fix partial writes

wider loads
deduplicate repacking code
2026-08-31 17:17:44 +02:00 · 2026-05-15 16:00:57 +02:00 · 2026-05-15 15:22:57 +02:00 · 2026-05-15 13:25:49 +02:00 · 2026-05-15 13:10:19 +02:00 · 2026-05-15 12:11:01 +02:00
1184 changed files with 27296 additions and 53357 deletions
@@ -5,9 +5,6 @@
 # Define the CANN base image for easier version updates later
 ARG CHIP_TYPE=910b
 ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.5.0-${CHIP_TYPE}-openeuler24.03-py3.11
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 # ==============================================================================
 # BUILD STAGE
@@ -58,7 +55,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full && \
    cp build/bin/* /app/full/ && \
    cp *.py /app/full/ && \
-    cp -r conversion /app/full/ && \
    cp -r gguf-py /app/full/ && \
    cp -r requirements /app/full/ && \
    cp requirements.txt /app/full/
@@ -71,19 +67,6 @@ RUN mkdir -p /app/full && \
 # ==============================================================================
 FROM ${CANN_BASE_IMAGE} AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 # -- Install runtime dependencies --
 RUN yum install -y libgomp curl && \
    yum clean all && \
@@ -1,7 +1,4 @@
 ARG UBUNTU_VERSION=24.04
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 FROM ubuntu:$UBUNTU_VERSION AS build

@@ -30,7 +27,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -39,19 +35,6 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ubuntu:$UBUNTU_VERSION AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN apt-get update \
    && apt-get install -y libgomp1 curl \
    && apt autoremove -y \
@@ -6,10 +6,6 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER

 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build

 # CUDA architecture to build for (defaults to all supported archs)
@@ -36,7 +32,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -45,19 +40,6 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN apt-get update \
    && apt-get install -y libgomp1 curl \
    && apt autoremove -y \
@@ -1,7 +1,4 @@
 ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 ## Build Image

@@ -36,7 +33,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -44,19 +40,6 @@ RUN mkdir -p /app/full \

 FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 ARG IGC_VERSION=v2.20.5
 ARG IGC_VERSION_FULL=2_2.20.5+19972
 ARG COMPUTE_RUNTIME_VERSION=25.40.35563.10
@@ -1,7 +1,4 @@
 ARG ASCEND_VERSION=8.5.0-910b-openeuler22.03-py3.10
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 FROM ascendai/cann:$ASCEND_VERSION AS build

@@ -31,20 +28,6 @@ RUN echo "Building with static libs" && \

 # TODO: use image with NNRT
 FROM ascendai/cann:$ASCEND_VERSION AS runtime
-
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 COPY --from=build /app/build/bin/llama-cli /app/build/bin/llama-completion /

 ENV LC_ALL=C.utf8
@@ -6,10 +6,6 @@ ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_V

 ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build

 # MUSA architecture to build for (defaults to all supported archs)
@@ -41,7 +37,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -50,19 +45,6 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_MUSA_RUN_CONTAINER} AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN apt-get update \
    && apt-get install -y libgomp1 curl \
    && apt autoremove -y \
@@ -3,7 +3,6 @@
  glibc,
  config,
  stdenv,
-  stdenvNoCC,
  runCommand,
  cmake,
  ninja,
@@ -20,8 +19,6 @@
  openssl,
  shaderc,
  spirv-headers,
-  nodejs,
-  importNpmLock,
  useBlas ?
    builtins.all (x: !x) [
      useCuda
@@ -133,31 +130,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
    src = lib.cleanSource ../../.;
  };

-  # Builds the webui locally, taking care not to require updating any sha256 hash.
-  webui = stdenvNoCC.mkDerivation {
-    pname = "webui";
-    version = llamaVersion;
-    src = lib.cleanSource ../../tools/ui;
-
-    nativeBuildInputs = [
-      nodejs
-      importNpmLock.linkNodeModulesHook
-    ];
-
-    # no sha256 required when using buildNodeModules
-    npmDeps = importNpmLock.buildNodeModules {
-      npmRoot = ../../tools/ui;
-      inherit nodejs;
-    };
-
-    installPhase = ''
-      LLAMA_UI_OUT_DIR=$out npm run build --offline
-    '';
-  };
-
-  postPatch = lib.optionalString useWebUi ''
-    cp -r ${finalAttrs.webui} tools/ui/dist
-    chmod -R u+w tools/ui/dist
+  postPatch = ''
  '';

  # With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015,
@@ -18,10 +18,6 @@ ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
 ARG http_proxy=
 ARG https_proxy=

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-
 ## Build Image
 FROM ubuntu:${UBUNTU_VERSION} AS build

@@ -81,7 +77,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/ReleaseOV/bin/* /app/full/ \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -93,18 +88,6 @@ FROM ubuntu:${UBUNTU_VERSION} AS base
 # Pass proxy args to runtime stage
 ARG http_proxy
 ARG https_proxy
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE

 RUN apt-get update \
    && apt-get install -y libgomp1 libtbb12 curl wget ocl-icd-libopencl1 \
@@ -7,10 +7,6 @@ ARG AMDGPU_VERSION=7.2.1
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-
 ### Build image
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build

@@ -53,7 +49,6 @@ RUN mkdir -p /app/lib \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -62,19 +57,6 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_ROCM_DEV_CONTAINER} AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN apt-get update \
    && apt-get install -y libgomp1 curl \
    && apt autoremove -y \
@@ -1,8 +1,5 @@
 ARG GCC_VERSION=15.2.0
 ARG UBUNTU_VERSION=24.04
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 ### Build Llama.cpp stage
 FROM gcc:${GCC_VERSION} AS build
@@ -37,7 +34,6 @@ RUN --mount=type=cache,target=/root/.ccache \

 COPY *.py             /opt/llama.cpp/bin
 COPY .devops/tools.sh /opt/llama.cpp/bin
-COPY conversion       /opt/llama.cpp/conversion

 COPY gguf-py          /opt/llama.cpp/gguf-py
 COPY requirements.txt /opt/llama.cpp/gguf-py
@@ -48,28 +44,14 @@ COPY requirements     /opt/llama.cpp/gguf-py/requirements
 FROM scratch AS collector

 # Copy llama.cpp binaries and libraries
-COPY --from=build /opt/llama.cpp/bin        /llama.cpp/bin
-COPY --from=build /opt/llama.cpp/lib        /llama.cpp/lib
-COPY --from=build /opt/llama.cpp/gguf-py    /llama.cpp/gguf-py
-COPY --from=build /opt/llama.cpp/conversion /llama.cpp/conversion
+COPY --from=build /opt/llama.cpp/bin     /llama.cpp/bin
+COPY --from=build /opt/llama.cpp/lib     /llama.cpp/lib
+COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py


 ### Base image
 FROM ubuntu:${UBUNTU_VERSION} AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
    apt update -y && \
@@ -109,7 +91,6 @@ RUN curl https://sh.rustup.rs -sSf | bash -s -- -y

 COPY --from=collector /llama.cpp/bin /app
 COPY --from=collector /llama.cpp/gguf-py /app/gguf-py
-COPY --from=collector /llama.cpp/conversion /app/conversion

 RUN pip install --no-cache-dir --break-system-packages \
        -r /app/gguf-py/requirements.txt
@@ -1,7 +1,4 @@
 ARG UBUNTU_VERSION=26.04
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A

 FROM ubuntu:$UBUNTU_VERSION AS build

@@ -26,7 +23,6 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
    && cp *.py /app/full \
-    && cp -r conversion /app/full \
    && cp -r gguf-py /app/full \
    && cp -r requirements /app/full \
    && cp requirements.txt /app/full \
@@ -35,19 +31,6 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ubuntu:$UBUNTU_VERSION AS base

-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
 RUN apt-get update \
    && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
    libglvnd0 libgl1 libglx0 libegl1 libgles2 \
@@ -1,101 +0,0 @@
-ARG UBUNTU_VERSION=24.04
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-
-FROM ubuntu:$UBUNTU_VERSION AS build
-
-RUN apt-get update && \
-    apt-get install -y gcc-13 g++-13 build-essential git cmake libssl-dev libomp-dev libnuma-dev python3 ca-certificates
-
-ENV CC=gcc-13 CXX=g++-13
-
-WORKDIR /app
-
-COPY . .
-
-RUN cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_ZENDNN=ON && \
-    cmake --build build -j $(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
-
-RUN mkdir -p /app/full \
-    && cp build/bin/* /app/full \
-    && cp *.py /app/full \
-    && cp -r conversion /app/full \
-    && cp -r gguf-py /app/full \
-    && cp -r requirements /app/full \
-    && cp requirements.txt /app/full \
-    && cp .devops/tools.sh /app/full/tools.sh
-
-## Base image
-FROM ubuntu:$UBUNTU_VERSION AS base
-
-ARG BUILD_DATE=N/A
-ARG APP_VERSION=N/A
-ARG APP_REVISION=N/A
-ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
-ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
-LABEL org.opencontainers.image.created=$BUILD_DATE \
-      org.opencontainers.image.version=$APP_VERSION \
-      org.opencontainers.image.revision=$APP_REVISION \
-      org.opencontainers.image.title="llama.cpp" \
-      org.opencontainers.image.description="LLM inference in C/C++" \
-      org.opencontainers.image.url=$IMAGE_URL \
-      org.opencontainers.image.source=$IMAGE_SOURCE
-
-RUN apt-get update \
-    && apt-get install -y libgomp1 libnuma1 curl \
-    && apt autoremove -y \
-    && apt clean -y \
-    && rm -rf /tmp/* /var/tmp/* \
-    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
-    && find /var/cache -type f -delete
-
-COPY --from=build /app/lib/ /app
-
-### Full
-FROM base AS full
-
-COPY --from=build /app/full /app
-
-WORKDIR /app
-
-RUN apt-get update \
-    && apt-get install -y \
-    git \
-    python3 \
-    python3-pip \
-    python3-wheel \
-    && pip install --break-system-packages --upgrade setuptools \
-    && pip install --break-system-packages -r requirements.txt \
-    && apt autoremove -y \
-    && apt clean -y \
-    && rm -rf /tmp/* /var/tmp/* \
-    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
-    && find /var/cache -type f -delete
-
-ENTRYPOINT ["/app/tools.sh"]
-
-### Light, CLI only
-FROM base AS light
-
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
-
-WORKDIR /app
-
-ENTRYPOINT [ "/app/llama-cli" ]
-
-### Server, Server only
-FROM base AS server
-
-ENV LLAMA_ARG_HOST=0.0.0.0
-
-COPY --from=build /app/full/llama-server /app
-
-WORKDIR /app
-
-HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
-
-ENTRYPOINT [ "/app/llama-server" ]
@@ -45,7 +45,7 @@ insert_final_newline = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset

-[tools/ui/**]
+[tools/server/webui/**]
 indent_style = unset
 indent_size = unset
 end_of_line = unset
@@ -100,8 +100,8 @@ body:
      label: Relevant log output
      description: >
          Please copy and paste any relevant log output, including the command that you entered and any generated text.
-          For very long logs (thousands of lines), please upload them as files instead; the `--log-file` CLI argument can be used for this purpose.
-          On Linux you can alternatively redirect the console output of any command into a file by appending ` > llama.log 2>&1` to your command.
+          For very long logs (thousands of lines), preferably upload them as files instead.
+          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
      value: |
        <details>
        <summary>Logs</summary>
@@ -88,8 +88,8 @@ body:
      description: >
          If applicable, please copy and paste any relevant log output, including any generated text.
          If you are encountering problems specifically with the `llama_params_fit` module, always upload `--verbose` logs as well.
-          For very long logs (thousands of lines), please upload them as files instead; the `--log-file` CLI argument can be used for this purpose.
-          On Linux you can alternatively redirect the console output of any command into a file by appending ` > llama.log 2>&1` to your command.
+          For very long logs (thousands of lines), please upload them as files instead.
+          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
      value: |
        <details>
        <summary>Logs</summary>
@@ -1,22 +0,0 @@
-name: "ccache-clear"
-description: "Delete all GitHub Actions caches matching a key prefix"
-inputs:
-  key:
-    description: "Cache key prefix to match and delete"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Clear caches
-      shell: bash
-      run: |
-        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: ${{ inputs.key }}"
-          exit 0
-        fi
-        while read -r id key; do
-          echo "Deleting cache: $id ($key)"
-          gh cache delete "$id"
-        done <<< "$CACHES"
@@ -15,6 +15,6 @@ runs:
      id: setup
      uses: ./.github/actions/unarchive-tar
      with:
-        url: https://github.com/spacemit-com/toolchain/releases/download/v${{ inputs.version }}/spacemit-toolchain-linux-glibc-x86_64-v${{ inputs.version }}.tar.xz
+        url: https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v${{ inputs.version }}.tar.xz
        path: ${{ inputs.path }}
        strip: 1
@@ -24,4 +24,4 @@ runs:
      run: |
        mkdir -p ${{ inputs.path }}
        cd ${{ inputs.path }}
-        curl --no-progress-meter -L ${{ inputs.url }} | tar -${{ inputs.type }}x --strip-components=${{ inputs.strip }}
+        curl --no-progress-meter ${{ inputs.url }} | tar -${{ inputs.type }}x --strip-components=${{ inputs.strip }}
@@ -96,34 +96,3 @@ runs:
          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
          echo "CUDA_PATH_V13_1=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-    - name: Install Cuda Toolkit 13.3
-      if: ${{ inputs.cuda_version == '13.3' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.3.33-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.3.29-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.3.33-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.3.33-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.5.1.27-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.3.33-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.3.29-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.3.27-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.3.27-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.3.3.1-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_crt-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_cudart-windows-x86_64-13.3.29-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvcc-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvrtc-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\libcublas-windows-x86_64-13.5.1.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\libnvvm-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvtx-windows-x86_64-13.3.29-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_profiler_api-windows-x86_64-13.3.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\visual_studio_integration-windows-x86_64-13.3.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cccl-windows-x86_64-13.3.3.3.1-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
-          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
@@ -73,10 +73,10 @@ android:
    - changed-files:
        - any-glob-to-any-file:
            - examples/llama.android/**
-server/ui:
+server/webui:
    - changed-files:
        - any-glob-to-any-file:
-            - tools/ui/**
+            - tools/server/webui/**
 server:
    - changed-files:
        - any-glob-to-any-file:
@@ -22,9 +22,9 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
  ubuntu-24-llguidance:
@@ -31,7 +31,7 @@ jobs:
  android-ndk-snapdragon:
    runs-on: ubuntu-latest
    container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.7'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
    defaults:
      run:
        shell: bash
@@ -61,7 +61,7 @@ jobs:
  linux-iot-snapdragon:
    runs-on: ubuntu-latest
    container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.7'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.1'
    defaults:
      run:
        shell: bash
@@ -27,12 +27,12 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  default:
+  android:
    runs-on: ubuntu-latest

    steps:
@@ -58,7 +58,7 @@ jobs:
          cd examples/llama.android
          ./gradlew build --no-daemon

-  ndk:
+  android-ndk:
    runs-on: ubuntu-latest
    container:
      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
@@ -73,11 +73,6 @@ jobs:
          fetch-depth: 0
          lfs: false

-      - name: Dependencies
-        run: |
-          apt-get update
-          apt-get install -y build-essential
-
      - name: Build
        id: ndk_build
        run: |
@@ -91,59 +86,3 @@ jobs:
        with:
          name: llama-cpp-android-arm64-cpu
          path: pkg-adb/llama.cpp
-
-  arm64:
-    runs-on: ubuntu-latest
-
-    env:
-      NDK_VERSION: "29.0.14206865"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      # note : disabled to spare some cache space (https://github.com/ggml-org/llama.cpp/pull/23789)
-      #        for some reason, the ccache does not improve the build time in this case
-      # example:
-      #   cache off: https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78160400831
-      #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
-      #
-      #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.21
-      #  with:
-      #    key: android-ubuntu-arm64
-      #    evict-old-files: 1d
-      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Set up JDK
-        uses: actions/setup-java@v5
-        with:
-          java-version: 17
-          distribution: temurin
-
-      - name: Setup Android SDK
-        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
-        with:
-          log-accepted-android-sdk-licenses: false
-
-      - name: Install NDK
-        run: |
-          sdkmanager "ndk;${{ env.NDK_VERSION }}"
-          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
-            -DANDROID_ABI=arm64-v8a \
-            -DANDROID_PLATFORM=android-28 \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_NATIVE=OFF \
-            -DGGML_CPU_ALL_VARIANTS=ON \
-            -DGGML_OPENMP=OFF \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
@@ -32,12 +32,12 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  macos-latest-arm64:
+  macOS-latest-ios:
    runs-on: macos-latest

    steps:
@@ -48,7 +48,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: apple-arm64
+          key: macOS-latest-ios
          evict-old-files: 1d
          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

@@ -56,58 +56,18 @@ jobs:
        id: cmake_build
        run: |
          sysctl -a
-          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
+          cmake -B build -G Xcode \
            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=OFF \
-            -DGGML_METAL_SHADER_DEBUG=ON \
-            -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
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main -E "test-llama-archs" --verbose --timeout 900
-
-  macos-latest-x64:
-    runs-on: macos-15-intel
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: apple-x64
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
-          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
-          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

  macos-latest-ios-xcode:
    runs-on: macos-latest
@@ -129,7 +89,6 @@ jobs:
            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
@@ -156,7 +115,7 @@ jobs:
          xcodebuild -downloadPlatform iOS
          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

-  macos-latest-tvos:
+  macOS-latest-tvos:
    runs-on: macos-latest

    steps:
@@ -164,11 +123,10 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      # TODO: this likely does not do anything - if yes, remove it
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: apple-tvos
+          key: macOS-latest-tvos
          evict-old-files: 1d
          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

@@ -180,7 +138,6 @@ jobs:
            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
@@ -190,7 +147,7 @@ jobs:
            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

-  macos-latest-visionos:
+  macOS-latest-visionos:
    runs-on: macos-latest

    steps:
@@ -198,14 +155,6 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      # TODO: this likely does not do anything - if yes, remove it
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: apple-visionos
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
      - name: Build
        id: cmake_build
        run: |
@@ -214,7 +163,6 @@ jobs:
            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
@@ -224,7 +172,7 @@ jobs:
            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

-  macos-latest-swift:
+  macOS-latest-swift:
    runs-on: macos-latest
    needs: macos-latest-ios-xcode

@@ -237,11 +185,10 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      # TODO: this likely does not do anything - if yes, remove it
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: apple-swift
+          key: macOS-latest-swift
          evict-old-files: 1d
          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

@@ -259,7 +206,6 @@ jobs:
            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TOOLS=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
@@ -28,7 +28,7 @@ jobs:
        id: cache-sdk
        with:
          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}

      - name: Setup Vulkan SDK
        if: steps.cache-sdk.outputs.cache-hit != 'true'
@@ -54,7 +54,7 @@ jobs:
  #      id: cache-toolchain
  #      with:
  #        path: ./spacemit_toolchain
-  #        key: cache-gha-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+  #        key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}

  #    - name: Setup SpacemiT Toolchain
  #      if: steps.cache-toolchain.outputs.cache-hit != 'true'
@@ -81,7 +81,7 @@ jobs:
        id: cache-openvino
        with:
          path: ./openvino_toolkit
-          key: cache-gha-openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}

      - name: Setup OpenVINO Toolkit
        if: steps.cache-openvino.outputs.cache-hit != 'true'
@@ -108,7 +108,7 @@ jobs:
        id: cache-rocm
        with:
          path: C:\Program Files\AMD\ROCm
-          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}

      - name: Setup ROCm
        if: steps.cache-rocm.outputs.cache-hit != 'true'
@@ -29,76 +29,74 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  openEuler-latest-cann:
-#    defaults:
-#      run:
-#        shell: bash -el {0}
-#    strategy:
-#      matrix:
-#        arch: [x86, aarch64]
-#        chip_type: ['910b', '310p']
-#        build: ['Release']
-#        use_acl_graph: ['on', 'off']
-#        exclude:
-#          # 310P does not support USE_ACL_GRAPH=on
-#          - chip_type: '310p'
-#            use_acl_graph: 'on'
-#    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-#    steps:
-#      - name: Checkout
-#        uses: actions/checkout@v6
-#        with:
-#          fetch-depth: 0
-#
-#      - name: Free up disk space
-#        uses: ggml-org/free-disk-space@v1.3.1
-#        with:
-#          tool-cache: true
-#
-#      - name: Set container image
-#        id: cann-image
-#        run: |
-#          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
-#          echo "image=${image}" >> "${GITHUB_OUTPUT}"
-#
-#      - name: Pull container image
-#        run: docker pull "${{ steps.cann-image.outputs.image }}"
-#
-#      - name: Build
-#        env:
-#          BUILD_TYPE: ${{ matrix.build }}
-#          SOC_TYPE: ascend${{ matrix.chip_type }}
-#          USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
-#        run: |
-#          HOST_UID=$(id -u)
-#          HOST_GID=$(id -g)
-#
-#          docker run --rm \
-#            -v "${PWD}:/workspace" \
-#            -w /workspace \
-#            -e SOC_TYPE=${SOC_TYPE} \
-#            -e BUILD_TYPE=${BUILD_TYPE} \
-#            -e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
-#            "${{ steps.cann-image.outputs.image }}" \
-#            bash -lc '
-#              set -e
-#              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
-#              yum clean all && rm -rf /var/cache/yum
-#              git config --global --add safe.directory "/workspace"
-#              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
-#              cmake -S . -B build \
-#                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
-#                  -DGGML_CANN=on \
-#                  -DSOC_TYPE=${SOC_TYPE} \
-#                  -DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
-#              cmake --build build -j $(nproc)
-#
-#              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
-#            '
+  openEuler-latest-cann:
+    defaults:
+      run:
+        shell: bash -el {0}
+    strategy:
+      matrix:
+        arch: [x86, aarch64]
+        chip_type: ['910b', '310p']
+        build: ['Release']
+        use_acl_graph: ['on', 'off']
+        exclude:
+          # 310P does not support USE_ACL_GRAPH=on
+          - chip_type: '310p'
+            use_acl_graph: 'on'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
+      - name: Set container image
+        id: cann-image
+        run: |
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
+          echo "image=${image}" >> "${GITHUB_OUTPUT}"
+
+      - name: Pull container image
+        run: docker pull "${{ steps.cann-image.outputs.image }}"
+
+      - name: Build
+        env:
+          BUILD_TYPE: ${{ matrix.build }}
+          SOC_TYPE: ascend${{ matrix.chip_type }}
+          USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
+        run: |
+          HOST_UID=$(id -u)
+          HOST_GID=$(id -g)
+
+          docker run --rm \
+            -v "${PWD}:/workspace" \
+            -w /workspace \
+            -e SOC_TYPE=${SOC_TYPE} \
+            -e BUILD_TYPE=${BUILD_TYPE} \
+            -e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
+            "${{ steps.cann-image.outputs.image }}" \
+            bash -lc '
+              set -e
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
+              yum clean all && rm -rf /var/cache/yum
+              git config --global --add safe.directory "/workspace"
+              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+              cmake -S . -B build \
+                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DGGML_CANN=on \
+                  -DSOC_TYPE=${SOC_TYPE} \
+                  -DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
+              cmake --build build -j $(nproc)
+
+              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
+            '
@@ -5,23 +5,23 @@ on:

 jobs:
  linux:
-    runs-on: [self-hosted, Linux, CPU]
+    runs-on: ubuntu-slim
    steps:
      - uses: actions/checkout@v6
        with:
          fetch-depth: 0

+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential tcl cmake
+
      - name: Build
        run: |
          PREFIX="$(pwd)"/inst
-          cmake -S . -B build \
-                -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_OPENSSL=OFF \
-                -DLLAMA_BUILD_TESTS=OFF \
-                -DLLAMA_BUILD_TOOLS=OFF \
-                -DLLAMA_BUILD_EXAMPLES=OFF \
-                -DLLAMA_BUILD_APP=OFF \
-                -DCMAKE_BUILD_TYPE=Release
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_OPENSSL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
          cmake --build build --config Release
          cmake --install build --prefix "$PREFIX" --config Release

@@ -1,215 +0,0 @@
-name: CI (cpu)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-cpu.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-cpu.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-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:
-  build-cmake-pkg:
-    uses: ./.github/workflows/build-cmake-pkg.yml
-
-  ubuntu:
-    strategy:
-      matrix:
-        include:
-          - build: 'x64'
-            os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-24.04-arm
-
-    runs-on: ${{ matrix.os }}
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cpu-${{ matrix.os }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build Dependencies
-        id: build_depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-            python3 python3-pip python3-dev python3-wheel \
-            libjpeg-dev build-essential libssl-dev \
-            git-lfs
-
-      - name: Toolchain workaround (GCC 14)
-        if: ${{ contains(matrix.os, 'ubuntu-24.04') }}
-        run: |
-          sudo apt-get install -y gcc-14 g++-14
-          echo "CC=gcc-14" >> "$GITHUB_ENV"
-          echo "CXX=g++-14" >> "$GITHUB_ENV"
-
-      - name: Python Dependencies
-        id: python_depends
-        run: |
-          export PIP_BREAK_SYSTEM_PACKAGES="1"
-          python3 -m pip install --upgrade pip setuptools
-          pip3 install ./gguf-py
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Test llama2c conversion
-        id: llama2c_test
-        run: |
-          cd build
-          echo "Fetch tokenizer"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
-          echo "Fetch llama2c model"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
-          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
-          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
-
-  windows:
-    runs-on: windows-2025
-
-    env:
-      OPENBLAS_VERSION: 0.3.23
-      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.313.2
-
-    strategy:
-      matrix:
-        include:
-          - build: 'x64-cpu-static'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
-          - build: 'x64-openblas'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'x64-vulkan'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
-          - build: 'arm64'
-            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cpu-windows-2025-${{ matrix.build }}
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'x64-openblas' }}
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
-          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/openblas
-          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
-          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
-
-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == '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: |
-          choco install ninja
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -S . -B build ${{ matrix.defines }} `
-            -DLLAMA_BUILD_BORINGSSL=ON
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: Add libopenblas.dll
-        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'x64-openblas' }}
-        run: |
-          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
-          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
-
-      - name: Test
-        id: cmake_test
-        if: ${{ matrix.arch == 'x64' }}
-        run: |
-          cd build
-          ctest -L main -C Release --verbose --timeout 900
-
-      # TODO: disabled for now, consider adding tests for all CPU variants instead
-      # - name: Test (Intel SDE)
-      #   id: cmake_test_sde
-      #   if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
-      #   run: |
-      #     curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
-      #     # for some weird reason windows tar doesn't like sde tar.xz
-      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
-      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
-      #     $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
-      #     cd build
-      #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
-      #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
@@ -277,7 +277,7 @@ jobs:

    env:
      # Make sure this is in sync with build-cache.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.2.4"
+      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"

    steps:
      - uses: actions/checkout@v6
@@ -287,7 +287,7 @@ jobs:
      #  id: cache-toolchain
      #  with:
      #    path: ./spacemit_toolchain
-      #    key: cache-gha-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+      #    key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}

      - name: Setup SpacemiT Toolchain
        #if: steps.cache-toolchain.outputs.cache-hit != 'true'
@@ -1,134 +0,0 @@
-name: CI (CUDA, ubuntu)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-cuda-ubuntu.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-cuda-ubuntu.yml',
-      'ggml/src/ggml-cuda/**'
-    ]
-
-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:
-  cuda:
-    runs-on: ubuntu-24.04
-    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Install dependencies
-        env:
-          DEBIAN_FRONTEND: noninteractive
-        run: |
-          apt update
-          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cuda-ubuntu-24.04-cuda
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build with CMake
-        # 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: |
-          cmake -S . -B build -G Ninja \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DCMAKE_CUDA_ARCHITECTURES=89-real \
-            -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
-            -DGGML_NATIVE=OFF \
-            -DGGML_CUDA=ON \
-            -DGGML_CUDA_CUB_3DOT2=ON
-          cmake --build build
-
-  hip:
-    runs-on: ubuntu-22.04
-    container: rocm/dev-ubuntu-22.04:6.1.2
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cuda-ubuntu-22.04-hip
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build with native CMake HIP support
-        id: cmake_build
-        run: |
-          cmake -B build -S . \
-            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
-            -DGPU_TARGETS="gfx1030" \
-            -DGGML_HIP=ON
-          cmake --build build --config Release -j $(nproc)
-
-  musa:
-    runs-on: ubuntu-22.04
-    container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          apt-get update
-          apt-get install -y build-essential git cmake libssl-dev
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: cuda-ubuntu-22.04-musa
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build with native CMake MUSA support
-        id: cmake_build
-        run: |
-          cmake -B build -S . \
-            -DGGML_MUSA=ON
-          time cmake --build build --config Release -j $(nproc)
@@ -1,162 +0,0 @@
-name: CI (CUDA, windows)
-
-# TODO: this workflow is only triggered manually because it is very heavy on the CI
-#       when we provision dedicated windows runners, we can enable it for pushes too
-# note: running this workflow manually will populate the ccache for the release builds
-#       this can be used before merging a PR to speed up the release workflow
-on:
-  workflow_dispatch: # allows manual triggering
-
-# note: this will run in queue with the release workflow
-concurrency:
-  group: release
-  queue: max
-
-env:
-  GH_TOKEN: ${{ github.token }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  cuda:
-    runs-on: windows-2022
-
-    permissions:
-      actions: write
-
-    strategy:
-      matrix:
-        cuda: ['12.4', '13.3']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
-
-      - name: Install Cuda Toolkit
-        uses: ./.github/actions/windows-setup-cuda
-        with:
-          cuda_version: ${{ matrix.cuda }}
-
-      - name: Install Ninja
-        id: install_ninja
-        run: |
-          choco install ninja
-
-      - name: Build
-        id: cmake_build
-        shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
-        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
-          cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
-            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
-            -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
-          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
-
-  hip:
-    runs-on: windows-2022
-
-    permissions:
-      actions: write
-
-    env:
-      # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-
-    strategy:
-      matrix:
-        include:
-          # sync with release.yml
-          - name: "radeon"
-            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - 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'
-        uses: ./.github/actions/windows-setup-rocm
-        with:
-          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
-
-      - 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
-          }
-          & $clangPath.FullName --version
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          # TODO: this build does not match the build in release.yml, so we use a different cache key
-          #       ideally, the builds should match, similar to the CUDA build above so that we would be able
-          #       to populate the ccache for the release with manual runs of this workflow
-          #key: release-windows-2022-x64-hip-${{ env.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_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 `
-            -DROCM_DIR="${env:HIP_PATH}" `
-            -DGGML_HIP=ON `
-            -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.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
-          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
@@ -1,150 +0,0 @@
-name: CI (ibm)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-ibm.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-ibm.yml',
-      'ggml/src/ggml-cpu/**'
-    ]
-
-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-s390x:
-    runs-on: ubuntu-24.04-s390x
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build Dependencies
-        id: build_depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-            python3 python3-pip python3-dev python3-wheel \
-            libjpeg-dev build-essential libssl-dev \
-            git-lfs
-
-      - name: Toolchain workaround (GCC 14)
-        run: |
-          sudo apt-get install -y gcc-14 g++-14
-          echo "CC=gcc-14" >> "$GITHUB_ENV"
-          echo "CXX=g++-14" >> "$GITHUB_ENV"
-
-      - name: Python Dependencies
-        id: python_depends
-        run: |
-          export PIP_BREAK_SYSTEM_PACKAGES="1"
-          python3 -m pip install --upgrade pip setuptools
-          pip3 install ./gguf-py
-
-      - name: Swap Endianness
-        id: endianness
-        run: |
-          for f in models/*.gguf; do
-            echo YES | python3 gguf-py/gguf/scripts/gguf_convert_endian.py $f big
-          done
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Test llama2c (s390x)
-        id: llama2c_test_s390x
-        run: |
-          cd build
-          echo "Fetch llama2c big-endian model"
-          wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
-          ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
-
-  ubuntu-24-ppc64le:
-    runs-on: ubuntu-24.04-ppc64le
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build Dependencies
-        id: build_depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends \
-            python3 python3-pip python3-dev python3-wheel \
-            libjpeg-dev build-essential libssl-dev \
-            git-lfs
-
-      - name: Toolchain workaround (GCC 14)
-        run: |
-          sudo apt-get install -y gcc-14 g++-14
-          echo "CC=gcc-14" >> "$GITHUB_ENV"
-          echo "CXX=g++-14" >> "$GITHUB_ENV"
-
-      - name: Python Dependencies
-        id: python_depends
-        run: |
-          export PIP_BREAK_SYSTEM_PACKAGES="1"
-          python3 -m pip install --upgrade pip setuptools
-          pip3 install ./gguf-py
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Test llama2c conversion
-        id: llama2c_test
-        run: |
-          cd build
-          echo "Fetch tokenizer"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
-          echo "Fetch llama2c model"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
-          ./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
@@ -15,9 +15,9 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
  windows-msys2:
@@ -37,7 +37,7 @@ jobs:
      #- name: ccache
      #  uses: ggml-org/ccache-action@v1.2.16
      #  with:
-      #    key: msys-windows-2025-x64
+      #    key: windows-msys2
      #    variant: ccache
      #    evict-old-files: 1d
      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -1,82 +0,0 @@
-name: CI (opencl)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-opencl.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cl'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-opencl.yml',
-      'ggml/src/ggml-opencl/**'
-    ]
-
-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:
-  windows-2025-opencl-adreno:
-    runs-on: windows-2025
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: opencl-windows-2025-x64
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Install Ninja
-        id: install_ninja
-        run: |
-          choco install ninja
-
-      - name: Install OpenCL Headers and Libs
-        id: install_opencl
-        run: |
-          git clone https://github.com/KhronosGroup/OpenCL-Headers
-          cd OpenCL-Headers
-          cmake -B build `
-            -DBUILD_TESTING=OFF `
-            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
-            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
-            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
-          cmake --build build --target install
-          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
-          cd OpenCL-ICD-Loader
-          cmake -B build-arm64-release `
-            -A arm64 `
-            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
-            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
-          cmake --build build-arm64-release --target install --config release
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON -DLLAMA_BUILD_BORINGSSL=ON
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
@@ -29,18 +29,30 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
  ubuntu-24-openvino:
-    runs-on: [self-hosted, Linux, Intel, OpenVINO]
+    name: ubuntu-24-openvino-${{ matrix.openvino_device }}

    concurrency:
-      group: openvino-gpu-${{ github.head_ref || github.ref }}
+      group: openvino-${{ matrix.variant }}-${{ github.head_ref || github.ref }}
      cancel-in-progress: false

+    strategy:
+      matrix:
+        include:
+          - variant: cpu
+            runner: '"ubuntu-24.04"'
+            openvino_device: "CPU"
+          - variant: gpu
+            runner: '["self-hosted","Linux","Intel","OpenVINO"]'
+            openvino_device: "GPU"
+
+    runs-on: ${{ fromJSON(matrix.runner) }}
+
    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
      OPENVINO_VERSION_MAJOR: "2026.0"
@@ -51,6 +63,14 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

+      - name: ccache
+        if: runner.environment == 'github-hosted'
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
      - name: Dependencies
        id: depends
        run: |
@@ -58,7 +78,16 @@ jobs:
          sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
          sudo apt-get install -y ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd

+      - name: Use OpenVINO Toolkit Cache
+        if: runner.environment == 'github-hosted'
+        uses: actions/cache@v5
+        id: cache-openvino
+        with:
+          path: ./openvino_toolkit
+          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+
      - name: Setup OpenVINO Toolkit
+        if: steps.cache-openvino.outputs.cache-hit != 'true'
        uses: ./.github/actions/linux-setup-openvino
        with:
          path: ./openvino_toolkit
@@ -80,17 +109,12 @@ jobs:
            -DGGML_OPENVINO=ON
          time cmake --build build/ReleaseOV --config Release -j $(nproc)

-      - name: Test (CPU)
-        id: cmake_test_cpu
+      - name: Test
+        id: cmake_test
        # 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" --verbose --timeout 2000
-
-      - name: Test (GPU)
-        id: cmake_test_gpu
-        # TODO: fix and re-enable the `test-llama-archs` test below
-        run: |
-          cd ${{ github.workspace }}
-          export GGML_OPENVINO_DEVICE=GPU
+          if [ "${{ matrix.openvino_device }}" = "GPU" ]; then
+            export GGML_OPENVINO_DEVICE=GPU
+          fi
          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
@@ -29,84 +29,11 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  ubuntu-cpu-riscv64-native:
-    runs-on: ubuntu-24.04-riscv
-
-    steps:
-      - name: Install dependencies
-        run: |
-          # Install necessary packages
-          sudo apt-get update
-          sudo apt-get install -y libssl-dev
-
-          # Set gcc-14 and g++-14 as the default compilers
-          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
-          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
-
-          git lfs install
-
-      - name: Check environment
-        run: |
-          uname -a
-          gcc --version
-          g++ --version
-          ldd --version
-          cmake --version
-          rustc --version
-          env
-          echo "nproc=$(nproc)"
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
-      #- name: ccache
-      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
-      #  with:
-      #    key: riscv-ubuntu-native
-      #    evict-old-files: 1d
-      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_OPENMP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=ON \
-            -DLLAMA_BUILD_TOOLS=ON \
-            -DLLAMA_BUILD_TESTS=ON \
-            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
-            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-            -DGGML_RPC=ON \
-            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
-
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Test llama2c conversion
-        id: llama2c_test
-        run: |
-          cd build
-          echo "Fetch tokenizer"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
-          echo "Fetch llama2c model"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
-          ./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
-
  ubuntu-riscv64-native-sanitizer:
    runs-on: ubuntu-24.04-riscv

@@ -135,13 +62,12 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
-      #- name: ccache
-      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
-      #  with:
-      #    key: riscv-ubuntu-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
-      #    evict-old-files: 1d
-      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+        with:
+          key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

      - name: Build
        id: cmake_build
@@ -1,66 +0,0 @@
-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
@@ -22,65 +22,66 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  ctest:
-    runs-on: [self-hosted, X64, CPU, Linux]
+  ubuntu-latest-sanitizer:
+    runs-on: ubuntu-latest

    continue-on-error: true

    strategy:
      matrix:
        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug]

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v6

-      # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
-      - name: Build (undefined)
-        id: cmake_build_undefined
-        if: ${{ matrix.sanitizer == 'UNDEFINED' }}
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Debug \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-latest-sanitizer-${{ matrix.sanitizer }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

-          cmake --build build --config Debug -j $(nproc)
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libssl-dev

      - name: Build
        id: cmake_build
-        if: ${{ matrix.sanitizer == 'ADDRESS' }}
+        if: ${{ matrix.sanitizer != 'THREAD' }}
        run: |
          cmake -B build \
-            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}

-          cmake --build build --config RelWithDebInfo -j $(nproc)
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)

      - name: Build (no OpenMP)
        id: cmake_build_no_openmp
        if: ${{ matrix.sanitizer == 'THREAD' }}
        run: |
          cmake -B build \
-            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
            -DGGML_OPENMP=OFF

-          cmake --build build --config RelWithDebInfo -j $(nproc)
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)

      - name: Test
        id: cmake_test
-        # skip run in Debug - very slow
-        if: ${{ matrix.sanitizer != 'UNDEFINED' }}
        run: |
          cd build
-          ctest -L main -E tokenizer --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
@@ -50,12 +50,27 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  gpu-cuda:
+  determine-tag:
+    name: Determine tag name
+    runs-on: ubuntu-slim
+    outputs:
+      tag_name: ${{ steps.tag.outputs.name }}
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+  ggml-ci-nvidia-cuda:
+    needs: determine-tag
    runs-on: [self-hosted, Linux, NVIDIA]

    steps:
@@ -65,11 +80,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          nvidia-smi
-          GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

-  gpu-vulkan-nvidia-cm:
+  ggml-ci-nvidia-vulkan-cm:
+    needs: determine-tag
    runs-on: [self-hosted, Linux, NVIDIA]

    steps:
@@ -79,11 +97,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

-  gpu-vulkan-nvidia-cm2:
+  ggml-ci-nvidia-vulkan-cm2:
+    needs: determine-tag
    runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]

    steps:
@@ -93,12 +114,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

-  gpu-webgpu-nvidia:
-    runs-on: [self-hosted, Linux, NVIDIA, X64]
+  ggml-ci-nvidia-webgpu:
+    runs-on: [self-hosted, Linux, NVIDIA]

    steps:
      - name: Clone
@@ -124,10 +147,10 @@ jobs:
          GG_BUILD_WEBGPU=1 \
          GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
          GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
-            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+            bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

  # TODO: provision AMX-compatible machine
-  #cpu-amx:
+  #ggml-ci-cpu-amx:
  #  runs-on: [self-hosted, Linux, CPU, AMX]

  #  steps:
@@ -138,10 +161,10 @@ jobs:
  #    - name: Test
  #      id: ggml-ci
  #      run: |
-  #        bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  #        bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

  # TODO: provision AMD GPU machine
-  # amd-vulkan:
+  # ggml-ci-amd-vulkan:
  #   runs-on: [self-hosted, Linux, AMD]

  #   steps:
@@ -153,10 +176,10 @@ jobs:
  #       id: ggml-ci
  #       run: |
  #         vulkaninfo --summary
-  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

  # TODO: provision AMD GPU machine
-  # amd-rocm:
+  # ggml-ci-amd-rocm:
  #   runs-on: [self-hosted, Linux, AMD]

  #   steps:
@@ -168,9 +191,10 @@ jobs:
  #       id: ggml-ci
  #       run: |
  #         amd-smi static
-  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp

-  gpu-metal:
+  ggml-ci-mac-metal:
+    needs: determine-tag
    runs-on: [self-hosted, macOS, ARM64]

    steps:
@@ -180,10 +204,13 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

-  gpu-webgpu-apple:
+  ggml-ci-mac-webgpu:
+    needs: determine-tag
    runs-on: [self-hosted, macOS, ARM64]

    steps:
@@ -206,11 +233,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

-  gpu-vulkan-apple:
+  ggml-ci-mac-vulkan:
+    needs: determine-tag
    runs-on: [self-hosted, macOS, ARM64]

    steps:
@@ -220,11 +250,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          vulkaninfo --summary
          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

-  gpu-vulkan-intel-linux:
+  ggml-ci-linux-intel-vulkan:
+    needs: determine-tag
    runs-on: [self-hosted, Linux, Intel]

    steps:
@@ -236,11 +269,14 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          vulkaninfo --summary
          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

-  gpu-vulkan-intel-windows:
+  ggml-ci-win-intel-vulkan:
+    needs: determine-tag
    runs-on: [self-hosted, Windows, X64, Intel]

    steps:
@@ -255,13 +291,15 @@ jobs:
          MSYSTEM: UCRT64
          CHERE_INVOKING: 1
          PATH: C:\msys64\ucrt64\bin;C:\msys64\usr\bin;C:\Windows\System32;${{ env.PATH }}
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          vulkaninfo --summary
          # Skip python related tests with GG_BUILD_LOW_PERF=1 since Windows MSYS2 UCRT64 currently fails to create
          # a valid python environment for testing
          LLAMA_FATAL_WARNINGS=OFF GG_BUILD_NINJA=1 GG_BUILD_VULKAN=1 GG_BUILD_LOW_PERF=1 ./ci/run.sh ./results/llama.cpp ./mnt/llama.cpp

-  gpu-openvino-low-perf:
+  ggml-ci-intel-openvino-gpu-low-perf:
+    needs: determine-tag
    runs-on: [self-hosted, Linux, Intel, OpenVINO]

    concurrency:
@@ -293,99 +331,8 @@ jobs:

      - name: Test
        id: ggml-ci
+        env:
+          HF_WEBUI_VERSION: ${{ needs.determine-tag.outputs.tag_name }}
        run: |
          source ./openvino_toolkit/setupvars.sh
-          GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  cpu-x64-high-perf:
-    runs-on: [self-hosted, Linux, X64]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  cpu-arm64-high-perf-graviton4:
-    runs-on: ah-ubuntu_22_04-c8g_8x
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          set -euxo pipefail
-          sudo apt-get update
-          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
-          apt-get install -y \
-          build-essential \
-          python3-venv \
-          gpg \
-          wget \
-          time \
-          git-lfs
-
-          git lfs install
-
-          # install the latest cmake
-          sudo install -d /usr/share/keyrings
-          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
-            | gpg --dearmor \
-            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
-          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
-            | sudo tee /etc/apt/sources.list.d/kitware.list
-          sudo apt-get update
-          sudo apt-get install -y cmake
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-
-  cpu-arm64-graviton4-kleidiai:
-    runs-on: ah-ubuntu_22_04-c8g_8x
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          set -euxo pipefail
-          sudo apt-get update
-          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
-          apt-get install -y \
-          build-essential \
-          python3-venv \
-          gpg \
-          wget \
-          time \
-          git-lfs
-
-          git lfs install
-
-          # install the latest cmake
-          sudo install -d /usr/share/keyrings
-          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
-            | gpg --dearmor \
-            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
-          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
-            | sudo tee /etc/apt/sources.list.d/kitware.list
-          sudo apt-get update
-          sudo apt-get install -y cmake
-
-      - name: Test
-        id: ggml-ci
-        run: |
-          GG_BUILD_KLEIDIAI=1 \
-          GG_BUILD_EXTRA_TESTS_0=1 \
-          bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
@@ -29,134 +29,132 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:

-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  ubuntu-24-sycl:
-#    strategy:
-#      matrix:
-#        build: [fp32]
-#        include:
-#          - build: fp32
-#            fp16: OFF
-#
-#    runs-on: ubuntu-24.04
-#
-#    env:
-#      ONEAPI_ROOT: /opt/intel/oneapi/
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#      LEVEL_ZERO_VERSION: "1.28.2"
-#      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
-#
-#    continue-on-error: true
-#
-#    steps:
-#      - uses: actions/checkout@v6
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          cd /tmp
-#          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-#          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
-#
-#      - name: Install Level Zero SDK
-#        shell: bash
-#        run: |
-#          cd /tmp
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
-#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
-#          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
-#
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: sycl-ubuntu-24-${{ matrix.build }}
-#          evict-old-files: 1d
-#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          source /opt/intel/oneapi/setvars.sh
-#          cmake -B build \
-#            -G "Ninja" \
-#            -DCMAKE_BUILD_TYPE=Release \
-#            -DGGML_SYCL=ON \
-#            -DCMAKE_C_COMPILER=icx \
-#            -DCMAKE_CXX_COMPILER=icpx \
-#            -DLLAMA_OPENSSL=OFF \
-#            -DGGML_NATIVE=OFF \
-#            -DGGML_SYCL_F16=${{ matrix.fp16 }}
-#          time cmake --build build --config Release -j $(nproc)
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON

-# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
-#       in order to enable it again, we have to provision dedicated runners  to run it
-#  windows-latest-sycl:
-#    runs-on: windows-2022
-#
-#    defaults:
-#      run:
-#        shell: bash
-#
-#    env:
-#      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
-#      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-#      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
-#      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: Use oneAPI Installation Cache
-#        uses: actions/cache@v5
-#        id: cache-sycl
-#        with:
-#          path: ${{ env.ONEAPI_ROOT }}
-#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-#
-#      - name: Download & Install oneAPI
-#        shell: bash
-#        if: steps.cache-sycl.outputs.cache-hit != 'true'
-#        run: |
-#          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-#
-#      - name: Install Level Zero SDK
-#        shell: pwsh
-#        run: |
-#          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
-#          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
-#          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
-#
-#      - name: ccache
-#        uses: ggml-org/ccache-action@v1.2.21
-#        with:
-#          key: sycl-windows-latest
-#          variant: ccache
-#          evict-old-files: 1d
-#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-#
-#      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-#
-#      - name: Build
-#        id: cmake_build
-#        run:  examples/sycl/win-build-sycl.bat
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      LEVEL_ZERO_VERSION: "1.28.2"
+      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: Install Level Zero SDK
+        shell: bash
+        run: |
+          cd /tmp
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-sycl-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+  windows-latest-sycl:
+    runs-on: windows-2022
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Install Level Zero SDK
+        shell: pwsh
+        run: |
+          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
+          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
+          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-latest-sycl
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat
@@ -31,49 +31,12 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
-  ubuntu-arm64:
-    runs-on: ubuntu-24.04-arm
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev spirv-headers libssl-dev ninja-build
-          echo "CC=gcc-14" >> "$GITHUB_ENV"
-          echo "CXX=g++-14" >> "$GITHUB_ENV"
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: vulkan-ubuntu-24.04-arm-new
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Configure
-        id: cmake_configure
-        run: |
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_VULKAN=ON
-
-      - name: Build
-        id: cmake_build
-        run: |
-          time cmake --build build -j $(nproc)
-
-  ubuntu-llvmpipe:
+  ubuntu-24-vulkan-llvmpipe:
    runs-on: ubuntu-24.04

    steps:
@@ -81,6 +44,13 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-vulkan-llvmpipe
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
      - name: Dependencies
        id: depends
        run: |
@@ -98,7 +68,7 @@ jobs:
        id: cache-sdk
        with:
          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}

      - name: Setup Vulkan SDK
        if: steps.cache-sdk.outputs.cache-hit != 'true'
@@ -107,13 +77,6 @@ jobs:
          path: ./vulkan_sdk
          version: ${{ env.VULKAN_SDK_VERSION }}

-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: vulkan-ubuntu-24.04-llvmpipe
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
      - name: Build
        id: cmake_build
        run: |
@@ -1,173 +0,0 @@
-name: CI (webgpu)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-webgpu.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.wgsl'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-webgpu.yml',
-      'ggml/src/ggml-webgpu/**'
-    ]
-
-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:
-  macos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: webgpu-macos-latest
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v20260317.182325"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Build
-        id: cmake_build
-        run: |
-          export CMAKE_PREFIX_PATH=dawn
-          cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
-          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-  ubuntu:
-    runs-on: ubuntu-24.04
-
-    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
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo add-apt-repository -y ppa:kisak/kisak-mesa
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers \
-            libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v20260317.182325"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Build
-        id: cmake_build
-        run: |
-          export Dawn_DIR=dawn/lib64/cmake/Dawn
-          cmake -B build \
-            -DGGML_WEBGPU=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          # 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)
@@ -19,7 +19,7 @@ on:

 jobs:
  check-vendor:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim

    steps:
      - name: Checkout
@@ -15,7 +15,7 @@ concurrency:

 jobs:
  model-naming:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    steps:
      - uses: actions/checkout@v6
      - name: Check model naming conventions
@@ -11,11 +11,6 @@ name: Publish Docker image

 on:
  workflow_dispatch: # allows manual triggering
-    inputs:
-      skip_s390x:
-        description: "Skip the s390x build target (useful for fast test runs that do not need the IBM Z runner)"
-        type: boolean
-        default: false
  schedule:
    # Rebuild daily rather than on every push because it is expensive
    - cron: '12 4 * * *'
@@ -69,8 +64,6 @@ jobs:
      - name: Generate build and merge matrices
        id: matrices
        shell: bash
-        env:
-          SKIP_S390X: ${{ inputs.skip_s390x || 'false' }}
        run: |
          set -euo pipefail

@@ -93,11 +86,6 @@ jobs:
          ]
          JSON

-          if [ "${SKIP_S390X}" = "true" ]; then
-            jq 'map(select(.platforms != "linux/s390x"))' build-matrix.json > build-matrix.json.tmp
-            mv build-matrix.json.tmp build-matrix.json
-          fi
-
          BUILD_MATRIX="$(jq -c . build-matrix.json)"
          MERGE_MATRIX="$(jq -c '
            reduce .[] as $entry ({}; .[$entry.tag] |= (
@@ -144,7 +132,6 @@ jobs:
        config: ${{ fromJSON(needs.prepare_matrices.outputs.build_matrix) }}
    steps:
      - name: Check out the repo
-        id: checkout
        uses: actions/checkout@v6
        with:
          fetch-depth: 0
@@ -200,10 +187,6 @@ jobs:
        env:
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'

-      - name: Get build date
-        id: build_date
-        run: echo "date=$(date -u +"%Y-%m-%dT%H:%M:%SZ")" >> $GITHUB_OUTPUT
-
      - name: Free Disk Space (Ubuntu)
        if: ${{ matrix.config.free_disk_space == true }}
        uses: ggml-org/free-disk-space@v1.3.1
@@ -228,26 +211,13 @@ jobs:
        with:
          context: .
          platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
          file: ${{ matrix.config.dockerfile }}
          target: full
          provenance: false
          build-args: |
-            BUILD_DATE=${{ steps.build_date.outputs.date }}
-            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
-            APP_REVISION=${{ steps.checkout.outputs.commit }}
-            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
-            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          annotations: |
-            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
-            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
-            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
-            manifest:org.opencontainers.image.title=llama.cpp
-            manifest:org.opencontainers.image.description=LLM inference in C/C++
-            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
-            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
          # using github experimental cache
          #cache-from: type=gha
          #cache-to: type=gha,mode=max
@@ -265,26 +235,13 @@ jobs:
        with:
          context: .
          platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
          file: ${{ matrix.config.dockerfile }}
          target: light
          provenance: false
          build-args: |
-            BUILD_DATE=${{ steps.build_date.outputs.date }}
-            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
-            APP_REVISION=${{ steps.checkout.outputs.commit }}
-            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
-            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          annotations: |
-            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
-            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
-            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
-            manifest:org.opencontainers.image.title=llama.cpp
-            manifest:org.opencontainers.image.description=LLM inference in C/C++
-            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
-            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
          # using github experimental cache
          #cache-from: type=gha
          #cache-to: type=gha,mode=max
@@ -302,26 +259,13 @@ jobs:
        with:
          context: .
          platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
          file: ${{ matrix.config.dockerfile }}
          target: server
          provenance: false
          build-args: |
-            BUILD_DATE=${{ steps.build_date.outputs.date }}
-            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
-            APP_REVISION=${{ steps.checkout.outputs.commit }}
-            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
-            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          annotations: |
-            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
-            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
-            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
-            manifest:org.opencontainers.image.title=llama.cpp
-            manifest:org.opencontainers.image.description=LLM inference in C/C++
-            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
-            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
          # using github experimental cache
          #cache-from: type=gha
          #cache-to: type=gha,mode=max
@@ -386,15 +330,10 @@ jobs:

    steps:
      - name: Check out the repo
-        id: checkout
        uses: actions/checkout@v6
        with:
          fetch-depth: 0

-      - name: Get build date
-        id: build_date
-        run: echo "date=$(date -u +"%Y-%m-%dT%H:%M:%SZ")" >> $GITHUB_OUTPUT
-
      - name: Download digest metadata
        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
        with:
@@ -422,8 +361,6 @@ jobs:
          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
          PREFIX="${IMAGE_REPO}:"
          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
-          BUILD_DATE="${{ steps.build_date.outputs.date }}"
-          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
          TAGS="${{ matrix.config.tag }}"
          ARCHES="${{ matrix.config.arches }}"
          DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -475,21 +412,11 @@ jobs:
                  refs+=("${IMAGE_REPO}@${digest}")
              done

-              local annotations=(
-                  --annotation "index:org.opencontainers.image.created=${BUILD_DATE}"
-                  --annotation "index:org.opencontainers.image.version=${SRC_TAG}"
-                  --annotation "index:org.opencontainers.image.revision=${COMMIT_SHA}"
-                  --annotation "index:org.opencontainers.image.title=llama.cpp"
-                  --annotation "index:org.opencontainers.image.description=LLM inference in C/C++"
-                  --annotation "index:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}"
-                  --annotation "index:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}"
-              )
-
              echo "Creating ${merged_tag} from ${refs[*]}"
-              docker buildx imagetools create "${annotations[@]}" --tag "${merged_tag}" "${refs[@]}"
+              docker buildx imagetools create --tag "${merged_tag}" "${refs[@]}"

              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
-              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
+              docker buildx imagetools create --tag "${merged_versioned_tag}" "${refs[@]}"
          }

          for tag in $TAGS; do
@@ -15,7 +15,7 @@ concurrency:

 jobs:
  editorconfig:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    steps:
      - uses: actions/checkout@v6
      - uses: editorconfig-checker/action-editorconfig-checker@840e866d93b8e032123c23bac69dece044d4d84c # v2.2.0
@@ -28,9 +28,9 @@ concurrency:
 env:
  GGML_NLOOP: 3
  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1

 jobs:
  ubuntu-22-hip-quality-check:
@@ -50,7 +50,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: hip-quality-check-ubuntu-22.04
+          key: ubuntu-22-hip-quality-check
          evict-old-files: 1d
          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

@@ -3,16 +3,16 @@ name: Check Pre-Tokenizer Hashes
 on:
    push:
        paths:
-            - 'conversion/base.py'
+            - 'convert_hf_to_gguf.py'
            - 'convert_hf_to_gguf_update.py'
    pull_request:
        paths:
-            - 'conversion/base.py'
+            - 'convert_hf_to_gguf.py'
            - 'convert_hf_to_gguf_update.py'

 jobs:
    pre-tokenizer-hashes:
-        runs-on: [self-hosted, fast]
+        runs-on: ubuntu-slim

        steps:
        - name: Checkout repository
@@ -30,16 +30,16 @@ jobs:

        - name: Update pre-tokenizer hashes
          run: |
-              cp conversion/base.py /tmp
+              cp convert_hf_to_gguf.py /tmp
              .venv/bin/python convert_hf_to_gguf_update.py --check-missing

        - name: Check if committed pre-tokenizer hashes matches generated version
          run: |
-              if ! diff -q conversion/base.py /tmp/base.py; then
-                  echo "Model pre-tokenizer hashes (in conversion/base.py) do not match generated hashes (from convert_hf_to_gguf_update.py)."
-                  echo "To fix: run ./convert_hf_to_gguf_update.py and commit the updated conversion/base.py along with your changes"
+              if ! diff -q convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py; then
+                  echo "Model pre-tokenizer hashes (in convert_hf_to_gguf.py) do not match generated hashes (from convert_hf_to_gguf_update.py)."
+                  echo "To fix: run ./convert_hf_to_gguf_update.py and commit the updated convert_hf_to_gguf.py along with your changes"
                  echo "Differences found:"
-                  diff conversion/base.py /tmp/base.py || true
+                  diff convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py || true
                  exit 1
              fi
              echo "Model pre-tokenizer hashes are up to date."
@@ -20,7 +20,7 @@ concurrency:

 jobs:
  python-check-requirements:
-    runs-on: [self-hosted, CPU, fast]
+    runs-on: ubuntu-slim
    name: check-requirements
    steps:
      - name: Check out source repository
@@ -21,7 +21,7 @@ concurrency:

 jobs:
  flake8-lint:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    name: Lint
    steps:
      - name: Check out source repository
@@ -22,7 +22,7 @@ concurrency:

 jobs:
  python-type-check:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
    name: python type-check
    steps:
      - name: Check out source repository
@@ -26,10 +26,10 @@ on:
    ]

 env:
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  LLAMA_ARG_LOG_VERBOSITY: 10
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
@@ -37,7 +37,7 @@ concurrency:

 jobs:
  server:
-    runs-on: [self-hosted, CPU, Linux, llama-server]
+    runs-on: ubuntu-latest

    strategy:
      matrix:
@@ -46,19 +46,19 @@ jobs:
      fail-fast: false

    steps:
-      #- name: Dependencies
-      #  id: depends
-      #  run: |
-      #    sudo apt-get update
-      #    sudo apt-get -y install \
-      #      build-essential \
-      #      xxd \
-      #      git \
-      #      cmake \
-      #      curl \
-      #      wget \
-      #      language-pack-en \
-      #      libssl-dev
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get -y install \
+            build-essential \
+            xxd \
+            git \
+            cmake \
+            curl \
+            wget \
+            language-pack-en \
+            libssl-dev

      - name: Clone
        id: checkout
@@ -67,13 +67,6 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
      - name: Build
        id: cmake_build
        run: |
@@ -29,17 +29,22 @@ on:
    ]

 env:
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  LLAMA_ARG_LOG_VERBOSITY: 10
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
  cancel-in-progress: true

 jobs:
+  webui-build:
+    name: Build WebUI
+    uses: ./.github/workflows/webui-build.yml
+
  server-metal:
+    needs: webui-build
    runs-on: [self-hosted, llama-server, macOS, ARM64]

    name: server-metal (${{ matrix.wf_name }})
@@ -67,12 +72,11 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download WebUI build artifact
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: webui-build
+          path: tools/server/public/

      - name: Build
        id: cmake_build
@@ -91,106 +95,42 @@ jobs:
          export ${{ matrix.extra_args }}
          pytest -v -x -m "not slow"

-  server-cuda:
-    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
-
-    name: server-cuda (${{ matrix.wf_name }})
-    strategy:
-      matrix:
-        build_type: [Release]
-        wf_name: ["GPUx1"]
-        include:
-          - build_type: Release
-            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-            wf_name:    "GPUx1, backend-sampling"
-      fail-fast: false
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
-        run: |
-          cd tools/server/tests
-          python3 -m venv venv
-          source venv/bin/activate
-          pip install -r requirements.txt
-          export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
-
-  server-kleidiai:
-    runs-on: ah-ubuntu_22_04-c8g_8x
-
-    name: server-kleidiai (${{ matrix.wf_name }})
-    strategy:
-      matrix:
-        include:
-          - build_type: Release
-            extra_build_flags: "-DGGML_CPU_KLEIDIAI=ON"
-            extra_args: ""
-            wf_name:    "CPUx1, kleidiai"
-      fail-fast: false
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Dependencies
-        id: depends
-        run: |
-          set -euxo pipefail
-          sudo apt-get update
-          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
-          apt-get install -y \
-           build-essential \
-           libssl-dev \
-           python3-venv \
-           gpg \
-           wget \
-           time \
-           git-lfs
-
-          git lfs install
-
-          # install the latest cmake
-          sudo install -d /usr/share/keyrings
-          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
-           | gpg --dearmor \
-           | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
-          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
-           | sudo tee /etc/apt/sources.list.d/kitware.list
-          sudo apt-get update
-          sudo apt-get install -y cmake
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build -DGGML_SCHED_NO_REALLOC=ON ${{ matrix.extra_build_flags }}
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
-        run: |
-          cd tools/server/tests
-          python3 -m venv venv
-          source venv/bin/activate
-          pip install -r requirements.txt
-          export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
+  # TODO: provision CUDA runner
+  #  server-cuda:
+  #    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
+  #
+  #    name: server-cuda (${{ matrix.wf_name }})
+  #    strategy:
+  #      matrix:
+  #        build_type: [Release]
+  #        wf_name: ["GPUx1"]
+  #        include:
+  #          - build_type: Release
+  #            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+  #            wf_name:    "GPUx1, backend-sampling"
+  #      fail-fast: false
+  #
+  #    steps:
+  #      - name: Clone
+  #        id: checkout
+  #        uses: actions/checkout@v6
+  #        with:
+  #          fetch-depth: 0
+  #          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+  #
+  #      - name: Build
+  #        id: cmake_build
+  #        run: |
+  #          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+  #          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+  #
+  #      - name: Tests
+  #        id: server_integration_tests
+  #        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+  #        run: |
+  #          cd tools/server/tests
+  #          python3 -m venv venv
+  #          source venv/bin/activate
+  #          pip install -r requirements.txt
+  #          export ${{ matrix.extra_args }}
+  #          pytest -v -x -m "not slow"
@@ -1,4 +1,4 @@
-name: UI
+name: Server WebUI

 on:
  workflow_dispatch:
@@ -11,39 +11,37 @@ on:
    branches:
      - master
    paths: [
-      '.github/workflows/ui.yml',
-      '.github/workflows/ui-build.yml',
-      'tools/ui/**.*',
+      '.github/workflows/server-webui.yml',
+      'tools/server/webui/**.*',
      'tools/server/tests/**.*'
    ]
  pull_request:
    types: [opened, synchronize, reopened]
    paths: [
-      '.github/workflows/ui.yml',
-      '.github/workflows/ui-build.yml',
-      'tools/ui/**.*',
+      '.github/workflows/server-webui.yml',
+      'tools/server/webui/**.*',
      'tools/server/tests/**.*'
    ]

 env:
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  LLAMA_ARG_LOG_VERBOSITY: 10
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
  cancel-in-progress: true

 jobs:
-  ui-build:
-    name: Build static output
-    uses: ./.github/workflows/ui-build.yml
+  webui-build:
+    name: Build WebUI
+    uses: ./.github/workflows/webui-build.yml

-  ui-checks:
-    name: Checks
-    needs: ui-build
-    runs-on: ubuntu-latest
+  webui-checks:
+    name: WebUI Checks
+    needs: webui-build
+    runs-on: ubuntu-24.04-arm
    continue-on-error: true
    steps:
      - name: Checkout code
@@ -58,44 +56,44 @@ jobs:
        with:
          node-version: "24"
          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          cache-dependency-path: "tools/server/webui/package-lock.json"

      - name: Install dependencies
        id: setup
        if: ${{ steps.node.conclusion == 'success' }}
        run: npm ci
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run type checking
        if: ${{ always() && steps.setup.conclusion == 'success' }}
        run: npm run check
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run linting
        if: ${{ always() && steps.setup.conclusion == 'success' }}
        run: npm run lint
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Install Playwright browsers
        id: playwright
        if: ${{ always() && steps.setup.conclusion == 'success' }}
        run: npx playwright install --with-deps
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run Client tests
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run test:client
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run Unit tests
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run test:unit
-        working-directory: tools/ui
+        working-directory: tools/server/webui

  e2e-tests:
    name: E2E Tests
-    needs: ui-build
-    runs-on: ubuntu-latest
+    needs: webui-build
+    runs-on: ubuntu-24.04-arm
    steps:
      - name: Checkout code
        uses: actions/checkout@v6
@@ -109,36 +107,36 @@ jobs:
        with:
          node-version: "24"
          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          cache-dependency-path: "tools/server/webui/package-lock.json"

      - name: Install dependencies
        id: setup
        if: ${{ steps.node.conclusion == 'success' }}
        run: npm ci
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Build application
        if: ${{ always() && steps.setup.conclusion == 'success' }}
        run: npm run build
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Install Playwright browsers
        id: playwright
        if: ${{ always() && steps.setup.conclusion == 'success' }}
        run: npx playwright install --with-deps
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Build Storybook
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run build-storybook
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run UI tests
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run test:ui -- --testTimeout=60000
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Run E2E tests
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run test:e2e
-        working-directory: tools/ui
+        working-directory: tools/server/webui
@@ -44,18 +44,37 @@ on:
    ]

 env:
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  LLAMA_ARG_LOG_VERBOSITY: 10
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
  cancel-in-progress: true

 jobs:
-  ubuntu:
-    runs-on: ubuntu-24.04-arm
+  webui-build:
+    name: Build WebUI
+    uses: ./.github/workflows/webui-build.yml
+
+  server:
+    needs: webui-build
+    runs-on: ubuntu-latest
+
+    name: server (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        build_type: [Release]
+        wf_name: ["default"]
+        include:
+          - build_type: Release
+            extra_args: ""
+            wf_name:    "default"
+          - build_type: Release
+            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "backend-sampling"
+      fail-fast: false

    steps:
      - name: Dependencies
@@ -79,19 +98,19 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
+      - name: Download WebUI build artifact
+        uses: actions/download-artifact@v7
        with:
-          key: server-ubuntu-24.04-arm
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          name: webui-build
+          path: tools/server/public/

      - name: Build
        id: cmake_build
        run: |
          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
            -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config Release -j $(nproc) --target llama-server
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server

      - name: Python setup
        id: setup_python
@@ -102,34 +121,23 @@ jobs:

      - name: Tests
        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
        run: |
          cd tools/server/tests
+          export ${{ matrix.extra_args }}
          pytest -v -x -m "not slow"

      - name: Slow tests
        id: server_integration_tests_slow
-        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
          cd tools/server/tests
+          export ${{ matrix.extra_args }}
          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
-          pytest -v -x -m "not slow"
-
-      - name: Slow tests (Backend sampling)
-        id: server_integration_tests_slow_backend_sampling
-        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
-        run: |
-          cd tools/server/tests
-          export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 pytest -v -x
-
-  windows:
-    runs-on: windows-2025
+  server-windows:
+    needs: webui-build
+    runs-on: windows-2022

    steps:
      - name: Clone
@@ -139,24 +147,17 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
+      - name: Download WebUI build artifact
+        uses: actions/download-artifact@v7
        with:
-          key: server-windows-2025-x64
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          name: webui-build
+          path: tools/server/public/

      - name: Build
        id: cmake_build
-        shell: cmd
        run: |
-          cmake -B build -G "Ninja Multi-Config" ^
-            -DCMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake ^
-            -DCMAKE_BUILD_TYPE=Release ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_SCHED_NO_REALLOC=ON
-          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% --target llama-server
+          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server

      - name: Python setup
        id: setup_python
@@ -167,6 +168,7 @@ jobs:

      - name: Tests
        id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
          cd tools/server/tests
          $env:PYTHONIOENCODING = ":replace"
@@ -174,7 +176,7 @@ jobs:

      - name: Slow tests
        id: server_integration_tests_slow
-        if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
          cd tools/server/tests
          $env:SLOW_TESTS = "1"
@@ -1,43 +0,0 @@
-name: UI Build (self-hosted)
-
-on:
-  workflow_call:
-
-jobs:
-  build:
-    runs-on: [self-hosted, fast]
-    env:
-      BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v6
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
-      - name: Install dependencies
-        run: npm ci
-        working-directory: tools/ui
-
-      - name: Build application
-        run: npm run build
-        working-directory: tools/ui
-
-      - name: Generate checksums
-        run: |
-          cd tools/ui/dist
-          for f in *; do
-            sha256sum "$f" | awk '{print $1, $2}' >> checksums.txt
-          done
-
-      - name: Upload built UI
-        uses: actions/upload-artifact@v6
-        with:
-          name: ui-build
-          path: tools/ui/dist/
-          retention-days: 1
@@ -1,118 +0,0 @@
-name: UI (self-hosted)
-
-# these are the same as ui.yml, but with self-hosted runners
-# the runners come with pre-installed Playwright browsers version: 1.56.1
-# the jobs are much lighter because they don't need to install node and playwright browsers
-
-on:
-  workflow_dispatch:
-    inputs:
-      sha:
-        description: 'Commit SHA1 to build'
-        required: false
-        type: string
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/ui-self-hosted.yml',
-      '.github/workflows/ui-build-self-hosted.yml',
-      'tools/ui/**.*',
-      'tools/server/tests/**.*'
-    ]
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/ui-self-hosted.yml',
-      '.github/workflows/ui-build-self-hosted.yml',
-      'tools/ui/**.*',
-      'tools/server/tests/**.*'
-    ]
-
-env:
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  LLAMA_ARG_LOG_VERBOSITY: 10
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  ui-build:
-    name: Build static output
-    uses: ./.github/workflows/ui-build-self-hosted.yml
-
-  ui-checks:
-    name: Checks
-    needs: ui-build
-    runs-on: [self-hosted, PLAYWRIGHT]
-    continue-on-error: true
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Install dependencies
-        id: setup
-        run: npm ci
-        working-directory: tools/ui
-
-      - name: Run type checking
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run check
-        working-directory: tools/ui
-
-      - name: Run linting
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run lint
-        working-directory: tools/ui
-
-      - name: Run Client tests
-        if: ${{ always() }}
-        run: npm run test:client
-        working-directory: tools/ui
-
-      - name: Run Unit tests
-        if: ${{ always() }}
-        run: npm run test:unit
-        working-directory: tools/ui
-
-  e2e-tests:
-    name: E2E Tests
-    needs: ui-build
-    runs-on: [self-hosted, PLAYWRIGHT]
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Install dependencies
-        id: setup
-        run: npm ci
-        working-directory: tools/ui
-
-      - name: Build application
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run build
-        working-directory: tools/ui
-
-      - name: Build Storybook
-        if: ${{ always() }}
-        run: npm run build-storybook
-        working-directory: tools/ui
-
-      - name: Run UI tests
-        if: ${{ always() }}
-        run: npm run test:ui -- --testTimeout=60000
-        working-directory: tools/ui
-
-      - name: Run E2E tests
-        if: ${{ always() }}
-        run: npm run test:e2e
-        working-directory: tools/ui
@@ -3,20 +3,18 @@ name: Update Operations Documentation
 on:
    push:
        paths:
-            - '.github/workflows/update-ops-docs.yml'
            - 'docs/ops.md'
            - 'docs/ops/**'
            - 'scripts/create_ops_docs.py'
    pull_request:
        paths:
-            - '.github/workflows/update-ops-docs.yml'
            - 'docs/ops.md'
            - 'docs/ops/**'
            - 'scripts/create_ops_docs.py'

 jobs:
    update-ops-docs:
-        runs-on: [self-hosted, fast, ARM64]
+        runs-on: ubuntu-slim

        steps:
        - name: Checkout repository
@@ -1,10 +1,11 @@
-name: UI Build
+name: Build WebUI

 on:
  workflow_call:

 jobs:
  build:
+    name: Build WebUI
    runs-on: ubuntu-slim
    env:
      BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -18,26 +19,26 @@ jobs:
        with:
          node-version: "24"
          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          cache-dependency-path: "tools/server/webui/package-lock.json"

      - name: Install dependencies
        run: npm ci
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Build application
        run: npm run build
-        working-directory: tools/ui
+        working-directory: tools/server/webui

      - name: Generate checksums
        run: |
-          cd tools/ui/dist
+          cd tools/server/public
          for f in *; do
            sha256sum "$f" | awk '{print $1, $2}' >> checksums.txt
          done

-      - name: Upload built UI
+      - name: Upload built webui
        uses: actions/upload-artifact@v6
        with:
-          name: ui-build
-          path: tools/ui/dist/
+          name: webui-build
+          path: tools/server/public/
          retention-days: 1
@@ -1,4 +1,4 @@
-name: UI Publish
+name: WebUI Publish

 on:
  workflow_call:
@@ -13,20 +13,15 @@ on:
        required: true

 jobs:
-  build:
-    name: Build static output
-    uses: ./.github/workflows/ui-build.yml
-
  publish:
-    name: Publish UI Static Output
-    needs: build
-    runs-on: ubuntu-slim
+    name: Publish WebUI Static Output
+    runs-on: ubuntu-24.04-arm

    permissions:
      contents: read

    env:
-      HF_BUCKET_NAME: ${{ vars.HF_BUCKET_UI_STATIC_OUTPUT }}
+      HF_BUCKET_NAME: ${{ vars.HF_BUCKET_WEBUI_STATIC_OUTPUT }}

    steps:
      - name: Checkout code
@@ -34,11 +29,11 @@ jobs:
        with:
          fetch-depth: 1

-      - name: Download UI build artifact
+      - name: Download WebUI build artifact
        uses: actions/download-artifact@v7
        with:
-          name: ui-build
-          path: tools/ui/dist/
+          name: webui-build
+          path: tools/server/public/

      - name: Install Hugging Face Hub CLI
        run: pip install -U huggingface_hub
@@ -49,12 +44,12 @@ jobs:
      - name: Sync built files to Hugging Face bucket (version tag)
        run: |
          # Upload the built files to the Hugging Face bucket under the release version
-          hf buckets sync tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/${{ inputs.version_tag }} --delete --quiet
+          hf buckets sync tools/server/public hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/${{ inputs.version_tag }} --delete --quiet

      - name: Sync built files to Hugging Face bucket (latest)
        run: |
          # Also upload to the 'latest' directory for fallback downloads
-          hf buckets sync tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/latest --delete --quiet
+          hf buckets sync tools/server/public hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/latest --delete --quiet

      - name: Verify upload
        run: |
@@ -54,6 +54,7 @@
 /tmp/
 /autogen-*.md
 /common/build-info.cpp
+/tools/server/public

 # Deprecated

@@ -92,12 +93,10 @@
 !/examples/sycl/*.bat
 !/examples/sycl/*.sh

-# Server Web UI temporary files (+ legacy directory)
+# Server Web UI temporary files

 /tools/server/webui/node_modules
 /tools/server/webui/dist
-/tools/ui/node_modules
-/tools/ui/dist

 # Python

@@ -1,7 +1,7 @@
 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.
+- By very precise and concise when writing code, comments, explanations, etc.
 - 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
@@ -16,15 +16,12 @@ Pull requests (PRs):
 - New branch names are prefixed with "gg/"
 - Before opening a pull request, ask the user to confirm the description
 - When creating a pull request, look for the repository's PR template and follow it
- For the AI usage disclosure section, write "YES. llama.cpp + pi + [MODEL]"
- Ask the user to tell you what model was used and write it in place of [MODEL]
+- For the AI usage disclosure section, write "YES. llama.cpp + pi"
 - Always create the pull requests in draft mode

 Commits:
 - On every commit that you make, include a "Assisted-by: llama.cpp:local pi" tag
 - Do not explicitly set the git author in commits - rely on the default git config
- Always use `--no-gpg-sign` when committing
- Never `git push` without explicit confirmation from the user

 Resources (read on demand):
 - [CONTRIBUTING.md](CONTRIBUTING.md)
@@ -104,16 +104,14 @@ option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})

 # extra artifacts
-option(LLAMA_BUILD_TESTS     "llama: build tests"                                                                ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_TOOLS     "llama: build tools"                                                                ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES  "llama: build examples"                                                             ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER    "llama: build server example"                                                       ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_APP       "llama: build the unified binary"                                                   ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       ON)
-option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
-
-option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
-option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
+option(LLAMA_BUILD_TESTS            "llama: build tests"                                                                            ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_TOOLS            "llama: build tools"                                                                            ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES         "llama: build examples"                                                                         ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER           "llama: build server example"                                                                   ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_WEBUI            "llama: build the embedded Web UI for server"                                                   ON)
+option(LLAMA_USE_PREBUILT_WEBUI     "llama: use prebuilt WebUI from HF Bucket when available (requires LLAMA_BUILD_WEBUI=ON)"       ON)
+option(LLAMA_TOOLS_INSTALL          "llama: install tools"                                                                          ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(LLAMA_TESTS_INSTALL          "llama: install tests"                                                                          ON)

 # 3rd party libs
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
@@ -218,8 +216,17 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
    add_subdirectory(tools)
 endif()

-if (LLAMA_BUILD_APP)
-    add_subdirectory(app)
+# Automatically add all files from the 'licenses' directory
+file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
+
+foreach(FILE_PATH ${EXTRA_LICENSES})
+    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
+    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
+    license_add_file("${NAME}" "${FILE_PATH}")
+endforeach()
+
+if (LLAMA_BUILD_COMMON)
+    license_generate(llama-common)
 endif()

 #
@@ -264,6 +271,18 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
              ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)

+install(
+    FILES convert_hf_to_gguf.py
+    PERMISSIONS
+        OWNER_READ
+        OWNER_WRITE
+        OWNER_EXECUTE
+        GROUP_READ
+        GROUP_EXECUTE
+        WORLD_READ
+        WORLD_EXECUTE
+    DESTINATION ${CMAKE_INSTALL_BINDIR})
+
 configure_file(cmake/llama.pc.in
        "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
        @ONLY)
@@ -15,7 +15,7 @@
 # ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
 # ggml-org/llama-mtmd       : ngxson
 # ggml-org/llama-server     : ggerganov, ngxson, allozaur, angt, ServeurpersoCom
-# ggml-org/llama-ui           : allozaur
+# ggml-org/llama-webui      : allozaur

 /.devops/*.Dockerfile                   @ngxson
 /.github/actions/                       @ggml-org/ci
@@ -26,7 +26,6 @@
 /common/fit.*                           @JohannesGaessler
 /common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
-/conversion/                            @CISC
 /convert_*.py                           @CISC
 /docs/backend/snapdragon/               @ggml-org/ggml-hexagon
 /examples/batched.swift/                @ggerganov
@@ -49,6 +48,7 @@
 /examples/parallel/                     @ggerganov
 /examples/passkey/                      @ggerganov
 /examples/retrieval/                    @ggerganov
+/examples/save-load-state/              @ggerganov
 /examples/speculative-simple/           @ggerganov
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
@@ -107,7 +107,7 @@
 /tools/rpc/                             @ggml-org/ggml-rpc
 /tools/server/*                         @ggml-org/llama-server # no subdir
 /tools/server/tests/                    @ggml-org/llama-server
-/tools/ui/                              @ggml-org/llama-ui
+/tools/server/webui/                    @ggml-org/llama-webui
 /tools/tokenize/                        @ggerganov
 /tools/tts/                             @ggerganov
 /vendor/                                @ggerganov
@@ -63,7 +63,6 @@ After submitting your PR:
 - Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
 - Let other maintainers merge their own PRs
 - 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)

 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
@@ -27,7 +27,6 @@ LLM inference in C/C++
 - 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/).

 ----

@@ -281,7 +280,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Metal](docs/build.md#metal-build) | Apple Silicon |
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
-| [SYCL](docs/backend/SYCL.md) | Intel GPU |
+| [SYCL](docs/backend/SYCL.md) | Intel and Nvidia 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 |
@@ -291,7 +290,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [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 |
+| [WebGPU [In Progress]](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 |
@@ -12,16 +12,16 @@

 ## Reporting a vulnerability

-> [!IMPORTANT]
-> The private security disclosure program is disabled until further notice. Please submit patches with fixes directly to the repo as public PRs. Emails will be ignored.
-
 If you have discovered a security vulnerability in this project that falls inside the [covered topics](#covered-topics), please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.

 Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).

 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.

-### Requirements
+> [!IMPORTANT]
+> For collaborators: if you are interested in helping out with reviewing private security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
+
+## Requirements

 Before submitting your report, ensure you meet the following requirements:

@@ -31,7 +31,7 @@ Before submitting your report, ensure you meet the following requirements:

 Maintainers reserve the right to close the report if these requirements are not fulfilled.

-### Covered Topics
+## Covered Topics

 Only vulnerabilities that fall within these parts of the project are considered valid. For problems falling outside of this list, please report them as issues.

@@ -1,31 +0,0 @@
-set(TARGET llama-app)
-
-add_executable(${TARGET} llama.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)
-
-target_link_libraries(${TARGET} PRIVATE
-    llama-server-impl
-    llama-cli-impl
-    llama-completion-impl
-    llama-bench-impl
-    llama-batched-bench-impl
-    llama-fit-params-impl
-    llama-quantize-impl
-    llama-perplexity-impl
-)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-# Automatically add all files from the 'licenses' directory
-file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
-
-foreach(FILE_PATH ${EXTRA_LICENSES})
-    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
-    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
-    license_add_file("${NAME}" "${FILE_PATH}")
-endforeach()
-
-license_generate(${TARGET})
-
-if(LLAMA_TOOLS_INSTALL)
-    install(TARGETS ${TARGET} RUNTIME)
-endif()
@@ -1,127 +0,0 @@
-#include "build-info.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <string>
-#include <vector>
-
-// embedded data generated by cmake
-extern const char * LICENSES[];
-
-// visible
-int llama_server(int argc, char ** argv);
-int llama_cli(int argc, char ** argv);
-
-// hidden
-int llama_completion(int argc, char ** argv);
-int llama_bench(int argc, char ** argv);
-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);
-
-// hands the update over to the install script, which downloads and swaps the binary
-static int llama_update(int argc, char ** argv) {
-    (void) argc;
-    (void) argv;
-
-#if defined(_WIN32)
-    return system("powershell -NoProfile -ExecutionPolicy Bypass -Command \"irm https://llama.app/install.ps1 | iex\"");
-#else
-    return system("curl -fsSL https://llama.app/install.sh | sh");
-#endif
-}
-
-static const char * progname;
-
-static int help(int argc, char ** argv);
-static int version(int argc, char ** argv);
-static int licenses(int argc, char ** argv);
-
-struct command {
-    const char * name;
-    const char * desc;
-    std::vector<std::string> aliases;
-    bool hidden;
-    int (*func)(int, char **);
-};
-
-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",                 {},           false, llama_update       },
-    {"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            },
-    {"licenses",      "Show third-party licenses",                          {"credits"},  false, licenses           },
-    {"help",          "Show available commands",                            {},           false, help               },
-};
-
-static int version(int argc, char ** argv) {
-    printf("%s\n", llama_build_info());
-    return 0;
-}
-
-static int licenses(int argc, char ** argv) {
-    for (int i = 0; LICENSES[i]; ++i) {
-        printf("%s\n", LICENSES[i]);
-    }
-    return 0;
-}
-
-static int help(int argc, char ** argv) {
-    const bool show_all = argc >= 2 && std::string(argv[1]) == "all";
-
-    printf("Usage: %s <command> [options]\n\nAvailable commands:\n", progname);
-
-    for (const auto & cmd : cmds) {
-        if (show_all || !cmd.hidden) {
-            printf("  %-15s %s\n", cmd.name, cmd.desc);
-        }
-    }
-    printf("\n");
-
-    if (!show_all) {
-        printf("Run '%s help all' to show additional commands.\n", progname);
-    }
-    printf("Run '%s <command> --help' for command-specific usage.\n", progname);
-
-    return 0;
-}
-
-static bool matches(const std::string & arg, const command & cmd) {
-    if (arg == cmd.name) {
-        return true;
-    }
-    for (const auto & alias : cmd.aliases) {
-        if (arg == alias) {
-            return true;
-        }
-    }
-    return false;
-}
-
-int main(int argc, char ** argv) {
-    progname = argv[0];
-
-    const std::string arg = argc >= 2 ? argv[1] : "help";
-
-    for (const auto & cmd : cmds) {
-        if (matches(arg, cmd)) {
-            // keep cmd.name so the router's child processes re-invoke correctly
-#ifdef _WIN32
-            _putenv_s("LLAMA_APP_CMD", cmd.name);
-#else
-            setenv("LLAMA_APP_CMD", cmd.name, 1);
-#endif
-            return cmd.func(argc - 1, argv + 1);
-        }
-    }
-
-    fprintf(stderr, "error: unknown command '%s'\n", arg.c_str());
-    return 1;
-}
@@ -7,8 +7,6 @@ VISIONOS_MIN_OS_VERSION=1.0
 TVOS_MIN_OS_VERSION=16.4

 BUILD_SHARED_LIBS=OFF
-LLAMA_BUILD_APP=OFF
-LLAMA_BUILD_COMMON=OFF
 LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
@@ -33,8 +31,6 @@ COMMON_CMAKE_ARGS=(
    -DCMAKE_XCODE_ATTRIBUTE_STRIP_INSTALLED_PRODUCT=NO
    -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
    -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
-    -DLLAMA_BUILD_APP=${LLAMA_BUILD_APP}
-    -DLLAMA_BUILD_COMMON=${LLAMA_BUILD_COMMON}
    -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
    -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
    -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
@@ -418,7 +414,7 @@ cmake -B build-ios-sim -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -S .
-cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-ios-sim --config Release -- -quiet

 echo "Building for iOS devices..."
 cmake -B build-ios-device -G Xcode \
@@ -432,7 +428,7 @@ cmake -B build-ios-device -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -S .
-cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-ios-device --config Release -- -quiet

 echo "Building for macOS..."
 cmake -B build-macos -G Xcode \
@@ -443,7 +439,7 @@ cmake -B build-macos -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -S .
-cmake --build build-macos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-macos --config Release -- -quiet

 echo "Building for visionOS..."
 cmake -B build-visionos -G Xcode \
@@ -458,7 +454,7 @@ cmake -B build-visionos -G Xcode \
    -DLLAMA_OPENSSL=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
    -S .
-cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-visionos --config Release -- -quiet

 echo "Building for visionOS simulator..."
 cmake -B build-visionos-sim -G Xcode \
@@ -473,7 +469,7 @@ cmake -B build-visionos-sim -G Xcode \
    -DLLAMA_OPENSSL=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
    -S .
-cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-visionos-sim --config Release -- -quiet

 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
 echo "Building for tvOS simulator..."
@@ -489,7 +485,7 @@ cmake -B build-tvos-sim -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -S .
-cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-tvos-sim --config Release -- -quiet

 echo "Building for tvOS devices..."
 cmake -B build-tvos-device -G Xcode \
@@ -504,7 +500,7 @@ cmake -B build-tvos-device -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -S .
-cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
+cmake --build build-tvos-device --config Release -- -quiet

 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."
@@ -66,8 +66,6 @@ fi

 if [ ! -z ${GG_BUILD_METAL} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
-else
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF"
 fi

 if [ ! -z ${GG_BUILD_CUDA} ]; then
@@ -116,12 +114,9 @@ fi
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"

+    # if on Mac, disable METAL
    if [[ "$OSTYPE" == "darwin"* ]]; then
-        MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION="/usr/local/lib/cmake/vulkan"
-        MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION="${MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION}/SPIRV-Headers/SPIRV-HeadersConfig.cmake"
-        if [[ -f "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" || -h "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" ]]; then
-            CMAKE_EXTRA="${CMAKE_EXTRA} -DSPIRV-Headers_DIR=${MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION}/SPIRV-Headers"
-        fi
+        CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
    fi

    # Build shared libs on Windows
@@ -132,7 +127,7 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
 fi

 if [ ! -z ${GG_BUILD_WEBGPU} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1 -DGGML_METAL=OFF -DGGML_BLAS=OFF"

    if [ ! -z "${GG_BUILD_WEBGPU_DAWN_PREFIX}" ]; then
        if [ -z "${CMAKE_PREFIX_PATH}" ]; then
@@ -166,8 +161,6 @@ fi

 if [ ! -z ${GG_BUILD_BLAS} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=${GG_BUILD_BLAS_VENDOR:-OpenBLAS}"
-else
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=OFF"
 fi

 if [ ! -z ${GG_BUILD_OPENVINO} ]; then
@@ -239,7 +232,7 @@ function gg_run_ctest_debug {
    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time cmake --build . --config Debug -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log

-    (time ctest -C Debug --output-on-failure -L main -E "test-opt|test-backend-ops|test-llama-archs" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest -C Debug --output-on-failure -L main -E "test-opt|test-backend-ops" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log

    set +e
 }
@@ -462,10 +455,10 @@ function gg_run_qwen3_0_6b {

    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

    function check_ppl {
        qnt="$1"
@@ -701,8 +694,8 @@ function gg_sum_test_backend_ops_cpu {

 ## main

-export LLAMA_ARG_LOG_PREFIX=1
-export LLAMA_ARG_LOG_TIMESTAMPS=1
+export LLAMA_LOG_PREFIX=1
+export LLAMA_LOG_TIMESTAMPS=1

 if [ -z ${GG_BUILD_LOW_PERF} ]; then
    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
@@ -7,7 +7,7 @@ set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)

 set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
 set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
-set(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")

 find_package(ggml REQUIRED HINTS ${LLAMA_LIB_DIR}/cmake)

@@ -4,6 +4,7 @@
 #include "chat.h"
 #include "common.h"
 #include "download.h"
+#include "hf-cache.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
@@ -50,6 +51,8 @@

 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083

+extern const char * LICENSES[];
+
 using json = nlohmann::ordered_json;
 using namespace common_arg_utils;

@@ -334,13 +337,11 @@ static bool common_params_handle_remote_preset(common_params & params, llama_exa
 struct handle_model_result {
    bool found_mmproj = false;
    common_params_model mmproj;
-
-    bool found_mtp = false;
-    common_params_model mtp;
 };

 static handle_model_result common_params_handle_model(struct common_params_model & model,
-                                                      const common_download_opts & opts) {
+                                                      const std::string          & bearer_token,
+                                                      bool                         offline) {
    handle_model_result result;

    if (!model.docker_repo.empty()) {
@@ -352,9 +353,10 @@ static handle_model_result common_params_handle_model(struct common_params_model
            model.hf_file = model.path;
            model.path = "";
        }
-        common_download_opts hf_opts = opts;
-        hf_opts.download_mmproj = true; // also look for mmproj when downloading hf model
-        auto download_result = common_download_model(model, hf_opts);
+        common_download_opts opts;
+        opts.bearer_token = bearer_token;
+        opts.offline = offline;
+        auto download_result = common_download_model(model, opts, true);

        if (download_result.model_path.empty()) {
            throw std::runtime_error("failed to download model from Hugging Face");
@@ -367,11 +369,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
            result.found_mmproj = true;
            result.mmproj.path  = download_result.mmproj_path;
        }
-
-        if (!download_result.mtp_path.empty()) {
-            result.found_mtp = true;
-            result.mtp.path  = download_result.mtp_path;
-        }
    } else if (!model.url.empty()) {
        if (model.path.empty()) {
            auto f = string_split<std::string>(model.url, '#').front();
@@ -379,6 +376,9 @@ static handle_model_result common_params_handle_model(struct common_params_model
            model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
        }

+        common_download_opts opts;
+        opts.bearer_token = bearer_token;
+        opts.offline = offline;
        auto download_result = common_download_model(model, opts);
        if (download_result.model_path.empty()) {
            throw std::runtime_error("failed to download model from " + model.url);
@@ -435,49 +435,23 @@ static bool parse_bool_value(const std::string & value) {
 // CLI argument parsing functions
 //

-bool common_params_handle_models(common_params & params, llama_example curr_ex) {
-    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
-                                         params.speculative.types.end(),
-                                         COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-
-    common_download_opts opts;
-    opts.bearer_token  = params.hf_token;
-    opts.offline       = params.offline;
-    opts.skip_download = params.skip_download;
-    opts.download_mtp  = spec_type_draft_mtp;
-
-    try {
-        auto res = common_params_handle_model(params.model, opts);
-        if (params.no_mmproj) {
-            params.mmproj = {};
-        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
-            // optionally, handle mmproj model when -hf is specified
-            params.mmproj = res.mmproj;
-        }
-        // only download mmproj if the current example is using it
-        for (const auto & ex : mmproj_examples) {
-            if (curr_ex == ex) {
-                common_params_handle_model(params.mmproj, opts);
-                break;
-            }
-        }
-
-        // when --spec-type mtp is set and no draft model was provided explicitly,
-        // fall back to the MTP head discovered alongside the -hf model
-        if (spec_type_draft_mtp && res.found_mtp &&
-            params.speculative.draft.mparams.path.empty() &&
-            params.speculative.draft.mparams.hf_repo.empty() &&
-            params.speculative.draft.mparams.url.empty()) {
-            params.speculative.draft.mparams.path = res.mtp.path;
-        }
-        common_params_handle_model(params.speculative.draft.mparams, opts);
-        common_params_handle_model(params.vocoder.model,             opts);
-        return true;
-    } catch (const common_skip_download_exception &) {
-        return false;
-    } catch (const std::exception &) {
-        throw;
+void common_params_handle_models(common_params & params, llama_example curr_ex) {
+    auto res = common_params_handle_model(params.model, params.hf_token, params.offline);
+    if (params.no_mmproj) {
+        params.mmproj = {};
+    } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
+        // optionally, handle mmproj model when -hf is specified
+        params.mmproj = res.mmproj;
    }
+    // only download mmproj if the current example is using it
+    for (const auto & ex : mmproj_examples) {
+        if (curr_ex == ex) {
+            common_params_handle_model(params.mmproj,    params.hf_token, params.offline);
+            break;
+        }
+    }
+    common_params_handle_model(params.speculative.draft.mparams, params.hf_token, params.offline);
+    common_params_handle_model(params.vocoder.model,             params.hf_token, params.offline);
 }

 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
@@ -542,11 +516,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
                throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
            }
            if (!seen_args.insert(arg).second) {
-                const bool skip = (arg == "--spec-type");
-
-                if (!skip) {
-                    LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
-                }
+                LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
            }
            auto & tmp = arg_to_options[arg];
            auto opt = *tmp.first;
@@ -595,6 +565,12 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
    // parse the first time to get -hf option (used for remote preset)
    parse_cli_args();

+    // TODO: Remove later
+    try {
+        hf_cache::migrate_old_cache_to_hf_cache(params.hf_token, params.offline);
+    } catch (const std::exception & e) {
+        LOG_WRN("HF cache migration failed: %s\n", e.what());
+    }
    // export_graph_ops loads only metadata
    const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;

@@ -903,11 +879,7 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
            throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
        }
        if (!seen_args.insert(arg).second) {
-            const bool skip = (arg == "--spec-type");
-
-            if (!skip) {
-                LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
-            }
+            LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
        }
        auto opt = *arg_to_options[arg];
        std::string val;
@@ -1041,9 +1013,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    // we define here to make sure it's included in llama-gen-docs
    if (ex == LLAMA_EXAMPLE_COMPLETION) {
        params.use_jinja = false;   // disable jinja by default
+
    } else if (ex == LLAMA_EXAMPLE_MTMD) {
        params.use_jinja = false;   // disable jinja by default
        params.sampling.temp = 0.2; // lower temp by default for better quality
+
    } else if (ex == LLAMA_EXAMPLE_SERVER) {
        params.n_parallel = -1;     // auto by default
    }
@@ -1064,6 +1038,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        sampler_type_names.pop_back(); // remove last semicolon
    }

+
    /**
     * filter options by example
     * rules:
@@ -1077,6 +1052,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    };

+
    add_opt(common_arg(
        {"-h", "--help", "--usage"},
        "print usage and exit",
@@ -1093,6 +1069,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            exit(0);
        }
    ));
+    add_opt(common_arg(
+        {"--license"},
+        "show source code license and dependencies",
+        [](common_params &) {
+            for (int i = 0; LICENSES[i]; ++i) {
+                printf("%s\n", LICENSES[i]);
+            }
+            exit(0);
+        }
+    ));
    add_opt(common_arg(
        {"-cl", "--cache-list"},
        "show list of models in cache",
@@ -1326,15 +1312,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
    add_opt(common_arg(
-        {"-cms", "--checkpoint-min-step"}, "N",
-        string_format("minimum spacing between context checkpoints in tokens (default: %d, 0 = no minimum)", params.checkpoint_min_step),
+        {"-cpent", "--checkpoint-every-n-tokens"}, "N",
+        string_format("create a checkpoint every n tokens during prefill (processing), -1 to disable (default: %d)", params.checkpoint_every_nt),
        [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("checkpoint-min-step must be non-negative");
-            }
-            params.checkpoint_min_step = value;
+            params.checkpoint_every_nt = value;
        }
-    ).set_env("LLAMA_ARG_CHECKPOINT_MIN_SPACING_NT").set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_env("LLAMA_ARG_CHECKPOINT_EVERY_NT").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
    add_opt(common_arg(
        {"-cram", "--cache-ram"}, "N",
        string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)"
@@ -2804,7 +2787,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, int value) {
            params.embd_normalize = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_DEBUG}));
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_DEBUG}));
    add_opt(common_arg(
        {"--embd-output-format"}, "FORMAT",
        "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix, \"raw\" = plain whitespace-delimited output (one embedding per line)",
@@ -2861,64 +2844,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.api_prefix = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
-    // Deprecated: use --ui-config instead (kept for backward compat)
    add_opt(common_arg(
        {"--webui-config"}, "JSON",
-        "[DEPRECATED: use --ui-config] JSON that provides default WebUI settings (overrides WebUI defaults)",
+        "JSON that provides default WebUI settings (overrides WebUI defaults)",
        [](common_params & params, const std::string & value) {
-            params.ui_config_json = value;
            params.webui_config_json = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG"));
-
-    add_opt(common_arg(
-        {"--ui-config"}, "JSON",
-        "JSON that provides default UI settings (overrides UI defaults)",
-        [](common_params & params, const std::string & value) {
-            params.ui_config_json = value;
-            params.webui_config_json = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_CONFIG"));
-
-    // Deprecated: use --ui-config-file instead (kept for backward compat)
    add_opt(common_arg(
        {"--webui-config-file"}, "PATH",
-        "[DEPRECATED: use --ui-config-file] JSON file that provides default WebUI settings (overrides WebUI defaults)",
+        "JSON file that provides default WebUI settings (overrides WebUI defaults)",
        [](common_params & params, const std::string & value) {
-            params.ui_config_json = read_file(value);
-            params.webui_config_json = params.ui_config_json;
+            params.webui_config_json = read_file(value);
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG_FILE"));
-
-    add_opt(common_arg(
-        {"--ui-config-file"}, "PATH",
-        "JSON file that provides default UI settings (overrides UI defaults)",
-        [](common_params & params, const std::string & value) {
-            params.ui_config_json = read_file(value);
-            params.webui_config_json = params.ui_config_json;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_CONFIG_FILE"));
-
-    // Deprecated: use --ui-mcp-proxy instead (kept for backward compat)
    add_opt(common_arg(
        {"--webui-mcp-proxy"},
        {"--no-webui-mcp-proxy"},
-        "[DEPRECATED: use --ui-mcp-proxy/--no-ui-mcp-proxy] experimental: whether to enable MCP CORS proxy",
+        string_format("experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: %s)", params.webui_mcp_proxy ? "enabled" : "disabled"),
        [](common_params & params, bool value) {
-            params.ui_mcp_proxy = value;
            params.webui_mcp_proxy = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_MCP_PROXY"));
-
-    add_opt(common_arg(
-        {"--ui-mcp-proxy"},
-        {"--no-ui-mcp-proxy"},
-        "experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)",
-        [](common_params & params, bool value) {
-            params.ui_mcp_proxy = value;
-            params.webui_mcp_proxy = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_MCP_PROXY"));
    add_opt(common_arg(
        {"--tools"}, "TOOL1,TOOL2,...",
        "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
@@ -2928,26 +2875,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.server_tools = parse_csv_row(value);
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
-    // Deprecated: use --ui/--no-ui instead (kept for backward compat)
    add_opt(common_arg(
        {"--webui"},
        {"--no-webui"},
-        "[DEPRECATED: use --ui/--no-ui] whether to enable the Web UI",
+        string_format("whether to enable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
        [](common_params & params, bool value) {
-            params.ui = value;
            params.webui = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI"));
-
-    add_opt(common_arg(
-        {"--ui"},
-        {"--no-ui"},
-        string_format("whether to enable the Web UI (default: %s)", params.ui ? "enabled" : "disabled"),
-        [](common_params & params, bool value) {
-            params.ui = value;
-            params.webui = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI"));
    add_opt(common_arg(
        {"--embedding", "--embeddings"},
        string_format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"),
@@ -2990,7 +2925,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            }
            key_file.close();
        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_KEY_FILE"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
    add_opt(common_arg(
        {"--ssl-key-file"}, "FNAME",
        "path to file a PEM-encoded SSL private key",
@@ -3018,7 +2953,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                params.default_template_kwargs[item.key()] = item.value().dump();
            }
        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_CHAT_TEMPLATE_KWARGS"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
    add_opt(common_arg(
        {"-to", "--timeout"}, "N",
        string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -3319,7 +3254,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params &, const std::string & value) {
            common_log_set_file(common_log_main(), value.c_str());
        }
-    ).set_env("LLAMA_ARG_LOG_FILE"));
+    ).set_env("LLAMA_LOG_FILE"));
    add_opt(common_arg(
        {"--log-colors"}, "[on|off|auto]",
        "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
@@ -3336,7 +3271,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                    string_format("error: unknown value for --log-colors: '%s'\n", value.c_str()));
            }
        }
-    ).set_env("LLAMA_ARG_LOG_COLORS"));
+    ).set_env("LLAMA_LOG_COLORS"));
    add_opt(common_arg(
        {"-v", "--verbose", "--log-verbose"},
        "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
@@ -3351,7 +3286,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.offline = true;
        }
-    ).set_env("LLAMA_ARG_OFFLINE"));
+    ).set_env("LLAMA_OFFLINE"));
    add_opt(common_arg(
        {"-lv", "--verbosity", "--log-verbosity"}, "N",
        string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n"
@@ -3359,14 +3294,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            " - 1: error\n"
            " - 2: warning\n"
            " - 3: info\n"
-            " - 4: trace (more info)\n"
-            " - 5: debug\n"
+            " - 4: debug\n"
            "(default: %d)\n", params.verbosity),
        [](common_params & params, int value) {
            params.verbosity = value;
            common_log_set_verbosity_thold(value);
        }
-    ).set_env("LLAMA_ARG_LOG_VERBOSITY"));
+    ).set_env("LLAMA_LOG_VERBOSITY"));
    add_opt(common_arg(
        {"--log-prefix"},
        {"--no-log-prefix"},
@@ -3586,15 +3520,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.speculative.draft.p_min = std::stof(value);
        }
    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_P_MIN"));
-    add_opt(common_arg(
-        {"--spec-draft-backend-sampling"},
-        {"--no-spec-draft-backend-sampling"},
-        string_format("offload draft sampling to the backend (default: %s)",
-                      params.speculative.draft.backend_sampling ? "enabled" : "disabled"),
-        [](common_params & params, bool value) {
-            params.speculative.draft.backend_sampling = value;
-        }
-    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING"));
    add_opt(common_arg(
        {"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",
        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
@@ -3635,9 +3560,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        string_format("comma-separated list of types of speculative decoding to use (default: %s)\n",
            common_speculative_type_name_str(params.speculative.types).c_str()),
        [](common_params & params, const std::string & value) {
-            const auto types_str = string_split<std::string>(value, ',');
-            auto types = common_speculative_types_from_names(types_str);
-            params.speculative.types.insert(params.speculative.types.end(), types.begin(), types.end());
+            const auto enabled_types = string_split<std::string>(value, ',');
+            params.speculative.types = common_speculative_types_from_names(enabled_types);
        }
    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_TYPE"));
    add_opt(common_arg(
@@ -4126,16 +4050,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        {"--spec-default"},
        string_format("enable default speculative decoding config"),
        [](common_params & params) {
-            params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
+            params.speculative.types = { COMMON_SPECULATIVE_TYPE_NGRAM_MOD };
            params.speculative.ngram_mod.n_match = 24;
            params.speculative.ngram_mod.n_min = 48;
            params.speculative.ngram_mod.n_max = 64;
-
-            // TODO: not sure if this is a good config - explore more settings and potentially enable it
-            //params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
-            //params.speculative.ngram_map_k4v.size_n = 8;
-            //params.speculative.ngram_map_k4v.size_m = 24;
-            //params.speculative.ngram_map_k4v.min_hits = 2;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));

@@ -129,11 +129,8 @@ 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);

-// 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);
+// Populate model paths (main model, mmproj, etc) from -hf if necessary
+void common_params_handle_models(common_params & params, llama_example curr_ex);

 // 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);
@@ -43,33 +43,11 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
                                                  const autoparser &              autoparser) {
    // Create the result structure
    common_chat_params data;
-    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens  = autoparser.preserved_tokens;
+    data.prompt           = common_chat_template_direct_apply(tmpl, inputs);
+    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens = autoparser.preserved_tokens;

-    std::string parser_generation_prompt = data.generation_prompt;
-
-    if (inputs.continue_final_message != COMMON_CHAT_CONTINUATION_NONE && !inputs.continue_msg.empty()) {
-        // Build up generation prompt manually
-        const auto & msg = inputs.continue_msg;
-
-        if (!autoparser.reasoning.start.empty()) {
-            data.generation_prompt = data.generation_prompt.substr(0, data.generation_prompt.find(autoparser.reasoning.start));
-            data.generation_prompt += autoparser.reasoning.start + msg.reasoning_content;
-            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-                data.generation_prompt += autoparser.reasoning.end;
-            }
-        }
-
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = autoparser.build_parser(inputs, parser_generation_prompt);
+    auto parser = autoparser.build_parser(inputs);
    data.parser = parser.save();

    // Build grammar if tools are present
@@ -109,7 +87,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
    return data;
 }

-common_peg_arena autoparser::build_parser(const generation_params & inputs, const std::string & generation_prompt) const {
+common_peg_arena autoparser::build_parser(const generation_params & inputs) const {
    if (!analysis_complete) {
        throw std::invalid_argument("Cannot call build_parser on autoparser without performing analysis first, call analyze_template(...)");
    }
@@ -143,7 +121,7 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs, cons
        } else {
            parser = content.build_parser(ctx);
        }
-        return pure_content ? p.prefix(generation_prompt, reasoning.start) + parser : p.prefix(generation_prompt, reasoning.start) << parser;
+        return pure_content ? p.prefix(inputs.generation_prompt, reasoning.start) + parser : p.prefix(inputs.generation_prompt, reasoning.start) << parser;
    });
 }

@@ -310,8 +310,6 @@ std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segm

 namespace autoparser {

-static const std::string ERR_TMPL = "#**ERROR**#";
-
 std::string apply_template(const common_chat_template & tmpl, const template_params & params) {
    generation_params tmpl_params;
    tmpl_params.messages              = params.messages;
@@ -328,7 +326,7 @@ std::string apply_template(const common_chat_template & tmpl, const template_par
        return common_chat_template_direct_apply(tmpl, tmpl_params);
    } catch (const std::exception & e) {
        LOG_DBG("Template application failed: %s\n", e.what());
-        return ERR_TMPL;
+        return "";
    }
 }

@@ -349,7 +347,7 @@ std::optional<compare_variants_result> compare_variants(
    std::string output_B = apply_template(tmpl, params_B);

    // Check for template application failures
-    if (output_A == ERR_TMPL || output_B == ERR_TMPL) {
+    if (output_A.empty() || output_B.empty()) {
        return std::nullopt;
    }

@@ -60,21 +60,16 @@ struct generation_params {
    common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_AUTO;
    bool                                  stream              = true;
    std::string                           grammar;
-    bool                                  add_generation_prompt  = false;
-    common_chat_continuation              continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
-    common_chat_msg                       continue_msg;
-    bool                                  enable_thinking        = true;
-    std::chrono::system_clock::time_point now                    = std::chrono::system_clock::now();
+    bool                                  add_generation_prompt = false;
+    bool                                  enable_thinking       = true;
+    std::chrono::system_clock::time_point now                   = std::chrono::system_clock::now();
+    std::string                           generation_prompt;
    json                                  extra_context;
    bool                                  add_bos       = false;
    bool                                  add_eos       = false;
    bool                                  is_inference  = true;
    bool                                  add_inference = false;
    bool                                  mark_input    = true;  // whether to mark input strings in the jinja context
-
-    bool has_continuation() const {
-        return continue_final_message != COMMON_CHAT_CONTINUATION_NONE && !continue_msg.empty();
-    }
 };

 // ============================================================================
@@ -377,8 +372,6 @@ struct analyze_tools : analyze_base {

 struct autoparser {
    jinja::caps          jinja_caps;
-    std::string          user_start;
-    std::string          assistant_start;
    analyze_reasoning    reasoning;
    analyze_content      content;
    analyze_tools        tools;
@@ -389,15 +382,11 @@ struct autoparser {

    autoparser() = default;

-    // Find the starting marker for the user message and assistant message
-    std::string detect_user_start_marker(const common_chat_template & tmpl);
-    std::string detect_assistant_start_marker(const common_chat_template & tmpl);
-
    // Run full differential analysis on a template
    void analyze_template(const common_chat_template & tmpl);

    // Build the PEG parser for this template
-    common_peg_arena build_parser(const generation_params & inputs, const std::string & generation_prompt) const;
+    common_peg_arena build_parser(const generation_params & inputs) const;

  private:
    // Collect tokens from entire analysis to preserve
@@ -8,9 +8,6 @@
 #include "peg-parser.h"

 #include <algorithm>
-#include <cctype>
-#include <ostream>
-#include <sstream>

 #define ANSI_RESET  "\033[0m"
 #define ANSI_PURPLE "\033[1m\x1b[38;5;126m"
@@ -26,7 +23,6 @@ static const std::string FUN_SECOND = "SSS_SECOND_FUN_S";
 static const std::string ARG_FIRST = "AA_ARG_FST_AA";
 static const std::string ARG_SECOND = "BB_ARG_SND_BB";
 static const std::string USER_MSG = "U_USER_MSG Hello END_U";
-static const std::string USER_MSG_TWO = "V_USER_MSG Hello END_V";
 static const std::string ASSISTANT_MSG = "A_ASST_MSG I can help END_A";
 static const std::string THINKING_CONTENT = "REASON_PART I am thinking END_R";
 static const std::string CALL_ID_001 = "call00001";
@@ -75,7 +71,6 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
              analysis.content.end   = "<|END_OF_TURN_TOKEN|>";
              analysis.preserved_tokens.push_back("<|CHATBOT_TOKEN|>");
              analysis.preserved_tokens.push_back("<|END_OF_TURN_TOKEN|>");
-              analysis.user_start = "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>";
              LOG_DBG(ANSI_ORANGE "[Patch: Cohere Command R+]\n" ANSI_RESET);
          }
      },
@@ -113,59 +108,7 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
              analysis.tools.function.close        = "```";
              LOG_DBG(ANSI_ORANGE "[Patch: DeepSeek-R1-Distill-Qwen]\n" ANSI_RESET);
          }
-      },
-      // Nemotron Nano v2
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("<SPECIAL_10>") != std::string::npos && tmpl.src.find("<SPECIAL_11>") != std::string::npos &&
-              tmpl.src.find("<SPECIAL_12>") != std::string::npos && tmpl.src.find("<TOOL_RESPONSE>") != std::string::npos) {
-
-              analysis.tools.format.mode           = tool_format::JSON_NATIVE;
-              analysis.tools.format.section_start  = "";
-              analysis.tools.format.section_end    = "";
-              analysis.tools.format.per_call_start = "<TOOLCALL>";
-              analysis.tools.format.per_call_end   = "</TOOLCALL>";
-              analysis.content.mode                = content_mode::PLAIN;
-              analysis.content.start               = "";
-              analysis.content.end                 = "";
-              analysis.reasoning.mode              = reasoning_mode::TAG_BASED;
-              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>");
-              analysis.preserved_tokens.push_back("</TOOLCALL>");
-              LOG_DBG(ANSI_ORANGE "[Patch: Nemotron Nano v2]\n" ANSI_RESET);
-          }
-      },
-      // Fireworks
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("{%- set system_prompt = '<|start_header_id|>' + 'system' + '<|end_header_id|>\\n\\n'"
-            " + message['content'] | trim + '\\n' + system_prompt_suffix + '<|eot_id|>' -%}") != std::string::npos) {
-              analysis.assistant_start             = "<|start_header_id|>assistant<|end_header_id|>";
-              analysis.user_start                  = "<|start_header_id|>user<|end_header_id|>";
-              LOG_DBG(ANSI_ORANGE "[Patch: Fireworks v2]\n" ANSI_RESET);
-          }
-      },
-      // Solar Open
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("<|begin|>assistant<|think|><|end|>") != std::string::npos) {
-              analysis.assistant_start             = "<|begin|>assistant";
-              LOG_DBG(ANSI_ORANGE "[Patch: Solar Open]\n" ANSI_RESET);
-          }
-      },
-      // Apriel 1.6
-      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
-          if (tmpl.src.find("if not loop.last and '[BEGIN FINAL RESPONSE]' in asst_text") != std::string::npos) {
-              analysis.user_start                  = "<|begin_user|>";
-              analysis.assistant_start             = "<|begin_assistant|>";
-              LOG_DBG(ANSI_ORANGE "[Patch: Apriel 1.6]\n" ANSI_RESET);
-          }
-      },
-
+      }
    });

 // Common JSON structures
@@ -223,8 +166,6 @@ void autoparser::analyze_template(const common_chat_template & tmpl) {
    reasoning = analyze_reasoning(tmpl, jinja_caps.supports_tool_calls);
    content = analyze_content(tmpl, reasoning);
    tools = analyze_tools(jinja_caps.supports_tool_calls ? analyze_tools(tmpl, jinja_caps, reasoning) : analyze_tools());
-    assistant_start = detect_assistant_start_marker(tmpl);
-    user_start = detect_user_start_marker(tmpl);
    collect_preserved_tokens();

    for (auto & workaround : workarounds) {
@@ -232,8 +173,6 @@ void autoparser::analyze_template(const common_chat_template & tmpl) {
    }

    LOG_DBG("\n--- Reasoning & Content Structure ---\n");
-    LOG_DBG("user_msg_start: %s\n", user_start.c_str());
-    LOG_DBG("assistant_msg_start: %s\n", assistant_start.c_str());
    LOG_DBG("reasoning_mode: %s\n", mode_to_str(reasoning.mode).c_str());
    LOG_DBG("reasoning_start: '%s'\n", reasoning.start.c_str());
    LOG_DBG("reasoning_end: '%s'\n", reasoning.end.c_str());
@@ -306,120 +245,6 @@ void autoparser::collect_preserved_tokens() {
    add_token(tools.call_id.suffix);
 }

-std::string autoparser::detect_assistant_start_marker(const common_chat_template & tmpl) {
-    json user_msg = json{
-        { "role",    "user"   },
-        { "content", USER_MSG }
-    };
-
-    json assistant_no_reasoning = json{
-        { "role",    "assistant"   },
-        { "content", ASSISTANT_MSG }
-    };
-
-    template_params params;
-    params.messages              = json::array({ user_msg });
-    params.add_generation_prompt = false;
-    params.enable_thinking       = true;
-
-    auto comparison = compare_variants(
-        tmpl, params, [&](template_params & p) {
-            p.messages = json::array({ user_msg, assistant_no_reasoning });
-        }
-    );
-
-    if (!comparison) {
-        LOG_DBG(ANSI_ORANGE "%s: Template application failed, skipping assistant start detection\n" ANSI_RESET, __func__);
-        return "";
-    }
-
-    auto usermsg = comparison->diff.right;
-    if (usermsg.find(ASSISTANT_MSG) == std::string::npos) {
-        LOG_DBG(ANSI_ORANGE "%s: Did not find assistant message in assistant message block, skipping detection\n" ANSI_RESET, __func__);
-    }
-
-    auto ast_prefix = usermsg.substr(0, usermsg.find(ASSISTANT_MSG));
-    if (!reasoning.start.empty() && ast_prefix.find(trim_whitespace(reasoning.start)) != std::string::npos) {
-        ast_prefix = ast_prefix.substr(0, ast_prefix.find(trim_whitespace(reasoning.start)));
-    }
-    if (!reasoning.end.empty() && ast_prefix.find(trim_whitespace(reasoning.end)) != std::string::npos) {
-        ast_prefix = ast_prefix.substr(0, ast_prefix.find(trim_whitespace(reasoning.end)));
-    }
-    return trim_whitespace(ast_prefix);
-}
-
-std::string autoparser::detect_user_start_marker(const common_chat_template & tmpl) {
-    json user_msg = json{
-        { "role",    "user"   },
-        { "content", USER_MSG }
-    };
-
-    json assistant = json{
-        { "role",    "assistant"   },
-        { "content", ASSISTANT_MSG }
-    };
-
-    json user_msg_two = json{
-        { "role",    "user"       },
-        { "content", USER_MSG_TWO }
-    };
-
-    template_params params;
-    params.messages              = json::array({});
-    params.add_generation_prompt = false;
-    params.enable_thinking       = true;
-
-    auto comparison = compare_variants(
-        tmpl, params, [&](template_params & p) {
-            p.messages = json::array({ user_msg });
-        }
-    );
-
-    if (!comparison) {
-        LOG_DBG(ANSI_ORANGE "%s: Template application failed, unsupported empty messages? trying complex variant\n" ANSI_RESET, __func__);
-        params.messages = json::array({ user_msg_two, assistant });
-        comparison = compare_variants(
-            tmpl, params, [&](template_params & p) {
-                p.messages = json::array({ user_msg_two, assistant, user_msg });
-            }
-        );
-        if (!comparison) {
-            LOG_DBG(ANSI_ORANGE "%s: Template application failed for reserve variant, aborting\n" ANSI_RESET, __func__);
-            return "";
-        }
-    }
-
-    auto usermsg = comparison->diff.right;
-    if (usermsg.find(USER_MSG) == std::string::npos) {
-        LOG_DBG(ANSI_ORANGE "%s: Did not find user message in user message block, aborting detection\n" ANSI_RESET, __func__);
-    }
-
-    if (usermsg.find(ASSISTANT_MSG) != std::string::npos) {
-        usermsg = usermsg.substr(usermsg.find(ASSISTANT_MSG) + ASSISTANT_MSG.size());
-    }
-
-    auto candidate = usermsg.substr(0, usermsg.find(USER_MSG));
-    auto candidate_split = segmentize_markers(candidate);
-    std::stringstream result;
-    bool encountered_marker = false;
-    for (const auto & mrk : candidate_split) {
-        std::string lower_mrk = std::string(mrk.value);
-        std::transform(lower_mrk.begin(), lower_mrk.end(), lower_mrk.begin(),
-            [](unsigned char c) { return std::tolower(c); });
-        // heuristic to weed out potential end markers, but only at the start
-        if (mrk.type == segment_type::MARKER && !encountered_marker &&
-            (lower_mrk.find("end") != std::string::npos || lower_mrk.find("close") != std::string::npos)) {
-            continue;
-        }
-        if (mrk.type == segment_type::TEXT && !encountered_marker && trim_whitespace(mrk.value).empty()) {
-            continue;
-        }
-        encountered_marker |= mrk.type == segment_type::MARKER;
-        result << mrk.value;
-    }
-    return trim_whitespace(result.str());
-}
-
 analyze_reasoning::analyze_reasoning(const common_chat_template & tmpl, bool supports_tools)
    : analyze_base(tmpl) {
    LOG_DBG(ANSI_PURPLE "=== Starting differential analysis ===\n" ANSI_RESET);
@@ -358,7 +358,35 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
            if (is_potential_container) {
                value_content = normalize_container_value(value_content);
            }
-            value_to_add += value_content;
+
+            // Try to parse as JSON value (number, bool, null, object, array)
+            try {
+                ordered_json parsed = ordered_json::parse(value_content);
+                if (parsed.is_string()) {
+                    // Don't add closing quote yet (added by arg_close) for monotonic streaming
+                    std::string escaped = parsed.dump();
+                    if (!escaped.empty() && escaped.back() == '"') {
+                        escaped.pop_back();
+                    }
+                    value_to_add          = escaped;
+                    closing_quote_pending = true;
+                } else {
+                    // Non-string values: use raw content to preserve whitespace for monotonicity
+                    value_to_add = value_content;
+                }
+            } catch (...) {
+                if (node.is_partial && is_potential_container) {
+                    // Partial container: pass through the already-normalized content
+                    value_to_add = value_content;
+                } else {
+                    // Not valid JSON - treat as string value
+                    if (!closing_quote_pending) {
+                        value_to_add          = "\"";
+                        closing_quote_pending = true;
+                    }
+                    value_to_add += escape_json_string_inner(value_content);
+                }
+            }
        }

        args_target() += value_to_add;
@@ -785,7 +813,7 @@ common_peg_parser common_chat_peg_builder::prefix(const std::string & s, const s
    if (delimiter.empty()) {
        return literal(s);
    }
-    return literal(s.substr(0, s.find(delimiter)));
+    return literal(s.substr(0, s.rfind(delimiter)));
 }

 common_peg_parser common_chat_peg_builder::optspace(const std::string & tag) {
@@ -90,7 +90,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {

    // Use for schema-declared string types - won't be treated as potential JSON container
    common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
-    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
+    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_VALUE, p)); }


    // Return a parser that parses the prefix of a string, up to a given delimiter.
@@ -70,65 +70,6 @@ static bool has_content_or_tool_calls(const common_chat_msg & msg) {
    return !msg.content.empty() || !msg.tool_calls.empty();
 }

-std::string common_chat_msg::render_content(const std::string & delimiter) const {
-    if (!content.empty() && !content_parts.empty()) {
-        throw std::runtime_error("Cannot specify both content and content_parts");
-    }
-    if (!content.empty()) {
-        return content;
-    }
-
-    std::string text;
-    for (const auto & part : content_parts) {
-        if (part.type == "text") {
-            if (!text.empty()) {
-                text += delimiter;
-            }
-            text += part.text;
-        }
-    }
-    return text;
-}
-
-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims) {
-    if (delims.empty() || prompt.empty()) {
-        return {};
-    }
-
-    auto parser = build_peg_parser([&](common_peg_parser_builder & p) {
-        std::vector<std::string>       all_delims;
-        std::vector<common_peg_parser> tagged_messages;
-
-        all_delims.reserve(delims.size());
-        tagged_messages.reserve(delims.size());
-        for (const auto & d : delims) {
-            all_delims.push_back(d.delimiter);
-        }
-
-        auto any_delim = p.until_one_of(all_delims);
-        for (const auto & d : delims) {
-            tagged_messages.push_back(p.tag(d.role, p.literal(d.delimiter) + any_delim));
-        }
-
-        return any_delim + p.zero_or_more(p.choice(tagged_messages)) + p.end();
-    });
-
-    common_peg_parse_context ctx(prompt);
-    const auto result = parser.parse(ctx);
-    if (!result.success()) {
-        return {};
-    }
-
-    std::vector<common_chat_msg_span> spans;
-    ctx.ast.visit(result, [&](const common_peg_ast_node & node) {
-        if (!node.tag.empty()) {
-            spans.push_back({ node.tag, node.start, node.end - node.start });
-        }
-    });
-
-    return spans;
-}
-
 json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
    if (!content.empty() && !content_parts.empty()) {
        throw std::runtime_error("Cannot specify both content and content_parts");
@@ -510,22 +451,6 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
    return result;
 }

-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
-    if (value.is_boolean() && value.get<bool>()) {
-        return COMMON_CHAT_CONTINUATION_AUTO;
-    }
-    if (value.is_string()) {
-        auto value_str = value.get<std::string>();
-        if (value_str == "reasoning_content") {
-            return COMMON_CHAT_CONTINUATION_REASONING;
-        }
-        if (value_str == "content") {
-            return COMMON_CHAT_CONTINUATION_CONTENT;
-        }
-    }
-    return COMMON_CHAT_CONTINUATION_NONE;
-}
-
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
    if (use_jinja) {
        try {
@@ -886,36 +811,6 @@ std::string common_chat_template_direct_apply(
    return common_chat_template_direct_apply_impl(tmpl, inputs, std::nullopt, std::nullopt, std::nullopt);
 }

-static std::string common_chat_template_generation_prompt_impl(
-    const common_chat_template & tmpl,
-    const autoparser::generation_params & inputs,
-    const std::optional<json> & messages_override = std::nullopt,
-    const std::optional<json> & tools_override = std::nullopt,
-    const std::optional<json> & additional_context = std::nullopt) {
-
-    auto adjusted_messages = messages_override ? *messages_override : inputs.messages;
-
-    autoparser::generation_params params = inputs;
-    params.add_generation_prompt = false;
-    params.continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
-    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
-    params.add_generation_prompt = true;
-    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
-
-    size_t prefix_len = 0;
-    size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
-    while (prefix_len < min_size && no_gen_prompt[prefix_len] == gen_prompt[prefix_len]) {
-        prefix_len++;
-    }
-    return gen_prompt.substr(prefix_len);
-}
-
-std::string common_chat_template_generation_prompt(
-    const common_chat_template & tmpl,
-    const autoparser::generation_params & inputs) {
-    return common_chat_template_generation_prompt_impl(tmpl, inputs, std::nullopt, std::nullopt, std::nullopt);
-}
-
 static common_chat_params common_chat_params_init_ministral_3(const common_chat_template &    tmpl,
                                                              const autoparser::generation_params & inputs) {
    common_chat_params data;
@@ -968,7 +863,6 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
    data.thinking_start_tag = "[THINK]";
    data.thinking_end_tag   = "[/THINK]";
    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.preserved_tokens  = {
        "[THINK]",
@@ -977,19 +871,8 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
        "[ARGS]",
    };

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "[THINK]" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "[/THINK]" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.eps();
+        auto generation_prompt = p.prefix(inputs.generation_prompt, "[THINK]");
        auto reasoning =
            extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();

@@ -1080,15 +963,6 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
    }

    data.prompt            = prompt;
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-    data.message_spans = common_chat_split_by_role(prompt, {
-        { "assistant", "<|start|>assistant" },
-        { "user",      "<|start|>user"      },
-        { "system",    "<|start|>developer" },
-        { "system",    "<|start|>system"    },
-        { "tool",      "<|start|>functions" },
-    });
-
    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking = true;

@@ -1098,18 +972,6 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
        "<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
    };

-    // Adjust prompt for continuation
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|start|>assistant<|channel|>analysis<|message|>" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<|end|><|start|>assistant<|channel|>final<|message|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
@@ -1218,21 +1080,14 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
    common_chat_params data;

    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);

    if (inputs.add_generation_prompt && string_ends_with(data.prompt, "<turn|>\n")) {
        // This may happen if the model generates content + tool_call, the
        // template does not add the model's next turn and confuses the model
        // from emitting its proper reasoning token sequence.
-        data.generation_prompt = "<|turn>model\n";
-        data.prompt += data.generation_prompt;
+        data.prompt += "<|turn>model\n";
    }

-    data.message_spans = common_chat_split_by_role(data.prompt, {
-        { "user",      "<|turn>user\n"  },
-        { "assistant", "<|turn>model\n" },
-    });
-
    data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
    data.supports_thinking  = true;
    data.thinking_start_tag = "<|channel>thought";
@@ -1246,25 +1101,13 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
        "<|turn>",
    };

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = string_ends_with(data.prompt, "<turn|>\n") ? "<|turn>model\n" : "";
-        data.generation_prompt += "<|channel>thought\n" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<channel|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;

    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start = p.rule("start", p.optional(p.literal("<|turn>model\n")));
+        auto start = p.rule("start", p.prefix(inputs.generation_prompt, "<|channel>"));

        if (extract_reasoning) {
            p.rule("thought", p.literal("<|channel>thought") + p.space() + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
@@ -1381,22 +1224,15 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
                                                                   const autoparser::generation_params & inputs) {
    common_chat_params data;

-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens  = {
+    data.prompt           = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens = {
        ">>>all",
    };

    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-        data.generation_prompt = "<|start_header_id|>assistant<|end_header_id|>\n\n>>>all\n" + msg.render_content();
-        data.prompt += data.generation_prompt;
-    }
-
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        // Functionary v3.2 format:
        // - Normal content: >>>all\n{content}
@@ -1408,7 +1244,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
        // When no tools, content goes until end
        auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
        auto content_until_end  = p.literal("all\n") + p.content(p.rest());
-        auto generation_prompt  = p.literal("<|start_header_id|>assistant<|end_header_id|>\n\n>>>");
+        auto generation_prompt  = p.literal(inputs.generation_prompt);

        // If no tools or tool_choice is NONE, just parse content
        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
@@ -1482,10 +1318,9 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
                                                          const autoparser::generation_params & inputs) {
    common_chat_params data;

-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
    data.preserved_tokens  = {
        "<|tool_calls_section_begin|>",
        "<|tool_calls_section_end|>",
@@ -1508,22 +1343,10 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp

    const std::string THINK_START = "<think>";
    const std::string THINK_END   = "</think>";
-    const std::string GEN_PROMPT  = "<|im_assistant|>assistant<|im_middle|>";

    data.thinking_start_tag = THINK_START;
    data.thinking_end_tag   = THINK_END;

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        // Kimi K2 Thinking format:
        // - Reasoning: <think>{reasoning}</think>
@@ -1543,7 +1366,7 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
        auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
            p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
            p.optional(p.literal(THINK_END))) : p.eps();
-        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);


        // Content only parser (no tools)
@@ -1619,7 +1442,6 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
    common_chat_params data;

    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking = true;
    data.preserved_tokens  = {
@@ -1639,24 +1461,12 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
    const std::string TOOL_CALL_END   = "<|tool_call_end|>";
    const std::string THINK_START     = "<think>";
    const std::string THINK_END       = "</think>";
-    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";

    data.thinking_start_tag = THINK_START;
    data.thinking_end_tag   = THINK_END;

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
        auto end = p.end();

        auto reasoning = p.eps();
@@ -1711,7 +1521,6 @@ static common_chat_params common_chat_params_init_lfm2_5(const common_chat_templ
    common_chat_params data;

    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking = true;
    data.preserved_tokens  = {
@@ -1727,24 +1536,12 @@ static common_chat_params common_chat_params_init_lfm2_5(const common_chat_templ

    const std::string THINK_START     = "<think>";
    const std::string THINK_END       = "</think>";
-    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";

    data.thinking_start_tag = THINK_START;
    data.thinking_end_tag   = THINK_END;

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
        auto end = p.end();

        auto reasoning = p.eps();
@@ -1795,7 +1592,6 @@ static common_chat_params common_chat_params_init_gigachat_v3(
    common_chat_params data;

    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking = false;
    data.preserved_tokens  = {
@@ -1803,12 +1599,6 @@ static common_chat_params common_chat_params_init_gigachat_v3(
        "<|role_sep|>\n",
    };

-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-        data.generation_prompt = "assistant<|role_sep|>\n" + msg.render_content();
-        data.prompt += data.generation_prompt;
-    }
-
    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
    const auto *tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
@@ -1844,7 +1634,7 @@ static common_chat_params common_chat_params_init_gigachat_v3(
            ret = p.content(p.rest());
        }

-        return p.literal("assistant<|role_sep|>\n") + ret;
+        return p.literal(inputs.generation_prompt) + ret;
    });

    data.parser = parser.save();
@@ -1872,13 +1662,12 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
                                                                 const autoparser::generation_params & inputs) {
    common_chat_params data;

-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
    data.thinking_start_tag = "<think>";
    data.thinking_end_tag   = "</think>";
-    data.preserved_tokens   = {
+    data.preserved_tokens  = {
        "｜DSML｜",
        "<think>",
        "</think>",
@@ -1898,21 +1687,9 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
    const std::string INVOKE_END   = "</" + DSML + "invoke>";
    const std::string PARAM_START  = "<" + DSML + "parameter";
    const std::string PARAM_END    = "</" + DSML + "parameter>";
-    const std::string GEN_PROMPT   = "<｜Assistant｜>";
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }

    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
        auto end = p.end();

        auto reasoning = p.eps();
@@ -2339,6 +2116,21 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
    return std::nullopt;
 }

+static std::string common_chat_templates_generation_prompt(const common_chat_template & tmpl, const autoparser::generation_params & inputs) {
+    autoparser::generation_params params = inputs;
+    params.add_generation_prompt = false;
+    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params);
+    params.add_generation_prompt = true;
+    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params);
+
+    size_t prefix_len = 0;
+    size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
+    while (prefix_len < min_size && no_gen_prompt[prefix_len] == gen_prompt[prefix_len]) {
+        prefix_len++;
+    }
+    return gen_prompt.substr(prefix_len);
+}
+
 static common_chat_params common_chat_templates_apply_jinja(const struct common_chat_templates *        tmpls,
                                                            const struct common_chat_templates_inputs & inputs) {
    autoparser::generation_params params;
@@ -2357,27 +2149,6 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
    params.add_bos               = tmpls->add_bos;
    params.add_eos               = tmpls->add_eos;

-    params.continue_final_message = inputs.continue_final_message;
-    if (params.continue_final_message != COMMON_CHAT_CONTINUATION_NONE) {
-        params.add_generation_prompt = false;
-
-        if (!inputs.messages.empty()) {
-            // Render messages[:-1] and store continuation message separately
-            params.continue_msg = inputs.messages.back();
-            params.messages.erase(params.messages.size() - 1);
-        }
-
-        if (params.continue_final_message == COMMON_CHAT_CONTINUATION_AUTO && !inputs.messages.empty()) {
-            // Resolve based on message content
-            params.continue_final_message = COMMON_CHAT_CONTINUATION_CONTENT;
-            if (!params.continue_msg.reasoning_content.empty() &&
-                params.continue_msg.content.empty() &&
-                params.continue_msg.content_parts.empty()) {
-                params.continue_final_message = COMMON_CHAT_CONTINUATION_REASONING;
-            }
-        }
-    }
-
    if (src.find("<|channel|>") == std::string::npos) {
        // map developer to system for all models except for GPT-OSS
        workaround::map_developer_role_to_system(params.messages);
@@ -2398,6 +2169,8 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        workaround::func_args_not_string(params.messages);
    }

+    params.generation_prompt = common_chat_templates_generation_prompt(tmpl, params);
+
    params.extra_context = common_chat_extra_context();
    for (auto el : inputs.chat_template_kwargs) {
        params.extra_context[el.first] = json::parse(el.second);
@@ -2427,16 +2200,17 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        auto params_copy               = params;
        params_copy.reasoning_format   = COMMON_REASONING_FORMAT_NONE;
        data.prompt                    = common_chat_template_direct_apply_impl(tmpl, params_copy);
-        data.generation_prompt         = common_chat_template_generation_prompt_impl(tmpl, params);
        data.format                    = COMMON_CHAT_FORMAT_PEG_NATIVE;
-        auto parser                    = build_chat_peg_parser([&data](common_chat_peg_builder &p) {
-            return p.literal(data.generation_prompt) << p.content(p.rest());
+        data.generation_prompt         = params.generation_prompt;
+        auto parser                    = build_chat_peg_parser([&params](common_chat_peg_builder &p) {
+            return p.prefix(params.generation_prompt) << p.content(p.rest());
        });
        data.parser                    = parser.save();
        return data;
    }

    if (auto result = common_chat_try_specialized_template(tmpl, src, params)) {
+        result->generation_prompt = params.generation_prompt;
        return *result;
    }

@@ -2445,24 +2219,12 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        struct autoparser::autoparser autoparser;
        autoparser.analyze_template(tmpl);
        auto auto_params = autoparser::peg_generator::generate_parser(tmpl, params, autoparser);
-
-        std::vector<common_chat_msg_delimiter> delimiters;
-        if (!autoparser.assistant_start.empty()) {
-            delimiters.push_back({ "assistant", autoparser.assistant_start });
-        }
-        if (!autoparser.user_start.empty()) {
-            delimiters.push_back({ "user", autoparser.user_start });
-        }
-
-        if (!delimiters.empty()) {
-            auto_params.message_spans = common_chat_split_by_role(auto_params.prompt, delimiters);
-        }
-
        auto_params.supports_thinking = autoparser.reasoning.mode != autoparser::reasoning_mode::NONE;
        if (auto_params.supports_thinking) {
            auto_params.thinking_start_tag = trim_whitespace(autoparser.reasoning.start);
            auto_params.thinking_end_tag   = trim_whitespace(autoparser.reasoning.end);
        }
+        auto_params.generation_prompt = params.generation_prompt;
        common_peg_arena arena;
        arena.load(auto_params.parser);
        LOG_DBG("%s: generated parser:\n%s\n\nparser generation prompt: %s\n", __func__, arena.dump(arena.root()).c_str(), auto_params.generation_prompt.c_str());
@@ -89,8 +89,6 @@ struct common_chat_msg {

    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;

-    std::string render_content(const std::string & delimiter = "\n\n") const;
-
    bool empty() const {
        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() &&
               tool_name.empty() && tool_call_id.empty();
@@ -143,17 +141,6 @@ struct common_chat_msg_diff {
    }
 };

-struct common_chat_msg_span {
-    std::string role;
-    std::size_t pos = 0;
-    std::size_t len = 0;
-};
-
-struct common_chat_msg_delimiter {
-    std::string role;
-    std::string delimiter;
-};
-
 struct common_chat_tool {
    std::string name;
    std::string description;
@@ -177,22 +164,12 @@ enum common_chat_format {
    COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };

-
-// Continuation method provided via `continue_final_message`
-enum common_chat_continuation {
-    COMMON_CHAT_CONTINUATION_NONE,
-    COMMON_CHAT_CONTINUATION_AUTO,
-    COMMON_CHAT_CONTINUATION_REASONING,
-    COMMON_CHAT_CONTINUATION_CONTENT,
-};
-
 struct common_chat_templates_inputs {
    std::vector<common_chat_msg>          messages;
    std::string                           grammar;
    std::string                           json_schema;
-    bool                                  add_generation_prompt  = true;
-    common_chat_continuation              continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
-    bool                                  use_jinja              = true;
+    bool                                  add_generation_prompt = true;
+    bool                                  use_jinja             = true;
    // Parameters below only supported when use_jinja is true
    std::vector<common_chat_tool>         tools;
    common_chat_tool_choice               tool_choice         = COMMON_CHAT_TOOL_CHOICE_AUTO;
@@ -219,7 +196,6 @@ struct common_chat_params {
    std::vector<std::string>            preserved_tokens;
    std::vector<std::string>            additional_stops;
    std::string                         parser;
-    std::vector<common_chat_msg_span>   message_spans;
 };

 // per-message parsing syntax
@@ -231,8 +207,6 @@ struct common_chat_parser_params {
    bool                    reasoning_in_content = false;
    std::string             generation_prompt;
    bool                    parse_tool_calls     = true;
-    bool                    is_continuation      = false;
-    bool                    echo                 = false;  // Include assistant prefilled msg in output
    bool                    debug                = false;  // Enable debug output for PEG parser
    common_peg_arena        parser               = {};
    common_chat_parser_params() = default;
@@ -293,8 +267,6 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::or

 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);

-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
-
 // DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);

@@ -307,16 +279,11 @@ std::string common_chat_template_direct_apply(
    const common_chat_template & tmpl,
    const autoparser::generation_params & inputs);

-std::string common_chat_template_generation_prompt(
-    const common_chat_template &          tmpl,
-    const autoparser::generation_params & inputs);
-
 std::optional<common_chat_params> common_chat_try_specialized_template(
        const common_chat_template &          tmpl,
        const std::string &                   src,
        autoparser::generation_params & params);

-
 // specialized per-task preset
 struct common_chat_prompt_preset {
    std::string system;
@@ -324,6 +291,3 @@ struct common_chat_prompt_preset {
 };

 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
-
-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims);
-
@@ -7,7 +7,6 @@
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
-#include "speculative.h"
 #include "unicode.h"

 #include <algorithm>
@@ -373,7 +372,7 @@ void common_init() {
    llama_log_set(common_log_default_callback, NULL);
 }

-void common_params_print_info(const common_params & params, bool print_devices) {
+void common_params_print_info(const common_params & params) {
 #ifdef NDEBUG
    const char * build_type = "";
 #else
@@ -382,16 +381,12 @@ void common_params_print_info(const common_params & params, bool print_devices)
    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);

    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) {
-        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);
-            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);
-        }
+    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);
+        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);
    }
    LOG_INF("%s\n", common_params_get_system_info(params).c_str());
 }
@@ -445,27 +440,6 @@ std::string string_strip(const std::string & str) {
    return str.substr(start, end - start);
 }

-std::string string_lcs(std::string_view a, std::string_view b) {
-    if (a.empty() || b.empty()) return {};
-
-    std::vector<std::vector<size_t>> dp(a.size() + 1, std::vector<size_t>(b.size() + 1, 0));
-    size_t best_len = 0;
-    size_t best_end_a = 0;
-
-    for (size_t i = 1; i <= a.size(); ++i) {
-        for (size_t j = 1; j <= b.size(); ++j) {
-            if (a[i - 1] == b[j - 1]) {
-                dp[i][j] = dp[i - 1][j - 1] + 1;
-                if (dp[i][j] > best_len) {
-                    best_len = dp[i][j];
-                    best_end_a = i;
-                }
-            }
-        }
-    }
-    return std::string(a.substr(best_end_a - best_len, best_len));
-}
-
 std::string string_get_sortable_timestamp() {
    using clock = std::chrono::system_clock;

@@ -1181,7 +1155,7 @@ struct common_init_result::impl {
    std::vector<llama_sampler_seq_config> samplers_seq_config;
 };

-common_init_result::common_init_result(common_params & params, bool model_only) :
+common_init_result::common_init_result(common_params & params) :
    pimpl(new impl{}) {
    auto mparams = common_model_params_to_llama(params);
    auto cparams = common_context_params_to_llama(params);
@@ -1194,7 +1168,7 @@ common_init_result::common_init_result(common_params & params, bool model_only)
            params.tensor_buft_overrides.data(),
            params.fit_params_target.data(),
            params.fit_params_min_ctx,
-            params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
    }

    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -1204,10 +1178,6 @@ common_init_result::common_init_result(common_params & params, bool model_only)

    pimpl->model.reset(model);

-    if (model_only) {
-        return;
-    }
-
    const llama_vocab * vocab = llama_model_get_vocab(model);

    // load and optionally apply lora adapters
@@ -1311,8 +1281,8 @@ std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
    return pimpl->lora;
 }

-common_init_result_ptr common_init_from_params(common_params & params, bool model_only) {
-    common_init_result_ptr res(new common_init_result(params, model_only));
+common_init_result_ptr common_init_from_params(common_params & params) {
+    common_init_result_ptr res(new common_init_result(params));

    llama_model * model = res->model();
    if (model == NULL) {
@@ -1320,10 +1290,6 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
        return res;
    }

-    if (model_only) {
-        return res;
-    }
-
    llama_context * lctx = res->context();
    if (lctx == NULL) {
        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
@@ -1387,7 +1353,7 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
    }

    if (params.warmup) {
-        LOG_INF("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
+        LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);

        llama_set_warmup(lctx, true);

@@ -1469,12 +1435,6 @@ common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
        goto done;
    }

-    if (llama_n_rs_seq(ctx) > 0) {
-        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)) {
        LOG_TRC("%s: the context does not support partial sequence removal\n", __func__);
@@ -1489,23 +1449,6 @@ done:
    return res;
 }

-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());
-    }
-}
-
-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);
-}
-
-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_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
    std::vector<llama_adapter_lora *> loras;
    std::vector<float> scales;
@@ -1562,8 +1505,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &

    cparams.n_ctx             = params.n_ctx;
    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_batch           = params.n_batch;
    cparams.n_ubatch          = params.n_ubatch;
    cparams.n_threads         = params.cpuparams.n_threads;
@@ -2133,11 +2074,3 @@ void common_prompt_checkpoint::load_dft(
        GGML_ABORT("checkpoint size mismatch: expected %zu, got %zu\n", data_dft.size(), n);
    }
 }
-
-void common_prompt_checkpoint::clear_tgt() {
-    data_tgt.clear();
-}
-
-void common_prompt_checkpoint::clear_dft() {
-    data_dft.clear();
-}
@@ -13,7 +13,6 @@
 #include <string_view>
 #include <vector>
 #include <map>
-#include <algorithm>

 #if defined(_WIN32) && !defined(_WIN32_WINNT)
 #define _WIN32_WINNT 0x0A00
@@ -160,7 +159,6 @@ enum common_speculative_type {
    COMMON_SPECULATIVE_TYPE_NONE,          // no speculative decoding
    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_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
@@ -299,13 +297,11 @@ struct common_params_model {

 // draft-model-based speculative decoding parameters
 struct common_params_speculative_draft {
-    int32_t n_max = 3; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min = 0;  // minimum number of draft tokens to use for speculative decoding

-    float p_split = 0.1f; // speculative decoding split probability
-    float p_min   = 0.0f; // minimum speculative decoding probability (greedy)
-
-    bool backend_sampling = true; // offload draft sampling to the backend (default: on)
+    float p_split = 0.1f;  // speculative decoding split probability
+    float p_min   = 0.75f; // minimum speculative decoding probability (greedy)

    common_params_model mparams;

@@ -359,14 +355,6 @@ struct common_params_speculative {
    bool has_dft() const {
        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;
-        });
-
-        return needs_rs_seq ? draft.n_max : 0u;
-    }
 };

 struct common_params_vocoder {
@@ -431,7 +419,6 @@ struct common_params {
    int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
    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 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)
@@ -480,7 +467,7 @@ struct common_params {

    std::set<std::string> model_alias;     // model aliases                                                 // NOLINT
    std::set<std::string> model_tags;      // model tags (informational, not used for routing)              // NOLINT
-    std::string hf_token             = ""; // HF token (aka bearer token)                                   // NOLINT
+    std::string hf_token             = ""; // HF token                                                      // NOLINT
    std::string prompt               = "";                                                                  // NOLINT
    std::string system_prompt        = "";                                                                  // NOLINT
    std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
@@ -508,7 +495,6 @@ 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
@@ -589,14 +575,14 @@ struct common_params {
    // server params
    int32_t port                = 8080;          // server listens on this network port
    bool    reuse_port          = false;         // allow multiple sockets to bind to the same port
-    int32_t timeout_read        = 3600;          // http read timeout in seconds
+    int32_t timeout_read        = 600;           // http read timeout in seconds
    int32_t timeout_write       = timeout_read;  // http write timeout in seconds
    int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
    int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
    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 = 256;   // minimum spacing between context checkpoints
+    int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
    int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.

    std::string hostname      = "127.0.0.1";
@@ -618,17 +604,15 @@ struct common_params {

    std::map<std::string, std::string> default_template_kwargs;

-    // UI configs
-    bool ui = true;
-
-    // Deprecated: use ui, ui_mcp_proxy, ui_config_json instead
-    bool webui = ui;
+    // webui configs
+#ifdef LLAMA_WEBUI_DEFAULT_ENABLED
+    bool webui = LLAMA_WEBUI_DEFAULT_ENABLED != 0;
+#else
+    bool webui = true; // default to enabled when not set
+#endif
    bool webui_mcp_proxy = false;
    std::string webui_config_json;

-    bool ui_mcp_proxy = false;
-    std::string ui_config_json;
-
    // "advanced" endpoints are disabled by default for better security
    bool endpoint_slots   = true;
    bool endpoint_props   = false; // only control POST requests, not GET
@@ -706,7 +690,7 @@ struct common_params {
 // initializes the logging system and prints info about the build
 void common_init();

-void common_params_print_info(const common_params & params, bool print_devices = true);
+void common_params_print_info(const common_params & params);
 std::string common_params_get_system_info(const common_params & params);

 bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
@@ -733,7 +717,6 @@ std::string string_format(const char * fmt, ...);

 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
-std::string string_lcs(std::string_view a, std::string_view b);

 std::string string_join(const std::vector<std::string> & values, const std::string & separator);
 std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
@@ -858,7 +841,7 @@ struct common_sampler;

 // note: defines the model, context, samplers, ets. lifetimes
 struct common_init_result {
-    common_init_result(common_params & params, bool model_only = false);
+    common_init_result(common_params & params);
    ~common_init_result();

    llama_model * model();
@@ -876,7 +859,7 @@ private:

 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);
+common_init_result_ptr common_init_from_params(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);
@@ -893,20 +876,15 @@ std::string common_get_model_endpoint();
 //

 enum common_context_seq_rm_type {
-    COMMON_CONTEXT_SEQ_RM_TYPE_NO           = 0, // seq_rm not supported (e.g. no memory module)
-    COMMON_CONTEXT_SEQ_RM_TYPE_PART         = 1, // can seq_rm partial sequences
-    COMMON_CONTEXT_SEQ_RM_TYPE_FULL         = 2, // can seq_rm full sequences only
-    COMMON_CONTEXT_SEQ_RM_TYPE_RS = 3, // can seq_rm partial sequences, bounded by n_rs_seq
+    COMMON_CONTEXT_SEQ_RM_TYPE_NO   = 0, // seq_rm not supported (e.g. no memory module)
+    COMMON_CONTEXT_SEQ_RM_TYPE_PART = 1, // can seq_rm partial sequences
+    COMMON_CONTEXT_SEQ_RM_TYPE_FULL = 2, // can seq_rm full sequences only
 };

 // check if the llama_context can remove sequences
 // note: clears the memory of the context
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);

-// 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
@@ -1088,7 +1066,4 @@ struct common_prompt_checkpoint {
            llama_context * ctx,
            llama_seq_id seq_id,
            llama_state_seq_flags flags) const;
-
-    void clear_tgt();
-    void clear_dft();
 };
@@ -292,10 +292,6 @@ 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
@@ -361,10 +357,6 @@ static int common_download_file_single_online(const std::string & url,
            LOG_DBG("%s: using cached file (same etag): %s\n", __func__, path.c_str());
            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;
@@ -574,11 +566,8 @@ static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
    return result;
 }

-// pick the best sibling GGUF whose filename contains `keyword` (e.g. "mmproj" / "mtp"),
-// 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) {
+static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
+                                          const std::string        & model) {
    hf_cache::hf_file best;
    size_t best_depth = 0;
    int best_diff = 0;
@@ -590,20 +579,20 @@ static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,

    for (const auto & f : files) {
        if (!string_ends_with(f.path, ".gguf") ||
-            f.path.find(keyword) == std::string::npos) {
+            f.path.find("mmproj") == std::string::npos) {
            continue;
        }

-        auto sib_parts = string_split<std::string>(f.path, '/');
-        auto sib_dir = sib_parts.end() - 1;
+        auto mmproj_parts = string_split<std::string>(f.path, '/');
+        auto mmproj_dir = mmproj_parts.end() - 1;

        auto [_, dir] = std::mismatch(model_parts.begin(), model_dir,
-                                      sib_parts.begin(), sib_dir);
-        if (dir != sib_dir) {
+                                      mmproj_parts.begin(), mmproj_dir);
+        if (dir != mmproj_dir) {
            continue;
        }

-        size_t depth = dir - sib_parts.begin();
+        size_t depth = dir - mmproj_parts.begin();
        auto bits = extract_quant_bits(f.path);
        auto diff = std::abs(bits - model_bits);

@@ -617,16 +606,6 @@ static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
    return best;
 }

-static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
-                                          const std::string        & model) {
-    return find_best_sibling(files, model, "mmproj");
-}
-
-static hf_cache::hf_file find_best_mtp(const hf_cache::hf_files & files,
-                                       const std::string        & model) {
-    return find_best_sibling(files, model, "mtp-");
-}
-
 static bool gguf_filename_is_model(const std::string & filepath) {
    if (!string_ends_with(filepath, ".gguf")) {
        return false;
@@ -638,8 +617,7 @@ 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("imatrix") == std::string::npos;
 }

 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -695,13 +673,11 @@ struct hf_plan {
    hf_cache::hf_file primary;
    hf_cache::hf_files model_files;
    hf_cache::hf_file mmproj;
-    hf_cache::hf_file mtp;
 };

 static hf_plan get_hf_plan(const common_params_model  & model,
                           const common_download_opts & opts,
-                           bool download_mmproj,
-                           bool download_mtp) {
+                           bool download_mmproj) {
    hf_plan plan;
    hf_cache::hf_files all;

@@ -747,10 +723,6 @@ static hf_plan get_hf_plan(const common_params_model  & model,
        plan.mmproj = find_best_mmproj(all, primary.path);
    }

-    if (download_mtp) {
-        plan.mtp = find_best_mtp(all, primary.path);
-    }
-
    return plan;
 }

@@ -783,26 +755,22 @@ static std::vector<download_task> get_url_tasks(const common_params_model & mode
 }

 common_download_model_result common_download_model(const common_params_model  & model,
-                                                   const common_download_opts & opts) {
+                                                   const common_download_opts & opts,
+                                                   bool download_mmproj) {
    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);
+        hf = get_hf_plan(model, opts, download_mmproj);
        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 {
@@ -814,22 +782,18 @@ common_download_model_result common_download_model(const common_params_model  &
        return result;
    }

-    std::vector<std::future<int>> futures;
+    std::vector<std::future<bool>> futures;
    for (const auto & task : tasks) {
        futures.push_back(std::async(std::launch::async,
            [&task, &opts, is_hf]() {
-                return common_download_file_single(task.url, task.path, opts, is_hf);
+                int status = common_download_file_single(task.url, task.path, opts, is_hf);
+                return is_http_status_ok(status);
            }
        ));
    }

    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) {
+        if (!f.get()) {
            return {};
        }
    }
@@ -843,10 +807,6 @@ common_download_model_result common_download_model(const common_params_model  &
        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;
    }
@@ -986,8 +946,7 @@ 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("mmproj") != std::string::npos) {
            continue;
        }
        if (seen.insert(f.repo_id + ":" + split.tag).second) {
@@ -52,9 +52,6 @@ struct common_download_opts {
    std::string bearer_token;
    common_header_list headers;
    bool offline = 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;
 };

@@ -62,12 +59,6 @@ struct common_download_opts {
 struct common_download_model_result {
    std::string model_path;
    std::string mmproj_path;
-    std::string mtp_path;
-};
-
-// 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") {}
 };

 // Download model from HuggingFace repo or URL
@@ -92,12 +83,12 @@ struct common_skip_download_exception : public std::runtime_error {
 // 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)
+// returns result with model_path and mmproj_path (empty on failure)
 common_download_model_result common_download_model(
    const common_params_model & model,
-    const common_download_opts & opts = {}
+    const common_download_opts & opts = {},
+    bool download_mmproj = false
 );

 // returns list of cached models
@@ -105,7 +96,6 @@ std::vector<common_cached_model_info> common_list_cached_models();

 // 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,
@@ -26,7 +26,7 @@ class common_params_fit_exception : public std::runtime_error {
    using std::runtime_error::runtime_error;
 };

-std::vector<llama_device_memory_data> common_get_device_memory_data(
+static std::vector<llama_device_memory_data> common_get_device_memory_data(
        const char * path_model,
        const llama_model_params * mparams,
        const llama_context_params * cparams,
@@ -1,11 +1,6 @@
 #pragma once

 #include "ggml.h"
-#include "ggml-backend.h"
-#include "llama.h"
-#include "../src/llama-ext.h"
-
-#include <vector>

 enum common_params_fit_status {
    COMMON_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
@@ -35,14 +30,3 @@ void common_fit_print(
                struct llama_context_params * cparams);

 void common_memory_breakdown_print(const struct llama_context * ctx);
-
-// Load a model + context with no_alloc and return the per-device memory breakdown.
-std::vector<llama_device_memory_data> common_get_device_memory_data(
-                                  const char   * path_model,
-        const struct llama_model_params         * mparams,
-        const struct llama_context_params       * cparams,
-        std::vector<ggml_backend_dev_t>         & devs,
-                                      uint32_t  & hp_ngl,
-                                      uint32_t  & hp_n_ctx_train,
-                                      uint32_t  & hp_n_expert,
-                           enum ggml_log_level    log_level);
@@ -11,6 +11,7 @@
 #include <filesystem>
 #include <fstream>
 #include <atomic>
+#include <regex> // migration only
 #include <string>
 #include <string_view>
 #include <stdexcept>
@@ -335,9 +336,15 @@ hf_files get_repo_files(const std::string & repo_id,
                if (item["lfs"].contains("oid") && item["lfs"]["oid"].is_string()) {
                    file.oid = item["lfs"]["oid"].get<std::string>();
                }
+                if (item["lfs"].contains("size") && item["lfs"]["size"].is_number()) {
+                    file.size = item["lfs"]["size"].get<size_t>();
+                }
            } else if (item.contains("oid") && item["oid"].is_string()) {
                file.oid = item["oid"].get<std::string>();
            }
+            if (file.size == 0 && item.contains("size") && item["size"].is_number()) {
+                file.size = item["size"].get<size_t>();
+            }

            if (!file.oid.empty() && !is_valid_oid(file.oid)) {
                LOG_WRN("%s: skip invalid oid: %s\n", __func__, file.oid.c_str());
@@ -495,4 +502,271 @@ std::string finalize_file(const hf_file & file) {
    return file.final_path;
 }

+// delete everything after this line, one day
+
+// copied from download.cpp without the tag part
+struct gguf_split_info {
+    std::string prefix; // tag included
+    int index;
+    int count;
+};
+
+static gguf_split_info get_gguf_split_info(const std::string & path) {
+    static const std::regex re_split("^(.+)-([0-9]{5})-of-([0-9]{5})$", std::regex::icase);
+    std::smatch m;
+
+    std::string prefix = path;
+    if (!string_remove_suffix(prefix, ".gguf")) {
+        return {};
+    }
+
+    int index = 1;
+    int count = 1;
+
+    if (std::regex_match(prefix, m, re_split)) {
+        index = std::stoi(m[2].str());
+        count = std::stoi(m[3].str());
+        prefix = m[1].str();
+    }
+
+    return {std::move(prefix), index, count};
+}
+
+static std::pair<std::string, std::string> parse_manifest_name(std::string & filename) {
+    static const std::regex re(R"(^manifest=([^=]+)=([^=]+)=.*\.json$)");
+    std::smatch match;
+    if (std::regex_match(filename, match, re)) {
+        return {match[1].str(), match[2].str()};
+    }
+    return {};
+}
+
+static std::string make_old_cache_filename(const std::string & owner,
+                                           const std::string & repo,
+                                           const std::string & filename) {
+    auto result = owner + "_" + repo + "_" + filename;
+    string_replace_all(result, "/", "_");
+    return result;
+}
+
+struct migrate_file {
+    std::string path;
+    std::string sha256;
+    size_t size;
+    fs::path old_path;
+    fs::path etag_path;
+    const hf_file * file;
+};
+
+using migrate_files = std::vector<migrate_file>;
+
+static bool collect_file(const fs::path    & old_cache,
+                         const std::string & owner,
+                         const std::string & repo,
+                         const std::string & path,
+                         const std::string & sha256,
+                         const hf_files    & files,
+                         migrate_files     & to_migrate) {
+
+    const hf_file * file = nullptr;
+
+    for (const auto & f : files) {
+        if (f.path == path) {
+            file = &f;
+            break;
+        }
+    }
+
+    std::string old_filename = make_old_cache_filename(owner, repo, path);
+    fs::path old_path = old_cache / old_filename;
+    fs::path etag_path = old_path.string() + ".etag";
+
+    if (!fs::exists(old_path)) {
+        if (file && fs::exists(file->final_path)) {
+            return true;
+        }
+        LOG_WRN("%s: %s not found in old cache or HF cache\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (!file) {
+        LOG_WRN("%s: %s not found in current repo\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (!sha256.empty() && !file->oid.empty() && sha256 != file->oid) {
+        LOG_WRN("%s: %s is not up to date (sha256 mismatch)\n", __func__, old_filename.c_str());
+        return false;
+    }
+
+    if (file->size > 0) {
+        size_t size = fs::file_size(old_path);
+        if (size != file->size) {
+            LOG_WRN("%s: %s has wrong size %zu (expected %zu)\n", __func__, old_filename.c_str(), size, file->size);
+            return false;
+        }
+    }
+
+    to_migrate.push_back({path, sha256, file->size, old_path, etag_path, file});
+    return true;
+}
+
+static bool collect_files(const fs::path    & old_cache,
+                          const std::string & owner,
+                          const std::string & repo,
+                          const nl::json    & node,
+                          const hf_files    & files,
+                          migrate_files     & to_migrate) {
+
+    if (!node.contains("rfilename") ||
+        !node.contains("lfs")       ||
+        !node["lfs"].contains("sha256")) {
+        return true;
+    }
+
+    std::string path = node["rfilename"];
+    std::string sha256 = node["lfs"]["sha256"];
+
+    auto split = get_gguf_split_info(path);
+
+    if (split.count <= 1) {
+        return collect_file(old_cache, owner, repo, path, sha256, files, to_migrate);
+    }
+
+    std::vector<std::pair<std::string, std::string>> splits;
+
+    for (const auto & f : files) {
+        auto split_f = get_gguf_split_info(f.path);
+        if (split_f.count == split.count && split_f.prefix == split.prefix) {
+            // sadly the manifest only provides the sha256 of the first file (index == 1)
+            // the rest will be verified using the size...
+            std::string f_sha256 = (split_f.index == 1) ? sha256 : "";
+            splits.emplace_back(f.path, f_sha256);
+        }
+    }
+
+    if ((int)splits.size() != split.count) {
+        LOG_WRN("%s: expected %d split files but found %d in repo\n", __func__, split.count, (int)splits.size());
+        return false;
+    }
+
+    for (const auto & [f_path, f_sha256] : splits) {
+        if (!collect_file(old_cache, owner, repo, f_path, f_sha256, files, to_migrate)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool migrate_file(const migrate_file & file) {
+    std::error_code ec;
+
+    fs::path new_path(file.file->local_path);
+    fs::create_directories(new_path.parent_path(), ec);
+
+    if (!fs::exists(new_path, ec)) {
+        fs::rename(file.old_path, new_path, ec);
+        if (ec) {
+            fs::copy_file(file.old_path, new_path, ec);
+            if (ec) {
+                LOG_ERR("%s: failed to move/copy %s: %s\n", __func__, file.old_path.string().c_str(), ec.message().c_str());
+                return false;
+            }
+        }
+        fs::remove(file.old_path, ec);
+    }
+    fs::remove(file.etag_path, ec);
+
+    std::string filename = finalize_file(*file.file);
+    LOG_INF("%s: migrated %s -> %s\n", __func__, file.old_path.filename().string().c_str(), filename.c_str());
+    return true;
+}
+
+void migrate_old_cache_to_hf_cache(const std::string & token, bool offline) {
+    fs::path old_cache = fs_get_cache_directory();
+    if (!fs::exists(old_cache)) {
+        return;
+    }
+
+    if (offline) {
+        LOG_WRN("%s: skipping migration in offline mode (will run when online)\n", __func__);
+        return; // -hf is not going to work
+    }
+
+    bool warned = false;
+
+    for (const auto & entry : fs::directory_iterator(old_cache)) {
+        if (!entry.is_regular_file()) {
+            continue;
+        }
+        auto filename = entry.path().filename().string();
+        auto [owner, repo] = parse_manifest_name(filename);
+
+        if (owner.empty() || repo.empty()) {
+            continue;
+        }
+
+        if (!warned) {
+            warned = true;
+            LOG_WRN("================================================================================\n"
+                    "WARNING: Migrating cache to HuggingFace cache directory\n"
+                    "  Old cache: %s\n"
+                    "  New cache: %s\n"
+                    "This one-time migration moves models previously downloaded with -hf\n"
+                    "from the legacy llama.cpp cache to the standard HuggingFace cache.\n"
+                    "Models downloaded with --model-url are not affected.\n"
+                    "================================================================================\n",
+                    old_cache.string().c_str(), get_cache_directory().string().c_str());
+        }
+
+        auto repo_id = owner + "/" + repo;
+        auto files = get_repo_files(repo_id, token);
+
+        if (files.empty()) {
+            LOG_WRN("%s: could not get repo files for %s, skipping\n", __func__, repo_id.c_str());
+            continue;
+        }
+
+        migrate_files to_migrate;
+        bool ok = true;
+
+        try {
+            std::ifstream manifest(entry.path());
+            auto json = nl::json::parse(manifest);
+            for (const char * key : {"ggufFile", "mmprojFile"}) {
+                if (json.contains(key)) {
+                    if (!collect_files(old_cache, owner, repo, json[key], files, to_migrate)) {
+                        ok = false;
+                        break;
+                    }
+                }
+            }
+        } catch (const std::exception & e) {
+            LOG_WRN("%s: failed to parse manifest %s: %s\n", __func__, filename.c_str(), e.what());
+            continue;
+        }
+
+        if (!ok) {
+            LOG_WRN("%s: migration skipped: one or more files failed validation\n", __func__);
+            continue;
+        }
+
+        for (const auto & file : to_migrate) {
+            if (!migrate_file(file)) {
+                ok = false;
+                break;
+            }
+        }
+
+        if (!ok) {
+            LOG_WRN("%s: migration failed: could not migrate all files\n", __func__);
+            continue;
+        }
+
+        LOG_INF("%s: migration complete, deleting manifest: %s\n", __func__, entry.path().string().c_str());
+        fs::remove(entry.path());
+    }
+}
+
 } // namespace hf_cache
@@ -14,6 +14,7 @@ struct hf_file {
    std::string final_path;
    std::string oid;
    std::string repo_id;
+    size_t size = 0; // only for the migration
 };

 using hf_files = std::vector<hf_file>;
@@ -29,4 +30,7 @@ hf_files get_cached_files(const std::string & repo_id = {});
 // Create snapshot path (link or move/copy) and return it
 std::string finalize_file(const hf_file & file);

+// TODO: Remove later
+void migrate_old_cache_to_hf_cache(const std::string & token, bool offline = false);
+
 } // namespace hf_cache
@@ -471,7 +471,7 @@ void common_ngram_map_draft(common_ngram_map & map,
        sum_occur += curr_occur;
    }

-    LOG_DBG("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
+    LOG_INF("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
            key_offset,
            max_occur, sum_occur, slot_max,
            curr_key.values[0].value_idx, curr_key.values[0].value_num,
@@ -482,7 +482,7 @@ void common_ngram_map_draft(common_ngram_map & map,
    // Print the tokens of the four values (if idx != 0), use LOG_INF
    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
        if (curr_key.values[v].value_idx != 0) {
-            LOG_DBG("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
+            LOG_INF("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
        }
    }

@@ -500,7 +500,7 @@ void common_ngram_map_draft(common_ngram_map & map,
        draft.push_back(inp[match_pos + n + i]);
    }

-    LOG_DBG("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
+    LOG_INF("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
            key_offset, slot_max,
            curr_key.key_num, draft.size());

@@ -1,7 +1,5 @@
 #include "ngram-mod.h"

-#include <algorithm>
-
 //
 // common_ngram_mod
 //
@@ -171,12 +171,22 @@ static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
    ctx->force_pos = 0;
 }

+// forward declaration for use in clone
 static struct llama_sampler * common_reasoning_budget_init_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);

-static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl);
+static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const common_reasoning_budget_ctx *) smpl->ctx;
+    return common_reasoning_budget_init_state(
+        ctx->vocab,
+        ctx->start_matcher.tokens,
+        ctx->end_matcher.tokens,
+        ctx->forced_tokens,
+        ctx->budget,
+        ctx->state);
+}

 static void common_reasoning_budget_free(struct llama_sampler * smpl) {
    delete (common_reasoning_budget_ctx *) smpl->ctx;
@@ -195,15 +205,6 @@ static struct llama_sampler_i common_reasoning_budget_i = {
    /* .backend_set_input = */ nullptr,
 };

-static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
-    const auto * ctx = (const common_reasoning_budget_ctx *) smpl->ctx;
-
-    return llama_sampler_init(
-        /* .iface = */ &common_reasoning_budget_i,
-        /* .ctx   = */ new common_reasoning_budget_ctx(*ctx)
-    );
-}
-
 static struct llama_sampler * common_reasoning_budget_init_state(
        const struct llama_vocab             * vocab,
        const std::vector<llama_token>       & start_tokens,
@@ -247,24 +248,3 @@ common_reasoning_budget_state common_reasoning_budget_get_state(const struct lla
    }
    return ((const common_reasoning_budget_ctx *)smpl->ctx)->state;
 }
-
-bool common_reasoning_budget_force(struct llama_sampler * smpl) {
-    if (!smpl) {
-        return false;
-    }
-
-    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
-
-    // only a sampler that is actively counting down the budget may be forced;
-    // any other state (idle, already forcing/waiting, or done) is left untouched
-    if (ctx->state != REASONING_BUDGET_COUNTING) {
-        return false;
-    }
-
-    ctx->state = REASONING_BUDGET_FORCING;
-    ctx->force_pos = 0;
-    ctx->end_matcher.reset();
-    LOG_INF("reasoning-budget: forced into forcing state (manual transition)\n");
-
-    return true;
-}
@@ -40,7 +40,3 @@ struct llama_sampler * common_reasoning_budget_init(
        common_reasoning_budget_state    initial_state = REASONING_BUDGET_IDLE);

 common_reasoning_budget_state common_reasoning_budget_get_state(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);
@@ -661,14 +661,6 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
    return llama_sampler_get_seed(gsmpl->chain);
 }

-bool common_sampler_reasoning_budget_force(struct common_sampler * gsmpl) {
-    if (!gsmpl) {
-        return false;
-    }
-
-    return common_reasoning_budget_force(gsmpl->rbudget);
-}
-
 // helpers

 llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
@@ -87,9 +87,6 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample

 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);

-// force the reasoning budget sampler (if any) to begin forcing its end sequence now.
-bool common_sampler_reasoning_budget_force(struct common_sampler * gsmpl);
-
 // helpers

 // access the internal list of current candidate tokens
@@ -3,7 +3,6 @@
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "../src/llama-ext.h" // staging API: llama_set_embeddings_pre_norm / llama_get_embeddings_pre_norm_ith (used by MTP)
 #include "log.h"
 #include "ngram-cache.h"
 #include "ngram-map.h"
@@ -24,7 +23,6 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
    {"none",          COMMON_SPECULATIVE_TYPE_NONE},
    {"draft-simple",  COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE},
    {"draft-eagle3",  COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3},
-    {"draft-mtp",     COMMON_SPECULATIVE_TYPE_DRAFT_MTP},
    {"ngram-simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
    {"ngram-map-k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
    {"ngram-map-k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
@@ -32,18 +30,6 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
    {"ngram-cache",   COMMON_SPECULATIVE_TYPE_NGRAM_CACHE}
 };

-static std::string common_speculative_get_devices_str(const std::vector<ggml_backend_dev_t> & devices) {
-    std::string result;
-    for (size_t i = 0; i < devices.size(); i++) {
-        if (devices[i] == nullptr) {
-            continue;
-        }
-        if (!result.empty()) result += ", ";
-        result += ggml_backend_dev_name(devices[i]);
-    }
-    return result.empty() ? "default" : result;
-}
-
 struct common_speculative_config {
    common_speculative_type type;
    common_params_speculative params;
@@ -156,13 +142,7 @@ struct common_speculative_impl {

    virtual void draft(common_speculative_draft_params_vec & dparams) = 0;

-    virtual void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) = 0;
-
-    // true if this implementation requires the target context to extract post-norm embeddings
-    virtual bool need_embd() const = 0;
-
-    // true if this implementation requires the target context to extract pre-norm embeddings
-    virtual bool need_embd_pre_norm() const { return false; }
+    virtual void accept(llama_seq_id seq_id, uint16_t n_accepted) = 0;
 };

 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -179,16 +159,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        auto * ctx_dft = this->params.ctx_dft;
        auto * ctx_tgt = this->params.ctx_tgt;

-        LOG_INF("%s: adding speculative implementation 'draft-simple'\n", __func__);
-        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
-        LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__,
-                this->params.n_gpu_layers,
-                ggml_type_name(this->params.cache_type_k),
-                ggml_type_name(this->params.cache_type_v),
-                ctx_tgt ? "yes" : "no",
-                ctx_dft ? "yes" : "no",
-                common_speculative_get_devices_str(this->params.devices).c_str());
-
        batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);

        // TODO: optimize or pass from outside?
@@ -365,24 +335,16 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
        }
    }

-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
    //common_params_speculative_eagle3 params;

-    common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq)
-    {
-        LOG_INF("%s: adding speculative implementation 'draft-eagle3'\n", __func__);
-        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, params.draft.n_max, params.draft.n_min, params.draft.p_min);
-    }
+    common_speculative_impl_draft_eagle3(const common_params_speculative & /*params*/, uint32_t n_seq)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq) {}

    void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
        // noop
@@ -397,384 +359,9 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
        // TODO: implement
    }

-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
-};
-
-struct common_speculative_impl_draft_mtp : public common_speculative_impl {
-    common_params_speculative_draft params; // reuses the draft-model params slot (ctx_tgt/ctx_dft)
-
-    llama_batch batch;
-
-    std::vector<common_sampler_ptr> smpls;
-
-    // backend sampler chain per seq, attached to ctx_dft
-    std::vector<llama_sampler *> backend_chains;
-
-    int32_t n_embd = 0;
-
-    // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
-    // The last h-row of one process() call needs the first token of the NEXT
-    // call to pair with, so it's stashed here until that next call fires.
-    std::vector<std::vector<float>> pending_h;   // [n_seq][n_embd]
-
-    std::vector<int32_t> i_batch_beg;
-    std::vector<int32_t> i_batch_end;
-
-    // Hidden rows from the most recent target verification batch, grouped by seq.
-    // Row 0 corresponds to the sampled token, row N to the Nth accepted draft token.
-    std::vector<std::vector<float>> verify_h;
-    std::vector<int32_t> verify_h_rows;
-
-    // Per-seq draft length from the last draft() call, used in accept() to
-    // roll back ctx_dft's recurrent state past the AR draft's redundant
-    // pre-advancement before process() mirrored the verify batch.
-    std::vector<uint16_t> last_n_drafted;
-
-    common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
-        , params(params.draft)
-    {
-        auto * ctx_tgt = this->params.ctx_tgt;
-        auto * ctx_dft = this->params.ctx_dft;
-        GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");
-
-        n_embd = llama_model_n_embd(llama_get_model(ctx_dft));
-
-        LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
-        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
-        LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__,
-                this->params.n_gpu_layers,
-                ggml_type_name(this->params.cache_type_k),
-                ggml_type_name(this->params.cache_type_v),
-                ctx_tgt ? "yes" : "no",
-                ctx_dft ? "yes" : "no",
-                common_speculative_get_devices_str(this->params.devices).c_str());
-
-        const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
-        batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd, /*n_seq_max=*/ 1);
-        // llama_batch_init allocates only one of token/embd; MTP needs both.
-        // TODO: fix, how to call without malloc
-        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_b);
-
-        smpls.resize(n_seq);
-        for (auto & s : smpls) {
-            common_params_sampling sparams;
-            sparams.no_perf  = false;
-            sparams.top_k    = 10;
-            sparams.samplers = { COMMON_SAMPLER_TYPE_TOP_K };
-            s.reset(common_sampler_init(llama_get_model(ctx_dft), sparams));
-        }
-
-        // offload draft sampling to the backend
-        backend_chains.assign(n_seq, nullptr);
-        if (this->params.backend_sampling) {
-            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-                llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
-                llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
-
-                if (!llama_set_sampler(ctx_dft, seq_id, chain)) {
-                    LOG_WRN("%s: backend offload failed for seq_id=%d; using CPU sampler\n", __func__, (int) seq_id);
-                    llama_sampler_free(chain);
-                    chain = nullptr;
-                }
-                backend_chains[seq_id] = chain;
-            }
-        }
-
-        llama_set_embeddings_pre_norm(ctx_tgt, true, /*masked*/ false);
-        llama_set_embeddings_pre_norm(ctx_dft, true, /*masked*/ true);
-
-        pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
-
-        i_batch_beg.assign(n_seq, -1);
-        i_batch_end.assign(n_seq, -1);
-
-        verify_h.assign(n_seq, {});
-        verify_h_rows.assign(n_seq, 0);
-
-        last_n_drafted.assign(n_seq, 0);
-    }
-
-    ~common_speculative_impl_draft_mtp() override {
-        auto * ctx_dft = this->params.ctx_dft;
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) backend_chains.size(); ++seq_id) {
-            if (backend_chains[seq_id] == nullptr) {
-                continue;
-            }
-            if (ctx_dft) {
-                llama_set_sampler(ctx_dft, seq_id, nullptr);
-            }
-            llama_sampler_free(backend_chains[seq_id]);
-        }
-        backend_chains.clear();
-
-        if (batch.token != nullptr) {
-            free(batch.token);
-            batch.token = nullptr;
-        }
-        llama_batch_free(batch);
-    }
-
-    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
-        const int32_t N = (int32_t) prompt.size();
-        if (N <= 0) {
-            return;
-        }
-        auto * ctx_dft = this->params.ctx_dft;
-        const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
-        if (pos_max < N - 1) {
-            LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - "
-                    "process() hook may not have run on every prefill ubatch "
-                    "(need_embd / logits=1 on every prompt position?). "
-                    "Drafts may degrade.\n",
-                    __func__, (int) pos_max, N - 1);
-        }
-    }
-
-    bool process(const llama_batch & batch_in) override {
-        if (batch_in.n_tokens <= 0) {
-            return true;
-        }
-
-        // TODO: how to make it work with vision tokens?
-        if (batch_in.token == nullptr || batch_in.embd != nullptr) {
-            return true;
-        }
-
-        const int32_t n_tokens = batch_in.n_tokens;
-
-        // remember the frist and last batch index for each sequence
-        std::fill(i_batch_beg.begin(), i_batch_beg.end(), -1);
-        std::fill(i_batch_end.begin(), i_batch_end.end(), -1);
-
-        for (int k = 0; k < n_tokens; ++k) {
-            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-                GGML_ASSERT(batch_in.n_seq_id[k] == 1);
-
-                if (batch_in.seq_id[k][0] == seq_id) {
-                    i_batch_end[seq_id] = k;
-                    if (i_batch_beg[seq_id] < 0) {
-                        i_batch_beg[seq_id] = k;
-                    }
-                }
-            }
-        }
-
-        auto * ctx_tgt = this->params.ctx_tgt;
-        auto * ctx_dft = this->params.ctx_dft;
-
-        const size_t row_bytes = (size_t) n_embd * sizeof(float);
-
-        common_batch_clear(batch);
-
-        for (int k = 0; k < n_tokens; ++k) {
-            common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
-        }
-
-        // shift the tgt embeddings to the right by one position
-        // assumes that the tokens in the batch are sequential for each sequence
-        // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
-        //                                                       ^--- this is a problem
-        // TODO:this is generally true, but would be nice to assert it
-        {
-            const float * h_tgt = llama_get_embeddings_pre_norm(ctx_tgt);
-            std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
-
-            //{
-            //    // string with seq_ids in the batch
-            //    std::stringstream ss;
-            //    for (int i = 0; i < n_tokens; ++i) {
-            //        ss << batch_in.seq_id[i][0] << ",";
-            //    }
-            //    LOG_WRN("%s: batch_in.seq_id = %s\n", __func__, ss.str().c_str());
-            //}
-        }
-
-        // fill the pending embeddings from a previous run
-        auto set_h = [&](int idx, const float * h_row) {
-            std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
-        };
-
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            if (i_batch_beg[seq_id] < 0) {
-                continue;
-            }
-
-            set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
-        }
-
-        const int32_t rc = llama_decode(ctx_dft, batch);
-        if (rc != 0) {
-            LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
-            return false;
-        }
-
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            if (i_batch_end[seq_id] < 0) {
-                continue;
-            }
-
-            const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1;
-            verify_h_rows[seq_id] = n_rows;
-            verify_h[seq_id].resize((size_t) n_rows * n_embd);
-
-            for (int32_t i = 0; i < n_rows; ++i) {
-                const float * h = llama_get_embeddings_pre_norm_ith(ctx_tgt, i_batch_beg[seq_id] + i);
-                std::memcpy(verify_h[seq_id].data() + (size_t) i * n_embd, h, row_bytes);
-            }
-
-            std::memcpy(pending_h[seq_id].data(),
-                    verify_h[seq_id].data() + (size_t) (n_rows - 1) * n_embd, row_bytes);
-        }
-
-        return true;
-    }
-
-    void draft(common_speculative_draft_params_vec & dparams) override {
-        auto & ctx_dft = params.ctx_dft;
-
-        common_batch_clear(batch);
-
-        // keep track of which sequences are still drafting
-        int n_drafting = 0;
-        std::vector<bool> drafting(n_seq);
-
-        const float * h_row = nullptr;
-        const size_t row_bytes = (size_t) n_embd * sizeof(float);
-
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            auto & dp = dparams[seq_id];
-
-            if (!dp.drafting) {
-                continue;
-            }
-
-            n_drafting++;
-            drafting[seq_id] = true;
-            common_sampler_reset(smpls[seq_id].get());
-
-            common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
-
-            h_row = pending_h[seq_id].data();
-            std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
-        }
-
-        int ret = llama_decode(ctx_dft, batch);
-        if (ret != 0) {
-            LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
-            return;
-        }
-
-        int i = 0;
-
-        while (n_drafting > 0) {
-            int i_batch = 0;
-
-            common_batch_clear(batch);
-
-            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-                if (!drafting[seq_id]) {
-                    continue;
-                }
-
-                auto * smpl = smpls[seq_id].get();
-
-                common_sampler_sample(smpl, ctx_dft, i_batch, true);
-                h_row = llama_get_embeddings_pre_norm_ith(ctx_dft, i_batch);
-                ++i_batch;
-
-                const auto * cur_p = common_sampler_get_candidates(smpl, true);
-
-                for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
-                    LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p,
-                            common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
-                }
-
-                // add drafted token for each sequence
-                const llama_token id = cur_p->data[0].id;
-
-                // only collect very high-confidence draft tokens
-                if (cur_p->data[0].p < params.p_min) {
-                    drafting[seq_id] = false;
-                    n_drafting--;
-
-                    continue;
-                }
-
-                common_sampler_accept(smpl, id, true);
-
-                auto & dp = dparams.at(seq_id);
-                auto & result = *dp.result;
-
-                result.push_back(id);
-
-                if (params.n_max <= (int) result.size()) {
-                    drafting[seq_id] = false;
-                    n_drafting--;
-                    continue;
-                }
-
-                common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
-                std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
-            }
-
-            if (batch.n_tokens == 0) {
-                break;
-            }
-
-            // evaluate the drafted tokens on the draft model
-            ret = llama_decode(ctx_dft, batch);
-            if (ret != 0) {
-                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
-                break;
-            }
-
-            ++i;
-        }
-
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            auto & dp = dparams[seq_id];
-            if (!dp.drafting) {
-                continue;
-            }
-
-            if (dp.result->size() < (size_t) params.n_min) {
-                dp.result->clear();
-            }
-
-            last_n_drafted[seq_id] = (uint16_t) dp.result->size();
-        }
-    }
-
-    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool /*is_other*/) override {
-        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
-            return;
-        }
-
-        const int32_t n_rows = verify_h_rows[seq_id];
-        if (n_rows <= 0) {
-            return;
-        }
-
-        const int32_t i_h = std::min<int32_t>(n_accepted, n_rows - 1);
-        const size_t row_bytes = (size_t) n_embd * sizeof(float);
-        std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
-    }
-
-    bool need_embd() const override {
-        return false;
-    }
-
-    bool need_embd_pre_norm() const override {
-        return true;
-    }
 };

 // state of self-speculation (simple implementation, not ngram-map)
@@ -789,12 +376,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
            common_ngram_simple_config config)
        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq)
        , params(params.ngram_simple)
-        , config(config)
-    {
-        LOG_INF("%s: adding speculative implementation 'ngram-simple'\n", __func__);
-        LOG_INF("%s: - size_n=%d, size_m=%d, min_hits=%d\n", __func__,
-                this->params.size_n, this->params.size_m, this->params.min_hits);
-    }
+        , config(config) {}

    void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
        // noop
@@ -818,31 +400,26 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
        }
    }

-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
+    common_params_speculative_ngram_map params;
+
    // n_seq configs
    std::vector<common_ngram_map> config;

    common_speculative_impl_ngram_map_k(
+            const common_params_speculative & params,
            const common_ngram_map & config,
            uint32_t n_seq)
        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, n_seq)
-    {
+        , params(params.ngram_map_k) {
        for (uint32_t i = 0; i < n_seq; i++) {
            this->config.push_back(config);
        }
-
-        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(this->type).c_str());
-        LOG_INF("%s: - size_key=%d, size_value=%d, key_only=%d, min_hits=%d\n", __func__,
-                config.size_key, config.size_value, config.key_only, config.min_hits);
    }

    void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
@@ -869,19 +446,11 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
        }
    }

-    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) override {
+    void accept(llama_seq_id seq_id, uint16_t n_accepted) override {
        GGML_ASSERT((seq_id < (llama_seq_id) config.size()));

-        if (is_other) {
-            return;
-        }
-
        common_ngram_map_accept(config[seq_id], n_accepted);
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_mod : public common_speculative_impl {
@@ -897,7 +466,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
        // the last position in the prompt that was added to the ngram container
        size_t i_last = 0;

-        // length of the last drafted n-gram (number of tokens returned by draft)
+        // length of the last drafted n‑gram (number of tokens returned by draft)
        size_t n_draft_last = 0;

        // consecutive accept rounds with low acceptance fraction (< 0.5)
@@ -915,11 +484,8 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
        , verbose(std::getenv("LLAMA_TRACE") != nullptr) {
        static_assert(sizeof(llama_token) == sizeof(common_ngram_mod::entry_t));

-        LOG_INF("%s: adding speculative implementation 'ngram-mod'\n", __func__);
-        LOG_INF("%s: - n_match=%d, n_max=%d, n_min=%d\n", __func__,
-                this->params.n_match, this->params.n_max, this->params.n_min);
-        LOG_INF("%s: - mod size=%zu (%.3f MB)\n", __func__,
-                mod.size(), (float)(mod.size_bytes())/1024/1024);
+        LOG_INF("%s: initialized ngram_mod with n_match=%d, size=%zu (%.3f MB)\n", __func__,
+                this->params.n_match, mod.size(), (float)(mod.size_bytes())/1024/1024);

        if (this->params.n_match < 16) {
            LOG_WRN("%s: ngram_mod n_match=%d is too small - poor quality is possible, "
@@ -1009,7 +575,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
        }
        result.resize(result.size() - n);

-        // store length of drafted n-gram for later acceptance analysis
+        // store length of drafted n‑gram for later acceptance analysis
        sinfo.n_draft_last = result.size();
    }

@@ -1031,21 +597,17 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
        }
    }

-    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) override {
-        if (is_other) {
-            return;
-        }
-
+    void accept(llama_seq_id seq_id, uint16_t n_accepted) override {
        auto & sinfo = sinfos[seq_id];

        // compute acceptance fraction if we have a recorded draft length
        if (sinfo.n_draft_last > 0) {
            const double f_acc = (double)n_accepted / (double)sinfo.n_draft_last;
-            if (f_acc < 0.25) {
+            if (f_acc < 0.5) {
                sinfo.n_low++;
-                if (sinfo.n_low >= 5) {
+                if (sinfo.n_low >= 3) {
                    if (verbose) {
-                        LOG_WRN("%s: low acceptance streak (%d) - resetting ngram_mod\n", __func__, sinfo.n_low);
+                        LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, sinfo.n_low);
                    }

                    mod.reset();
@@ -1057,10 +619,6 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
            }
        }
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative_impl_ngram_cache : public common_speculative_impl {
@@ -1095,12 +653,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
        , save_dynamic(save_dynamic)
        , save_static(save_static)
    {
-        LOG_INF("%s: adding speculative implementation 'ngram-cache'\n", __func__);
-        LOG_INF("%s: - n_draft=%d, cache_static=%s, cache_dynamic=%s\n", __func__,
-                n_draft,
-                path_static.empty() ? "none" : path_static.c_str(),
-                path_dynamic.empty() ? "none" : path_dynamic.c_str());
-
        sinfos.resize(n_seq);

        if (!path_static.empty()) {
@@ -1197,13 +749,9 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
        }
    }

-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
        // noop
    }
-
-    bool need_embd() const override {
-        return false;
-    }
 };

 struct common_speculative {
@@ -1272,7 +820,6 @@ std::string common_speculative_type_to_str(common_speculative_type type) {
        case COMMON_SPECULATIVE_TYPE_NONE:          return "none";
        case COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE:  return "draft-simple";
        case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:  return "draft-eagle3";
-        case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:     return "draft-mtp";
        case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram-simple";
        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram-map-k";
        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram-map-k4v";
@@ -1317,40 +864,6 @@ static uint32_t common_get_enabled_speculative_configs(const std::vector<common_
    return result;
 }

-int32_t common_speculative_n_max(const common_params_speculative * spec) {
-    int32_t n_max = 0;
-
-    for (const auto type : spec->types) {
-        switch (type) {
-            case COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE:
-            case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:
-            case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:
-                n_max = std::max(n_max, std::max(0, spec->draft.n_max));
-                break;
-            case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:
-                n_max = std::max(n_max, (int32_t) spec->ngram_simple.size_m);
-                break;
-            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
-                n_max = std::max(n_max, (int32_t) spec->ngram_map_k.size_m);
-                break;
-            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V:
-                n_max = std::max(n_max, (int32_t) spec->ngram_map_k4v.size_m);
-                break;
-            case COMMON_SPECULATIVE_TYPE_NGRAM_MOD:
-                n_max = std::max(n_max, std::max(0, spec->ngram_mod.n_max));
-                break;
-            case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:
-                n_max = std::max(n_max, (int32_t) 8);
-                break;
-            case COMMON_SPECULATIVE_TYPE_NONE:
-            case COMMON_SPECULATIVE_TYPE_COUNT:
-                break;
-        }
-    }
-
-    return n_max;
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -1359,9 +872,11 @@ common_speculative * common_speculative_init(common_params_speculative & params,
    {
        uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);

+        bool has_draft_model_path = !params.draft.mparams.path.empty();
+
        bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
+        // bool has_mtp = false; // TODO: add MTP here
        bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
-        bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;

        bool has_ngram_cache   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_CACHE));
        bool has_ngram_simple  = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE));
@@ -1370,7 +885,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
        bool has_ngram_mod     = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD));

        // when adding a new type - update here the logic above
-        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 9);
+        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 8);

        // this list here defines the priority of the speculators
        // the one with highest priority are listed first
@@ -1391,20 +906,29 @@ common_speculative * common_speculative_init(common_params_speculative & params,
        if (has_ngram_cache) {
            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
        }
+        if (has_draft_simple) {
+            if (!has_draft_model_path) {
+                LOG_WRN("%s: draft model is not specified - cannot use 'draft' type\n", __func__);
+                has_draft_simple = false;
+            }
+        } else if (has_draft_model_path) {
+            LOG_WRN("%s: draft model is specified but 'draft' speculative type is not explicitly enabled - enabling it\n", __func__);
+            has_draft_simple = true;
+        }
+
        if (has_draft_simple) {
            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, params));
        }
+        // TODO: add MTP here
        if (has_draft_eagle3) {
            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
        }
-        if (has_mtp) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
-        }
    }

    std::vector<std::unique_ptr<common_speculative_impl>> impls = {};

    for (const common_speculative_config & config : configs) {
+        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(config.type).c_str());
        switch (config.type) {
            case COMMON_SPECULATIVE_TYPE_NONE:
                break;
@@ -1416,10 +940,6 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                impls.push_back(std::make_unique<common_speculative_impl_draft_eagle3>(config.params, n_seq));
                break;
            }
-            case COMMON_SPECULATIVE_TYPE_DRAFT_MTP: {
-                impls.push_back(std::make_unique<common_speculative_impl_draft_mtp>(config.params, n_seq));
-                break;
-            }
            case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
                common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple);

@@ -1438,16 +958,11 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                impls.push_back(std::move(state));
                break;
            }
-            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: {
-                impls.push_back(
-                        std::make_unique<common_speculative_impl_ngram_map_k>(
-                            get_common_ngram_map(config.type, config.params.ngram_map_k), n_seq));
-                break;
-            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
                impls.push_back(
                        std::make_unique<common_speculative_impl_ngram_map_k>(
-                            get_common_ngram_map(config.type, config.params.ngram_map_k4v), n_seq));
+                            config.params, get_common_ngram_map(config.type, config.params.ngram_map_k), n_seq));
                break;
            }
            case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
@@ -1525,34 +1040,6 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
    return result;
 }

-bool common_speculative_need_embd(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
-bool common_speculative_need_embd_pre_norm(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd_pre_norm()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
 void common_speculative_draft(common_speculative * spec) {
    if (spec == nullptr) {
        return;
@@ -1635,6 +1122,10 @@ void common_speculative_draft(common_speculative * spec) {
 }

 void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
+    if (n_accepted == 0) {
+        return;
+    }
+
    common_speculative_impl * impl = spec->impl_last[seq_id];

    GGML_ASSERT(impl);
@@ -1646,16 +1137,9 @@ void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, u
            impl->n_acc_tokens += n_accepted;
        }

-        impl->accept(seq_id, n_accepted, false);
+        impl->accept(seq_id, n_accepted);
        impl->n_call_accept++;
    }
-
-    // accept with the rest of the implementations, using is_other == true
-    for (auto & impl_other : spec->impls) {
-        if (impl_other.get() != impl) {
-            impl_other->accept(seq_id, n_accepted, true);
-        }
-    }
 }

 void common_speculative_print_stats(const common_speculative * spec) {
@@ -1675,7 +1159,7 @@ void common_speculative_print_stats(const common_speculative * spec) {
            str_perf = "";
        }

-        LOG_INF("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s\n",
+        LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
                common_speculative_type_to_str(impl->type).c_str(),
                impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
                impl->n_gen_drafts,
@@ -20,9 +20,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);

-// return the max number of draft tokens based on the speculative parameters
-int32_t common_speculative_n_max(const common_params_speculative * spec);
-
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);

 void common_speculative_free(common_speculative * spec);
@@ -56,12 +53,6 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);

-// true if any implementation requires target post-norm embeddings to be extracted
-bool common_speculative_need_embd(common_speculative * spec);
-
-// true if any implementation requires target pre-norm embeddings to be extracted
-bool common_speculative_need_embd_pre_norm(common_speculative * spec);
-
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Ruben Ortlam	938872e93f	fix partial writes	2026-05-15 16:00:57 +02:00
Ruben Ortlam	ff6ad60994	wider loads	2026-05-15 15:22:57 +02:00
Ruben Ortlam	13a55c8e50	deduplicate repacking code	2026-05-15 13:25:49 +02:00
Ruben Ortlam	57fb74fba3	add q4_1, q8_0, iq4_nl repacking	2026-05-15 13:10:19 +02:00
Ruben Ortlam	6906f78189	replace malloc/free with thread_local memory	2026-05-15 12:11:01 +02:00
Ruben Ortlam	b64f294cbf	add missing repacking functions	2026-05-15 12:04:08 +02:00
Ruben Ortlam	b4e2621de8	add mxfp4 repacking	2026-05-15 11:58:13 +02:00
Ruben Ortlam	b1243aa933	fix double semicolon	2026-05-15 11:23:59 +02:00
Ruben Ortlam	5c1e95c901	add coopmat2 support	2026-05-15 11:23:58 +02:00
Ruben Ortlam	c285bb9838	vulkan: repack q4_0 into aligned arrays	2026-05-15 11:20:02 +02:00