp
-    json.cpp
-    json.h
     llguidance.cpp
     log.cpp
     log.h
@@ -96,16 +96,14 @@ add_library(${TARGET}
     peg-parser.h
     preset.cpp
     preset.h
+    regex-partial.cpp
     reasoning-budget.cpp
     reasoning-budget.h
+    regex-partial.h
     sampling.cpp
     sampling.h
     speculative.cpp
     speculative.h
-    subproc.cpp
-    subproc.h
-    trie.cpp
-    trie.h
     unicode.cpp
     unicode.h
     jinja/lexer.cpp
@@ -123,19 +121,14 @@ add_library(${TARGET}
     )
 
 set_target_properties(${TARGET} PROPERTIES
-    VERSION ${LLAMA_VERSION_BASE}
-    SOVERSION ${LLAMA_VERSION_MAJOR}
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
     MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )
 
-target_include_directories(${TARGET} PUBLIC .)
-target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
+target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
-if (LLAMA_SUBPROCESS)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_SUBPROCESS)
-endif()
-
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 

common/arg.h

@@ -8,12 +8,10 @@
 #include <string>
 #include <vector>
 #include <cstring>
-#include <memory>
 
 // pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP    "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT       "__PRESET_STOP_TIMEOUT"
-#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
 
 //
 // CLI argument parsing
@@ -124,9 +122,6 @@ struct common_params_context {
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
-// load all backends and print the list of available (non-CPU) devices to stdout
-void common_print_available_devices();
-
 // parse input arguments from CLI into a map
 bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
 
@@ -135,20 +130,14 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
 // see: https://github.com/ggml-org/llama.cpp/issues/18163
 void common_params_add_preset_options(std::vector<common_arg> & args);
 
-struct common_models_handler {
-    common_download_hf_plan plan;
-    common_download_hf_plan plan_spec;
-    common_download_opts opts;
-};
-
-// initialize downloading opts and hf_plan if needed, but does not download anything yet
-common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex);
-
-// check if the model is a preset repo (i.e. has a preset file)
-bool common_models_handler_is_preset_repo(const common_models_handler & handler);
-
-// download and update params with the downloaded model path
-void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback = nullptr);
+// populate model paths (main model, mmproj, etc) from -hf if necessary
+// return true if the model is ready to use
+// throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
+// if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
+bool common_params_handle_models(
+    common_params & params,
+    llama_example curr_ex,
+    common_download_callback * callback = nullptr);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/build-info.cpp.in

@@ -29,7 +29,7 @@ const char * llama_build_info(void) {
     return s.c_str();
 }
 
-void llama_print_build_info(const char * llama_version) {
-    fprintf(stderr, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
-    fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
+void llama_print_build_info(void) {
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, llama_build_number(), llama_commit());
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, llama_compiler(), llama_build_target());
 }

common/build-info.h

@@ -8,4 +8,4 @@ const char * llama_compiler(void);
 const char * llama_build_target(void);
 const char * llama_build_info(void);
 
-void llama_print_build_info(const char *);
+void llama_print_build_info(void);

common/chat-auto-parser-generator.cpp

@@ -5,12 +5,13 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <stdexcept>
 #include <string>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 // Helper to iterate over tools/functions
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
@@ -46,8 +47,6 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
     data.generation_prompt = common_chat_template_generation_prompt(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.preserved_tokens  = autoparser.preserved_tokens;
-    data.additional_stops.insert(data.additional_stops.end(),
-        autoparser.additional_stops.begin(), autoparser.additional_stops.end());
 
     std::string parser_generation_prompt = data.generation_prompt;
 
@@ -148,8 +147,7 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs, cons
         } else {
             parser = content.build_parser(ctx);
         }
-        const std::string reasoning_start = trim_whitespace(reasoning.start);
-        return pure_content ? p.prefix(generation_prompt, reasoning_start) + parser : p.prefix(generation_prompt, reasoning_start) << parser;
+        return pure_content ? p.prefix(generation_prompt, reasoning.start) + parser : p.prefix(generation_prompt, reasoning.start) << parser;
     });
 }
 
@@ -263,10 +261,6 @@ common_peg_parser analyze_tools::build_func_parser(common_chat_peg_builder & p,
     bool              matched_atomic = false;
     common_peg_parser func_parser    = p.eps();
 
-    if (!function.args_separator.empty()) {
-        open = open + p.space() + p.literal(function.args_separator);
-    }
-
     if (!function.name_suffix.empty()) {
         func_parser    = open + call_id_section + p.space() + args;
         matched_atomic = true;
@@ -287,13 +281,7 @@ common_peg_parser analyze_tools::build_func_parser(common_chat_peg_builder & p,
         // we only emit tool_close when we can actually see the closing marker. This prevents
         // premature closing during partial parsing when we've seen e.g. "</" which could be
         // either "</tool_call>" (end) or "<arg_key>" prefix that failed to match.
-        // Laguna (v4): the model may emit whitespace between the last </arg_value> and
-        // </tool_call> even though the template renders them tight. Tolerate optional
-        // leading space in the close lookahead so the tool call still closes.
-        auto close_peek = arguments.tolerate_intertag_whitespace
-                              ? p.peek(p.space() + p.literal(format.per_call_end))
-                              : p.peek(p.literal(format.per_call_end));
-        func_parser = func_parser + p.tool_close(close_peek);
+        func_parser = func_parser + p.tool_close(p.peek(p.literal(format.per_call_end)));
     } else {
         func_parser = func_parser + p.tool_close(p.space());  // force this to process tool closing callbacks in mapper
     }
@@ -390,7 +378,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
 
         std::set<std::string> required;
         if (params.contains("required")) {
-            required = params.at("required").get<std::set<std::string>>();
+            params.at("required").get_to(required);
         }
 
         auto schema_info = common_schema_info();

common/chat-auto-parser-helpers.cpp

@@ -4,11 +4,14 @@
 #include "chat-peg-parser.h"
 #include "chat.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <cctype>
 #include <numeric>
 
+using json = nlohmann::ordered_json;
+
 std::string trim_whitespace(const std::string & str) {
     size_t start = 0;
     while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {

common/chat-auto-parser.h

@@ -4,7 +4,7 @@
 #include "common.h"
 #include "jinja/caps.h"
 #include "peg-parser.h"
-#include "json.h"
+#include "nlohmann/json.hpp"
 
 #include <chrono>
 #include <optional>
@@ -12,7 +12,7 @@
 #include <utility>
 #include <vector>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 class common_chat_peg_builder;
 
@@ -192,10 +192,9 @@ struct tool_format_analysis {
 };
 
 struct tool_function_analysis {
-    std::string name_prefix;     // e.g., "<function=", "\"name\": \"", "functions."
-    std::string name_suffix;     // e.g., ">", "\"", ":0"
-    std::string args_separator;  // e.g., "<tool_sep>" (marker between function name and arguments)
-    std::string close;           // e.g., "</function>", "" (for tag-based)
+    std::string name_prefix;  // e.g., "<function=", "\"name\": \"", "functions."
+    std::string name_suffix;  // e.g., ">", "\"", ":0"
+    std::string close;        // e.g., "</function>", "" (for tag-based)
 };
 
 struct tool_arguments_analysis {
@@ -206,7 +205,6 @@ struct tool_arguments_analysis {
     std::string value_prefix;  // e.g., "", "<arg_value>", ""
     std::string value_suffix;  // e.g., "</param>", "</arg_value>", ""
     std::string separator;     // e.g., "", "\n", ","
-    bool tolerate_intertag_whitespace = false; // Laguna: accept optional whitespace between arg tags
 };
 
 struct tool_id_analysis {
@@ -389,7 +387,6 @@ struct autoparser {
 
     // Preserved tokens for tokenizer (union of all non-empty markers)
     std::vector<std::string> preserved_tokens;
-    std::vector<std::string> additional_stops;  // literal stop strings (e.g. Laguna </assistant>) caught however tokenized
 
     autoparser() = default;
 

common/chat-diff-analyzer.cpp

@@ -4,11 +4,11 @@
 #include "chat.h"
 #include "common.h"
 #include "log.h"
+#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <algorithm>
 #include <cctype>
-#include <numeric>
 #include <ostream>
 #include <sstream>
 
@@ -17,7 +17,7 @@
 #define ANSI_ORANGE "\033[1m\x1b[38;5;214m"
 #define ANSI_RED    "\033[1m\x1b[38;5;196m"
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 namespace autoparser {
 
@@ -124,16 +124,16 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
               analysis.tools.format.section_end    = "";
               analysis.tools.format.per_call_start = "<TOOLCALL>";
               analysis.tools.format.per_call_end   = "</TOOLCALL>";
-              analysis.tools.format.tools_array_wrapped = true;
               analysis.content.mode                = content_mode::PLAIN;
               analysis.content.start               = "";
               analysis.content.end                 = "";
               analysis.reasoning.mode              = reasoning_mode::TAG_BASED;
-              analysis.reasoning.start             = "<think>\n";
+              analysis.reasoning.start             = "<think>\n\n";
               analysis.reasoning.end               = "</think>";
               analysis.assistant_start             = "<SPECIAL_11>Assistant";
               analysis.user_start                  = "<SPECIAL_11>User";
               analysis.preserved_tokens.clear();
+              analysis.preserved_tokens.push_back("<SPECIAL_12>");
               analysis.preserved_tokens.push_back("<SPECIAL_11>");
               analysis.preserved_tokens.push_back("</think>");
               analysis.preserved_tokens.push_back("<TOOLCALL>");
@@ -173,34 +173,6 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
               LOG_DBG(ANSI_ORANGE "[Patch: JSON name/parameters tool instruction]\n" ANSI_RESET);
           }
       },
-      // Laguna (poolside) - the v4 chat template renders reasoning and tool-arg
-      // delimiters with formatting whitespace ("<think>\n", "</arg_value>\n") that
-      // the model does not emit, so the inferred delimiters carry a spurious
-      // newline and never match the model output. Trim to the bare tag. (v8
-      // renders without the whitespace, so this is a no-op there.)
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("laguna_glm_thinking") != std::string::npos) {
-              analysis.reasoning.start              = trim_whitespace(analysis.reasoning.start);
-              analysis.reasoning.end                = trim_whitespace(analysis.reasoning.end);
-              analysis.tools.arguments.value_prefix = trim_whitespace(analysis.tools.arguments.value_prefix);
-              analysis.tools.arguments.value_suffix = trim_whitespace(analysis.tools.arguments.value_suffix);
-              analysis.tools.arguments.separator    = trim_whitespace(analysis.tools.arguments.separator);
-              analysis.tools.arguments.tolerate_intertag_whitespace = true;
-              // The CONTROL/eot </assistant> token only halts generation when emitted as the
-              // single token; after tool calls the model can spell it out as text tokens.
-              // A literal stop string catches it either way.
-              analysis.additional_stops.push_back("</assistant>");
-              LOG_DBG(ANSI_ORANGE "[Patch: Laguna]\n" ANSI_RESET);
-          }
-      },
-      // Bailing V3
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("Bailing V3 chat template") != std::string::npos) {
-              analysis.tools.arguments.value_suffix = trim_whitespace(analysis.tools.arguments.value_suffix);
-              analysis.tools.arguments.tolerate_intertag_whitespace = true;
-              LOG_DBG(ANSI_ORANGE "[Patch: Bailing V3]\n" ANSI_RESET);
-          }
-      },
 
     });
 
@@ -287,7 +259,6 @@ void autoparser::analyze_template(const common_chat_template & tmpl) {
     LOG_DBG("per_call_end: '%s'\n", tools.format.per_call_end.c_str());
     LOG_DBG("func_name_prefix: '%s'\n", tools.function.name_prefix.c_str());
     LOG_DBG("func_name_suffix: '%s'\n", tools.function.name_suffix.c_str());
-    LOG_DBG("func_args_separator: '%s'\n", tools.function.args_separator.c_str());
     LOG_DBG("func_close: '%s'\n", tools.function.close.c_str());
     LOG_DBG("call_id_prefix: '%s'\n", tools.call_id.prefix.c_str());
     LOG_DBG("call_id_suffix: '%s'\n", tools.call_id.suffix.c_str());
@@ -331,7 +302,6 @@ void autoparser::collect_preserved_tokens() {
     add_token(tools.format.per_call_end);
     add_token(tools.function.name_prefix);
     add_token(tools.function.name_suffix);
-    add_token(tools.function.args_separator);
     add_token(tools.function.close);
     add_token(tools.arguments.start);
     add_token(tools.arguments.end);
@@ -929,7 +899,7 @@ void analyze_tools::analyze_tool_call_format_json_native(const std::string & cle
     int  json_end       = clean_haystack.find_last_of('}');
     std::string cut     = clean_haystack.substr(json_start, json_end - json_start + 1);
     json call_struct    = json::parse(cut);
-    auto register_field = [&](const std::string & prefix, const common_json_entry & subel) {
+    auto register_field = [&](const std::string & prefix, const nlohmann::detail::iteration_proxy_value<json::iterator> & subel) {
         if (subel.value().is_string() && std::string(subel.value()).find("call0000") != std::string::npos) {
             format.id_field = !prefix.empty() ? prefix + "." + subel.key() : subel.key();
         } else if (subel.value().is_string() && std::string(subel.value()) == fun_name_needle) {
@@ -1081,23 +1051,6 @@ void analyze_tools::check_per_call_markers() {
         format.section_start.clear();
         format.section_end.clear();
     }
-
-    if (!format.per_call_end.empty()) {
-        auto count_occurrences = [](const std::string & haystack, const std::string & needle) {
-            size_t count = 0;
-            for (size_t pos = haystack.find(needle); pos != std::string::npos;
-                 pos = haystack.find(needle, pos + needle.size())) {
-                count++;
-            }
-            return count;
-        };
-        size_t calls_one = count_occurrences(one_vs_two->output_A, format.per_call_end);
-        size_t calls_two = count_occurrences(one_vs_two->output_B, format.per_call_end);
-        if (calls_one > 0 && calls_one == calls_two) {
-            format.section_end = format.per_call_end;
-            format.per_call_end.clear();
-        }
-    }
 }
 
 void analyze_tools::extract_function_markers() {
@@ -1179,17 +1132,6 @@ void analyze_tools::extract_function_markers() {
             auto suf_result = suffix_parser.parse_and_extract(diff.suffix);
             if (suf_result.result.success()) {
                 function.name_suffix += suf_result.tags["ext"];
-
-                auto arg_start = [&](common_peg_parser_builder &p) {
-                    return p.marker() + p.space() + p.choice({ p.literal(ARG_FIRST), p.literal(ARG_SECOND) });
-                };
-                auto sep_parser = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-                    return p.tag("sep", p.zero_or_more(p.negate(arg_start(p)) + p.any())) + arg_start(p);
-                });
-                auto sep_result = sep_parser.parse_and_extract(diff.suffix.substr(suf_result.tags["ext"].size()));
-                if (sep_result.result.success()) {
-                    function.args_separator = trim_whitespace(sep_result.tags["sep"]);
-                }
             }
         }
 

common/chat-peg-parser.cpp

@@ -4,10 +4,9 @@
 #include "ggml.h"
 #include "peg-parser.h"
 
-#include <cstdint>
-#include <functional>
+#include <nlohmann/json.hpp>
 
-using ordered_json = common_json;
+using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
@@ -236,43 +235,6 @@ common_peg_parser common_chat_peg_builder::tag_with_safe_content(const std::stri
     return zero_or_more(choice({ p, content_chunk }));
 }
 
-common_peg_parser common_chat_peg_builder::permute(const std::string &                    rule_prefix,
-                                                   const std::vector<common_peg_parser> & parsers) {
-    if (parsers.empty()) {
-        return eps();
-    }
-
-    if (parsers.size() == 1 || parsers.size() > COMMON_CHAT_MAX_PERMUTE) {
-        return sequence(parsers);
-    }
-
-    std::map<uint32_t, common_peg_parser>      rules;
-    std::function<common_peg_parser(uint32_t)> remaining_of;
-
-    remaining_of = [&](uint32_t remaining) -> common_peg_parser {
-        if (remaining == 0) {
-            return eps();
-        }
-
-        auto cached = rules.find(remaining);
-        if (cached != rules.end()) {
-            return cached->second;
-        }
-
-        auto alternatives = choice();
-        for (size_t i = 0; i < parsers.size(); i++) {
-            const uint32_t bit = 1u << i;
-            if (remaining & bit) {
-                alternatives |= parsers[i] + remaining_of(remaining & ~bit);
-            }
-        }
-
-        return rules.emplace(remaining, rule(rule_prefix + "-" + std::to_string(remaining), alternatives)).first->second;
-    };
-
-    return remaining_of((1u << parsers.size()) - 1);
-}
-
 std::string & common_chat_peg_mapper::args_target() {
     return (current_tool && !current_tool->name.empty()) ? current_tool->arguments : args_buffer;
 }
@@ -592,7 +554,9 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
 
                 // Full argument: name="value" or name=value
                 auto arg_rule = tool_arg(
-                    tool_arg_open(tool_arg_name(arg_name_parser) + literal("=")) +
+                    tool_arg_open(eps()) +
+                    tool_arg_name(arg_name_parser) +
+                    literal("=") +
                     arg_value_parser +
                     tool_arg_close(eps())
                 );
@@ -1092,141 +1056,3 @@ void common_chat_peg_gemma4_mapper::visit(const common_peg_ast_arena & arena, co
         visit(arena, child_id);
     }
 }
-
-static void minimax_m3_collect(const common_peg_ast_arena &     arena,
-                               const common_peg_ast_node &      node,
-                               const std::string &              tag,
-                               std::vector<common_peg_ast_id> & out) {
-    for (auto child_id : node.children) {
-        const auto & child = arena.get(child_id);
-        if (child.tag == tag) {
-            out.push_back(child_id);
-        } else {
-            minimax_m3_collect(arena, child, tag, out);
-        }
-    }
-}
-
-static common_peg_ast_id minimax_m3_value_of(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
-    for (auto child_id : node.children) {
-        const auto & tag = arena.get(child_id).tag;
-        if (tag == common_chat_peg_builder::TOOL_ARG_VALUE ||
-            tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE ||
-            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT ||
-            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
-            return child_id;
-        }
-    }
-    return COMMON_PEG_INVALID_AST_ID;
-}
-
-static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed);
-
-static std::string minimax_m3_member_to_json(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
-    auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_ARG_NAME);
-    if (name_id == COMMON_PEG_INVALID_AST_ID) {
-        return "";
-    }
-
-    return ordered_json(arena.get(name_id).text).dump() + ":" +
-           minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, node), !node.is_partial);
-}
-
-static std::string minimax_m3_container_to_json(const common_peg_ast_arena & arena,
-                                                const common_peg_ast_node & node,
-                                                bool                        is_object,
-                                                bool                        closed) {
-    const std::string tag = is_object ? common_chat_peg_builder::TOOL_ARG
-                                      : common_chat_peg_minimax_m3_mapper::TOOL_ARG_ITEM;
-
-    std::vector<common_peg_ast_id> entries;
-    minimax_m3_collect(arena, node, tag, entries);
-
-    std::string result = is_object ? "{" : "[";
-
-    bool add_comma = false;
-    for (auto entry_id : entries) {
-        const auto & entry = arena.get(entry_id);
-
-        std::string text;
-        if (is_object) {
-            text = minimax_m3_member_to_json(arena, entry);
-        } else {
-            text = minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, entry), !entry.is_partial);
-        }
-
-        if (text.empty()) {
-            continue;
-        }
-
-        if (add_comma) {
-            result += ",";
-        }
-        add_comma = true;
-        result += text;
-    }
-
-    if (closed) {
-        result += is_object ? "}" : "]";
-    }
-    return result;
-}
-
-static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed) {
-    if (id == COMMON_PEG_INVALID_AST_ID) {
-        return "";
-    }
-
-    const auto & node = arena.get(id);
-
-    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT) {
-        return minimax_m3_container_to_json(arena, node, /* is_object = */ true, closed);
-    }
-
-    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
-        return minimax_m3_container_to_json(arena, node, /* is_object = */ false, closed);
-    }
-
-    if (node.tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE) {
-        return "\"" + escape_json_string_inner(std::string(node.text)) + (closed ? "\"" : "");
-    }
-
-    // Numbers and booleans are written verbatim by the template
-    return std::string(node.text);
-}
-
-void common_chat_peg_minimax_m3_mapper::from_ast(const common_peg_ast_arena &    arena,
-                                                 const common_peg_parse_result & result) {
-    for (const auto & node : result.nodes) {
-        visit(arena, node);
-    }
-}
-
-void common_chat_peg_minimax_m3_mapper::visit(const common_peg_ast_arena & arena, common_peg_ast_id id) {
-    const auto & node = arena.get(id);
-
-    if (node.tag == common_chat_peg_builder::REASONING) {
-        result.reasoning_content += std::string(node.text);
-        return;
-    }
-
-    if (node.tag == common_chat_peg_builder::CONTENT) {
-        result.content += std::string(node.text);
-        return;
-    }
-
-    if (node.tag == common_chat_peg_builder::TOOL) {
-        auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_NAME);
-        if (name_id != COMMON_PEG_INVALID_AST_ID) {
-            common_chat_tool_call call;
-            call.name      = std::string(arena.get(name_id).text);
-            call.arguments = minimax_m3_container_to_json(arena, node, /* is_object = */ true, !node.is_partial);
-            result.tool_calls.push_back(call);
-        }
-        return;
-    }
-
-    for (auto child_id : node.children) {
-        visit(arena, child_id);
-    }
-}

common/chat-peg-parser.h

@@ -40,23 +40,9 @@ class common_chat_peg_gemma4_mapper : public common_chat_peg_mapper {
     void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
 };
 
-class common_chat_peg_minimax_m3_mapper : public common_chat_peg_mapper {
-  public:
-    static constexpr const char * TOOL_ARG_OBJECT = "tool-arg-object";
-    static constexpr const char * TOOL_ARG_ARRAY  = "tool-arg-array";
-    static constexpr const char * TOOL_ARG_ITEM   = "tool-arg-item";
-
-    common_chat_peg_minimax_m3_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
-    virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
-  private:
-    void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
-};
-
 struct content_structure;
 struct tool_call_structure;
 
-constexpr size_t COMMON_CHAT_MAX_PERMUTE = 6;
-
 class common_chat_peg_builder : public common_peg_parser_builder {
   public:
     // Tag constants (from former common_chat_peg_base_builder)
@@ -107,9 +93,6 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
 
 
-    // Matches every parser exactly once, in any order.
-    common_peg_parser permute(const std::string & rule_prefix, const std::vector<common_peg_parser> & parsers);
-
     // Return a parser that parses the prefix of a string, up to a given delimiter.
     common_peg_parser prefix(const std::string & s, const std::string & delimiter = {});
 
@@ -128,7 +111,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const common_json &   tools,
+                                          const nlohmann::ordered_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -143,13 +126,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const common_json &             tools,
+                                                 const nlohmann::ordered_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const common_json & tools,
+    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
                                               bool                           parallel_tool_calls,
                                               bool                           allow_json_literals);
 
@@ -158,19 +141,19 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser python_or_json_value();
 
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const common_json & tools,
+    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
                                                        const std::string &            args_key,
                                                        const std::string &            effective_args_key,
                                                        const std::string &            call_id_key,
                                                        const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const common_json & tools,
+    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
                                                    const std::string &            effective_name_key,
                                                    const std::string &            effective_args_key,
                                                    const std::string &            call_id_key,
                                                    const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const common_json &   tools,
+    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,

common/chat.h

@@ -8,7 +8,7 @@
 #include "jinja/runtime.h"
 #include "jinja/caps.h"
 
-#include "json.h"
+#include "nlohmann/json_fwd.hpp"
 
 #include <chrono>
 #include <functional>
@@ -17,6 +17,7 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
+using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 
@@ -86,7 +87,7 @@ struct common_chat_msg {
     std::string                               tool_name;
     std::string                               tool_call_id;
 
-    common_json to_json_oaicompat(bool concat_typed_text = false) const;
+    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     std::string render_content(const std::string & delimiter = "\n\n") const;
 
@@ -142,75 +143,15 @@ struct common_chat_msg_diff {
     }
 };
 
-enum common_chat_role {
-    COMMON_CHAT_ROLE_UNKNOWN,
-    COMMON_CHAT_ROLE_SYSTEM,
-    COMMON_CHAT_ROLE_ASSISTANT,
-    COMMON_CHAT_ROLE_USER,
-    COMMON_CHAT_ROLE_TOOL
-};
-
-common_chat_role common_chat_role_from_string(const std::string & role);
-const char *     common_chat_role_to_string(common_chat_role role);
-
 struct common_chat_msg_span {
-    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
+    std::string role;
     std::size_t pos = 0;
     std::size_t len = 0;
-
-    bool valid() const {
-        return role != COMMON_CHAT_ROLE_UNKNOWN;
-    }
-};
-
-struct common_chat_msg_spans {
-    std::vector<common_chat_msg_span> spans;
-
-    void add(common_chat_role role, size_t pos, size_t len) {
-        spans.push_back({ role, pos, len });
-    }
-
-    bool is_user_start(int32_t pos) const {
-        for (auto it = spans.begin(); it != spans.end(); ++it) {
-            if (it->role == COMMON_CHAT_ROLE_USER && pos == (int32_t) it->pos) {
-                return true;
-            }
-        }
-        return false;
-    }
-
-    int32_t last_user_message_pos() const {
-        for (auto it = spans.rbegin(); it != spans.rend(); ++it) {
-            if (it->role == COMMON_CHAT_ROLE_USER) {
-                return (int32_t) it->pos;
-            }
-        }
-        return -1;
-    }
 };
 
 struct common_chat_msg_delimiter {
-    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
-    std::string      delimiter;
-    llama_tokens     tokens = {};
-};
-
-struct common_chat_msg_delimiters {
-    std::vector<common_chat_msg_delimiter> delimiters;
-
-    common_chat_msg_delimiters() = default;
-    common_chat_msg_delimiters(std::initializer_list<common_chat_msg_delimiter> delims) : delimiters(delims) {}
-
-    void add(common_chat_role role, const std::string & delimiter) {
-        delimiters.push_back({ role, delimiter });
-    }
-
-    void tokenize(const llama_vocab * vocab);
-
-    // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
-    common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
-
-    common_json to_json() const;
+    std::string role;
+    std::string delimiter;
 };
 
 struct common_chat_tool {
@@ -232,7 +173,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_PEG_SIMPLE,
     COMMON_CHAT_FORMAT_PEG_NATIVE,
     COMMON_CHAT_FORMAT_PEG_GEMMA4,
-    COMMON_CHAT_FORMAT_PEG_MINIMAX_M3,
 
     COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };
@@ -274,12 +214,12 @@ struct common_chat_params {
     std::string                         generation_prompt;
     bool                                supports_thinking    = false;
     std::string                         thinking_start_tag;  // e.g., "<think>"
-    std::vector<std::string>            thinking_end_tags;   // e.g., "</think>"
+    std::string                         thinking_end_tag;    // e.g., "</think>"
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
     std::string                         parser;
-    common_chat_msg_delimiters          message_delimiters;
+    std::vector<common_chat_msg_span>   message_spans;
 };
 
 // per-message parsing syntax
@@ -349,16 +289,16 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const common_json & messages);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const common_json & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
 
-common_chat_continuation common_chat_continuation_parse(const common_json & value);
+common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
 
 // DEPRECATED: only used in tests
-common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
@@ -385,4 +325,5 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-common_chat_msg_delimiters common_chat_msg_delimiters_parse(const common_json & delimiters);
+std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims);
+

common/common.cpp

@@ -55,10 +55,6 @@
 #include <pwd.h>
 #endif
 
-#if defined(_AIX)
-#include <sys/systemcfg.h>
-#endif
-
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -76,16 +72,7 @@ common_time_meas::~common_time_meas() {
 //
 
 int32_t common_cpu_get_num_physical_cores() {
-#if defined(_AIX)
-    int32_t logical_cpus = _system_configuration.ncpus;
-    int32_t smt_threads = _system_configuration.smt_threads;
-    if (smt_threads > 0) {
-        return static_cast<int32_t>(logical_cpus / smt_threads);
-    }
-    if (logical_cpus > 0) {
-        return static_cast<int32_t>(logical_cpus);
-    }
-#elif defined(__linux__)
+#ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
     std::unordered_set<std::string> siblings;
     for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
@@ -215,14 +202,6 @@ int32_t common_cpu_get_num_math() {
             }
         }
     }
-#elif defined(__powerpc64__) || defined(__powerpc__)
-    int32_t smt_factor = 1;
-    int phy_cpus = common_cpu_get_num_physical_cores();
-    int logical_cpus = sysconf(_SC_NPROCESSORS_ONLN);
-    if (phy_cpus > 0 && logical_cpus > phy_cpus) {
-        smt_factor = logical_cpus / phy_cpus;
-    }
-    return phy_cpus * std::min(smt_factor, 2);
 #endif
     return common_cpu_get_num_physical_cores();
 }
@@ -246,7 +225,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
     }
 
     if (!SetPriorityClass(GetCurrentProcess(), p)) {
-        COM_WRN("failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
+        LOG_WRN("failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
         return false;
     }
 
@@ -272,7 +251,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
     }
 
     if (setpriority(PRIO_PROCESS, 0, p) != 0) {
-        COM_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
+        LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
     return true;
@@ -305,14 +284,14 @@ void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_para
 
     if (n_set && n_set < cpuparams.n_threads) {
         // Not enough set bits, may experience performance issues.
-        COM_WRN("Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
+        LOG_WRN("Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
     }
 }
 
 bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) {
     size_t dash_loc = range.find('-');
     if (dash_loc == std::string::npos) {
-        COM_ERR("%s", "Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
+        LOG_ERR("Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
         return false;
     }
 
@@ -324,7 +303,7 @@ bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THRE
     } else {
         start_i = std::stoull(range.substr(0, dash_loc));
         if (start_i >= GGML_MAX_N_THREADS) {
-            COM_ERR("%s", "Start index out of bounds!\n");
+            LOG_ERR("Start index out of bounds!\n");
             return false;
         }
     }
@@ -334,7 +313,7 @@ bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THRE
     } else {
         end_i = std::stoull(range.substr(dash_loc + 1));
         if (end_i >= GGML_MAX_N_THREADS) {
-            COM_ERR("%s", "End index out of bounds!\n");
+            LOG_ERR("End index out of bounds!\n");
             return false;
         }
     }
@@ -354,7 +333,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
     }
 
     size_t num_digits = mask.length() - start_i;
-    num_digits = std::min<size_t>(num_digits, 128);
+    if (num_digits > 128) num_digits = 128;
 
     size_t end_i = num_digits + start_i;
 
@@ -369,7 +348,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
         } else if (c >= 'A' && c <= 'F') {
             id -= 'A' - 10;
         } else {
-            COM_ERR("Invalid hex character '%c' at position %d\n", c, int32_t(i));
+            LOG_ERR("Invalid hex character '%c' at position %d\n", c, int32_t(i));
             return false;
         }
 
@@ -400,21 +379,21 @@ void common_params_print_info(const common_params & params, bool print_devices)
 #else
     const char * build_type = " (debug)";
 #endif
-    COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
+    LOG_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 
-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
+    LOG_INF("log_info: verbosity = %d (adjust with the `-lv N` CLI arg)\n", common_log_get_verbosity_thold());
 
     // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
     if (print_devices) {
-        COM_TRC("%s", "device_info:\n");
+        LOG_INF("device_info:\n");
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             auto * dev = ggml_backend_dev_get(i);
             size_t free, total;
             ggml_backend_dev_memory(dev, &free, &total);
-            COM_TRC("  - %-8s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
+            LOG_INF("  - %-8s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
         }
     }
-    COM_TRC("%s\n", common_params_get_system_info(params).c_str());
+    LOG_INF("%s\n", common_params_get_system_info(params).c_str());
 }
 
 std::string common_params_get_system_info(const common_params & params) {
@@ -681,7 +660,7 @@ void string_process_escapes(std::string & input) {
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
     const char * sep = strchr(data, '=');
     if (sep == nullptr || sep - data >= 128) {
-        COM_ERR("%s: malformed KV override '%s'\n", __func__, data);
+        LOG_ERR("%s: malformed KV override '%s'\n", __func__, data);
         return false;
     }
     llama_model_kv_override kvo;
@@ -704,20 +683,20 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
         } else if (std::strcmp(sep, "false") == 0) {
             kvo.val_bool = false;
         } else {
-            COM_ERR("%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            LOG_ERR("%s: invalid boolean value for KV override '%s'\n", __func__, data);
             return false;
         }
     } else if (strncmp(sep, "str:", 4) == 0) {
         sep += 4;
         kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
         if (strlen(sep) > 127) {
-            COM_ERR("%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data);
+            LOG_ERR("%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data);
             return false;
         }
         strncpy(kvo.val_str, sep, 127);
         kvo.val_str[127] = '\0';
     } else {
-        COM_ERR("%s: invalid type for KV override '%s'\n", __func__, data);
+        LOG_ERR("%s: invalid type for KV override '%s'\n", __func__, data);
         return false;
     }
     overrides.emplace_back(std::move(kvo));
@@ -998,42 +977,24 @@ bool fs_is_directory(const std::string & path) {
     return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
 }
 
-std::string common_get_env(const std::string & name) {
-    const char * value = std::getenv(name.c_str());
-    return value == nullptr ? "" : value;
-}
-
-void common_set_env(const std::string & name, const std::string & value) {
-#if defined(_WIN32)
-    _putenv_s(name.c_str(), value.c_str());
-#else
-    if (value.empty()) {
-        unsetenv(name.c_str());
-    } else {
-        setenv(name.c_str(), value.c_str(), 1);
-    }
-#endif
-}
-
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
         // Make sure to add trailing slash
-        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
+        if (p.back() != DIRECTORY_SEPARATOR) {
             p += DIRECTORY_SEPARATOR;
         }
         return p;
     };
-    cache_directory = common_get_env("LLAMA_CACHE");
-    if (cache_directory.empty()) {
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
         defined(__OpenBSD__) || defined(__NetBSD__)
-        const std::string xdg_cache_home = common_get_env("XDG_CACHE_HOME");
-        const std::string home           = common_get_env("HOME");
-        if (!xdg_cache_home.empty()) {
-            cache_directory = xdg_cache_home;
-        } else if (!home.empty()) {
-            cache_directory = home + "/.cache/";
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else if (std::getenv("HOME")) {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
         } else {
 #if defined(__linux__)
             /* no $HOME is defined, fallback to getpwuid */
@@ -1048,16 +1009,9 @@ std::string fs_get_cache_directory() {
 #endif /* defined(__linux__) */
         }
 #elif defined(__APPLE__)
-        cache_directory = common_get_env("HOME");
-        if (cache_directory.empty()) {
-            throw std::runtime_error("Failed to find $HOME directory");
-        }
-        cache_directory += "/Library/Caches/";
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
-        cache_directory = common_get_env("LOCALAPPDATA");
-        if (cache_directory.empty()) {
-            throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
-        }
+        cache_directory = std::getenv("LOCALAPPDATA");
 #elif defined(__EMSCRIPTEN__)
         GGML_ABORT("not implemented on this platform");
 #else
@@ -1069,51 +1023,6 @@ std::string fs_get_cache_directory() {
     return ensure_trailing_slash(cache_directory);
 }
 
-std::string fs_get_config_directory() {
-    std::string config_directory = "";
-    auto ensure_trailing_slash = [](std::string p) {
-        if (p.empty() || p.back() != DIRECTORY_SEPARATOR) {
-            p += DIRECTORY_SEPARATOR;
-        }
-        return p;
-    };
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
-        defined(__OpenBSD__) || defined(__NetBSD__) || defined(__APPLE__)
-    const std::string xdg_config_home = common_get_env("XDG_CONFIG_HOME");
-    const std::string home            = common_get_env("HOME");
-    if (!xdg_config_home.empty()) {
-        config_directory = xdg_config_home;
-    } else if (!home.empty()) {
-        config_directory = home + "/.config/";
-    } else {
-#if defined(__linux__)
-        /* no $HOME is defined, fallback to getpwuid */
-        struct passwd *pw = getpwuid(getuid());
-        if ((!pw) || (!pw->pw_dir)) {
-            throw std::runtime_error("Failed to find $HOME directory");
-        }
-
-        config_directory = std::string(pw->pw_dir) + std::string("/.config/");
-#else
-        throw std::runtime_error("Failed to find $HOME directory");
-#endif
-    }
-#elif defined(_WIN32)
-    config_directory = common_get_env("APPDATA");
-    if (config_directory.empty()) {
-        throw std::runtime_error("Failed to find %APPDATA% directory");
-    }
-#elif defined(__EMSCRIPTEN__)
-    // caller decides what to do when there is no config directory
-    throw std::runtime_error("not implemented on this platform");
-#else
-#  error Unknown architecture
-#endif
-    config_directory = ensure_trailing_slash(config_directory);
-    config_directory += "llama.cpp";
-    return ensure_trailing_slash(config_directory);
-}
-
 std::string fs_get_cache_file(const std::string & filename) {
     GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
     std::string cache_directory = fs_get_cache_directory();
@@ -1275,8 +1184,6 @@ struct common_init_result::impl {
 
     // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
 
-    common_threadpools threadpools;
-
     llama_model_ptr   model;
     llama_context_ptr context;
 
@@ -1292,36 +1199,13 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     auto cparams = common_context_params_to_llama(params);
 
     if (params.fit_params) {
-        COM_TRC("%s", "fitting params to device memory ...\n");
-        COM_TRC("%s", "(for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n");
-
-        // the draft context is created from the same base params and follows the main context, fit both together
-        const bool has_draft = params.speculative.has_dft();
-        const bool spec_mtp  = std::find(params.speculative.types.begin(), params.speculative.types.end(),
-            COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-
-        common_params params_dft = common_base_params_to_speculative(params);
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-        auto cparams_dft = common_context_params_to_llama(params_dft);
-        if (spec_mtp) {
-            cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-        }
-        cparams_dft.n_rs_seq = 0;
-
-        const common_fit_extra_model extra = {
-            /*.path_model   =*/ params_dft.model.path.c_str(),
-            /*.mparams      =*/ &mparams_dft,
-            /*.cparams      =*/ &cparams_dft,
-            /*.shares_model =*/ !has_draft, // an MTP context runs on the weights of the main model
-        };
-
+        LOG_INF("%s: fitting params to device memory ...\n", __func__);
+        LOG_INF("%s: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n", __func__);
         common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
-            has_draft || spec_mtp ? &extra : nullptr,
             params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1343,7 +1227,8 @@ common_init_result::common_init_result(common_params & params, bool model_only)
         llama_adapter_lora_ptr lora;
         lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
         if (lora == nullptr) {
-            COM_ERR("failed to load lora adapter '%s'\n", la.path.c_str());
+            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
+            pimpl->model.reset(model);
             return;
         }
 
@@ -1361,14 +1246,14 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     common_init_sampler_from_model(model, params.sampling);
 
     if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        COM_WRN("%s", "vocab does not have an EOS token, ignoring --ignore-eos\n");
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
     }
 
     // initialize once
     for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
         if (llama_vocab_is_eog(vocab, i)) {
-            COM_TRC("added %s logit bias = %f\n", common_token_to_piece(vocab, i).c_str(), -INFINITY);
+            LOG_TRC("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
             params.sampling.logit_bias_eog.push_back({i, -INFINITY});
         }
     }
@@ -1380,6 +1265,16 @@ common_init_result::common_init_result(common_params & params, bool model_only)
                 params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
     }
 
+    //if (params.sampling.penalty_last_n == -1) {
+    //    LOG_TRC("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    //if (params.sampling.dry_penalty_last_n == -1) {
+    //    LOG_TRC("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    //}
+
     // init the backend samplers as part of the context creation
     pimpl->samplers.resize(cparams.n_seq_max);
     pimpl->samplers_seq_config.resize(cparams.n_seq_max);
@@ -1396,15 +1291,11 @@ common_init_result::common_init_result(common_params & params, bool model_only)
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        COM_ERR("failed to create context with model '%s'\n", params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         return;
     }
 
     pimpl->context.reset(lctx);
-
-    set_process_priority(params.cpuparams.priority);
-
-    pimpl->threadpools.init(lctx, params);
 }
 
 llama_model * common_init_result::model() {
@@ -1437,7 +1328,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 
     llama_model * model = res->model();
     if (model == NULL) {
-        COM_ERR("failed to load model '%s'\n", params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return res;
     }
 
@@ -1447,14 +1338,14 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 
     llama_context * lctx = res->context();
     if (lctx == NULL) {
-        COM_ERR("failed to create context with model '%s'\n", params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         return res;
     }
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
-        COM_WRN("%s", "KV cache shifting is not supported for this context, disabling KV cache shifting\n");
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
@@ -1483,7 +1374,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
         bool ok = true;
 
         if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            COM_WRN("%s", "vocab does not have a  BOS token, reranking will not work\n");
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
@@ -1492,10 +1383,10 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
         bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;
 
         if (!has_eos && !has_sep && !has_rerank_prompt) {
-            COM_WRN("%s", "vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n");
+            LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
             ok = false;
         } else if (!has_eos) {
-            COM_WRN("%s", "vocab does not have an EOS token, using SEP token as fallback\n");
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
         }
 
         if (!ok) {
@@ -1508,7 +1399,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
     }
 
     if (params.warmup) {
-        COM_TRC("%s", "warming up the model with an empty run - please wait ... (--no-warmup to disable)\n");
+        LOG_INF("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
         std::vector<llama_token> tmp;
         llama_token bos = llama_vocab_bos(vocab);
@@ -1551,32 +1442,18 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 common_init_result::~common_init_result() = default;
 
 std::string common_get_model_endpoint() {
-    std::string endpoint = common_get_env("MODEL_ENDPOINT");
-    if (endpoint.empty()) {
-        // the HF_ENDPOINT variable is respected for backward compatibility
-        endpoint = common_get_env("HF_ENDPOINT");
+    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
+    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
+    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
+    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
+    std::string model_endpoint = "https://huggingface.co/";
+    if (endpoint_env) {
+        model_endpoint = endpoint_env;
+        if (model_endpoint.back() != '/') {
+            model_endpoint += '/';
+        }
     }
-    if (endpoint.empty()) {
-        return "https://huggingface.co/";
-    }
-    if (endpoint.back() != '/') {
-        endpoint += '/';
-    }
-    return endpoint;
-}
-
-char * common_get_model_or_exit(int argc, char * argv[]) {
-    if (argc > 1) {
-        return argv[1];
-    }
-
-    char * path = getenv("LLAMACPP_TEST_MODELFILE");
-    if (!path || strlen(path) == 0) {
-        fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n\033[0m");
-        exit(EXIT_SUCCESS);
-    }
-
-    return path;
+    return model_endpoint;
 }
 
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
@@ -1596,20 +1473,20 @@ common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
 
     int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
     if (ret != 0) {
-        COM_ERR("llama_decode() failed: %d\n", ret);
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
         res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
         goto done;
     }
 
     if (llama_n_rs_seq(ctx) > 0) {
-        COM_TRC("%s", "the context supports bounded partial sequence removal\n");
+        LOG_INF("%s: the context supports bounded partial sequence removal\n", __func__);
         res = COMMON_CONTEXT_SEQ_RM_TYPE_RS;
         goto done;
     }
 
     // try to remove the last tokens
     if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
-        COM_TRC("%s", "the context does not support partial sequence removal\n");
+        LOG_TRC("%s: the context does not support partial sequence removal\n", __func__);
         res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
         goto done;
     }
@@ -1621,49 +1498,23 @@ done:
     return res;
 }
 
-static void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     if (!llama_memory_seq_rm(mem, seq_id, p0, p1)) {
         GGML_ABORT("%s", string_format("failed to remove sequence %d with p0=%d, p1=%d\n", seq_id, p0, p1).c_str());
     }
 }
 
-static void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_cp(mem, seq_id_src, seq_id_dst, p0, p1);
 }
 
-static void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_add(mem, seq_id, p0, p1, delta);
 }
 
-void common_memory::init(llama_context * ctx_tgt, llama_context * ctx_dft) {
-    this->ctx_tgt = ctx_tgt;
-    this->ctx_dft = ctx_dft;
-}
-
-void common_memory::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
-    common_context_seq_rm(ctx_tgt, seq_id, p0, p1);
-    if (ctx_dft) {
-        common_context_seq_rm(ctx_dft, seq_id, p0, p1);
-    }
-}
-
-void common_memory::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const {
-    common_context_seq_cp(ctx_tgt, seq_id_src, seq_id_dst, p0, p1);
-    if (ctx_dft) {
-        common_context_seq_cp(ctx_dft, seq_id_src, seq_id_dst, p0, p1);
-    }
-}
-
-void common_memory::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const {
-    common_context_seq_add(ctx_tgt, seq_id, p0, p1, delta);
-    if (ctx_dft) {
-        common_context_seq_add(ctx_dft, seq_id, p0, p1, delta);
-    }
-}
-
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;
@@ -1686,8 +1537,10 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.n_gpu_layers    = params.n_gpu_layers;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
-    mparams.load_mode       = params.load_mode;
     mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_direct_io   = params.use_direct_io;
+    mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;
     mparams.no_host         = params.no_host;
@@ -1709,7 +1562,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.progress_callback           = params.load_progress_callback;
     mparams.progress_callback_user_data = params.load_progress_callback_user_data;
     mparams.no_alloc                    = params.no_alloc;
-    mparams.load_mtp                    = std::find(params.speculative.types.begin(), params.speculative.types.end(), COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
 
     return mparams;
 }
@@ -1721,7 +1573,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_seq_max         = params.n_parallel;
     cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
     cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
-    cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
@@ -1753,10 +1604,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     return cparams;
 }
 
-//
-// Threadpool utils
-//
-
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params) {
     struct ggml_threadpool_params tpp;
 
@@ -1773,58 +1620,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const commo
     return tpp;
 }
 
-common_threadpools::~common_threadpools() {
-    if (!free_fn) {
-        return;
-    }
-    free_fn(threadpool);
-    free_fn(threadpool_batch);
-}
-
-void common_threadpools::init(llama_context * ctx, const common_params & params) {
-    GGML_ASSERT(!threadpool);
-    GGML_ASSERT(!threadpool_batch);
-
-    COM_INF("llama threadpool init, n_threads = %d\n", (int) params.cpuparams.n_threads);
-
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (!cpu_dev) {
-        COM_WRN("%s", "no CPU backend found\n");
-        return;
-    }
-    auto * reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
-    free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
-
-    struct ggml_threadpool_params tpp_batch =
-            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
-    struct ggml_threadpool_params tpp =
-            ggml_threadpool_params_from_cpu_params(params.cpuparams);
-
-    // each pool needs to match the respective n_threads exactly
-    // see: https://github.com/ggml-org/llama.cpp/pull/27138#issuecomment-5332307332
-    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
-        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
-        if (!threadpool_batch) {
-            COM_WRN("batch threadpool create failed : n_threads %d\n", tpp_batch.n_threads);
-            return;
-        }
-
-        // start the non-batch threadpool in the paused state
-        tpp.paused = true;
-    }
-
-    threadpool = ggml_threadpool_new_fn(&tpp);
-    if (!threadpool) {
-        COM_WRN("threadpool create failed : n_threads %d\n", tpp.n_threads);
-        free_fn(threadpool_batch);
-        threadpool_batch = nullptr;
-        return;
-    }
-
-    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
-}
-
 //
 // Batch utils
 //
@@ -2008,13 +1803,13 @@ static common_control_vector_data common_control_vector_load_one(const common_co
     };
     struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
     if (!ctx_gguf) {
-        COM_ERR("failed to load control vector file from %s\n", load_info.fname.c_str());
+        LOG_ERR("%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str());
         return result;
     }
 
     int32_t n_tensors = gguf_get_n_tensors(ctx_gguf);
     if (n_tensors == 0) {
-        COM_WRN("no direction tensors found in %s\n", load_info.fname.c_str());
+        LOG_WRN("%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str());
     }
 
     for (int i = 0; i < n_tensors; i++) {
@@ -2032,23 +1827,23 @@ static common_control_vector_data common_control_vector_load_one(const common_co
             }
         }
         if (layer_idx < 0) {
-            COM_ERR("invalid/unparsable direction tensor layer index in %s\n", load_info.fname.c_str());
+            LOG_ERR("%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
             result.n_embd = -1;
             break;
         } else if (layer_idx == 0) {
-            COM_ERR("invalid (zero) direction tensor layer index in %s\n", load_info.fname.c_str());
+            LOG_ERR("%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
             result.n_embd = -1;
             break;
         }
 
         struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
         if (tensor->type != GGML_TYPE_F32) {
-            COM_ERR("invalid (non-F32) direction tensor type in %s\n", load_info.fname.c_str());
+            LOG_ERR("%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str());
             result.n_embd = -1;
             break;
         }
         if (ggml_n_dims(tensor) != 1) {
-            COM_ERR("invalid (non-1D) direction tensor shape in %s\n", load_info.fname.c_str());
+            LOG_ERR("%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str());
             result.n_embd = -1;
             break;
         }
@@ -2056,7 +1851,7 @@ static common_control_vector_data common_control_vector_load_one(const common_co
         if (result.n_embd == -1) {
             result.n_embd = ggml_nelements(tensor);
         } else if (ggml_nelements(tensor) != result.n_embd) {
-            COM_ERR("direction tensor in %s does not match previous dimensions\n", load_info.fname.c_str());
+            LOG_ERR("%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str());
             result.n_embd = -1;
             break;
         }
@@ -2073,7 +1868,7 @@ static common_control_vector_data common_control_vector_load_one(const common_co
     }
 
     if (result.n_embd == -1) {
-        COM_WRN("skipping %s due to invalid direction tensors\n", load_info.fname.c_str());
+        LOG_WRN("%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str());
         result.data.clear();
     }
 
@@ -2094,7 +1889,7 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
             break;
         }
         if (result.n_embd != -1 && result.n_embd != cur.n_embd) {
-            COM_ERR("control vectors in %s does not match previous dimensions\n", info.fname.c_str());
+            LOG_ERR("%s: control vectors in %s does not match previous dimensions\n", __func__, info.fname.c_str());
             result.n_embd = -1;
             break;
         }
@@ -2110,7 +1905,7 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
     }
 
     if (result.n_embd == -1) {
-        COM_ERR("%s", "no valid control vector files passed\n");
+        LOG_ERR("%s: no valid control vector files passed\n", __func__);
         result.data.clear();
     }
 
@@ -2221,13 +2016,13 @@ bool common_prompt_batch_decode(
         // memory, so we can't just remove the last token from the memory and replay the last token which
         // is the reason for this logic.
         if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_tokens_before_last))) {
-            COM_ERR("%s", "failed to eval\n");
+            LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
         n_past += n_tokens_before_last;
 
         llama_state_save_file(ctx, state_path.data(), all_tokens.data(), all_tokens.size());
-        COM_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size());
+        LOG_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size());
 
         llama_token last_token = all_tokens.back();
         llama_batch batch = llama_batch_get_one(&last_token, 1);
@@ -2235,13 +2030,13 @@ bool common_prompt_batch_decode(
         batch.pos = &pos;
 
         if (llama_decode(ctx, batch)) {
-            COM_ERR("%s", "failed to eval last token\n");
+            LOG_ERR("%s : failed to eval last token\n", __func__);
             return false;
         }
         n_past++;
     } else {
         if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_new))) {
-            COM_ERR("%s", "failed to eval\n");
+            LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
         n_past += n_new;

common/common.h

@@ -6,7 +6,6 @@
 
 #include "ggml-opt.h"
 #include "ggml.h"
-#include "llama.h"
 
 #include <set>
 #include <sstream>
@@ -15,7 +14,6 @@
 #include <vector>
 #include <map>
 #include <algorithm>
-#include <fstream>
 
 #if defined(_WIN32) && !defined(_WIN32_WINNT)
 #define _WIN32_WINNT 0x0A00
@@ -27,13 +25,6 @@
 #define DIRECTORY_SEPARATOR '/'
 #endif // _WIN32
 
-#define COM_DBG(fmt, ...) LOG_DBG("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define COM_TRC(fmt, ...) LOG_TRC("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define COM_INF(fmt, ...) LOG_INF("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define COM_WRN(fmt, ...) LOG_WRN("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define COM_ERR(fmt, ...) LOG_ERR("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define COM_CNT(fmt, ...) LOG_CNT(""              fmt,               __VA_ARGS__)
-
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
@@ -105,8 +96,6 @@ enum llama_example {
     LLAMA_EXAMPLE_FIT_PARAMS,
     LLAMA_EXAMPLE_RESULTS,
     LLAMA_EXAMPLE_EXPORT_GRAPH_OPS,
-    LLAMA_EXAMPLE_DOWNLOAD,
-    LLAMA_EXAMPLE_TOKENIZE,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -172,8 +161,6 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE,  // standalone draft model speculative decoding
     COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3,  // Eagle3 speculative decoding
     COMMON_SPECULATIVE_TYPE_DRAFT_MTP,     // Multi-token prediction
-    COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH,  // DFlash speculative decoding
-    COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK,  // DSpark speculative decoding (DFlash + Markov head)
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding based on n-grams
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
@@ -235,14 +222,14 @@ struct common_params_sampling {
     float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
     float   dynatemp_range     = 0.00f;  // 0.0 = disabled
     float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty)
+    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float   penalty_repeat     = 1.00f;  // 1.0 = disabled
     float   penalty_freq       = 0.00f;  // 0.0 = disabled
     float   penalty_present    = 0.00f;  // 0.0 = disabled
     float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
     float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
     int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = 64;     // how many tokens to scan for repetitions (0 = disable penalty)
+    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
     float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
     float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
     int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
@@ -285,27 +272,31 @@ struct common_params_sampling {
 
     // reasoning budget sampler parameters
     // these are populated by the server/CLI based on chat template params
-    int32_t                   reasoning_budget_tokens   = -1;  // -1 = disabled, >= 0 = token budget
-    std::vector<llama_token>  reasoning_budget_start;          // start tag token sequence
-    std::vector<llama_tokens> reasoning_budget_end;            // end tag token sequences; the first tag is used as the forcing sequence
-    std::vector<llama_token>  reasoning_budget_forced;         // forced sequence (message + first end tag)
-    std::string               reasoning_budget_message;        // message injected before end tag when budget exhausted
-    bool                      reasoning_control = false;       // create the budget sampler on demand so reasoning can be ended at runtime
+    int32_t                  reasoning_budget_tokens   = -1;   // -1 = disabled, >= 0 = token budget
+    std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
+    std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
+    std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)
+    std::string              reasoning_budget_message;         // message injected before end tag when budget exhausted
+    bool                     reasoning_control = false;        // create the budget sampler on demand so reasoning can be ended at runtime
 
     bool backend_sampling = false;
 
+    bool has_logit_bias() const {
+        return !logit_bias.empty();
+    }
+
     // print the parameters into a string
     std::string print() const;
 };
 
 struct common_params_model {
-    std::string path        = ""; // model local path
-    std::string url         = ""; // model url to download
-    std::string hf_repo     = ""; // HF repo
-    std::string hf_file     = ""; // HF file
-    std::string docker_repo = ""; // Docker repo
+    std::string path        = ""; // model local path                                       // NOLINT
+    std::string url         = ""; // model url to download                                  // NOLINT
+    std::string hf_repo     = ""; // HF repo                                                // NOLINT
+    std::string hf_file     = ""; // HF file                                                // NOLINT
+    std::string docker_repo = ""; // Docker repo                                            // NOLINT
 
-    std::string get_name() const {
+    std::string get_name() {
         if (!hf_repo.empty()) {
             return hf_repo;
         }
@@ -314,10 +305,6 @@ struct common_params_model {
         }
         return path;
     }
-
-    bool empty() const {
-        return get_name().empty();
-    }
 };
 
 // draft-model-based speculative decoding parameters
@@ -380,18 +367,26 @@ struct common_params_speculative {
     common_params_speculative_ngram_cache ngram_cache;
 
     bool has_dft() const {
-        return !draft.mparams.empty();
+        return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
     }
 
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
-            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
+            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3;
         });
 
         return needs_rs_seq ? draft.n_max : 0u;
     }
 };
 
+struct common_params_vocoder {
+    struct common_params_model model;
+
+    std::string speaker_file; // speaker file path
+
+    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy
+};
+
 struct common_params_diffusion {
     int32_t steps         = 128;
     bool    visual_mode   = false;
@@ -447,7 +442,6 @@ struct common_params {
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
     int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
-    int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -473,11 +467,12 @@ struct common_params {
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
 
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;
 
+    int sampling_n_threads = -1; // number of threads for sampling, used by server
+
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
@@ -490,6 +485,7 @@ struct common_params {
 
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
+    struct common_params_vocoder     vocoder;
     struct common_params_diffusion   diffusion;
 
     struct common_params_model model;
@@ -525,6 +521,7 @@ struct common_params {
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
+    bool    skip_download              = false; // skip model file downloading
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -563,6 +560,9 @@ struct common_params {
     bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool use_mmap          = true;  // enable mmap to use filesystem cache
+    bool use_direct_io     = false; // read from disk without buffering
+    bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool no_kv_offload     = false; // disable KV offloading
@@ -581,10 +581,9 @@ struct common_params {
 
     // multimodal models (see tools/mtmd)
     struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;                 // use GPU for multimodal model
-    ggml_backend_dev_t mmproj_device = nullptr; // GPU device to use for multimodal model
-    bool no_mmproj = false;                     // explicitly disable multimodal model
-    std::vector<std::string> image;             // path to image file(s) ; TODO: change the name to "media"
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
+    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
     int mtmd_batch_max_tokens = 1024;
@@ -612,7 +611,7 @@ struct common_params {
     bool    cache_prompt        = true;  // whether to enable prompt caching
     bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
-    int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
+    int32_t checkpoint_min_step = 256;   // minimum spacing between context checkpoints
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
@@ -620,14 +619,6 @@ struct common_params {
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = true;                                                                                  // NOLINT
-
-    // server CORS params
-    std::string cors_origins = "*";
-    std::string cors_methods = "GET, POST, DELETE, OPTIONS";
-    std::string cors_headers = "*";
-    bool cors_credentials = true;
-    bool cors_origins_explicit = false; // for --agent option
-
     bool enable_chat_template = true;
     bool force_pure_content_parser = false;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
@@ -642,9 +633,6 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
-    // CLI params
-    std::string server_base; // if set, connect to this server instead of starting a new one
-
     // UI configs
     bool ui = true;
     bool ui_mcp_proxy = false;
@@ -657,11 +645,6 @@ struct common_params {
 
     // enable built-in tools
     std::vector<std::string> server_tools;
-    std::string server_tools_runtime;
-
-    // MCP server configs (Cursor-compatible JSON)
-    std::string mcp_servers_config;   // path to JSON file with MCP server definitions
-    std::string mcp_servers_json;     // inline JSON with MCP server definitions
 
     // router server configs
     std::string models_dir    = "";     // directory containing models for the router server
@@ -719,12 +702,6 @@ struct common_params {
     // batched-bench params
     bool batched_bench_output_jsonl = false;
 
-    // tokenize params
-    bool tokenize_ids        = false; // if true, only print the token IDs
-    bool tokenize_stdin      = false; // if true, read the prompt from stdin
-    bool tokenize_no_bos     = false; // if true, do not add the BOS token
-    bool tokenize_show_count = false; // if true, print the total token count
-
     // common params
     std::string out_file; // output filename for all example programs
     // optional callback for model loading progress and cancellation:
@@ -733,12 +710,6 @@ struct common_params {
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
     bool no_alloc = false; // Don't allocate model buffers
-
-    // TTS params
-    std::string tts_lang = "";
-    std::string tts_speaker_file = "";
-
-    bool is_gen_docs = false; // whether we are running inside llama-gen-docs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -863,15 +834,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
 bool glob_match(const std::string & pattern, const std::string & str);
 
-//
-// Environment utils
-//
-
-// portable environment access, an unset variable reads as an empty string
-// and setting an empty value unsets the variable
-std::string common_get_env(const std::string & name);
-void        common_set_env(const std::string & name, const std::string & value);
-
 //
 // Filesystem utils
 //
@@ -882,7 +844,6 @@ bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
-std::string fs_get_config_directory();
 
 struct common_file_info {
     std::string path;
@@ -930,8 +891,9 @@ using common_init_result_ptr = std::unique_ptr<common_init_result>;
 
 common_init_result_ptr common_init_from_params(common_params & params, bool model_only = false);
 
-struct llama_model_params   common_model_params_to_llama  (      common_params & params);
-struct llama_context_params common_context_params_to_llama(const common_params & params);
+struct llama_model_params     common_model_params_to_llama  (      common_params & params);
+struct llama_context_params   common_context_params_to_llama(const common_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
@@ -939,31 +901,6 @@ void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adap
 // model endpoint from env
 std::string common_get_model_endpoint();
 
-// for testing purposes
-char * common_get_model_or_exit(int, char*[]);
-
-//
-// Threadpool utils
-//
-
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
-
-struct common_threadpools {
-    common_threadpools() = default;
-    ~common_threadpools();
-
-    common_threadpools(const common_threadpools &) = delete;
-    common_threadpools & operator=(const common_threadpools &) = delete;
-
-    void init(llama_context * ctx, const common_params & params);
-
-private:
-    ggml_threadpool * threadpool       = nullptr;
-    ggml_threadpool * threadpool_batch = nullptr;
-
-    decltype(ggml_threadpool_free) * free_fn = nullptr;
-};
-
 //
 // Context utils
 //
@@ -979,17 +916,10 @@ enum common_context_seq_rm_type {
 // note: clears the memory of the context
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
 
-struct common_memory {
-    llama_context * ctx_tgt = nullptr;
-    llama_context * ctx_dft = nullptr;
-
-    void init(llama_context * ctx_tgt, llama_context * ctx_dft = nullptr);
-
-    // aborts execution on failure
-    void seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) const;
-    void seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const;
-    void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const;
-};
+// aborts execution on failure
+void common_context_seq_rm (llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1);
+void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta);
+void common_context_seq_cp (llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1);
 
 //
 // Batch utils
@@ -1137,9 +1067,6 @@ enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);
 struct common_prompt_checkpoint {
     int64_t n_tokens;
 
-    // (optional) id of the task that created the checkpoint
-    int id_task = -1;
-
     llama_pos pos_min;
     llama_pos pos_max;
 

common/download.cpp

@@ -5,7 +5,9 @@
 #include "log.h"
 #include "download.h"
 #include "hf-cache.h"
-#include "json.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
 
 #include <algorithm>
 #include <filesystem>
@@ -42,6 +44,8 @@
 #include <unistd.h>
 #endif
 
+using json = nlohmann::ordered_json;
+
 //
 // downloader
 //
@@ -288,6 +292,10 @@ static int common_download_file_single_online(const std::string & url,
 
     const bool file_exists = std::filesystem::exists(path);
 
+    if (!file_exists && opts.skip_download) {
+        return -2; // file is missing and download is disabled
+    }
+
     if (file_exists && skip_etag) {
         LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
         return 304; // 304 Not Modified - fake cached response
@@ -354,6 +362,9 @@ static int common_download_file_single_online(const std::string & url,
             return 304; // 304 Not Modified - fake cached response
         }
         // pass this point, the file exists but is different from the server version, so we need to redownload it
+        if (opts.skip_download) {
+            return -2; // special code to indicate that the download was skipped due to etag mismatch
+        }
         if (remove(path.c_str()) != 0) {
             LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
             return -1;
@@ -564,30 +575,16 @@ static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
 }
 
 // pick the best sibling GGUF whose filename contains `keyword` (e.g. "mmproj" / "mtp"),
-// preferring deeper shared directory prefix with the model, then exact `tag` match,
-// then closest quantization to the tag when given, or to the model otherwise
+// preferring deeper shared directory prefix with the model, then closest quantization
 static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
                                            const std::string        & model,
-                                           const std::string        & keyword,
-                                           const std::string        & tag = "") {
+                                           const std::string        & keyword) {
     hf_cache::hf_file best;
     size_t best_depth = 0;
     int best_diff = 0;
-    bool best_exact = false;
     bool found = false;
 
-    std::string tag_upper = tag;
-    for (char & c : tag_upper) {
-        c = (char) std::toupper((unsigned char) c);
-    }
-
-    int model_bits = 0;
-    if (!tag_upper.empty()) {
-        auto pos = tag_upper.find_first_of("0123456789");
-        model_bits = pos == std::string::npos ? 0 : std::stoi(tag_upper.substr(pos));
-    } else {
-        model_bits = extract_quant_bits(model);
-    }
+    auto model_bits = extract_quant_bits(model);
     auto model_parts = string_split<std::string>(model, '/');
     auto model_dir = model_parts.end() - 1;
 
@@ -610,19 +607,10 @@ static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
         auto bits = extract_quant_bits(f.path);
         auto diff = std::abs(bits - model_bits);
 
-        std::string path_upper = f.path;
-        for (char & c : path_upper) {
-            c = (char) std::toupper((unsigned char) c);
-        }
-        bool exact = !tag_upper.empty() && path_upper.find("-" + tag_upper + ".") != std::string::npos;
-
-        if (!found || depth > best_depth ||
-            (depth == best_depth && exact && !best_exact) ||
-            (depth == best_depth && exact == best_exact && diff < best_diff)) {
+        if (!found || depth > best_depth || (depth == best_depth && diff < best_diff)) {
             best = f;
             best_depth = depth;
             best_diff = diff;
-            best_exact = exact;
             found = true;
         }
     }
@@ -635,27 +623,8 @@ static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
 }
 
 static hf_cache::hf_file find_best_mtp(const hf_cache::hf_files & files,
-                                       const std::string        & model,
-                                       const std::string        & tag = "") {
-    return find_best_sibling(files, model, "mtp-", tag);
-}
-
-static hf_cache::hf_file find_best_eagle3(const hf_cache::hf_files & files,
-                                          const std::string        & model,
-                                          const std::string        & tag = "") {
-    return find_best_sibling(files, model, "eagle3-", tag);
-}
-
-static hf_cache::hf_file find_best_dflash(const hf_cache::hf_files & files,
-                                          const std::string        & model,
-                                          const std::string        & tag = "") {
-    return find_best_sibling(files, model, "dflash-", tag);
-}
-
-static hf_cache::hf_file find_best_dspark(const hf_cache::hf_files & files,
-                                          const std::string        & model,
-                                          const std::string        & tag = "") {
-    return find_best_sibling(files, model, "dspark-", tag);
+                                       const std::string        & model) {
+    return find_best_sibling(files, model, "mtp-");
 }
 
 static bool gguf_filename_is_model(const std::string & filepath) {
@@ -670,10 +639,7 @@ static bool gguf_filename_is_model(const std::string & filepath) {
 
     return filename.find("mmproj")  == std::string::npos &&
            filename.find("imatrix") == std::string::npos &&
-           filename.find("mtp-")    == std::string::npos &&
-           filename.find("eagle3-") == std::string::npos &&
-           filename.find("dflash-") == std::string::npos &&
-           filename.find("dspark-") == std::string::npos;
+           filename.find("mtp-")    == std::string::npos;
 }
 
 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -725,8 +691,19 @@ static void list_available_gguf_files(const hf_cache::hf_files & files) {
     }
 }
 
-common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts) {
-    common_download_hf_plan plan;
+struct hf_plan {
+    hf_cache::hf_file primary;
+    hf_cache::hf_files model_files;
+    hf_cache::hf_file mmproj;
+    hf_cache::hf_file mtp;
+    hf_cache::hf_file preset; // if set, only this file is downloaded
+};
+
+static hf_plan get_hf_plan(const common_params_model  & model,
+                           const common_download_opts & opts,
+                           bool download_mmproj,
+                           bool download_mtp) {
+    hf_plan plan;
     hf_cache::hf_files all;
 
     auto [repo, tag] = common_download_split_repo_tag(model.hf_repo);
@@ -765,77 +742,135 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
         }
     } else {
         primary = find_best_model(all, tag);
-        // a requested sidecar can resolve on its own, without a full model of the same tag
-        if (primary.path.empty() && !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3 && !opts.download_dspark) {
+        if (primary.path.empty()) {
             LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
             list_available_gguf_files(all);
             return plan;
         }
     }
 
-    if (!primary.path.empty()) {
-        plan.primary = primary;
-        plan.model_files = get_split_files(all, primary);
-    }
+    plan.primary = primary;
+    plan.model_files = get_split_files(all, primary);
 
-    if (opts.download_mmproj && !primary.path.empty()) {
+    if (download_mmproj) {
         plan.mmproj = find_best_mmproj(all, primary.path);
     }
-    if (opts.download_mtp) {
-        plan.mtp = find_best_mtp(all, primary.path, tag);
-    }
-    if (opts.download_dflash) {
-        plan.dflash = find_best_dflash(all, primary.path, tag);
-    }
-    if (opts.download_eagle3) {
-        plan.eagle3 = find_best_eagle3(all, primary.path, tag);
-    }
-    if (opts.download_dspark) {
-        plan.dspark = find_best_dspark(all, primary.path, tag);
-    }
 
-    if (primary.path.empty() &&
-        plan.mtp.local_path.empty() && plan.dflash.local_path.empty() && plan.eagle3.local_path.empty() && plan.dspark.local_path.empty()) {
-        LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
-        list_available_gguf_files(all);
+    if (download_mtp) {
+        plan.mtp = find_best_mtp(all, primary.path);
     }
 
     return plan;
 }
 
-void common_download_run_tasks(const std::vector<common_download_task> & tasks) {
+struct download_task {
+    std::string url;
+    std::string path;
+};
+
+static std::vector<download_task> get_url_tasks(const common_params_model & model) {
+    auto split = get_gguf_split_info(model.url);
+
+    if (split.count <= 1) {
+        return {{model.url, model.path}};
+    }
+
+    auto filename = split.prefix;
+    if (auto pos = split.prefix.rfind('/'); pos != std::string::npos) {
+        filename = split.prefix.substr(pos + 1);
+    }
+
+    auto parent_path = std::filesystem::path(model.path).parent_path();
+    auto prefix_path = (parent_path / filename).string();
+
+    std::vector<download_task> tasks;
+    for (int i = 1; i <= split.count; i++) {
+        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
+        tasks.push_back({split.prefix + suffix, prefix_path + suffix});
+    }
+    return tasks;
+}
+
+common_download_model_result common_download_model(const common_params_model  & model,
+                                                   const common_download_opts & opts) {
+    common_download_model_result result;
+    std::vector<download_task> tasks;
+    hf_plan hf;
+
+    bool download_mmproj = opts.download_mmproj;
+    bool download_mtp = opts.download_mtp;
+    bool is_hf = !model.hf_repo.empty();
+
+    if (is_hf) {
+        hf = get_hf_plan(model, opts, download_mmproj, download_mtp);
+        if (!hf.preset.path.empty()) {
+            // if preset.ini exists, only download that file alone
+            tasks.push_back({hf.preset.url, hf.preset.local_path});
+        } else {
+            for (const auto & f : hf.model_files) {
+                tasks.push_back({f.url, f.local_path});
+            }
+            if (!hf.mmproj.path.empty()) {
+                tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
+            }
+            if (!hf.mtp.path.empty()) {
+                tasks.push_back({hf.mtp.url, hf.mtp.local_path});
+            }
+        }
+    } else if (!model.url.empty()) {
+        tasks = get_url_tasks(model);
+    } else {
+        result.model_path = model.path;
+        return result;
+    }
+
+    if (tasks.empty()) {
+        return result;
+    }
+
     std::vector<std::future<int>> futures;
     for (const auto & task : tasks) {
         futures.push_back(std::async(std::launch::async,
-            [&task]() {
-                return common_download_file_single(task.url, task.local_path, task.opts, task.is_hf);
+            [&task, &opts, is_hf]() {
+                return common_download_file_single(task.url, task.path, opts, is_hf);
             }
         ));
     }
 
-    for (size_t i = 0; i < futures.size(); ++i) {
-        std::string url = tasks[i].url;
-        int status = futures[i].get();
+    for (auto & f : futures) {
+        int status = f.get();
+        if (status == -2 && opts.skip_download) {
+            throw common_skip_download_exception();
+        }
         bool is_ok = is_http_status_ok(status);
         if (!is_ok) {
-            throw std::runtime_error(string_format("Download '%s' failed with status code: %d", url.c_str(), status));
+            return {};
         }
     }
-}
 
-std::vector<std::string> common_download_get_all_parts(const std::string & url) {
-    auto split = get_gguf_split_info(url);
+    if (is_hf) {
+        if (!hf.preset.path.empty()) {
+            // if preset.ini is used, do not set other paths
+            result.preset_path = hf_cache::finalize_file(hf.preset);
+        } else {
+            for (const auto & f : hf.model_files) {
+                hf_cache::finalize_file(f);
+            }
+            result.model_path = hf.primary.final_path;
 
-    if (split.count <= 1) {
-        return {url};
+            if (!hf.mmproj.path.empty()) {
+                result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
+            }
+
+            if (!hf.mtp.path.empty()) {
+                result.mtp_path = hf_cache::finalize_file(hf.mtp);
+            }
+        }
+    } else {
+        result.model_path = model.path;
     }
 
-    std::vector<std::string> parts;
-    for (int i = 1; i <= split.count; i++) {
-        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
-        parts.push_back(split.prefix + suffix);
-    }
-    return parts;
+    return result;
 }
 
 //
@@ -852,8 +887,8 @@ static std::string common_docker_get_token(const std::string & repo) {
         throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
     }
 
-    std::string response_str(res.second.begin(), res.second.end());
-    common_json response = common_json::parse(response_str);
+    std::string            response_str(res.second.begin(), res.second.end());
+    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
 
     if (!response.contains("token")) {
         throw std::runtime_error("Docker registry token response missing 'token' field");
@@ -915,9 +950,9 @@ std::string common_docker_resolve_model(const std::string & docker) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
         }
 
-        std::string manifest_str(manifest_res.second.begin(), manifest_res.second.end());
-        common_json manifest = common_json::parse(manifest_str);
-        std::string gguf_digest;  // Find the GGUF layer
+        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
+        std::string            gguf_digest;  // Find the GGUF layer
         if (manifest.contains("layers")) {
             for (const auto & layer : manifest["layers"]) {
                 if (layer.contains("mediaType")) {
@@ -970,11 +1005,8 @@ std::vector<common_cached_model_info> common_list_cached_models() {
     for (const auto & f : files) {
         auto split = get_gguf_split_info(f.path);
         if (split.index != 1 || split.tag.empty() ||
-            split.prefix.find("mmproj")  != std::string::npos ||
-            split.prefix.find("mtp-")    != std::string::npos ||
-            split.prefix.find("eagle3-") != std::string::npos ||
-            split.prefix.find("dflash-") != std::string::npos ||
-            split.prefix.find("dspark-") != std::string::npos) {
+            split.prefix.find("mmproj") != std::string::npos ||
+            split.prefix.find("mtp-")   != std::string::npos) {
             continue;
         }
         if (seen.insert(f.repo_id + ":" + split.tag).second) {
@@ -985,26 +1017,6 @@ std::vector<common_cached_model_info> common_list_cached_models() {
     return result;
 }
 
-std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
-    auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
-
-    auto files = hf_cache::get_cached_files(repo);
-    if (files.empty()) {
-        return "";
-    }
-
-    if (!hf_file.empty()) {
-        for (const auto & f : files) {
-            if (f.path == hf_file) {
-                return f.local_path;
-            }
-        }
-        return "";
-    }
-
-    return find_best_model(files, tag).local_path;
-}
-
 bool common_download_remove(const std::string & hf_repo_with_tag) {
     namespace fs = std::filesystem;
 

common/download.h

@@ -1,10 +1,7 @@
 #pragma once
 
-#include "hf-cache.h"
-
 #include <string>
 #include <vector>
-#include <functional>
 
 struct common_params_model;
 
@@ -50,47 +47,66 @@ struct common_cached_model_info {
     }
 };
 
-// Options for common_download_file_single
+// Options for common_download_model and common_download_file_single
 struct common_download_opts {
     std::string bearer_token;
     common_header_list headers;
     bool offline = false;
-    bool download_mmproj  = false;
-    bool download_mtp     = false;
-    bool download_eagle3  = false;
-    bool download_dflash  = false;
-    bool download_dspark  = false;
+    bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
+    bool download_mmproj = false;
+    bool download_mtp = false;
     common_download_callback * callback = nullptr;
 };
 
-struct common_download_task {
-    common_download_opts opts;
-    std::string url;
-    std::string local_path;
-    std::function<void()> on_done;
-    bool is_hf = false;
-
-    common_download_task() = default;
-    common_download_task(hf_cache::hf_file f,
-            const common_download_opts & opts,
-            std::function<void()> on_done = nullptr)
-        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true) {}
+// Result of common_download_model
+struct common_download_model_result {
+    std::string model_path;
+    std::string mmproj_path;
+    std::string mtp_path;
+    std::string preset_path;
 };
 
-void common_download_run_tasks(const std::vector<common_download_task> & tasks);
+// throw if the file is missing or invalid (e.g. ETag check failed)
+struct common_skip_download_exception : public std::runtime_error {
+    common_skip_download_exception() : std::runtime_error("skip download") {}
+};
 
-// if url is a multi-part GGUF file, returns all parts, otherwise returns the single file
-std::vector<std::string> common_download_get_all_parts(const std::string & url);
+// Download model from HuggingFace repo or URL
+//
+// input (via model struct):
+// - model.hf_repo: HF repo with optional tag, see common_download_split_repo_tag
+// - model.hf_file: specific file in the repo (requires hf_repo)
+// - model.url: simple download (used if hf_repo is empty)
+// - model.path: local file path
+//
+// tag matching (for HF repos without model.hf_file):
+// - if tag is specified, searches for GGUF matching that quantization
+// - if no tag, searches for Q4_K_M, then Q4_0, then first available GGUF
+//
+// split GGUF: multi-part files like "model-00001-of-00003.gguf" are automatically
+// detected and all parts are downloaded
+//
+// caching:
+// - HF repos: uses HuggingFace cache
+// - URLs: uses ETag-based caching
+//
+// when opts.offline=true, no network requests are made
+// when download_mmproj=true, searches for mmproj in same directory as model or any parent directory
+// then with the closest quantization bits
+// when download_mtp=true, applies the same sibling search for an MTP-head GGUF
+//
+// returns result with model_path, mmproj_path and mtp_path (empty when not found / on failure)
+common_download_model_result common_download_model(
+    const common_params_model & model,
+    const common_download_opts & opts = {}
+);
 
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();
 
-// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
-// returns an empty string if the model is not present in the cache
-std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
-
 // download single file from url to local path
 // returns status code or -1 on error
+// returns -2 if the download was skipped due to ETag mismatch (file outdated, skip_download=true)
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
 int common_download_file_single(const std::string & url,
                                 const std::string & path,
@@ -107,15 +123,3 @@ std::string common_docker_resolve_model(const std::string & docker);
 // - if tag is present, removes only files matching that tag (and orphaned blobs)
 // returns true if anything was removed
 bool common_download_remove(const std::string & hf_repo_with_tag);
-
-struct common_download_hf_plan {
-    hf_cache::hf_file primary;
-    hf_cache::hf_files model_files;
-    hf_cache::hf_file mmproj;
-    hf_cache::hf_file mtp;
-    hf_cache::hf_file eagle3;
-    hf_cache::hf_file dflash;
-    hf_cache::hf_file dspark;
-    hf_cache::hf_file preset; // if set, only this file is downloaded
-};
-common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);

common/fit.cpp

@@ -54,7 +54,8 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
 
     llama_model_params mparams_copy = *mparams;
     mparams_copy.no_alloc  = true;
-    mparams_copy.load_mode = LLAMA_LOAD_MODE_NONE;
+    mparams_copy.use_mmap  = false;
+    mparams_copy.use_mlock = false;
 
     llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
     if (model == nullptr) {
@@ -137,9 +138,6 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
     }
 
     hp_ngl         = llama_model_n_layer(model);
-    if (mparams->load_mtp) {
-        hp_ngl    += llama_model_n_layer_nextn(model);
-    }
     hp_n_ctx_train = llama_model_n_ctx_train(model);
     hp_n_expert    = llama_model_n_expert(model);
 
@@ -178,7 +176,7 @@ common_device_memory_data_vec common_get_device_memory_data(
 static void common_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t * margins_s, uint32_t n_ctx_min, const common_fit_extra_model * extra, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
         throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
     }
@@ -191,92 +189,10 @@ static void common_params_fit_impl(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
-    // with non-unified kv, we need to take into account n_streams
-    // for example, if memory can hold more than model's trained context size, we must extend the n_ctx to hold enough n_streams
-    const uint32_t n_streams  = cparams->kv_unified ? 1 : std::max<uint32_t>(1, cparams->n_seq_max);
-    const bool     n_ctx_auto = cparams->n_ctx == 0;
-
-    dmds_t   dmds_extra;       // memory of the extra model, laid out on the devices of the main model
-    uint32_t n_ctx_extra = 0;  // context that memory was measured at
-
-    // the extra model competes for the same memory as the main model, add it to every measurement
-    // its memory is measured again whenever the context it follows changes
-    auto add_extra_memory = [&](dmds_t & dmds) {
-        if (extra == nullptr) {
-            return;
-        }
-
-        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
-            std::vector<ggml_backend_dev_t> devs_extra;
-            uint32_t ngl_extra = 0;
-            uint32_t nct_extra = 0;
-            uint32_t nex_extra = 0;
-
-            extra->cparams->n_ctx = cparams->n_ctx;
-
-            LOG_TRC("%s: getting device memory data for the extra model at a context size of %" PRIu32 ":\n",
-                __func__, cparams->n_ctx);
-
-            dmds_t measured;
-            try {
-                measured = common_get_device_memory_data_impl(
-                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
-            } catch (const std::runtime_error & e) {
-                // the extra model is optional, fit the main model alone rather than giving up
-                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
-                dmds_extra = dmds_t(devs.size() + 1);
-                n_ctx_extra = cparams->n_ctx;
-                return;
-            }
-
-            dmds_extra = dmds_t(devs.size() + 1);
-            dmds_extra.back().mb = measured.back().mb;
-            for (size_t je = 0; je < devs_extra.size(); je++) {
-                for (size_t id = 0; id < devs.size(); id++) {
-                    if (devs_extra[je] == devs[id]) {
-                        dmds_extra[id].mb.model   += measured[je].mb.model;
-                        dmds_extra[id].mb.context += measured[je].mb.context;
-                        dmds_extra[id].mb.compute += measured[je].mb.compute;
-                        break;
-                    }
-                }
-            }
-            if (extra->shares_model) {
-                for (llama_device_memory_data & dmd : dmds_extra) {
-                    dmd.mb.model = 0;
-                }
-            }
-
-            n_ctx_extra = cparams->n_ctx;
-        }
-
-        for (size_t id = 0; id < dmds.size(); id++) {
-            dmds[id].mb.model   += dmds_extra[id].mb.model;
-            dmds[id].mb.context += dmds_extra[id].mb.context;
-            dmds[id].mb.compute += dmds_extra[id].mb.compute;
-        }
-    };
-
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-
-    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
-    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
-    const uint32_t n_ctx_min_total = (uint32_t) std::min<uint64_t>(uint64_t(n_ctx_min) * n_streams, UINT32_MAX);
-
-    // llama_context would use only hp_nct in total for n_ctx == 0, resolve the context before measuring anything else:
-    if (n_ctx_auto) {
-        cparams->n_ctx = n_ctx_max;
-        if (n_streams > 1) {
-            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
-                __func__, n_ctx_max, n_streams);
-            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        }
-    }
-    add_extra_memory(dmds_full);
-
+    const dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
     const size_t nd = devs.size(); // number of devices
 
     std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
@@ -317,7 +233,7 @@ static void common_params_fit_impl(
         sum_projected_used = dmds_full.back().mb.total();
         sum_free           = dmds_full.back().total;
         sum_projected_free = sum_free - sum_projected_used;
-        LOG_TRC("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n",
+        LOG_INF("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n",
             __func__, sum_projected_used/MiB, sum_free/MiB);
         if (sum_projected_free >= margins[0]) {
             LOG_TRC("%s: will leave %" PRId64 " >= %" PRId64 " MiB of system memory, no changes needed\n",
@@ -389,8 +305,8 @@ static void common_params_fit_impl(
                     "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
                     __func__, -global_surplus/MiB);
             }
-            if (n_ctx_auto) {
-                if (n_ctx_max > n_ctx_min_total) {
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
                     int64_t sum_used_target = sum_free;
                     if (nd == 0) {
                         sum_used_target -= margins[0];
@@ -410,9 +326,8 @@ static void common_params_fit_impl(
                     }
 
                     int64_t sum_projected_used_min_ctx = 0;
-                    cparams->n_ctx = n_ctx_min_total;
-                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-                    add_extra_memory(dmds_min_ctx);
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
                     if (nd == 0) {
                         sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
                     } else {
@@ -422,16 +337,14 @@ static void common_params_fit_impl(
                     }
                     if (sum_used_target > sum_projected_used_min_ctx) {
                         // linear interpolation between minimum and maximum context size:
-                        cparams->n_ctx += (n_ctx_max - n_ctx_min_total) * (sum_used_target - sum_projected_used_min_ctx)
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
                             / (sum_projected_used - sum_projected_used_min_ctx);
-                        // round down context for CUDA backend, keep it divisible by the number of streams:
-                        const uint32_t align = 256 * n_streams;
-                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % align, n_ctx_min_total);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
 
-                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (n_ctx_max - n_ctx_min_total);
-                        const int64_t memory_reduction = (n_ctx_max - cparams->n_ctx) * bytes_per_ctx;
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd <= 1) {
                             LOG_TRC("%s: entire model can be fit by reducing context\n", __func__);
                             return;
@@ -440,14 +353,14 @@ static void common_params_fit_impl(
                     } else {
                         const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     if (n_ctx_min == UINT32_MAX) {
-                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, n_ctx_max);
+                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
                     } else {
                         LOG_TRC("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                            __func__, n_ctx_max, n_ctx_min_total);
+                            __func__, hp_nct, n_ctx_min);
                     }
                 }
             } else {
@@ -592,9 +505,8 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        dmds_t dmd_nl = common_get_device_memory_data_impl(
+        const dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmd_nl);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
         for (size_t id = 0; id < nd; id++) {
@@ -621,9 +533,8 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmds_cpu_moe);
 
         for (size_t id = 0; id < nd; id++) {
             global_surplus_cpu_moe += dmds_cpu_moe[id].free;
@@ -883,12 +794,11 @@ enum common_params_fit_status common_fit_params(
         llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t * margins,
         uint32_t n_ctx_min,
-        const common_fit_extra_model * extra,
         ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
     common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, extra, log_level);
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LOG_TRC("%s: successfully fit params to free device memory\n", __func__);
     } catch (const common_params_fit_exception & e) {
         LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());

common/fit.h

@@ -11,16 +11,6 @@ enum common_params_fit_status {
     COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
 };
 
-// a second model that shares the devices of the main model, e.g. a draft model
-//   - its context follows the context of the main model, so its memory is measured again whenever that context changes
-//   - shares_model tells the fit that the weights are already counted in the main model, as for an MTP context
-struct common_fit_extra_model {
-    const char * path_model;
-    llama_model_params * mparams;
-    llama_context_params * cparams;
-    bool shares_model;
-};
-
 // fits mparams and cparams to free device memory (assumes system memory is unlimited)
 //   - returns true if the parameters could be successfully modified to fit device memory
 //   - this function is NOT thread safe because it modifies the global llama logger state
@@ -34,7 +24,6 @@ common_params_fit_status common_fit_params(
    llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
                              size_t * margins,               // margins of memory to leave per device in bytes
                            uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
-      const common_fit_extra_model * extra,                  // model to fit alongside the main one, nullptr if there is none
                      ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout

common/hf-cache.cpp

@@ -4,7 +4,9 @@
 #include "common.h"
 #include "log.h"
 #include "http.h"
-#include "json.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
 
 #include <filesystem>
 #include <fstream>
@@ -13,6 +15,8 @@
 #include <string_view>
 #include <stdexcept>
 
+namespace nl = nlohmann;
+
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
@@ -191,8 +195,8 @@ static void safe_write_file(const fs::path & path, const std::string & data) {
     }
 }
 
-static common_json api_get(const std::string & url,
-                           const std::string & token) {
+static nl::json api_get(const std::string & url,
+                        const std::string & token) {
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
@@ -210,10 +214,10 @@ static common_json api_get(const std::string & url,
         auto body = res->body;
 
         if (res->status == 200) {
-            return common_json::parse(res->body);
+            return nl::json::parse(res->body);
         }
         try {
-            body = common_json::parse(res->body)["error"].get<std::string>();
+            body = nl::json::parse(res->body)["error"].get<std::string>();
         } catch (...) { }
 
         throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
@@ -276,7 +280,7 @@ static std::string get_repo_commit(const std::string & repo_id,
         safe_write_file(refs_path / name, commit);
         return commit;
 
-    } catch (const common_json_error & e) {
+    } catch (const nl::json::exception & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
@@ -354,7 +358,7 @@ hf_files get_repo_files(const std::string & repo_id,
 
             files.push_back(file);
         }
-    } catch (const common_json_error & e) {
+    } catch (const nl::json::exception & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());

common/http.h

@@ -2,16 +2,6 @@
 
 #include <cpp-httplib/httplib.h>
 
-#ifdef _WIN32
-#include <winsock2.h>
-#include <windows.h>
-#else
-#include <sys/socket.h>
-#include <netinet/in.h>
-#include <arpa/inet.h>
-#include <unistd.h>
-#endif
-
 struct common_http_url {
     std::string scheme;
     std::string user;
@@ -21,11 +11,6 @@ struct common_http_url {
     std::string path;
 };
 
-// bracket an IPv6 literal host for a URL authority (RFC 3986)
-static std::string common_http_format_host(const std::string & host) {
-    return host.find(':') != std::string::npos ? "[" + host + "]" : host;
-}
-
 static common_http_url common_http_parse_url(const std::string & url) {
     common_http_url parts;
     auto scheme_end = url.find("://");
@@ -64,28 +49,11 @@ static common_http_url common_http_parse_url(const std::string & url) {
         parts.path = "/";
     }
 
-    // split the authority into host and optional port, a bracketed IPv6 literal keeps its inner colons (RFC 3986)
-    std::string port_str;
-    if (!parts.host.empty() && parts.host.front() == '[') {
-        auto close = parts.host.find(']');
-        if (close == std::string::npos) {
-            throw std::runtime_error("invalid IPv6 URL authority: " + parts.host);
-        }
-        auto after = parts.host.substr(close + 1);
-        if (!after.empty() && after.front() == ':') {
-            port_str = after.substr(1);
-        }
-        parts.host = parts.host.substr(1, close - 1);
-    } else {
-        auto colon_pos = parts.host.find(':');
-        if (colon_pos != std::string::npos) {
-            port_str = parts.host.substr(colon_pos + 1);
-            parts.host = parts.host.substr(0, colon_pos);
-        }
-    }
+    auto colon_pos = parts.host.find(':');
 
-    if (!port_str.empty()) {
-        parts.port = std::stoi(port_str);
+    if (colon_pos != std::string::npos) {
+        parts.port = std::stoi(parts.host.substr(colon_pos + 1));
+        parts.host = parts.host.substr(0, colon_pos);
     } else if (parts.scheme == "http") {
         parts.port = 80;
     } else if (parts.scheme == "https") {
@@ -115,7 +83,7 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
     }
 #endif
 
-    httplib::Client cli(parts.scheme + "://" + common_http_format_host(parts.host) + ":" + std::to_string(parts.port));
+    httplib::Client cli(parts.scheme + "://" + parts.host + ":" + std::to_string(parts.port));
 
     if (!parts.user.empty()) {
         cli.set_basic_auth(parts.user, parts.password);
@@ -127,65 +95,5 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
 }
 
 static std::string common_http_show_masked_url(const common_http_url & parts) {
-    return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + common_http_format_host(parts.host) + parts.path;
-}
-
-static int common_http_get_free_port() {
-#ifdef _WIN32
-    WSADATA wsaData;
-    if (WSAStartup(MAKEWORD(2, 2), &wsaData) != 0) {
-        return -1;
-    }
-    typedef SOCKET native_socket_t;
-#define INVALID_SOCKET_VAL INVALID_SOCKET
-#define CLOSE_SOCKET(s) closesocket(s)
-#else
-    typedef int native_socket_t;
-#define INVALID_SOCKET_VAL -1
-#define CLOSE_SOCKET(s) close(s)
-#endif
-
-    native_socket_t sock = socket(AF_INET, SOCK_STREAM, 0);
-    if (sock == INVALID_SOCKET_VAL) {
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-    struct sockaddr_in serv_addr;
-    std::memset(&serv_addr, 0, sizeof(serv_addr));
-    serv_addr.sin_family = AF_INET;
-    serv_addr.sin_addr.s_addr = htonl(INADDR_ANY);
-    serv_addr.sin_port = htons(0);
-
-    if (bind(sock, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) != 0) {
-        CLOSE_SOCKET(sock);
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-#ifdef _WIN32
-    int namelen = sizeof(serv_addr);
-#else
-    socklen_t namelen = sizeof(serv_addr);
-#endif
-    if (getsockname(sock, (struct sockaddr*)&serv_addr, &namelen) != 0) {
-        CLOSE_SOCKET(sock);
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-    int port = ntohs(serv_addr.sin_port);
-
-    CLOSE_SOCKET(sock);
-#ifdef _WIN32
-    WSACleanup();
-#endif
-
-    return port;
+    return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
 }

common/imatrix-loader.cpp

@@ -102,8 +102,7 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) {
     const int64_t chunk_count_key = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT);
     const int64_t chunk_size_key  = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE);
 
-    if (datasets_key != -1 && gguf_get_kv_type(ctx_gguf, datasets_key) == GGUF_TYPE_ARRAY &&
-        gguf_get_arr_type(ctx_gguf, datasets_key) == GGUF_TYPE_STRING) {
+    if (datasets_key != -1 && gguf_get_arr_type(ctx_gguf, datasets_key) == GGUF_TYPE_STRING) {
         const int64_t n = gguf_get_arr_n(ctx_gguf, datasets_key);
         imatrix.datasets.reserve(imatrix.datasets.size() + n);
         for (int64_t i = 0; i < n; ++i) {
@@ -144,13 +143,6 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) {
             return false;
         }
 
-        if (in_sum2->type != GGML_TYPE_F32 || counts->type != GGML_TYPE_F32) {
-            LOG_ERR("%s: sums and counts for %s must be F32\n", __func__, name.c_str());
-            gguf_free(ctx_gguf);
-            ggml_free(ctx);
-            return false;
-        }
-
         auto & e = imatrix.entries[name];
 
         const int64_t nval    = ggml_nelements(in_sum2);

common/jinja/README.md

@@ -7,7 +7,7 @@ The implementation can be found in the `common/jinja` directory.
 ## Key Features
 
 - Input marking: security against special token injection
-- Decoupled from the JSON library: `common_json` is only used for JSON-to-internal type translation and is completely optional
+- Decoupled from `nlohmann::json`: this dependency is only used for JSON-to-internal type translation and is completely optional
 - Minimal primitive types: int, float, bool, string, array, object, none, undefined
 - Detailed logging: allow source tracing on error
 - Clean architecture: workarounds are applied to input data before entering the runtime (see `common/chat.cpp`)

common/jinja/caps.cpp

@@ -4,53 +4,34 @@
 
 // note: the json dependency is only for defining input in a convenient way
 // we can remove it in the future when we figure out a better way to define inputs using jinja::value
-#include "json.h"
+#include <nlohmann/json.hpp>
 
 #include <functional>
 #include <sstream>
 
 #define FILENAME "jinja-caps"
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 namespace jinja {
 
 using caps_json_fn = std::function<json()>;
-using caps_ctx_fn = std::function<void(context &)>;
-using caps_analyze_fn = std::function<void(context &, bool, value &, value &, const std::string &)>;
-
-void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) {
-    ctx.set_val("preserve_thinking",         mk_val<value_bool>(enabled));
-    ctx.set_val("clear_thinking",            mk_val<value_bool>(!enabled));
-    ctx.set_val("truncate_history_thinking", mk_val<value_bool>(!enabled));
-    ctx.set_val("drop_thinking",             mk_val<value_bool>(!enabled));
-}
-
-void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort) {
-    value var = mk_val<value_string>(effort); // bind to the same value for stats
-    ctx.set_val("reasoning_effort",   var);
-    ctx.set_val("reasoning_strength", var);
-}
+using caps_analyze_fn = std::function<void(bool, value &, value &)>;
 
 static void caps_try_execute(jinja::program & prog,
                              const caps_json_fn & messages_fn,
-                             const caps_ctx_fn & ctx_fn,
                              const caps_json_fn & tools_fn,
                              const caps_analyze_fn & analyze_fn) {
     context ctx;
     ctx.is_get_stats = true;
     jinja::global_from_json(ctx, json{
         {"messages", messages_fn()},
-        {"tools", tools_fn ? tools_fn() : json::array()},
+        {"tools", tools_fn()},
         {"bos_token", ""},
         {"eos_token", ""},
         {"add_generation_prompt", true}
     }, true);
 
-    if (ctx_fn) {
-        ctx_fn(ctx);
-    }
-
     auto messages = ctx.get_val("messages");
     auto tools = ctx.get_val("tools");
 
@@ -68,7 +49,7 @@ static void caps_try_execute(jinja::program & prog,
         // ignore exceptions during capability analysis
     }
 
-    analyze_fn(ctx, success, messages, tools, result);
+    analyze_fn(success, messages, tools);
 }
 
 // for debugging only
@@ -93,7 +74,6 @@ std::map<std::string, bool> caps::to_map() const {
         {"supports_parallel_tool_calls", supports_parallel_tool_calls},
         {"supports_system_role", supports_system_role},
         {"supports_preserve_reasoning", supports_preserve_reasoning},
-        {"supports_reasoning_effort", supports_reasoning_effort},
         {"supports_object_arguments", supports_object_arguments},
     };
 }
@@ -117,8 +97,6 @@ caps caps_get(jinja::program & prog) {
 
     JJ_DEBUG("%s\n", ">>> Running capability check: typed content");
 
-    static const std::string content_marker = "STRING_MARKER";
-
     // case: typed content support
     caps_try_execute(
         prog,
@@ -127,26 +105,24 @@ caps caps_get(jinja::program & prog) {
             return json::array({
                 {
                     {"role", "user"},
-                    {"content", content_marker}
+                    {"content", "content"}
                 }
             });
         },
-        nullptr, // ctx_fn
-        nullptr, // tools_fn
-        [&](context &, bool success, value & messages, value &, const std::string & rendered) {
+        [&]() {
+            // tools
+            return json{nullptr};
+        },
+        [&](bool success, value & messages, value &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
-            bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
-            if (used_as_array) {
+            if (has_op(content, "selectattr") || has_op(content, "array_access")) {
                 // accessed as an array
                 result.supports_typed_content = true;
             }
             if (!success) {
                 // failed to execute with content as string
                 result.supports_string_content = false;
-            } else if (used_as_array && rendered.find(content_marker) == std::string::npos) {
-                // edge case: string may be accessed for checking, but does not appear in the output
-                result.supports_string_content = false;
             }
         }
     );
@@ -169,9 +145,11 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        nullptr, // ctx_fn
-        nullptr, // tools_fn
-        [&](context &, bool, value & messages, value &, const std::string &) {
+        [&]() {
+            // tools
+            return json::array();
+        },
+        [&](bool, value & messages, value &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
             if (!content->stats.used) {
@@ -223,7 +201,6 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        nullptr, // ctx_fn
         [&]() {
             // tools
             return json::array({
@@ -247,7 +224,7 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        [&](context &, bool success, value & messages, value & tools, const std::string &) {
+        [&](bool success, value & messages, value & tools) {
             if (!success) {
                 return; // Nothing can be inferred
             }
@@ -316,7 +293,6 @@ caps caps_get(jinja::program & prog) {
                     },
                 });
             },
-            nullptr, // ctx_fn
             [&]() {
                 // tools
                 return json::array({
@@ -340,7 +316,7 @@ caps caps_get(jinja::program & prog) {
                     },
                 });
             },
-            [&](context &, bool success, value & messages, value & tools, const std::string &) {
+            [&](bool success, value & messages, value & tools) {
                 if (!success) {
                     result.supports_tool_calls = false;
                     result.supports_tools = false;
@@ -418,7 +394,6 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        nullptr, // ctx_fn
         [&]() {
             // tools
             return json::array({
@@ -442,7 +417,7 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        [&](context &, bool success, value & messages, value &, const std::string &) {
+        [&](bool success, value & messages, value & /*tools*/) {
             if (!success) {
                 result.supports_parallel_tool_calls = false;
                 return;
@@ -463,22 +438,11 @@ caps caps_get(jinja::program & prog) {
     JJ_DEBUG("%s\n", ">>> Running capability check: preserve reasoning");
 
     // case: preserve reasoning content in chat history
-    const std::string reasoning_placeholder = "<REASONING_CONTENT_PLACEHOLDER>";
     caps_try_execute(
         prog,
         [&]() {
             // messages
             return json::array({
-                {
-                    {"role", "user"},
-                    {"content", "User message"}
-                },
-                {
-                    {"role", "assistant"},
-                    {"content", "Assistant message"},
-                    // check of reasoning_content deeper in the history, not just the last assistant message
-                    {"reasoning_content", reasoning_placeholder}
-                },
                 {
                     {"role", "user"},
                     {"content", "User message"}
@@ -494,45 +458,19 @@ caps caps_get(jinja::program & prog) {
                 },
             });
         },
-        [&](context & ctx) {
-            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
-            caps_apply_preserve_reasoning(ctx, true);
+        [&]() {
+            // tools
+            return json::array();
         },
-        nullptr, // tools_fn
-        [&](context &, bool, value &, value &, const std::string & output) {
-            // note: we cannot use stats here because the reasoning_content may be used for "if" condition test, but not actually outputted in the final result
-            if (output.find(reasoning_placeholder) != std::string::npos) {
+        [&](bool, value & messages, value &) {
+            auto & content = messages->at(1)->at("reasoning_content");
+            caps_print_stats(content, "messages[1].reasoning_content");
+            if (content->stats.used) {
                 result.supports_preserve_reasoning = true;
             }
         }
     );
 
-    JJ_DEBUG("%s\n", ">>> Running capability check: reasoning effort");
-
-    // case: reasoning effort level
-    caps_try_execute(
-        prog,
-        [&]() {
-            // messages
-            return json::array({
-                {
-                    {"role", "user"},
-                    {"content", "User message"}
-                },
-            });
-        },
-        [&](context & ctx) {
-            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
-            caps_apply_reasoning_effort(ctx, "low");
-        },
-        nullptr, // tools_fn
-        [&](context & ctx, bool, value &, value &, const std::string &) {
-            value effort = ctx.get_val("reasoning_effort");
-            caps_print_stats(effort, "reasoning_effort");
-            result.supports_reasoning_effort = effort->stats.used;
-        }
-    );
-
     JJ_DEBUG("%s\n", result.to_string().c_str());
 
     return result;

common/jinja/caps.h

@@ -12,12 +12,7 @@ struct caps {
     bool supports_tool_calls = true;
     bool supports_system_role = true;
     bool supports_parallel_tool_calls = true;
-
-    // supports preserve reasoning trace in the full history, not just the last assistant message
-    bool supports_preserve_reasoning = false;
-
-    // supports reasoning effort levels
-    bool supports_reasoning_effort = false;
+    bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
 
     // one of the 2 content capabilities must be true
     bool supports_string_content = true;
@@ -34,7 +29,4 @@ struct caps {
 
 caps caps_get(jinja::program & prog);
 
-void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled);
-void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort);
-
 } // namespace jinja

common/jinja/runtime.cpp

@@ -263,7 +263,7 @@ value binary_expression::execute_impl(context & ctx) {
             return res;
         }
         for (int64_t i = 0; i < repeat; ++i) {
-            res->val_str.append(str);
+            res->val_str = res->val_str.append(str);
         }
         return res;
     }
@@ -954,50 +954,4 @@ value keyword_argument_expression::execute_impl(context & ctx) {
     return mk_val<value_kwarg>(k, v);
 }
 
-std::string runtime::debug_dump_program(const program & prog, const std::string & src) {
-    std::ostringstream oss;
-    size_t lvl = 0;
-    context ctx;
-    ctx.src.reset(new std::string(src));
-
-    auto indent = [](size_t lvl) -> std::string {
-        return std::string(lvl * 2, ' ');
-    };
-
-    ctx.visitor = [&](bool is_leaf, statement * node, std::vector<visitor_pair> children) {
-        oss << indent(lvl) << node->type() << ":\n";
-        lvl++;
-        if (is_leaf) {
-            const auto & pos = node->pos;
-            oss << indent(lvl) << "(leaf) at " << get_line_col(src, pos) << " in source:\n";
-            std::string snippet = peak_source(src, pos);
-            string_replace_all(snippet, "\n", "\n" + indent(lvl));
-            oss << indent(lvl) << snippet << "\n";
-        } else {
-            for (auto & [label, children_vec] : children) {
-                oss << indent(lvl) << label << ":\n";
-                lvl++;
-                if (children_vec.empty()) {
-                    oss << indent(lvl) << "<empty>\n\n";
-                } else {
-                    for (auto * child : children_vec) {
-                        if (!child) {
-                            continue;
-                        }
-                        child->visit(ctx);
-                    }
-                }
-                lvl--;
-            }
-        }
-        lvl--;
-    };
-
-    for (const auto & stmt : prog.body) {
-        stmt->visit(ctx);
-    }
-
-    return oss.str();
-}
-
 } // namespace jinja

common/jinja/runtime.h

@@ -47,19 +47,12 @@ const T * cast_stmt(const statement_ptr & ptr) {
 // not thread-safe
 void enable_debug(bool enable);
 
-// for visiting AST nodes
-// function signature: void(bool is_leaf, statement * node, pair of <label, children>)
-using visitor_pair = std::pair<std::string, std::vector<statement *>>;
-using visitor_fn = std::function<void(bool, statement *, std::vector<visitor_pair>)>;
-
 struct context {
     std::shared_ptr<std::string> src; // for debugging; use shared_ptr to avoid copying on scope creation
     std::time_t current_time; // for functions that need current time
 
     bool is_get_stats = false; // whether to collect stats
 
-    visitor_fn visitor;
-
     // src is optional, used for error reporting
     context(std::string src = "") : src(std::make_shared<std::string>(std::move(src))) {
         env = mk_val<value_object>();
@@ -106,15 +99,6 @@ private:
     value_object env;
 };
 
-// utils for visiting AST nodes
-static std::vector<statement *> stmts_to_ptr(const statements & stmts) {
-    std::vector<statement *> children;
-    for (const auto & stmt : stmts) {
-        children.push_back(stmt.get());
-    }
-    return children;
-}
-
 /**
  * Base class for all nodes in the AST.
  */
@@ -122,7 +106,6 @@ struct statement {
     size_t pos; // position in source, for debugging
     virtual ~statement() = default;
     virtual std::string type() const { return "Statement"; }
-    virtual void visit(context & ctx) { ctx.visitor(true, this, {}); }
 
     // execute_impl must be overridden by derived classes
     virtual value execute_impl(context &) { throw_exec_error(); }
@@ -183,13 +166,6 @@ struct if_statement : public statement {
 
     std::string type() const override { return "If"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"test", {test.get()}},
-            {"body", stmts_to_ptr(body)},
-            {"alternate", stmts_to_ptr(alternate)}
-        });
-    }
 };
 
 struct identifier;
@@ -214,14 +190,6 @@ struct for_statement : public statement {
 
     std::string type() const override { return "For"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"loopvar", {loopvar.get()}},
-            {"iterable", {iterable.get()}},
-            {"body", stmts_to_ptr(body)},
-            {"default_block", stmts_to_ptr(default_block)}
-        });
-    }
 };
 
 struct break_statement : public statement {
@@ -273,13 +241,6 @@ struct set_statement : public statement {
 
     std::string type() const override { return "Set"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"assignee", {assignee.get()}},
-            {"value", {val.get()}},
-            {"body", stmts_to_ptr(body)}
-        });
-    }
 };
 
 struct macro_statement : public statement {
@@ -295,13 +256,6 @@ struct macro_statement : public statement {
 
     std::string type() const override { return "Macro"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"name", {name.get()}},
-            {"args", stmts_to_ptr(args)},
-            {"body", stmts_to_ptr(body)}
-        });
-    }
 };
 
 struct comment_statement : public statement {
@@ -335,12 +289,6 @@ struct member_expression : public expression {
     }
     std::string type() const override { return "MemberExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"object", {object.get()}},
-            {"property", {property.get()}}
-        });
-    }
 };
 
 struct call_expression : public expression {
@@ -354,12 +302,6 @@ struct call_expression : public expression {
     }
     std::string type() const override { return "CallExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"callee", {callee.get()}},
-            {"args", stmts_to_ptr(args)}
-        });
-    }
 };
 
 /**
@@ -463,12 +405,6 @@ struct binary_expression : public expression {
     }
     std::string type() const override { return "BinaryExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"left", {left.get()}},
-            {"right", {right.get()}}
-        });
-    }
 };
 
 /**
@@ -495,12 +431,6 @@ struct filter_expression : public expression {
 
     std::string type() const override { return "FilterExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"operand", {operand.get()}},
-            {"filter", {filter.get()}}
-        });
-    }
 };
 
 struct filter_statement : public statement {
@@ -513,12 +443,6 @@ struct filter_statement : public statement {
     }
     std::string type() const override { return "FilterStatement"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"filter", {filter.get()}},
-            {"body", stmts_to_ptr(body)}
-        });
-    }
 };
 
 /**
@@ -544,12 +468,6 @@ struct select_expression : public expression {
         }
         return lhs->execute_impl(ctx);
     }
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"lhs", {lhs.get()}},
-            {"test", {test.get()}}
-        });
-    }
 };
 
 /**
@@ -568,12 +486,6 @@ struct test_expression : public expression {
     }
     std::string type() const override { return "TestExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"operand", {operand.get()}},
-            {"test", {test.get()}}
-        });
-    }
 };
 
 /**
@@ -589,11 +501,6 @@ struct unary_expression : public expression {
     }
     std::string type() const override { return "UnaryExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"argument", {argument.get()}}
-        });
-    }
 };
 
 struct slice_expression : public expression {
@@ -611,13 +518,6 @@ struct slice_expression : public expression {
     [[noreturn]] value execute_impl(context &) override {
         throw std::runtime_error("must be handled by MemberExpression");
     }
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"start_expr", {start_expr.get()}},
-            {"stop_expr", {stop_expr.get()}},
-            {"step_expr", {step_expr.get()}}
-        });
-    }
 };
 
 struct keyword_argument_expression : public expression {
@@ -631,12 +531,6 @@ struct keyword_argument_expression : public expression {
     }
     std::string type() const override { return "KeywordArgumentExpression"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"key", {key.get()}},
-            {"val", {val.get()}}
-        });
-    }
 };
 
 struct spread_expression : public expression {
@@ -645,11 +539,6 @@ struct spread_expression : public expression {
         chk_type<expression>(this->argument);
     }
     std::string type() const override { return "SpreadExpression"; }
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"argument", {argument.get()}}
-        });
-    }
 };
 
 struct call_statement : public statement {
@@ -664,13 +553,6 @@ struct call_statement : public statement {
     }
     std::string type() const override { return "CallStatement"; }
     value execute_impl(context & ctx) override;
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"call", {call.get()}},
-            {"caller_args", stmts_to_ptr(caller_args)},
-            {"body", stmts_to_ptr(body)}
-        });
-    }
 };
 
 struct ternary_expression : public expression {
@@ -693,13 +575,6 @@ struct ternary_expression : public expression {
             return false_expr->execute(ctx);
         }
     }
-    void visit(context & ctx) override {
-        ctx.visitor(false, this, {
-            {"condition", {condition.get()}},
-            {"true_expr", {true_expr.get()}},
-            {"false_expr", {false_expr.get()}}
-        });
-    }
 };
 
 struct raised_exception : public std::exception {
@@ -763,26 +638,16 @@ struct runtime {
         gather_string_parts_recursive(val, parts);
         // join consecutive parts with the same type
         auto & p = parts->val_str.parts;
-        if (p.empty()) {
-            return parts;
-        }
-        size_t w = 0;
-        for (size_t r = 1; r < p.size(); r++) {
-            if (p[w].is_input == p[r].is_input) {
-                p[w].val += p[r].val;
+        for (size_t i = 1; i < p.size(); ) {
+            if (p[i].is_input == p[i - 1].is_input) {
+                p[i - 1].val += p[i].val;
+                p.erase(p.begin() + i);
             } else {
-                w++;
-                if (w != r) {
-                    // the guard is needed, self-move leaves the string in an unspecified state
-                    p[w] = std::move(p[r]);
-                }
+                i++;
             }
         }
-        p.resize(w + 1);
         return parts;
     }
-
-    static std::string debug_dump_program(const program & prog, const std::string & src);
 };
 
 } // namespace jinja

common/jinja/string.cpp

@@ -103,7 +103,7 @@ void string::mark_input_based_on(const string & other) {
     }
 }
 
-string & string::append(const string & other) {
+string string::append(const string & other) {
     for (const auto & part : other.parts) {
         parts.push_back(part);
     }

common/jinja/string.h

@@ -47,7 +47,7 @@ struct string {
     // mark this string as input if other has ALL parts as input
     void mark_input_based_on(const string & other);
 
-    string & append(const string & other);
+    string append(const string & other);
 
     // in-place transformations
 

common/jinja/value.cpp

@@ -3,7 +3,7 @@
 #include "value.h"
 
 // for converting from JSON to jinja values
-#include "json.h"
+#include <nlohmann/json.hpp>
 
 #include <sstream>
 #include <string>
@@ -750,50 +750,11 @@ const func_builtins & value_string_t::get_builtins() const {
             res->val_str.mark_input_based_on(args.get_pos(0)->val_str);
             return res;
         }},
-        {"format", [](const func_args & args) -> value {
-            value val_input = args.get_pos(0);
-            if (!is_val<value_string>(val_input)) {
-                throw raised_exception("format() first argument must be a string");
-            }
-            const jinja::string & fmt = val_input->as_string();
-            const bool fmt_is_input = fmt.all_parts_are_input();
-
-            const std::string str = fmt.str();
-            jinja::string result;
-            std::string literal;
-            auto flush_literal = [&]() {
-                if (!literal.empty()) {
-                    result.parts.push_back({fmt_is_input, literal});
-                    literal.clear();
-                }
-            };
-
-            size_t arg_idx = 1; // positional args follow the format string
-            for (size_t i = 0; i < str.size(); ++i) {
-                if (str[i] != '{') {
-                    literal += str[i];
-                    continue;
-                }
-                if (i + 1 >= str.size() || str[i + 1] != '}') {
-                    throw not_implemented_exception("format() only supports simple '{}' placeholders");
-                }
-                ++i;
-                flush_literal();
-                const jinja::string arg_str = args.get_pos(arg_idx++)->as_string();
-                result.parts.insert(result.parts.end(), arg_str.parts.begin(), arg_str.parts.end());
-            }
-            flush_literal();
-            return mk_val<value_string>(result);
-        }},
         {"int", [](const func_args & args) -> value {
             value val_input   = args.get_pos(0);
             value val_default = args.get_kwarg_or_pos("default", 1);
             value val_base    = args.get_kwarg_or_pos("base",    2);
             const int base = val_base->is_undefined() ? 10 : val_base->as_int();
-            if (base != 0 && (base < 2 || base > 36)) {
-                // an out-of-range base makes std::stoi fail fast on the MSVC CRT instead of throwing
-                throw raised_exception("int() base must be 0 or between 2 and 36");
-            }
             if (is_val<value_string>(val_input) == false) {
                 throw raised_exception("int() first argument must be a string");
             }
@@ -1147,50 +1108,6 @@ const func_builtins & value_array_t::get_builtins() const {
             std::reverse(arr.begin(), arr.end());
             return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
-        {"min", [](const func_args & args) -> value {
-            args.ensure_count(1, 4);
-            args.ensure_vals<value_array>();
-            value val_case    = args.get_kwarg_or_pos("case_sensitive", 1);
-            value attribute   = args.get_kwarg_or_pos("attribute",      2);
-            if (!attribute->is_undefined()) {
-                throw not_implemented_exception("min: attribute not implemented");
-            }
-            // FIXME: min is currently always case sensitive
-            (void) val_case;
-            const auto & arr = args.get_pos(0)->as_array();
-            if (arr.empty()) {
-                return mk_val<value_undefined>();
-            }
-            value result = arr[0];
-            for (size_t i = 1; i < arr.size(); ++i) {
-                if (value_compare(arr[i], result, value_compare_op::lt)) {
-                    result = arr[i];
-                }
-            }
-            return result;
-        }},
-        {"max", [](const func_args & args) -> value {
-            args.ensure_count(1, 4);
-            args.ensure_vals<value_array>();
-            value val_case    = args.get_kwarg_or_pos("case_sensitive", 1);
-            value attribute   = args.get_kwarg_or_pos("attribute",      2);
-            if (!attribute->is_undefined()) {
-                throw not_implemented_exception("max: attribute not implemented");
-            }
-            // FIXME: max is currently always case sensitive
-            (void) val_case;
-            const auto & arr = args.get_pos(0)->as_array();
-            if (arr.empty()) {
-                return mk_val<value_undefined>();
-            }
-            value result = arr[0];
-            for (size_t i = 1; i < arr.size(); ++i) {
-                if (value_compare(arr[i], result, value_compare_op::gt)) {
-                    result = arr[i];
-                }
-            }
-            return result;
-        }},
         {"unique", array_unique_not_implemented},
     };
     return builtins;
@@ -1355,7 +1272,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const common_json & j, bool mark_input) {
+static value from_json(const nlohmann::ordered_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1452,7 +1369,7 @@ bool value_compare(const value & a, const value & b, value_compare_op op) {
 }
 
 template<>
-void global_from_json(context & ctx, const common_json & json_obj, bool mark_input) {
+void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bool mark_input) {
     // printf("global_from_json: %s\n" , json_obj.dump(2).c_str());
     if (json_obj.is_null() || !json_obj.is_object()) {
         throw std::runtime_error("global_from_json: input JSON value must be an object");

common/jinja/value.h

@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be common_json
+// Note: T_JSON can be nlohmann::ordered_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 

common/json-partial.cpp

@@ -0,0 +1,324 @@
+#include "json-partial.h"
+
+#include "log.h"
+
+#include <nlohmann/json.hpp>
+
+#include <string>
+#include <regex>
+
+using json = nlohmann::ordered_json;
+
+enum common_json_stack_element_type {
+    COMMON_JSON_STACK_ELEMENT_OBJECT,
+    COMMON_JSON_STACK_ELEMENT_KEY,
+    COMMON_JSON_STACK_ELEMENT_ARRAY,
+};
+
+struct common_json_stack_element {
+    common_json_stack_element_type type;
+    std::string key;
+};
+
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    std::string::const_iterator it = input.begin();
+    const auto end = input.end();
+    return common_json_parse(it, end, healing_marker, out);
+}
+
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    // // https://json.nlohmann.me/features/parsing/sax_interface/
+    struct json_error_locator : public nlohmann::json_sax<json> {
+        std::size_t position;
+        bool found_error;
+        std::string last_token;
+        std::string exception_message;
+        std::vector<common_json_stack_element> stack;
+
+        json_error_locator() : position(0), found_error(false) {}
+
+        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
+            this->position = position - 1;
+            this->found_error = true;
+            this->last_token = last_token;
+            this->exception_message = ex.what();
+            return false;
+        }
+        void close_value() {
+            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
+                stack.pop_back();
+            }
+        }
+        bool null() override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool boolean(bool) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_integer(number_integer_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_unsigned(number_unsigned_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_float(number_float_t, const string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool string(string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool binary(binary_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool start_object(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
+            return true;
+        }
+        bool end_object() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+        bool key(string_t & key) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
+            return true;
+        }
+        bool start_array(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
+            return true;
+        }
+        bool end_array() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+    };
+    json_error_locator err_loc;
+    auto start = it;
+    json::sax_parse(it, end, &err_loc);
+
+    if (err_loc.found_error) {
+        it = start;
+        auto temptative_end = it + err_loc.position;
+        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
+
+        auto input = std::string(it, temptative_end);
+        try {
+            out.json = json::parse(input);
+            // out.json = json::parse(it, temptative_end);
+            it = temptative_end;
+            return true;
+        } catch (const std::exception & ex) {
+            // No, needs healing.
+            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
+        }
+        auto can_parse = [](const std::string & str) {
+            try {
+                auto _ = json::parse(str); // NOLINT
+                return true;
+            } catch (const std::exception &) {
+                return false;
+            }
+        };
+        if (!healing_marker.empty() && !err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
+            if (last_non_sp_pos == std::string::npos) {
+                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+            }
+            auto last_non_sp_char = str[last_non_sp_pos];
+            // Used to detect stops on a number, which may not be complete.
+            auto was_maybe_number = [&]() {
+                if (!str.empty() && std::isspace(str.back())) {
+                    return false;
+                }
+                return std::isdigit(last_non_sp_char) ||
+                    last_non_sp_char == '.' ||
+                    last_non_sp_char == 'e' ||
+                    last_non_sp_char == 'E' ||
+                    last_non_sp_char == '-';
+            };
+
+            std::string closing;
+            for (size_t i = err_loc.stack.size(); i > 0; i--) {
+                auto & el = err_loc.stack[i - 1];
+                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                    closing += "}";
+                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                    closing += "]";
+                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
+                    throw std::runtime_error("Unexpected stack element type");
+                }
+            }
+
+            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
+            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
+
+            auto is_high_surrogate = [&](const std::string & s) {
+                // Check if a partial of a high surrogate (U+D800-U+DBFF)
+                return s.length() >= 4 &&
+                    s[0] == '\\' && s[1] == 'u' &&
+                    std::tolower(s[2]) == 'd' &&
+                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
+            };
+
+            // Initialize the unicode marker to a low surrogate to handle the edge case
+            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
+            // backslash (\)
+            std::string unicode_marker_padding = "udc00";
+            std::smatch last_unicode_seq;
+
+            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
+                std::smatch second_last_seq;
+                std::string prelude = str.substr(0, last_unicode_seq.position());
+
+                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
+                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
+
+                if (is_high_surrogate(last_unicode_seq.str())) {
+                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
+                    unicode_marker_padding += "\\udc00";
+                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
+                    if (is_high_surrogate(second_last_seq.str())) {
+                        // If this follows a high surrogate, pad it to be a low surrogate
+                        if (last_unicode_seq.length() == 2) {
+                            unicode_marker_padding = "dc00";
+                        } else if (last_unicode_seq.length() == 3) {
+                            unicode_marker_padding = "c00";
+                        } else {
+                            // The original unicode_marker_padding is already padded with 0s
+                        }
+                    }
+                }
+            }
+
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
+
+            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
+                // We're inside an object value
+                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
+                    // Was about to create an object value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + ": 1" + closing)) {
+                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
+                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
+                    // Was about to create an object
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an object value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an object value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an object value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
+                } else {
+                    // find last :
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+                    }
+                    // Cutting back to opening : for object value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
+                    // Was about to create an array value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an array value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an array value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an array value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
+                    // Had just finished a value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
+                } else {
+                    auto last_pos = str.find_last_of("[,");
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
+                    }
+                    // Cutting back to last [ or , for array value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
+                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\": 1" + closing)) {
+                    // Was inside an object key string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
+                    // Was inside an object key string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
+                    // Was inside an object key string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
+                } else {
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+                    }
+                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else {
+                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+            }
+            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
+        // handle unclosed top-level primitive
+        if (err_loc.position != 0 && !healing_marker.empty() && err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;
+            if (can_parse(str + "\"")) {
+                // Was inside an string
+                str += (out.healing_marker.json_dump_marker = magic_seed) + "\"";
+            } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"")) {
+                // Was inside an string after an escape
+                str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"";
+            } else {
+                // TODO: handle more unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
+                // fprintf(stderr, "Closing: TODO\n");
+                return false;
+            }
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
+        return false;
+    }
+    out.json = json::parse(it, end);
+    it = end;
+    return true;
+}

common/json-partial.h

@@ -0,0 +1,39 @@
+#pragma once
+
+// TODO: use json_fwd.hpp when possible
+#include <nlohmann/json.hpp>
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

common/json-schema-to-grammar.cpp

@@ -1,8 +1,9 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
-#include <limits>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -11,7 +12,7 @@
 #include <unordered_set>
 #include <vector>
 
-using json = common_json;
+using json = nlohmann::ordered_json;
 
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
@@ -277,9 +278,7 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
     {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
-static const int MAX_PATTERN_DEPTH = 100;
-
-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -310,32 +309,6 @@ static std::string format_literal(const std::string & literal) {
 
 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
 
-static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
-    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
-        return 0;
-    }
-    size_t n_hex = 0;
-    switch (pattern[pos + 1]) {
-        case 'x': n_hex = 2; break;
-        case 'u': n_hex = 4; break;
-        case 'U': n_hex = 8; break;
-        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
-            return 2;
-        default:
-            return 0;
-    }
-    if (pos + 2 + n_hex > pattern.length()) {
-        return 0;
-    }
-    for (size_t i = pos + 2; i < pos + 2 + n_hex; i++) {
-        char h = pattern[i];
-        if (!((h >= '0' && h <= '9') || (h >= 'a' && h <= 'f') || (h >= 'A' && h <= 'F'))) {
-            return 0;
-        }
-    }
-    return 2 + n_hex;
-}
-
 class common_schema_converter {
 private:
     friend class common_schema_info;
@@ -372,42 +345,16 @@ private:
         return string_join(rules, " | ");
     }
 
-    // thrown when the pattern is a valid regex with no grammar equivalent
-    struct unsupported_pattern : public std::runtime_error {
-        using std::runtime_error::runtime_error;
-    };
-
-    // thrown when the pattern is not a valid regex
-    struct invalid_pattern : public std::runtime_error {
-        using std::runtime_error::runtime_error;
-    };
-
     std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        auto rules_snapshot = _rules;
-        try {
-            return _pattern_to_rule(pattern, name);
-        } catch (const unsupported_pattern & err) {
-            // revert rules
-            _rules = std::move(rules_snapshot);
-            _warnings.push_back("pattern " + pattern + " is not supported (" + err.what() + "), accepting any string");
-            return _add_rule(name, _add_primitive("string", PRIMITIVE_RULES.at("string")));
-        } catch (const invalid_pattern & err) {
-            _rules = std::move(rules_snapshot);
-            _errors.push_back("Invalid pattern " + pattern + ": " + err.what());
+        if (!(pattern.front() == '^' && pattern.back() == '$')) {
+            _errors.push_back("Pattern must start with '^' and end with '$'");
             return "";
         }
-    }
-
-    std::string _pattern_to_rule(const std::string & pattern, const std::string & name) {
-        if (pattern.length() < 2 || pattern.front() != '^' || pattern.back() != '$') {
-            throw unsupported_pattern("not anchored with '^' and '$'");
-        }
         std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
         std::unordered_map<std::string, std::string> sub_rule_ids;
 
         size_t i = 0;
         size_t length = sub_pattern.length();
-        int paren_depth = 0;
 
         using literal_or_rule = std::pair<std::string, bool>;
         auto to_rule = [&](const literal_or_rule & ls) {
@@ -416,6 +363,7 @@ private:
             return is_literal ? "\"" + s + "\"" : s;
         };
         std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
+            size_t start = i;
             std::vector<literal_or_rule> seq;
 
             auto get_dot = [&]() {
@@ -472,42 +420,43 @@ private:
                         if (i + 1 < length && sub_pattern[i + 1] == ':') {
                             i += 2; // skip "?:" for non-capturing group, treat as regular group
                         } else {
-                            // lookaround, named group, inline flags, ...
-                            throw unsupported_pattern("unsupported group syntax");
+                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
+                            _warnings.push_back("Unsupported pattern syntax");
+                            // skip to matching ')' to avoid UB on empty seq
+                            int depth = 1;
+                            while (i < length && depth > 0) {
+                                if (sub_pattern[i] == '\\' && i + 1 < length) {
+                                    i += 2; // skip escaped character
+                                } else {
+                                    if (sub_pattern[i] == '(') depth++;
+                                    else if (sub_pattern[i] == ')') depth--;
+                                    i++;
+                                }
+                            }
+                            continue;
                         }
                     }
-                    paren_depth++;
-                    if (paren_depth > MAX_PATTERN_DEPTH) {
-                        throw unsupported_pattern("pattern nesting too deep");
-                    }
                     seq.emplace_back("(" + to_rule(transform()) + ")", false);
                 } else if (c == ')') {
                     i++;
-                    if (paren_depth == 0) {
-                        throw invalid_pattern("unbalanced parentheses");
+                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
+                        _errors.push_back("Unbalanced parentheses");
                     }
-                    paren_depth--;
                     return join_seq();
-                } else if (c == '^' || c == '$') {
-                    throw unsupported_pattern("anchor inside the pattern");
                 } else if (c == '[') {
                     std::string square_brackets = std::string(1, c);
                     i++;
                     while (i < length && sub_pattern[i] != ']') {
                         if (sub_pattern[i] == '\\') {
-                            auto escape_length = gbnf_escape_length(sub_pattern, i);
-                            if (escape_length == 0) {
-                                throw unsupported_pattern("unsupported escape in character class: " + sub_pattern.substr(i, 2));
-                            }
-                            square_brackets += sub_pattern.substr(i, escape_length);
-                            i += escape_length;
+                            square_brackets += sub_pattern.substr(i, 2);
+                            i += 2;
                         } else {
                             square_brackets += sub_pattern[i];
                             i++;
                         }
                     }
                     if (i >= length) {
-                        throw invalid_pattern("unterminated character class");
+                        _errors.push_back("Unbalanced square brackets");
                     }
                     square_brackets += ']';
                     i++;
@@ -516,9 +465,6 @@ private:
                     seq.emplace_back("|", false);
                     i++;
                 } else if (c == '*' || c == '+' || c == '?') {
-                    if (seq.empty()) {
-                        throw invalid_pattern("nothing to repeat");
-                    }
                     seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
                     i++;
                 } else if (c == '{') {
@@ -529,19 +475,18 @@ private:
                         i++;
                     }
                     if (i >= length) {
-                        throw unsupported_pattern("unterminated curly brackets");
+                        _errors.push_back("Unbalanced curly brackets");
                     }
                     curly_brackets += '}';
                     i++;
                     auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                     int min_times = 0;
                     int max_times = std::numeric_limits<int>::max();
-                    if (nums.size() != 1 && nums.size() != 2) {
-                        throw unsupported_pattern("wrong number of values in curly brackets");
-                    }
                     try {
                         if (nums.size() == 1) {
                             min_times = max_times = std::stoi(nums[0]);
+                        } else if (nums.size() != 2) {
+                            _errors.push_back("Wrong number of values in curly brackets");
                         } else {
                             if (!nums[0].empty()) {
                                 min_times = std::stoi(nums[0]);
@@ -550,11 +495,9 @@ private:
                                 max_times = std::stoi(nums[1]);
                             }
                         }
-                    } catch (const std::logic_error &) {
-                        throw unsupported_pattern("invalid number in curly brackets");
-                    }
-                    if (seq.empty()) {
-                        throw invalid_pattern("nothing to repeat");
+                    } catch (const std::invalid_argument & e) {
+                        _errors.push_back("Invalid number in curly brackets");
+                        return std::make_pair("", false);
                     }
                     auto &last = seq.back();
                     auto &sub = last.first;
@@ -580,22 +523,15 @@ private:
                         return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                     };
                     while (i < length) {
-                        if (sub_pattern[i] == '\\') {
-                            if (i == length - 1) {
-                                throw invalid_pattern("trailing backslash");
-                            }
+                        if (sub_pattern[i] == '\\' && i < length - 1) {
                             char next = sub_pattern[i + 1];
                             if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
                                 i++;
                                 literal += sub_pattern[i];
                                 i++;
                             } else {
-                                auto escape_length = gbnf_escape_length(sub_pattern, i);
-                                if (escape_length == 0) {
-                                    throw unsupported_pattern("unsupported escape: " + sub_pattern.substr(i, 2));
-                                }
-                                literal += sub_pattern.substr(i, escape_length);
-                                i += escape_length;
+                                literal += sub_pattern.substr(i, 2);
+                                i += 2;
                             }
                         } else if (sub_pattern[i] == '"') {
                             literal += "\\\"";
@@ -608,21 +544,14 @@ private:
                             break;
                         }
                     }
-                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
-                        throw unsupported_pattern(std::string("unsupported character: ") + c);
+                    if (!literal.empty()) {
+                        seq.emplace_back(literal, true);
                     }
-                    seq.emplace_back(literal, true);
                 }
             }
             return join_seq();
         };
-
-        auto rule = to_rule(transform());
-        if (paren_depth != 0) {
-            throw invalid_pattern("unbalanced parentheses");
-        }
-
-        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
+        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
     }
 
     /*
@@ -916,11 +845,7 @@ public:
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
         }
         if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
-            std::vector<json> alt_schemas;
-            for (const auto & alt : alts) {
-                alt_schemas.push_back(alt);
-            }
+            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
         }
         if (schema_type.is_array()) {
@@ -1114,7 +1039,7 @@ common_schema_info::~common_schema_info() = default;
 common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
 common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
 
-void common_schema_info::resolve_refs(common_json & schema) {
+void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
     impl_->resolve_refs(schema, "");
 }
 
@@ -1122,7 +1047,7 @@ void common_schema_info::resolve_refs(common_json & schema) {
 // Some models emit raw string values rather than JSON-encoded strings for string parameters.
 // If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
 // true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const common_json & schema) {
+bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
     std::unordered_set<std::string> visited_refs;
 
     std::function<bool(const json &)> check = [&](const json & s) -> bool {
@@ -1230,7 +1155,7 @@ bool common_schema_info::resolves_to_string(const common_json & schema) {
     return check(schema);
 }
 
-std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
+std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
         return "%llguidance {}\nstart: %json " + schema.dump();
@@ -1251,10 +1176,10 @@ std::string build_grammar(const std::function<void(const common_grammar_builder
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
+        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
             return converter.visit(schema, name == "root" ? "" : name);
         },
-        /* .resolve_refs = */ [&](common_json & schema) {
+        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
             converter.resolve_refs(schema, "");
         }
     };

common/json-schema-to-grammar.h

@@ -1,12 +1,12 @@
 #pragma once
 
-#include "json.h"
+#include <nlohmann/json_fwd.hpp>
 
 #include <functional>
 #include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const common_json & schema,
+std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);
 
 class common_schema_converter;
@@ -24,14 +24,14 @@ class common_schema_info {
     common_schema_info(common_schema_info &&) noexcept;
     common_schema_info & operator=(common_schema_info &&) noexcept;
 
-    void resolve_refs(common_json & schema);
-    bool resolves_to_string(const common_json & schema);
+    void resolve_refs(nlohmann::ordered_json & schema);
+    bool resolves_to_string(const nlohmann::ordered_json & schema);
 };
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const common_json &)> add_schema;
-    std::function<void(common_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
+    std::function<void(nlohmann::ordered_json &)> resolve_refs;
 };
 
 struct common_grammar_options {

common/json.cpp

@@ -1,437 +0,0 @@
-#include "json.h"
-
-#include "ggml.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
-
-#include <iterator>
-#include <new>
-#include <set>
-#include <unordered_map>
-#include <vector>
-
-using nlohmann::ordered_json;
-
-// a common_json is the backing value, so any value of a tree can be used as a common_json
-static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
-static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
-
-// runs fn and gives every error of the backing library as a common_json_error
-template <typename F>
-static decltype(auto) guard(F && fn) {
-    try {
-        return fn();
-    } catch (const ordered_json::exception & e) {
-        throw common_json_error(e.what());
-    }
-}
-
-static ordered_json & as_json(common_json * self) {
-    return *reinterpret_cast<ordered_json *>(self);
-}
-
-static const ordered_json & as_json(const common_json * self) {
-    return *reinterpret_cast<const ordered_json *>(self);
-}
-
-static common_json & as_common(ordered_json & json) {
-    return *reinterpret_cast<common_json *>(&json);
-}
-
-static const common_json & as_common(const ordered_json & json) {
-    return *reinterpret_cast<const common_json *>(&json);
-}
-
-static ordered_json to_json(const common_json_value & val) {
-    switch (val.type) {
-        case common_json_value::VAL_NULL:   return nullptr;
-        case common_json_value::VAL_BOOL:   return val.val_bool;
-        case common_json_value::VAL_INT:    return val.val_int;
-        case common_json_value::VAL_UINT:   return val.val_uint;
-        case common_json_value::VAL_DOUBLE: return val.val_double;
-        case common_json_value::VAL_STRING: return val.val_string;
-        case common_json_value::VAL_JSON:
-            // one owner means no one else can see this tree, so it is safe to move it out
-            // note: this makes a value single use, same as the json_ref of the backing library
-            if (val.val_json.use_count() == 1) {
-                return std::move(as_json(val.val_json.get()));
-            }
-            return as_json(val.val_json.get());
-    }
-
-    return nullptr;
-}
-
-common_json_value::common_json_value(const char * val) {
-    if (val) {
-        type       = VAL_STRING;
-        val_string = val;
-    } else {
-        type = VAL_NULL;
-    }
-}
-
-common_json_value::common_json_value(const common_json & val) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
-
-common_json_value::common_json_value(common_json && val) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
-
-template <typename T>
-common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::array();
-
-    for (const auto & val : vals) {
-        out.push_back(val);
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
-
-COMMON_JSON_SET(int)
-COMMON_JSON_SET(std::string)
-
-#undef COMMON_JSON_SET
-
-template <typename T>
-common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
-
-COMMON_JSON_MAP(bool)
-COMMON_JSON_MAP(std::string)
-
-#undef COMMON_JSON_MAP
-
-template <typename T>
-common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
-
-COMMON_JSON_UMAP(size_t)
-
-#undef COMMON_JSON_UMAP
-
-template <typename T>
-common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::array();
-
-    for (const auto & val : vals) {
-        out.push_back(val);
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
-// a vector value is usable only for the types below
-// note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
-
-COMMON_JSON_VEC(int)
-COMMON_JSON_VEC(unsigned char)
-COMMON_JSON_VEC(unsigned int)
-COMMON_JSON_VEC(long)
-COMMON_JSON_VEC(unsigned long)
-COMMON_JSON_VEC(long long)
-COMMON_JSON_VEC(unsigned long long)
-COMMON_JSON_VEC(float)
-COMMON_JSON_VEC(double)
-COMMON_JSON_VEC(std::string)
-COMMON_JSON_VEC(std::vector<float>)
-COMMON_JSON_VEC(common_json)
-
-#undef COMMON_JSON_VEC
-
-common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
-
-// null, same as the backing library
-// operator[] turns it into an object, push_back() into an array
-common_json::common_json() {
-    new (storage) ordered_json();
-}
-
-common_json::common_json(const common_json & other) {
-    new (storage) ordered_json(as_json(&other));
-}
-
-common_json::common_json(common_json && other) noexcept {
-    new (storage) ordered_json(std::move(as_json(&other)));
-}
-
-common_json::common_json(std::initializer_list<common_json_item> items) {
-    new (storage) ordered_json(ordered_json::object());
-
-    for (const auto & item : items) {
-        set(item);
-    }
-}
-
-common_json::common_json(const common_json_value & val) {
-    new (storage) ordered_json(to_json(val));
-}
-
-common_json::common_json(std::nullptr_t) {
-    new (storage) ordered_json(nullptr);
-}
-
-common_json & common_json::operator=(common_json other) noexcept {
-    as_json(this).swap(as_json(&other));
-
-    return *this;
-}
-
-common_json::~common_json() {
-    as_json(this).~basic_json();
-}
-
-common_json common_json::parse(const std::string & text) {
-    try {
-        // the assignment moves the parsed tree in, it does not copy
-        common_json out;
-        as_json(&out) = ordered_json::parse(text);
-        return out;
-    } catch (const std::exception & e) {
-        throw common_json_error(e.what());
-    }
-}
-
-common_json common_json::parse_no_throw(const std::string & text) {
-    common_json out;
-    as_json(&out) = ordered_json::parse(text, nullptr, false);
-    return out;
-}
-
-bool common_json::is_discarded() const {
-    return as_json(this).is_discarded();
-}
-
-common_json common_json::array() {
-    common_json out;
-    as_json(&out) = ordered_json::array();
-    return out;
-}
-
-common_json common_json::array(std::initializer_list<common_json_value> vals) {
-    common_json out;
-    ordered_json & arr = as_json(&out);
-    arr = ordered_json::array();
-
-    for (const auto & val : vals) {
-        arr.push_back(to_json(val));
-    }
-
-    return out;
-}
-
-common_json common_json::object() {
-    common_json out;
-    as_json(&out) = ordered_json::object();
-    return out;
-}
-
-common_json common_json::object(std::initializer_list<common_json_item> items) {
-    return common_json(items);
-}
-
-common_json common_json::make(const common_json_value & val) {
-    return common_json(val);
-}
-
-bool common_json::is_null()           const { return as_json(this).is_null(); }
-bool common_json::is_object()         const { return as_json(this).is_object(); }
-bool common_json::is_array()          const { return as_json(this).is_array(); }
-bool common_json::is_string()         const { return as_json(this).is_string(); }
-bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
-bool common_json::is_number()         const { return as_json(this).is_number(); }
-bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
-bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
-
-bool   common_json::empty() const { return as_json(this).empty(); }
-size_t common_json::size()  const { return as_json(this).size(); }
-
-bool common_json::contains(const std::string & key) const {
-    return as_json(this).contains(key);
-}
-
-bool common_json::operator==(const common_json_value & val) const {
-    // compare a tree in place, to_json() would copy it
-    if (val.type == common_json_value::VAL_JSON) {
-        return as_json(this) == as_json(val.val_json.get());
-    }
-    return as_json(this) == to_json(val);
-}
-
-bool common_json::operator!=(const common_json_value & val) const {
-    return !(*this == val);
-}
-
-common_json       & common_json::at(const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this).at(key)); }); }
-const common_json & common_json::at(const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
-common_json       & common_json::at(size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this).at(idx)); }); }
-const common_json & common_json::at(size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
-
-common_json       & common_json::operator[](const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this)[key]); }); }
-const common_json & common_json::operator[](const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
-common_json       & common_json::operator[](size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this)[idx]); }); }
-const common_json & common_json::operator[](size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
-
-common_json       & common_json::front()       { return as_common(as_json(this).front()); }
-const common_json & common_json::front() const { return as_common(as_json(this).front()); }
-common_json       & common_json::back()        { return as_common(as_json(this).back()); }
-const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
-
-void common_json::clear() {
-    as_json(this).clear();
-}
-
-void common_json::erase(const std::string & key) {
-    guard([&] { as_json(this).erase(key); });
-}
-
-void common_json::erase(size_t idx) {
-    guard([&] { as_json(this).erase(idx); });
-}
-
-void common_json::assign(const common_json_value & val) {
-    as_json(this) = to_json(val);
-}
-
-void common_json::set(const common_json_item & item) {
-    guard([&] { as_json(this)[item.key] = to_json(item.val); });
-}
-
-void common_json::push_back(const common_json_value & val) {
-    guard([&] { as_json(this).push_back(to_json(val)); });
-}
-
-void common_json::push_back(std::initializer_list<common_json_item> items) {
-    common_json val(items);
-
-    guard([&] { as_json(this).push_back(std::move(as_json(&val))); });
-}
-
-size_t common_json::count(const std::string & key) const {
-    return as_json(this).count(key);
-}
-
-void common_json::insert(const common_json & vals) {
-    guard([&] {
-        ordered_json & self = as_json(this);
-
-        self.insert(self.end(), as_json(&vals).begin(), as_json(&vals).end());
-    });
-}
-
-std::string common_json::dump(int indent) const {
-    return guard([&] { return as_json(this).dump(indent); });
-}
-
-std::string common_json::dump_safe(int indent) const {
-    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
-}
-
-// an array is indexed directly, an object needs a walk from the start
-common_json & common_json::iterator::operator*() const {
-    return guard([&]() -> common_json & {
-        ordered_json & j = as_json(node);
-
-        if (j.is_object()) {
-            return as_common(std::next(j.begin(), idx).value());
-        }
-        if (j.is_array()) {
-            return as_common(j[idx]);
-        }
-
-        // a plain value gives itself once, same as the backing library
-        return *node;
-    });
-}
-
-std::string common_json::iterator::key() const {
-    return guard([&] { return std::next(as_json(node).begin(), idx).key(); });
-}
-
-common_json::iterator common_json::begin() const {
-    return iterator(const_cast<common_json *>(this), 0);
-}
-
-common_json::iterator common_json::end() const {
-    return iterator(const_cast<common_json *>(this), size());
-}
-
-// the keys follow the backing library: the index for an array, "" for a plain value
-common_json::items_view::entry common_json::items_view::iterator::operator*() const {
-    return guard([&]() -> entry {
-        ordered_json & j = as_json(node);
-
-        if (j.is_object()) {
-            auto it = std::next(j.begin(), idx);
-
-            return { it.key(), as_common(it.value()) };
-        }
-        if (j.is_array()) {
-            return { std::to_string(idx), as_common(j[idx]) };
-        }
-
-        return { std::string(), *node };
-    });
-}
-
-common_json::items_view common_json::items() const {
-    return items_view(const_cast<common_json *>(this), size());
-}
-
-template <typename T> T common_json::get() const {
-    return guard([&] { return as_json(this).get<T>(); });
-}
-
-// the backing library cannot build a common_json, so this one is just a copy
-template <> common_json common_json::get<common_json>() const {
-    return *this;
-}
-
-// get<T>() is usable only for the types below
-
-#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
-
-COMMON_JSON_GET(bool)
-COMMON_JSON_GET(int)
-COMMON_JSON_GET(unsigned int)
-COMMON_JSON_GET(long)
-COMMON_JSON_GET(unsigned long)
-COMMON_JSON_GET(long long)
-COMMON_JSON_GET(unsigned long long)
-COMMON_JSON_GET(float)
-COMMON_JSON_GET(double)
-COMMON_JSON_GET(std::string)
-COMMON_JSON_GET(std::vector<float>)
-COMMON_JSON_GET(std::vector<std::string>)
-COMMON_JSON_GET(std::set<std::string>)
-COMMON_JSON_GET(std::vector<int>)
-COMMON_JSON_GET(std::vector<size_t>)
-COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
-
-#undef COMMON_JSON_GET

common/json.h

@@ -1,354 +0,0 @@
-#pragma once
-
-#include <cstddef>
-#include <cstdint>
-#include <initializer_list>
-#include <iterator>
-#include <map>
-#include <memory>
-#include <set>
-#include <stdexcept>
-#include <string>
-#include <string_view>
-#include <type_traits>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-// common_json, a thin wrapper around vendor json library
-// the underlay library is pimpl, we are using nlohmann::json for now
-//
-// many features of the library are deliberately left out, to keep this interface small and generic and to keep compile time down
-//
-// some main differences compared to nlohmann::json :
-// - object keys keep the order in which they are added
-// - errors are always throw as common_json_error
-// - obj.push_back({key, val}) is intentionally unsupported to avoid confusion with push_back on a vector; write it as obj[key] = val for clarity
-// - a braced pair in value position does not build, e.g. {"key", {"a", "b"}}; write array({"a", "b"}) where nlohmann made an array
-//
-// in doubt, search the code base for an existing usage example; do not add anything to this header unless absolutely necessary
-
-class common_json;
-
-// common_json_value holds a list of these, and each of them holds a value, so one must come first
-struct common_json_item;
-
-struct common_json_error : std::runtime_error {
-    using std::runtime_error::runtime_error;
-};
-
-// one value, tagged so that this header stays free of the backing library
-// note: a value that holds a tree is single use, the second use gives null
-struct common_json_value {
-    enum value_type {
-        VAL_NULL,
-        VAL_BOOL,
-        VAL_INT,
-        VAL_UINT,
-        VAL_DOUBLE,
-        VAL_STRING,
-        VAL_JSON,
-    };
-
-    value_type type = VAL_NULL;
-
-    union {
-        bool     val_bool;
-        int64_t  val_int;
-        uint64_t val_uint = 0;
-        double   val_double;
-    };
-
-    std::string                  val_string;
-    std::shared_ptr<common_json> val_json;
-
-    common_json_value(std::nullptr_t = nullptr) : type(VAL_NULL) {}
-    common_json_value(bool val) : type(VAL_BOOL), val_bool(val) {}
-    common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
-    // without this a string_view lands on the common_json ctor below and recurses
-    common_json_value(std::string_view val) : type(VAL_STRING), val_string(val) {}
-    common_json_value(const char * val);
-    common_json_value(const common_json & val);
-    common_json_value(common_json && val);
-    // only for the types instantiated in json.cpp, the rest fails at link time
-    template <typename T> common_json_value(const std::vector<T> & vals);
-    // a set becomes an array, in the set's own order
-    template <typename T> common_json_value(const std::set<T> & vals);
-    // a map becomes an object, keyed in the map's own order
-    template <typename T> common_json_value(const std::map<std::string, T> & vals);
-    template <typename T> common_json_value(const std::unordered_map<std::string, T> & vals);
-
-    // nested object, e.g. {"fn", {{"name", "x"}}}
-    // note: a nested pair {"a", "b"} does not build, use common_json::array({"a", "b"}) for an array
-    common_json_value(std::initializer_list<common_json_item> items);
-
-    template <typename T, typename std::enable_if<std::is_integral<T>::value && !std::is_same<T, bool>::value, int>::type = 0>
-    common_json_value(T val) : type(std::is_signed<T>::value ? VAL_INT : VAL_UINT) {
-        if (std::is_signed<T>::value) {
-            val_int = (int64_t) val;
-        } else {
-            val_uint = (uint64_t) val;
-        }
-    }
-
-    template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
-    common_json_value(T val) : type(VAL_DOUBLE), val_double((double) val) {}
-};
-
-struct common_json_item {
-    std::string       key;
-    common_json_value val;
-
-    template <typename T>
-    common_json_item(std::string key, T && val) :
-        key(std::move(key)), val(std::forward<T>(val)) {}
-
-    // a braced list cannot deduce T, so it needs its own overload
-    common_json_item(std::string key, std::initializer_list<common_json_item> items) :
-        key(std::move(key)), val(items) {}
-};
-
-// the types common_json_value holds on its own
-// anything else reaches its common_json ctor and recurses forever
-template <typename T> struct common_json_is_value : std::integral_constant<bool,
-    std::is_arithmetic<T>::value ||
-    std::is_same<T, std::nullptr_t>::value ||
-    std::is_same<T, std::string>::value ||
-    std::is_same<T, std::string_view>::value ||
-    std::is_same<T, char *>::value ||
-    std::is_same<T, const char *>::value ||
-    std::is_same<T, common_json>::value> {};
-
-template <typename T, typename A>
-struct common_json_is_value<std::vector<T, A>> : std::true_type {};
-
-template <typename T, typename C, typename A>
-struct common_json_is_value<std::set<T, C, A>> : std::true_type {};
-
-template <typename V, typename C, typename A>
-struct common_json_is_value<std::map<std::string, V, C, A>> : std::true_type {};
-
-template <typename V, typename H, typename E, typename A>
-struct common_json_is_value<std::unordered_map<std::string, V, H, E, A>> : std::true_type {};
-
-class common_json {
-  public:
-    common_json();
-    common_json(const common_json & other);
-    common_json(common_json && other) noexcept;
-    common_json(std::initializer_list<common_json_item> items);
-    common_json(const common_json_value & val);
-
-    // direct, a value would need two conversions in a row
-    common_json(std::nullptr_t);
-
-    // one step, so that "abc" or a vector can go straight into a common_json
-    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value &&
-                                                  !std::is_same<typename std::decay<T>::type, common_json_value>::value, int>::type = 0>
-    common_json(T && val) : common_json(common_json_value(std::forward<T>(val))) {
-        static_assert(common_json_is_value<typename std::decay<T>::type>::value,
-                      "no common_json_value ctor holds this type, add one instead of letting it recurse");
-    }
-
-    // by value, same as the backing library
-    // the right side is copied before the left side can invalidate it, e.g. msg["a"] = msg.at("b")
-    common_json & operator=(common_json other) noexcept;
-
-    ~common_json();
-
-    // throws common_json_error if the text is not valid JSON
-    static common_json parse(const std::string & text);
-
-    // gives a discarded value instead of throwing, check it with is_discarded()
-    static common_json parse_no_throw(const std::string & text);
-
-    bool is_discarded() const;
-
-    static common_json array();
-    static common_json array(std::initializer_list<common_json_value> vals);
-    static common_json object();
-    static common_json object(std::initializer_list<common_json_item> items);
-
-    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
-    static common_json make(const common_json_value & val);
-
-    bool is_null()    const;
-    bool is_object()  const;
-    bool is_array()   const;
-    bool is_string()  const;
-    bool is_boolean() const;
-    bool is_number()  const;
-    bool is_number_integer() const;
-    bool is_number_float()   const;
-
-    bool   empty() const;
-    size_t size()  const;
-
-    bool contains(const std::string & key) const;
-
-    bool operator==(const common_json_value & val) const;
-    bool operator!=(const common_json_value & val) const;
-
-    // at() throws common_json_error if the key is missing, operator[] adds a null value instead
-    // note: a const operator[] cannot add, it throws like at()
-    common_json       & at(const std::string & key);
-    const common_json & at(const std::string & key) const;
-    common_json       & at(size_t idx);
-    const common_json & at(size_t idx) const;
-
-    common_json       & operator[](const std::string & key);
-    const common_json & operator[](const std::string & key) const;
-    common_json       & operator[](const char * key)       { return (*this)[std::string(key)]; }
-    const common_json & operator[](const char * key) const { return (*this)[std::string(key)]; }
-    common_json       & operator[](int idx)       { return (*this)[to_idx(idx)]; }
-    const common_json & operator[](int idx) const { return (*this)[to_idx(idx)]; }
-    common_json       & operator[](size_t idx);
-    const common_json & operator[](size_t idx) const;
-
-    common_json       & front();
-    const common_json & front() const;
-    common_json       & back();
-    const common_json & back()  const;
-
-    void clear();
-
-    void erase(const std::string & key);
-    void erase(size_t idx);
-
-    // only for the types instantiated in json.cpp, the rest fails at link time
-    template <typename T> T get() const;
-
-    // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
-    // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
-    // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const { return get<std::string>(); }
-
-    template <typename T>
-    T value(const std::string & key, T def) const {
-        return contains(key) ? at(key).get<T>() : def;
-    }
-
-    std::string value(const std::string & key, const char * def) const {
-        return contains(key) ? at(key).get<std::string>() : std::string(def);
-    }
-
-    // a JSON default needs no get<T>(), it is already the right type
-    common_json value(const std::string & key, const common_json & def) const {
-        return contains(key) ? at(key) : def;
-    }
-
-    void assign(const common_json_value & val);
-    void set(const common_json_item & item);
-    void push_back(const common_json_value & val);
-
-    // appends one object, e.g. push_back({{"a", 1}})
-    void push_back(std::initializer_list<common_json_item> items);
-
-    // 1 if the key is there, 0 if not
-    size_t count(const std::string & key) const;
-
-    // appends every value of another array; inserting an array into itself throws
-    void insert(const common_json & vals);
-
-    // a common_json goes through the copy assignment above, everything else becomes a value
-    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
-    common_json & operator=(T && val) {
-        assign(common_json_value(std::forward<T>(val)));
-        return *this;
-    }
-
-    std::string dump(int indent = -1) const;
-
-    // same as dump(), but bad UTF-8 gets replaced instead of throwing
-    std::string dump_safe(int indent = -1) const;
-
-    // walks an array by index, or an object in insertion order
-    // a plain value gives itself once, same as the backing library
-    class iterator {
-      public:
-        using iterator_category = std::forward_iterator_tag;
-        using value_type        = common_json;
-        using difference_type   = std::ptrdiff_t;
-        using pointer           = common_json *;
-        using reference         = common_json &;
-
-        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
-
-        common_json & operator*() const;
-        common_json & value()     const { return **this; }
-        std::string   key()       const;
-
-        iterator & operator++() {
-            idx++;
-            return *this;
-        }
-
-        bool operator!=(const iterator & other) const { return idx != other.idx; }
-        bool operator==(const iterator & other) const { return idx == other.idx; }
-
-      private:
-        common_json * node;
-        size_t        idx;
-    };
-
-    iterator begin() const;
-    iterator end()   const;
-
-    // allows: for (const auto & [key, val] : obj.items())
-    class items_view {
-      public:
-        // the members are public, so an entry also works with structured bindings
-        struct entry {
-            std::string   k;
-            common_json & v;
-
-            const std::string & key()   const { return k; }
-            common_json &       value() const { return v; }
-        };
-
-        items_view(common_json * node, size_t n) : node(node), n(n) {}
-
-        class iterator {
-          public:
-            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
-
-            entry operator*() const;
-
-            iterator & operator++() {
-                idx++;
-                return *this;
-            }
-
-            bool operator!=(const iterator & other) const { return idx != other.idx; }
-
-          private:
-            common_json * node;
-            size_t        idx;
-        };
-
-        iterator begin() const { return iterator(node, 0); }
-        iterator end()   const { return iterator(node, n); }
-
-      private:
-        common_json * node;
-        size_t        n;
-    };
-
-    items_view items() const;
-
-  private:
-    // a negative index must not turn into a huge size_t
-    static size_t to_idx(int idx) {
-        if (idx < 0) {
-            throw common_json_error("negative array index");
-        }
-        return (size_t) idx;
-    }
-
-    // the backing value is built here, json.cpp checks that it fits
-    // it cannot be a pointer: a value inside a tree would then not be a common_json
-    // at() could then only give back a copy instead of a real reference
-    alignas(8) unsigned char storage[32];
-};
-
-using common_json_entry = common_json::items_view::entry;

common/llguidance.cpp

@@ -116,8 +116,6 @@ static llama_sampler_i llama_sampler_llg_i = {
     /* .backend_accept    = */ NULL,
     /* .backend_apply     = */ NULL,
     /* .backend_set_input = */ NULL,
-    /* .backend_reset     = */ NULL,
-    /* .copy_state        = */ NULL,
 };
 
 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,

common/ngram-map.cpp

@@ -125,16 +125,6 @@ void common_ngram_map_begin(
     LOG_DBG("%s: begin, idx_last_draft=%zu, new begin=%zu, #keys=%zu\n", __func__,
             map.idx_last_check, size_begin, map.keys.size());
 
-    size_t idx_begin_cleanup = map.size_last_begin;
-    if (idx_begin_cleanup > size_begin) {
-        if (size_begin > (size_t) map.size_key + map.size_value) {
-            idx_begin_cleanup = size_begin - map.size_key - map.size_value;
-        } else {
-            idx_begin_cleanup = 0;
-        }
-        LOG_INF("%s: shrink cleanup begin: %zu -> %zu\n", __func__, map.size_last_begin, idx_begin_cleanup);
-    }
-
     size_t count_map_entries_upd = 0;
     if (!map.key_map.empty() && size_begin < map.idx_last_check) {
         if (map.show_key_map_stats) {
@@ -160,23 +150,27 @@ void common_ngram_map_begin(
         // Update the map from hash to key index (clear outdated entries).
         for (size_t i = 0; i < map.key_map.size(); ++i) {
             uint32_t key_idx = map.key_map[i];
-            if (key_idx != 0 && key_idx >= idx_begin_cleanup) {
+            if (key_idx >= map.size_last_begin) {
                 map.key_map[i] = 0;
                 count_map_entries_upd++;
             }
         }
-        map.key_map_last_idx = (idx_begin_cleanup > 0) ? (uint32_t) (idx_begin_cleanup - 1) : 0;
+        map.key_map_last_idx = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
     }
 
     if (size_begin < map.idx_last_check && !map.keys.empty()) {
+        // The next token generation will start at index size_begin.
+        // The tokens between map.size_last_begin and size_begin are no longer valid.
+        //
+        // Refresh map: Remove all entries with index >= map.size_last_begin.
         size_t count_keys = map.keys.size();
         size_t count_keys_del = 0;
         size_t count_values_del = 0;
         for (int32_t i = map.keys.size() - 1; i >= 0; --i) {
             common_ngram_map_key & key = map.keys[i];
-            if (key.key_idx >= idx_begin_cleanup) {
+            if (key.key_idx >= map.size_last_begin) {
                 // Delete the key.
-                LOG_DBG("%s: delete key %d at index %zu (>= idx_begin_cleanup=%zu)\n", __func__, i, key.key_idx, idx_begin_cleanup);
+                LOG_DBG("%s: delete key %d at index %zu (>= size_last_begin=%zu)\n", __func__, i, key.key_idx, map.size_last_begin);
                 map.keys.erase(map.keys.begin() + i);
                 count_keys_del++;
                 continue;
@@ -188,7 +182,7 @@ void common_ngram_map_begin(
             // Check the indices of the values.
             for (int16_t j = COMMON_NGRAM_MAX_VALUES - 1; j >= 0; --j) {
                 common_ngram_map_value & value = key.values[j];
-                if (value.value_idx != 0 && value.value_idx >= idx_begin_cleanup) {
+                if (value.value_idx >= map.size_last_begin) {
                     // Delete the value.
                     count_values_del++;
 

common/peg-parser.cpp

@@ -3,13 +3,14 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "trie.h"
 #include "unicode.h"
 
 #include <algorithm>
+#include <deque>
 #include <initializer_list>
 #include <map>
 #include <memory>
+#include <nlohmann/json.hpp>
 #include <regex>
 #include <set>
 #include <stdexcept>
@@ -31,6 +32,154 @@ static bool is_hex_digit(const char c) {
     return (c >= '0' && c <= '9') || (c >= 'a' && c <= 'f') || (c >= 'A' && c <= 'F');
 }
 
+// Trie for matching multiple literals.
+// This is used in common_peg_until_parser and to build a GBNF exclusion grammar
+struct trie {
+    struct node {
+        std::map<uint32_t, size_t> children;  // Use uint32_t to store Unicode codepoints
+        bool is_word;
+    };
+
+    std::vector<node> nodes;
+
+    trie(const std::vector<std::string> & words) {
+      create_node(); // root node
+      for (const auto & w : words) {
+          insert(w);
+      }
+    }
+
+    enum match_result { NO_MATCH, PARTIAL_MATCH, COMPLETE_MATCH };
+
+    // Check if a delimiter starts at the given position
+    match_result check_at(std::string_view sv, size_t start_pos) const {
+        size_t current = 0; // Start at root
+        size_t pos = start_pos;
+
+        // LOG_DBG("%s: checking at pos %zu, sv='%s'\n", __func__, start_pos, std::string(sv).c_str());
+
+        while (pos < sv.size()) {
+            auto result = common_parse_utf8_codepoint(sv, pos);
+            if (result.status != utf8_parse_result::SUCCESS) {
+                break;
+            }
+
+            auto it = nodes[current].children.find(result.codepoint);
+            if (it == nodes[current].children.end()) {
+                // Can't continue matching
+                return match_result{match_result::NO_MATCH};
+            }
+
+            current = it->second;
+            pos += result.bytes_consumed;
+
+            // Check if we've matched a complete word
+            if (nodes[current].is_word) {
+                return match_result{match_result::COMPLETE_MATCH};
+            }
+        }
+
+        // Reached end of input while still in the trie (not at root)
+        if (current != 0) {
+            // We're in the middle of a potential match
+            return match_result{match_result::PARTIAL_MATCH};
+        }
+
+        // Reached end at root (no match)
+        return match_result{match_result::NO_MATCH};
+    }
+
+  private:
+    size_t create_node() {
+        size_t index = nodes.size();
+        nodes.emplace_back();
+        return index;
+    }
+
+    void insert(const std::string & word) {
+        size_t current = 0;
+        size_t pos     = 0;
+        while (pos < word.length()) {
+            auto result = common_parse_utf8_codepoint(word, pos);
+            if (result.status != utf8_parse_result::SUCCESS) {
+                break;
+            }
+
+            uint32_t ch = result.codepoint;
+            pos += result.bytes_consumed;
+
+            auto it = nodes[current].children.find(ch);
+            if (it == nodes[current].children.end()) {
+                size_t child = create_node();
+                nodes[current].children[ch] = child;
+                current = child;
+            } else {
+                current = it->second;
+            }
+        }
+        nodes[current].is_word = true;
+    }
+};
+
+// Aho-Corasick automaton
+struct aho_corasick {
+    trie                t;
+    std::vector<size_t> fail;      // failure links
+    std::vector<size_t> order;     // states in BFS order
+    std::vector<bool>   terminal;  // match states (directly or via a suffix link)
+    std::set<uint32_t>  alphabet;  // every character with a transition
+
+    aho_corasick(const std::vector<std::string> & strings) : t(strings) {
+        const auto & nodes = t.nodes;
+        const size_t n = nodes.size();
+
+        fail.assign(n, 0);
+        order.reserve(n);
+
+        std::deque<size_t> queue{ 0 };
+        while (!queue.empty()) {
+            size_t u = queue.front();
+            queue.pop_front();
+            order.push_back(u);
+            for (const auto & [ch, v] : nodes[u].children) {
+                if (u != 0) {
+                    size_t f = fail[u];
+                    while (f && nodes[f].children.find(ch) == nodes[f].children.end()) {
+                        f = fail[f];
+                    }
+                    auto it = nodes[f].children.find(ch);
+                    fail[v] = (it != nodes[f].children.end() && it->second != v) ? it->second : 0;
+                }
+                queue.push_back(v);
+            }
+        }
+
+        terminal.assign(n, false);
+        for (size_t u : order) {
+            terminal[u] = nodes[u].is_word || (u != 0 && terminal[fail[u]]);
+        }
+
+        for (const auto & node : nodes) {
+            for (const auto & [ch, v] : node.children) {
+                alphabet.insert(ch);
+            }
+        }
+    }
+
+    size_t num_states()          const { return t.nodes.size(); }
+    bool   is_terminal(size_t s) const { return terminal[s]; }
+
+    // follow failure links until a transition on `ch` exists.
+    size_t next(size_t state, uint32_t ch) const {
+        const auto & nodes = t.nodes;
+        while (state && nodes[state].children.find(ch) == nodes[state].children.end()) {
+            state = fail[state];
+        }
+        auto it = nodes[state].children.find(ch);
+        return it != nodes[state].children.end() ? it->second : 0;
+    }
+};
+
 static std::pair<uint32_t, size_t> parse_hex_escape(const std::string & str, size_t pos, int hex_count) {
     if (pos + hex_count > str.length()) {
         return {0, 0};
@@ -569,34 +718,23 @@ struct parser_executor {
     }
 
     static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
-        auto save = pos;
-
         ++pos; // consume '\'
         if (pos >= ctx.input.size()) {
             if (!ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
             }
-            pos = save; // suppress unmatched '\'
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
         }
 
         char c = ctx.input[pos];
-
         if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
             ++pos;
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
+        } else if (c == 'u') {
+            return handle_unicode_escape(ctx, start, pos);
+        } else {
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
         }
-
-        if (c == 'u') {
-            auto result = handle_unicode_escape(ctx, start, pos);
-            if (result.need_more_input()) {
-                pos = save; // suppress incomplete sequence
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
-            }
-            return result;
-        }
-
-        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
     }
 
     static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {
@@ -659,7 +797,7 @@ struct parser_executor {
     }
 
     common_peg_parse_result operator()(const common_peg_until_parser & p) const {
-        common_trie matcher(p.delimiters);
+        trie matcher(p.delimiters);
 
         // Scan input and check for delimiters
         size_t pos = start_pos;
@@ -686,12 +824,12 @@ struct parser_executor {
             // Check if a delimiter starts at this position
             auto match = matcher.check_at(ctx.input, pos);
 
-            if (match == common_trie::COMPLETE_MATCH) {
+            if (match == trie::COMPLETE_MATCH) {
                 // Found a complete delimiter, return everything before it
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
             }
 
-            if (match == common_trie::PARTIAL_MATCH) {
+            if (match == trie::PARTIAL_MATCH) {
                 // Found a partial match extending to end of input, return everything before it
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
             }
@@ -1119,8 +1257,8 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
-common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<nlohmann::ordered_json>(schema), raw}));
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1421,7 +1559,7 @@ static std::string gbnf_ac_grammar(
                                     const std::map<size_t, std::vector<uint32_t>> &,
                                     const std::vector<uint32_t> &,
                                     const std::function<std::string(size_t)> &)> & build_rule) {
-    common_aho_corasick ac(strings);
+    aho_corasick ac(strings);
 
     auto state_name = [&](size_t s) -> std::string {
         if (s == 0) {
@@ -1804,8 +1942,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     }
 }
 
-static common_json serialize_parser_variant(const common_peg_parser_variant & variant) {
-    using json = common_json;
+static nlohmann::json serialize_parser_variant(const common_peg_parser_variant & variant) {
+    using json = nlohmann::json;
 
     return std::visit([](const auto & p) -> json {
         using T = std::decay_t<decltype(p)>;
@@ -1859,7 +1997,7 @@ static common_json serialize_parser_variant(const common_peg_parser_variant & va
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : json(nullptr)},
+                {"schema", p.schema ? *p.schema : nullptr},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1887,19 +2025,19 @@ static common_json serialize_parser_variant(const common_peg_parser_variant & va
     }, variant);
 }
 
-common_json common_peg_arena::to_json() const {
-    auto parsers = common_json::array();
+nlohmann::json common_peg_arena::to_json() const {
+    auto parsers = nlohmann::json::array();
     for (const auto & parser : parsers_) {
         parsers.push_back(serialize_parser_variant(parser));
     }
-    return common_json{
+    return nlohmann::json{
         {"parsers", parsers},
         {"rules", rules_},
         {"root", root_}
     };
 }
 
-static common_peg_parser_variant deserialize_parser_variant(const common_json & j) {
+static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json & j) {
     if (!j.contains("type") || !j["type"].is_string()) {
         throw std::runtime_error("Parser variant JSON missing or invalid 'type' field");
     }
@@ -1968,9 +2106,9 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
         }
         common_peg_chars_parser parser;
         parser.pattern = j["pattern"];
-        parser.negated = j["negated"].get<bool>();
-        parser.min_count = j["min_count"].get<int>();
-        parser.max_count = j["max_count"].get<int>();
+        parser.negated = j["negated"];
+        parser.min_count = j["min_count"];
+        parser.max_count = j["max_count"];
         for (const auto & range_json : j["ranges"]) {
             if (!range_json.contains("start") || !range_json.contains("end")) {
                 throw std::runtime_error("char_range missing 'start' or 'end' field");
@@ -2006,7 +2144,7 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
         if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<common_json>(j["schema"]);
+            parser.schema = std::make_shared<nlohmann::ordered_json>(j["schema"]);
         }
         parser.raw = j["raw"].get<bool>();
         return parser;
@@ -2068,7 +2206,7 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
     throw std::runtime_error("Unknown parser type: " + type);
 }
 
-common_peg_arena common_peg_arena::from_json(const common_json & j) {
+common_peg_arena common_peg_arena::from_json(const nlohmann::json & j) {
     if (!j.contains("parsers") || !j["parsers"].is_array()) {
         throw std::runtime_error("JSON missing or invalid 'parsers' array");
     }
@@ -2108,7 +2246,7 @@ std::string common_peg_arena::save() const {
 }
 
 void common_peg_arena::load(const std::string & data) {
-    *this = from_json(common_json::parse(data));
+    *this = from_json(nlohmann::json::parse(data));
 }
 
 common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {

common/peg-parser.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "json.h"
+#include <nlohmann/json_fwd.hpp>
 
 #include <memory>
 #include <set>
@@ -245,7 +245,7 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<common_json> schema;
+    std::shared_ptr<nlohmann::ordered_json> schema;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -332,8 +332,8 @@ class common_peg_arena {
 
     std::string dump(common_peg_parser_id id) const;
 
-    common_json to_json() const;
-    static common_peg_arena from_json(const common_json & j);
+    nlohmann::json to_json() const;
+    static common_peg_arena from_json(const nlohmann::json & j);
 
     std::string save() const;
     void load(const std::string & data);
@@ -490,7 +490,7 @@ class common_peg_parser_builder {
 
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
-    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.
     // If trigger=true, marks this rule as an entry point for lazy grammar generation.

common/preset.cpp

@@ -7,7 +7,6 @@
 #include <fstream>
 #include <sstream>
 #include <filesystem>
-#include <regex>
 
 static std::string rm_leading_dashes(const std::string & str) {
     size_t pos = 0;
@@ -17,23 +16,6 @@ static std::string rm_leading_dashes(const std::string & str) {
     return str.substr(pos);
 }
 
-static std::string canonical_tag(const std::string & tag) {
-    static const std::regex re_tag("[-.]([A-Z0-9_]+)$", std::regex::icase);
-    std::smatch m;
-    if (std::regex_search(tag, m, re_tag)) {
-        std::string canon = m[1].str();
-        for (char & c : canon) {
-            c = (char) std::toupper((unsigned char) c);
-        }
-        return canon;
-    }
-    std::string upper = tag;
-    for (char & c : upper) {
-        c = (char) std::toupper((unsigned char) c);
-    }
-    return upper;
-}
-
 std::vector<std::string> common_preset::to_args(const std::string & bin_path) const {
     std::vector<std::string> args;
 
@@ -288,18 +270,11 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
 
     for (auto section : ini_data) {
         common_preset preset;
-        std::string section_name = section.first.empty() ? std::string(COMMON_PRESET_DEFAULT_NAME) : section.first;
-        if (section_name != "*" && section_name != COMMON_PRESET_DEFAULT_NAME) {
-            auto colon_idx = section_name.rfind(':');
-            if (colon_idx != std::string::npos) {
-                std::string tag       = section_name.substr(colon_idx + 1);
-                std::string canon_tag = canonical_tag(tag);
-                if (canon_tag != tag) {
-                    section_name = section_name.substr(0, colon_idx + 1) + canon_tag;
-                }
-            }
+        if (section.first.empty()) {
+            preset.name = COMMON_PRESET_DEFAULT_NAME;
+        } else {
+            preset.name = section.first;
         }
-        preset.name = section_name;
         LOG_DBG("loading preset: %s\n", preset.name.c_str());
         for (const auto & [key, value] : section.second) {
             if (key == "version") {
@@ -322,8 +297,6 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
                     preset.options[opt] = value;
                 }
                 LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
-            } else if (ignore_unknown_keys) {
-                LOG_WRN("ignoring option '%s' from %s: not supported by this program\n", key.c_str(), path.c_str());
             } else {
                 throw std::runtime_error(string_format(
                     "option '%s' not recognized in preset '%s'",
@@ -332,10 +305,6 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
             }
         }
 
-        if (preset.name == COMMON_PRESET_DEFAULT_NAME && preset.options.empty()) {
-            continue;
-        }
-
         if (preset.name == "*") {
             // handle global preset
             global = preset;
@@ -365,25 +334,8 @@ struct local_model {
     std::string name;
     std::string path;
     std::string path_mmproj;
-    std::string path_draft;
 };
 
-// TODO @ngxson: handle "eagle3-" when it's supported by common_speculative_types_from_gguf()
-static const char * draft_prefixes[] = { "mtp-", "dspark-", "dflash-" };
-
-static bool is_mmproj_file(const std::string & fname) {
-    return fname.find("mmproj") != std::string::npos;
-}
-
-static bool is_draft_file(const std::string & fname) {
-    for (const auto & prefix : draft_prefixes) {
-        if (fname.rfind(prefix, 0) == 0) {
-            return true;
-        }
-    }
-    return false;
-}
-
 common_presets common_preset_context::load_from_models_dir(const std::string & models_dir) const {
     if (!std::filesystem::exists(models_dir) || !std::filesystem::is_directory(models_dir)) {
         throw std::runtime_error(string_format("error: '%s' does not exist or is not a directory\n", models_dir.c_str()));
@@ -395,15 +347,10 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         common_file_info model_file;
         common_file_info first_shard_file;
         common_file_info mmproj_file;
-        common_file_info draft_file;
         for (const auto & file : files) {
             if (string_ends_with(file.name, ".gguf")) {
-                if (is_mmproj_file(file.name)) {
+                if (file.name.find("mmproj") != std::string::npos) {
                     mmproj_file = file;
-                } else if (is_draft_file(file.name)) {
-                    if (draft_file.path.empty()) {
-                        draft_file = file; // first sidecar found wins
-                    }
                 } else if (file.name.find("-00001-of-") != std::string::npos) {
                     first_shard_file = file;
                 } else {
@@ -415,8 +362,7 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         local_model model{
             /* name        */ name,
             /* path        */ first_shard_file.path.empty() ? model_file.path : first_shard_file.path,
-            /* path_mmproj */ mmproj_file.path, // can be empty
-            /* path_draft  */ draft_file.path   // can be empty
+            /* path_mmproj */ mmproj_file.path // can be empty
         };
         if (!model.path.empty()) {
             models.push_back(model);
@@ -428,17 +374,13 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         if (file.is_dir) {
             scan_subdir(file.path, file.name);
         } else if (string_ends_with(file.name, ".gguf")) {
-            if (is_mmproj_file(file.name) || is_draft_file(file.name)) {
-                continue; // companion file, cannot be loaded as a model on its own
-            }
             // single file model
             std::string name = file.name;
             string_replace_all(name, ".gguf", "");
             local_model model{
                 /* name        */ name,
                 /* path        */ file.path,
-                /* path_mmproj */ "",
-                /* path_draft  */ ""
+                /* path_mmproj */ ""
             };
             models.push_back(model);
         }
@@ -453,9 +395,6 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
         if (!model.path_mmproj.empty()) {
             preset.set_option(*this, "LLAMA_ARG_MMPROJ", model.path_mmproj);
         }
-        if (!model.path_draft.empty()) {
-            preset.set_option(*this, "LLAMA_ARG_SPEC_DRAFT_MODEL", model.path_draft);
-        }
         out[preset.name] = preset;
     }
 

common/preset.h

@@ -59,10 +59,6 @@ struct common_preset_context {
     bool filter_allowed_keys = false;
     std::set<std::string> allowed_keys;
 
-    // if true, options unknown to the current example are skipped instead of being an error
-    // used for config files shared by all binaries, where each binary only knows a subset of options
-    bool ignore_unknown_keys = false;
-
     // if only_remote_allowed is true, only accept whitelisted keys
     common_preset_context(llama_example ex);
 

common/reasoning-budget.cpp

@@ -1,52 +1,39 @@
 #include "reasoning-budget.h"
 #include "common.h"
-#include "trie.h"
 #include "unicode.h"
 
 #include "log.h"
 
-#include <algorithm>
 #include <cmath>
 #include <cstdint>
 #include <string>
 #include <vector>
 
 struct token_matcher {
-    std::vector<llama_tokens> seqs;
-    common_aho_corasick ac;
-    size_t state = 0;
+    std::vector<llama_token> tokens;
+    size_t pos = 0;
 
-    token_matcher(const std::vector<llama_tokens> & seqs) : seqs(collect(seqs)), ac(build_trie(this->seqs)) {}
+    bool advance(llama_token token) {
+        if (tokens.empty()) {
+            return false;
+        }
 
-    static std::vector<llama_tokens> collect(const std::vector<llama_tokens> & seqs) {
-        std::vector<llama_tokens> res;
-        for (const auto & seq : seqs) {
-            if (!seq.empty() && std::find(res.begin(), res.end(), seq) == res.end()) {
-                res.push_back(seq);
+        if (token == tokens[pos]) {
+            pos++;
+            if (pos >= tokens.size()) {
+                pos = 0;
+                return true;
+            }
+        } else {
+            pos = 0;
+            if (token == tokens[0]) {
+                pos = 1;
             }
         }
-        return res;
+        return false;
     }
 
-    static common_trie build_trie(const std::vector<llama_tokens> & seqs) {
-        common_trie t;
-        for (const auto & seq : seqs) {
-            t.insert(std::vector<uint32_t>(seq.begin(), seq.end()));
-        }
-        return t;
-    }
-
-    // returns the index into seqs of the longest sequence ending at this token, or -1
-    int32_t advance(llama_token token) {
-        state = ac.next(state, (uint32_t) token);
-        const int32_t p = ac.match_pattern(state);
-        if (p >= 0) {
-            state = 0;
-        }
-        return p;
-    }
-
-    void reset() { state = 0; }
+    void reset() { pos = 0; }
 };
 
 struct common_reasoning_budget_ctx {
@@ -54,7 +41,7 @@ struct common_reasoning_budget_ctx {
 
     token_matcher start_matcher;
     token_matcher end_matcher;
-    llama_tokens forced_tokens;
+    std::vector<llama_token> forced_tokens;
 
     int32_t budget;           // maximum tokens in reasoning block
     int32_t remaining;        // tokens remaining in budget
@@ -63,8 +50,6 @@ struct common_reasoning_budget_ctx {
 
     // for forcing
     size_t force_pos;         // next position in forced_tokens to force
-
-    int32_t end_match;        // index into end_matcher.seqs of the sequence that transitioned to DONE, -1 if none
 };
 
 static const char * common_reasoning_budget_name(const struct llama_sampler * /*smpl*/) {
@@ -77,15 +62,15 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
     switch (ctx->state) {
         case REASONING_BUDGET_IDLE:
         {
-            if (ctx->start_matcher.advance(token) >= 0) {
+            if (ctx->start_matcher.advance(token)) {
                 ctx->state = REASONING_BUDGET_COUNTING;
                 ctx->remaining = ctx->budget;
-                COM_TRC("activated, budget=%d tokens\n", ctx->budget);
+                LOG_INF("reasoning-budget: activated, budget=%d tokens\n", ctx->budget);
 
                 if (ctx->remaining <= 0) {
                     ctx->state = REASONING_BUDGET_FORCING;
                     ctx->force_pos = 0;
-                    COM_TRC("%s", "budget=0, forcing immediately\n");
+                    LOG_INF("reasoning-budget: budget=0, forcing immediately\n");
                 }
             }
             break;
@@ -93,11 +78,9 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
         case REASONING_BUDGET_COUNTING:
         case REASONING_BUDGET_WAITING_UTF8:
         {
-            const int32_t match = ctx->end_matcher.advance(token);
-            if (match >= 0) {
+            if (ctx->end_matcher.advance(token)) {
                 ctx->state = REASONING_BUDGET_DONE;
-                ctx->end_match = match;
-                COM_TRC("%s", "deactivated (natural end)\n");
+                LOG_INF("reasoning-budget: deactivated (natural end)\n");
                 break;
             }
 
@@ -112,7 +95,7 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
                     ctx->state = REASONING_BUDGET_FORCING;
                     ctx->force_pos = 0;
                     ctx->end_matcher.reset();
-                    COM_TRC("%s", "UTF-8 complete, now forcing end sequence\n");
+                    LOG_INF("reasoning-budget: UTF-8 complete, now forcing end sequence\n");
                 }
             } else if (ctx->state == REASONING_BUDGET_COUNTING) {
                 ctx->remaining--;
@@ -121,42 +104,36 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
                         ctx->state = REASONING_BUDGET_FORCING;
                         ctx->force_pos = 0;
                         ctx->end_matcher.reset();
-                        COM_TRC("%s", "budget exhausted, forcing end sequence\n");
+                        LOG_INF("reasoning-budget: budget exhausted, forcing end sequence\n");
                     } else {
                         ctx->state = REASONING_BUDGET_WAITING_UTF8;
                         ctx->end_matcher.reset();
-                        COM_TRC("%s", "budget exhausted, waiting for UTF-8 completion\n");
+                        LOG_INF("reasoning-budget: budget exhausted, waiting for UTF-8 completion\n");
                     }
                 }
             }
             break;
         }
         case REASONING_BUDGET_FORCING:
-        {
-            // track the end sequence within forced_tokens so it is also reported on DONE
-            const int32_t match = ctx->end_matcher.advance(token);
             ctx->force_pos++;
             if (ctx->force_pos >= ctx->forced_tokens.size()) {
                 ctx->state = REASONING_BUDGET_DONE;
-                ctx->end_match = match;
-                COM_TRC("%s", "forced sequence complete, done\n");
+                LOG_INF("reasoning-budget: forced sequence complete, done\n");
             }
             break;
-        }
         case REASONING_BUDGET_DONE:
             // Re-arm on a new start tag: some models emit multiple <think> blocks
             // per response, and each should get a fresh budget window.
-            if (ctx->start_matcher.advance(token) >= 0) {
+            if (ctx->start_matcher.advance(token)) {
                 ctx->state = REASONING_BUDGET_COUNTING;
                 ctx->remaining = ctx->budget;
                 ctx->end_matcher.reset();
-                ctx->end_match = -1;
-                COM_TRC("re-activated on new start tag, budget=%d tokens\n", ctx->budget);
+                LOG_INF("reasoning-budget: re-activated on new start tag, budget=%d tokens\n", ctx->budget);
 
                 if (ctx->remaining <= 0) {
                     ctx->state = REASONING_BUDGET_FORCING;
                     ctx->force_pos = 0;
-                    COM_TRC("%s", "budget=0, forcing immediately\n");
+                    LOG_INF("reasoning-budget: budget=0, forcing immediately\n");
                 }
             }
             break;
@@ -192,12 +169,11 @@ static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
     ctx->start_matcher.reset();
     ctx->end_matcher.reset();
     ctx->force_pos = 0;
-    ctx->end_match = -1;
 }
 
 static struct llama_sampler * common_reasoning_budget_init_state(
-        const struct llama_vocab * vocab, const std::vector<llama_tokens> & start_seqs,
-        const std::vector<llama_tokens> & end_seqs, const llama_tokens & forced_tokens,
+        const struct llama_vocab * vocab, const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens, const std::vector<llama_token> & forced_tokens,
         int32_t budget, common_reasoning_budget_state initial_state);
 
 static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl);
@@ -217,8 +193,6 @@ static struct llama_sampler_i common_reasoning_budget_i = {
     /* .backend_accept    = */ nullptr,
     /* .backend_apply     = */ nullptr,
     /* .backend_set_input = */ nullptr,
-    /* .backend_reset     = */ nullptr,
-    /* .copy_state        = */ nullptr,
 };
 
 static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
@@ -231,12 +205,12 @@ static struct llama_sampler * common_reasoning_budget_clone(const struct llama_s
 }
 
 static struct llama_sampler * common_reasoning_budget_init_state(
-        const struct llama_vocab        * vocab,
-        const std::vector<llama_tokens> & start_seqs,
-        const std::vector<llama_tokens> & end_seqs,
-        const llama_tokens              & forced_tokens,
-        int32_t                           budget,
-        common_reasoning_budget_state     initial_state) {
+        const struct llama_vocab             * vocab,
+        const std::vector<llama_token>       & start_tokens,
+        const std::vector<llama_token>       & end_tokens,
+        const std::vector<llama_token>       & forced_tokens,
+        int32_t                                budget,
+        common_reasoning_budget_state          initial_state) {
     // promote COUNTING with budget <= 0 to FORCING
     if (initial_state == REASONING_BUDGET_COUNTING && budget <= 0) {
         initial_state = REASONING_BUDGET_FORCING;
@@ -246,26 +220,25 @@ static struct llama_sampler * common_reasoning_budget_init_state(
         /* .iface = */ &common_reasoning_budget_i,
         /* .ctx   = */ new common_reasoning_budget_ctx {
             /* .vocab         = */ vocab,
-            /* .start_matcher = */ token_matcher(start_seqs),
-            /* .end_matcher   = */ token_matcher(end_seqs),
+            /* .start_matcher = */ { start_tokens, 0 },
+            /* .end_matcher   = */ { end_tokens, 0 },
             /* .forced_tokens = */ forced_tokens,
             /* .budget        = */ budget,
             /* .remaining     = */ budget,
             /* .state         = */ initial_state,
             /* .force_pos     = */ 0,
-            /* .end_match     = */ -1,
         }
     );
 }
 
 struct llama_sampler * common_reasoning_budget_init(
-        const struct llama_vocab        * vocab,
-        const std::vector<llama_tokens> & start_seqs,
-        const std::vector<llama_tokens> & end_seqs,
-        const llama_tokens              & forced_tokens,
-        int32_t                           budget,
-        common_reasoning_budget_state     initial_state) {
-    return common_reasoning_budget_init_state(vocab, start_seqs, end_seqs, forced_tokens, budget, initial_state);
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state) {
+    return common_reasoning_budget_init_state(vocab, start_tokens, end_tokens, forced_tokens, budget, initial_state);
 }
 
 common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl) {
@@ -275,19 +248,6 @@ common_reasoning_budget_state common_reasoning_budget_get_state(const struct lla
     return ((const common_reasoning_budget_ctx *)smpl->ctx)->state;
 }
 
-const llama_tokens * common_reasoning_budget_get_end_match(const struct llama_sampler * smpl) {
-    if (!smpl) {
-        return nullptr;
-    }
-
-    const auto * ctx = (const common_reasoning_budget_ctx *) smpl->ctx;
-    if (ctx->end_match < 0) {
-        return nullptr;
-    }
-
-    return &ctx->end_matcher.seqs[ctx->end_match];
-}
-
 bool common_reasoning_budget_force(struct llama_sampler * smpl) {
     if (!smpl) {
         return false;
@@ -304,7 +264,7 @@ bool common_reasoning_budget_force(struct llama_sampler * smpl) {
     ctx->state = REASONING_BUDGET_FORCING;
     ctx->force_pos = 0;
     ctx->end_matcher.reset();
-    COM_TRC("%s", "forced into forcing state (manual transition)\n");
+    LOG_INF("reasoning-budget: forced into forcing state (manual transition)\n");
 
     return true;
 }

common/reasoning-budget.h

@@ -2,8 +2,6 @@
 
 #include "llama.h"
 
-#include "common.h"
-
 #include <cstdint>
 #include <vector>
 
@@ -19,34 +17,30 @@ enum common_reasoning_budget_state {
 // reasoning block (e.g. between <think> and </think>).
 //
 // State machine: IDLE -> COUNTING -> WAITING_UTF8 -> FORCING -> DONE
-//   IDLE:         passthrough, watching for a start sequence
-//   COUNTING:     counting down remaining tokens, watching for a natural end sequence
+//   IDLE:         passthrough, watching for start_tokens sequence
+//   COUNTING:     counting down remaining tokens, watching for natural end_tokens
 //   WAITING_UTF8: budget exhausted, allowing tokens to complete a UTF-8 sequence
 //   FORCING:      forces forced_tokens token-by-token (all other logits -> -inf)
 //   DONE:         passthrough forever
 //
 // Parameters:
 //   vocab          - vocabulary (used for UTF-8 boundary detection; can be nullptr)
-//   start_seqs     - token sequences, any of which activates counting
-//   end_seqs       - token sequences, any of which naturally deactivates
+//   start_tokens   - token sequence that activates counting
+//   end_tokens     - token sequence for natural deactivation
 //   forced_tokens  - token sequence forced when budget expires
 //   budget         - max tokens allowed in the reasoning block
 //   initial_state  - initial state
 //
 struct llama_sampler * common_reasoning_budget_init(
-        const struct llama_vocab        * vocab,
-        const std::vector<llama_tokens> & start_seqs,
-        const std::vector<llama_tokens> & end_seqs,
-        const llama_tokens              & forced_tokens,
-        int32_t                           budget,
-        common_reasoning_budget_state     initial_state = REASONING_BUDGET_IDLE);
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state = REASONING_BUDGET_IDLE);
 
 common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl);
 
-// The end sequence that transitioned the sampler to DONE, or nullptr if none
-// was recorded. Cleared when a new start sequence re-arms the sampler.
-const llama_tokens * common_reasoning_budget_get_end_match(const struct llama_sampler * smpl);
-
 // Manually transition the reasoning budget sampler into the FORCING state.
 // Returns true if the transition occurred.
 bool common_reasoning_budget_force(struct llama_sampler * smpl);