Implement llama-pull tool

Complete llama-pull tool with documentation

Signed-off-by: Eric Curtin <eric.curtin@docker.com>

.clang-tidy

@@ -17,7 +17,6 @@ Checks: >
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,
-    -performance-enum-size,
     portability-*,
     -portability-simd-intrinsics,
     misc-*,

.devops/cann.Dockerfile

@@ -3,8 +3,7 @@
 # ==============================================================================
 
 # Define the CANN base image for easier version updates later
-ARG CHIP_TYPE=910b
-ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-${CHIP_TYPE}-openeuler24.03-py3.11
+ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
 
 # ==============================================================================
 # BUILD STAGE
@@ -12,8 +11,11 @@ ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-${CHIP_TYPE}-openeuler24.03-py3.
 # ==============================================================================
 FROM ${CANN_BASE_IMAGE} AS build
 
+# Define the Ascend chip model for compilation. Default is Ascend910B3
+ARG ASCEND_SOC_TYPE=Ascend910B3
+
 # -- Install build dependencies --
-RUN yum install -y gcc g++ cmake make git openssl-devel python3 python3-pip && \
+RUN yum install -y gcc g++ cmake make git libcurl-devel python3 python3-pip && \
     yum clean all && \
     rm -rf /var/cache/yum
 
@@ -34,22 +36,20 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 # For brevity, only core variables are listed here. You can paste the original ENV list here.
 
 # -- Build llama.cpp --
-# Use the passed CHIP_TYPE argument and add general build options
-ARG CHIP_TYPE
+# Use the passed ASCEND_SOC_TYPE argument and add general build options
 RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
     && \
     cmake -B build \
         -DGGML_CANN=ON \
         -DCMAKE_BUILD_TYPE=Release \
-        -DSOC_TYPE=ascend${CHIP_TYPE} \
-        -DUSE_ACL_GRAPH=ON \
+        -DSOC_TYPE=${ASCEND_SOC_TYPE} \
         . && \
     cmake --build build --config Release -j$(nproc)
 
 # -- Organize build artifacts for copying in later stages --
 # Create a lib directory to store all .so files
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 # Create a full directory to store all executables and Python scripts
 RUN mkdir -p /app/full && \
@@ -108,11 +108,11 @@ ENTRYPOINT ["/app/tools.sh"]
 # ENTRYPOINT ["/app/llama-server"]
 
 ### Target: light
-# Lightweight image containing only llama-cli and llama-completion
+# Lightweight image containing only llama-cli
 # ==============================================================================
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 ENTRYPOINT [ "/app/llama-cli" ]
 

.devops/cpu.Dockerfile

@@ -5,7 +5,7 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 ARG TARGETARCH
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -20,7 +20,7 @@ RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
     cmake --build build -j $(nproc)
 
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -68,7 +68,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.devops/cuda-new.Dockerfile

@@ -1,95 +0,0 @@
-ARG UBUNTU_VERSION=24.04
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=13.1.0
-# Target the CUDA build image
-ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-
-ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
-
-# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=default
-
-RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-    export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release -j$(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 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 ${BASE_CUDA_RUN_CONTAINER} AS base
-
-RUN apt-get update \
-    && apt-get install -y libgomp1 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" ]

.devops/cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 WORKDIR /app
 
@@ -25,7 +25,7 @@ RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -74,7 +74,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.devops/intel.Dockerfile

@@ -1,12 +1,12 @@
-ARG ONEAPI_VERSION=2025.2.2-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
 
 ## Build Image
 
-FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
-    apt-get install -y git libssl-dev
+    apt-get install -y git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -21,7 +21,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -31,7 +31,7 @@ RUN mkdir -p /app/full \
     && cp requirements.txt /app/full \
     && cp .devops/tools.sh /app/full/tools.sh
 
-FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS base
 
 RUN apt-get update \
     && apt-get install -y libgomp1 curl\
@@ -73,7 +73,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.devops/llama-cli-cann.Dockerfile

@@ -6,7 +6,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN yum install -y gcc g++ cmake make openssl-devel
+RUN yum install -y gcc g++ cmake make libcurl-devel
 ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
 ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
 ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
@@ -23,12 +23,11 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
     cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
-    cmake --build build --config Release --target llama-cli && \
-    cmake --build build --config Release --target llama-completion
+    cmake --build build --config Release --target llama-cli
 
 # TODO: use image with NNRT
 FROM ascendai/cann:$ASCEND_VERSION AS runtime
-COPY --from=build /app/build/bin/llama-cli /app/build/bin/llama-completion /
+COPY --from=build /app/build/bin/llama-cli /llama-cli
 
 ENV LC_ALL=C.utf8
 

.devops/llama-cpp-cuda.srpm.spec

@@ -37,7 +37,6 @@ make -j GGML_CUDA=1
 %install
 mkdir -p %{buildroot}%{_bindir}/
 cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
-cp -p llama-completion %{buildroot}%{_bindir}/llama-cuda-completion
 cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
 cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
 
@@ -69,7 +68,6 @@ rm -rf %{_builddir}/*
 
 %files
 %{_bindir}/llama-cuda-cli
-%{_bindir}/llama-cuda-completion
 %{_bindir}/llama-cuda-server
 %{_bindir}/llama-cuda-simple
 /usr/lib/systemd/system/llamacuda.service

.devops/llama-cpp.srpm.spec

@@ -39,7 +39,6 @@ make -j
 %install
 mkdir -p %{buildroot}%{_bindir}/
 cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
-cp -p llama-completion %{buildroot}%{_bindir}/llama-completion
 cp -p llama-server %{buildroot}%{_bindir}/llama-server
 cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
 
@@ -71,7 +70,6 @@ rm -rf %{_builddir}/*
 
 %files
 %{_bindir}/llama-cli
-%{_bindir}/llama-completion
 %{_bindir}/llama-server
 %{_bindir}/llama-simple
 /usr/lib/systemd/system/llama.service

.devops/musa.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc4.3.0
+ARG MUSA_VERSION=rc4.2.0
 # Target the MUSA build image
 ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
 
@@ -18,7 +18,7 @@ RUN apt-get update && \
     python3 \
     python3-pip \
     git \
-    libssl-dev \
+    libcurl4-openssl-dev \
     libgomp1
 
 WORKDIR /app
@@ -32,7 +32,7 @@ RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -81,7 +81,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.devops/nix/nixpkgs-instances.nix

@@ -4,7 +4,7 @@
   # the module `{ pkgs ... }: { /* config */ }` implicitly uses
   # `_module.args.pkgs` (defined in this case by flake-parts).
   perSystem =
-    { lib, system, ... }:
+    { system, ... }:
     {
       _module.args = {
         # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs
@@ -33,7 +33,7 @@
                 "CUDA EULA"
                 "cuDNN EULA"
               ]
-            ) (p.meta.licenses or (lib.toList p.meta.license));
+            ) (p.meta.licenses or [ p.meta.license ]);
         };
         # Ensure dependencies use ROCm consistently
         pkgsRocm = import inputs.nixpkgs {

.devops/nix/package-gguf-py.nix

@@ -3,7 +3,6 @@
   llamaVersion,
   numpy,
   tqdm,
-  requests,
   sentencepiece,
   pyyaml,
   poetry-core,
@@ -21,7 +20,6 @@ buildPythonPackage {
     tqdm
     sentencepiece
     pyyaml
-    requests
   ];
   src = lib.cleanSource ../../gguf-py;
   pythonImportsCheck = [

.devops/nix/package.nix

@@ -32,8 +32,8 @@
   useMpi ? false,
   useRocm ? config.rocmSupport,
   rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets,
+  enableCurl ? true,
   useVulkan ? false,
-  useRpc ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
 
   # It's necessary to consistently use backendStdenv when building with CUDA support,
@@ -128,6 +128,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   };
 
   postPatch = ''
+    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+      --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+      --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
   '';
 
   # With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015,
@@ -159,13 +163,15 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ++ optionals useMpi [ mpi ]
     ++ optionals useRocm rocmBuildInputs
     ++ optionals useBlas [ blas ]
-    ++ optionals useVulkan vulkanBuildInputs;
+    ++ optionals useVulkan vulkanBuildInputs
+    ++ optionals enableCurl [ curl ];
 
   cmakeFlags =
     [
       (cmakeBool "LLAMA_BUILD_SERVER" true)
       (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
       (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+      (cmakeBool "LLAMA_CURL" enableCurl)
       (cmakeBool "GGML_NATIVE" false)
       (cmakeBool "GGML_BLAS" useBlas)
       (cmakeBool "GGML_CUDA" useCuda)
@@ -173,7 +179,6 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       (cmakeBool "GGML_METAL" useMetalKit)
       (cmakeBool "GGML_VULKAN" useVulkan)
       (cmakeBool "GGML_STATIC" enableStatic)
-      (cmakeBool "GGML_RPC" useRpc)
     ]
     ++ optionals useCuda [
       (

.devops/nix/scope.nix

@@ -7,6 +7,13 @@
 
 let
   pythonPackages = python3.pkgs;
+  buildPythonPackage = pythonPackages.buildPythonPackage;
+  numpy = pythonPackages.numpy;
+  tqdm = pythonPackages.tqdm;
+  sentencepiece = pythonPackages.sentencepiece;
+  pyyaml = pythonPackages.pyyaml;
+  poetry-core = pythonPackages.poetry-core;
+  pytestCheckHook = pythonPackages.pytestCheckHook;
 in
 
 # We're using `makeScope` instead of just writing out an attrset
@@ -16,18 +23,17 @@ in
 lib.makeScope newScope (self: {
   inherit llamaVersion;
   gguf-py = self.callPackage ./package-gguf-py.nix {
-    inherit (pythonPackages)
+    inherit
+      buildPythonPackage
       numpy
       tqdm
       sentencepiece
+      poetry-core
       pyyaml
       pytestCheckHook
-      requests
-      buildPythonPackage
-      poetry-core
       ;
   };
-  python-scripts = self.callPackage ./python-scripts.nix { inherit (pythonPackages) buildPythonPackage poetry-core; };
+  python-scripts = self.callPackage ./python-scripts.nix { inherit buildPythonPackage poetry-core; };
   llama-cpp = self.callPackage ./package.nix { };
   docker = self.callPackage ./docker.nix { };
   docker-min = self.callPackage ./docker.nix { interactive = false; };

.devops/rocm.Dockerfile

@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04
 
 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=7.2
-ARG AMDGPU_VERSION=7.2
+ARG ROCM_VERSION=6.4
+ARG AMDGPU_VERSION=6.4
 
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@@ -11,12 +11,14 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
 # Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-7.2.0/reference/system-requirements.html
-# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/native_linux/native_linux_compatibility.html
-# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityryz/native_linux/native_linux_compatibility.html
+# gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
+# gfx906 is deprecated
+#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
 
-ARG ROCM_DOCKER_ARCH='gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201'
+ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
+#ARG ROCM_DOCKER_ARCH='gfx1151'
 
 # Set ROCm architectures
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
@@ -26,7 +28,7 @@ RUN apt-get update \
     build-essential \
     cmake \
     git \
-    libssl-dev \
+    libcurl4-openssl-dev \
     curl \
     libgomp1
 
@@ -34,17 +36,20 @@ WORKDIR /app
 
 COPY . .
 
+RUN git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
     cmake -S . -B build \
         -DGGML_HIP=ON \
         -DGGML_HIP_ROCWMMA_FATTN=ON \
+        -DCMAKE_HIP_FLAGS="-I$(pwd)/rocwmma/library/include/" \
         -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
         -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
         -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \
-    && find build -name "*.so*" -exec cp -P {} /app/lib \;
+    && find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -93,7 +98,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.devops/s390x.Dockerfile

@@ -1,126 +0,0 @@
-ARG GCC_VERSION=15.2.0
-ARG UBUNTU_VERSION=24.04
-
-### Build Llama.cpp stage
-FROM gcc:${GCC_VERSION} AS build
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
-    apt update -y && \
-    apt upgrade -y && \
-    apt install -y --no-install-recommends \
-        git cmake ccache ninja-build \
-        # WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
-        libopenblas-dev libssl-dev && \
-    rm -rf /var/lib/apt/lists/*
-
-WORKDIR /app
-COPY . .
-
-RUN --mount=type=cache,target=/root/.ccache \
-    --mount=type=cache,target=/app/build \
-    cmake -S . -B build -G Ninja \
-        -DCMAKE_BUILD_TYPE=Release \
-        -DCMAKE_C_COMPILER_LAUNCHER=ccache \
-        -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-        -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_NATIVE=OFF \
-        -DGGML_BACKEND_DL=ON \
-        -DGGML_CPU_ALL_VARIANTS=ON \
-        -DGGML_BLAS=ON \
-        -DGGML_BLAS_VENDOR=OpenBLAS && \
-    cmake --build build --config Release -j $(nproc) && \
-    cmake --install build --prefix /opt/llama.cpp
-
-COPY *.py             /opt/llama.cpp/bin
-COPY .devops/tools.sh /opt/llama.cpp/bin
-
-COPY gguf-py          /opt/llama.cpp/gguf-py
-COPY requirements.txt /opt/llama.cpp/gguf-py
-COPY requirements     /opt/llama.cpp/gguf-py/requirements
-
-
-### Collect all llama.cpp binaries, libraries and distro libraries
-FROM scratch AS collector
-
-# Copy llama.cpp binaries and libraries
-COPY --from=build /opt/llama.cpp/bin     /llama.cpp/bin
-COPY --from=build /opt/llama.cpp/lib     /llama.cpp/lib
-COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py
-
-
-### Base image
-FROM ubuntu:${UBUNTU_VERSION} AS base
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
-    apt update -y && \
-    apt install -y --no-install-recommends \
-        # WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
-        # See: https://github.com/ggml-org/llama.cpp/pull/15915#issuecomment-3317166506
-        curl libgomp1 libopenblas-dev && \
-    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 llama.cpp libraries
-COPY --from=collector /llama.cpp/lib /usr/lib/s390x-linux-gnu
-
-
-### Full
-FROM base AS full
-
-ENV PATH="/root/.cargo/bin:${PATH}"
-WORKDIR /app
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
-    apt update -y && \
-    apt install -y \
-        git cmake libjpeg-dev \
-        python3 python3-pip python3-dev && \
-    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
-
-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
-
-RUN pip install --no-cache-dir --break-system-packages \
-        -r /app/gguf-py/requirements.txt
-
-ENTRYPOINT [ "/app/tools.sh" ]
-
-
-### CLI Only
-FROM base AS light
-
-WORKDIR /llama.cpp/bin
-
-# Copy llama.cpp binaries and libraries
-COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
-
-ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
-
-
-### Server
-FROM base AS server
-
-ENV LLAMA_ARG_HOST=0.0.0.0
-
-WORKDIR /llama.cpp/bin
-
-# Copy llama.cpp binaries and libraries
-COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
-
-EXPOSE 8080
-
-ENTRYPOINT [ "/llama.cpp/bin/llama-server" ]

.devops/tools.sh

@@ -13,8 +13,6 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     exec ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     exec ./llama-cli "$@"
-elif [[ "$arg1" == '--run-legacy' || "$arg1" == '-l' ]]; then
-    exec ./llama-completion "$@"
 elif [[ "$arg1" == '--bench' || "$arg1" == '-b' ]]; then
     exec ./llama-bench "$@"
 elif [[ "$arg1" == '--perplexity' || "$arg1" == '-p' ]]; then
@@ -34,10 +32,8 @@ elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "
-    echo "  --run (-r): Run a model (chat) previously converted into ggml"
-    echo "              ex: -m /models/7B/ggml-model-q4_0.bin"
-    echo "  --run-legacy (-l): Run a model (legacy completion) previously converted into ggml"
-    echo "              ex: -m /models/7B/ggml-model-q4_0.bin -no-cnv -p \"Building a website can be done in 10 simple steps:\" -n 512"
+    echo "  --run (-r): Run a model previously converted into ggml"
+    echo "              ex: -m /models/7B/ggml-model-q4_0.bin -p \"Building a website can be done in 10 simple steps:\" -n 512"
     echo "  --bench (-b): Benchmark the performance of the inference for various parameters."
     echo "              ex: -m model.gguf"
     echo "  --perplexity (-p): Measure the perplexity of a model over a given text."

.devops/vulkan.Dockerfile

@@ -1,24 +1,42 @@
-ARG UBUNTU_VERSION=26.04
+ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
+# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
+
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget xz-utils
 
-# Install SSL and Vulkan SDK dependencies
-RUN apt install -y libssl-dev curl \
-    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc
+# Install Vulkan SDK
+ARG VULKAN_VERSION=1.4.321.1
+RUN ARCH=$(uname -m) && \
+    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
+    mkdir -p /opt/vulkan && \
+    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
+    mv /tmp/${ARCH}/* /opt/vulkan/ && \
+    rm -rf /tmp/*
+
+# Install cURL and Vulkan SDK dependencies
+RUN apt install -y libcurl4-openssl-dev curl \
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
+
+# Set environment variables
+ENV VULKAN_SDK=/opt/vulkan
+ENV PATH=$VULKAN_SDK/bin:$PATH
+ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
+ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
+ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
 
 # Build it
 WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \
-    find build -name "*.so*" -exec cp -P {} /app/lib \;
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
@@ -32,8 +50,7 @@ RUN mkdir -p /app/full \
 FROM ubuntu:$UBUNTU_VERSION AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
-    libglvnd0 libgl1 libglx0 libegl1 libgles2 \
+    && apt-get install -y libgomp1 curl libvulkan-dev \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -51,10 +68,8 @@ WORKDIR /app
 
 RUN apt-get update \
     && apt-get install -y \
-    build-essential \
     git \
     python3 \
-    python3-dev \
     python3-pip \
     python3-wheel \
     && pip install --break-system-packages --upgrade setuptools \
@@ -70,7 +85,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama-cli /app
 
 WORKDIR /app
 

.editorconfig

@@ -60,11 +60,3 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
-
-[benches/**]
-indent_style = unset
-indent_size = unset
-end_of_line = unset
-charset = unset
-trim_trailing_whitespace = unset
-insert_final_newline = unset

.gemini/settings.json

@@ -1 +0,0 @@
-{ "contextFileName": "AGENTS.md" }

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -8,8 +8,7 @@ body:
       value: >
         Thanks for taking the time to fill out this bug report!
         This issue template is intended for bug reports where the compilation of llama.cpp fails.
-        Before opening an issue, please confirm that the compilation still fails
-        after recreating the CMake build directory and with `-DGGML_CCACHE=OFF`.
+        Before opening an issue, please confirm that the compilation still fails with `-DGGML_CCACHE=OFF`.
         If the compilation succeeds with ccache disabled you should be able to permanently fix the issue
         by clearing `~/.cache/ccache` (on Linux).
   - type: textarea
@@ -41,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -11,7 +11,7 @@ body:
         (i.e. the generated text) are incorrect or llama.cpp crashes during model evaluation.
         If you encountered the issue while using an external UI (e.g. ollama),
         please reproduce your issue using one of the examples/binaries in this repository.
-        The `llama-completion` binary can be used for simple and reproducible model inference.
+        The `llama-cli` binary can be used for simple and reproducible model inference.
   - type: textarea
     id: version
     attributes:
@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true
@@ -74,12 +74,9 @@ body:
         Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
         that information would be very much appreciated by us.
-
-        If possible, please try to reproduce the issue using `llama-completion` with `-fit off`.
-        If you can only reproduce the issue with `-fit on`, please provide logs both with and without `--verbose`.
       placeholder: >
-        e.g. when I run llama-completion with `-fa on` I get garbled outputs for very long prompts.
-        With short prompts or `-fa off` it works correctly.
+        e.g. when I run llama-cli with -ngl 99 I get garbled outputs.
+        When I use -ngl 0 it works correctly.
         Here are the exact commands that I used: ...
     validations:
       required: true
@@ -98,18 +95,7 @@ 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), 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>
-        <!-- Copy-pasted short logs go into the "console" area here -->
-
-        ```console
-
-        ```
-        </details>
-
-        <!-- Long logs that you upload as files go here, outside the "console" area -->
+          This will be automatically formatted into code, so no need for backticks.
+      render: shell
     validations:
       required: true

.github/ISSUE_TEMPLATE/019-bug-misc.yml

@@ -85,19 +85,7 @@ body:
       label: Relevant log output
       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.
-          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
-      value: |
-        <details>
-        <summary>Logs</summary>
-        <!-- Copy-pasted short logs go into the "console" area here -->
-
-        ```console
-
-        ```
-        </details>
-
-        <!-- Long logs that you upload as files go here, outside the "console" area -->
+          This will be automatically formatted into code, so no need for backticks.
+      render: shell
     validations:
       required: false

.github/actions/install-exe/action.yml

@@ -1,36 +0,0 @@
-name: "Install exe"
-description: "Download and install exe"
-inputs:
-  url:
-    description: "URL of the exe installer"
-    required: true
-  args:
-    description: "Installer arguments"
-    required: true
-  timeout:
-    description: "Timeout (in ms)"
-    required: false
-    default: "600000"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install EXE
-      shell: pwsh
-      run: |
-        $ErrorActionPreference = "Stop"
-        write-host "Downloading Installer EXE"
-        Invoke-WebRequest -Uri "${{ inputs.url }}" -OutFile "${env:RUNNER_TEMP}\temp-install.exe"
-        write-host "Installing"
-        $proc = Start-Process "${env:RUNNER_TEMP}\temp-install.exe" -ArgumentList '${{ inputs.args }}' -NoNewWindow -PassThru
-        $completed = $proc.WaitForExit(${{ inputs.timeout }})
-        if (-not $completed) {
-            Write-Error "Installer timed out. Killing the process"
-            $proc.Kill()
-            exit 1
-        }
-        if ($proc.ExitCode -ne 0) {
-            Write-Error "Installer failed with exit code $($proc.ExitCode)"
-            exit 1
-        }
-        write-host "Completed installation"

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

@@ -1,20 +0,0 @@
-name: "Linux - Setup SpacemiT Toolchain"
-description: "Setup SpacemiT Toolchain for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "SpacemiT toolchain version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup SpacemiT Toolchain
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v${{ inputs.version }}.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1

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

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

.github/actions/unarchive-tar/action.yml

@@ -1,27 +0,0 @@
-name: "Unarchive tar"
-description: "Download and unarchive tar into directory"
-inputs:
-  url:
-    description: "URL of the tar archive"
-    required: true
-  path:
-    description: "Directory to unarchive into"
-    required: true
-  type:
-    description: "Compression type (tar option)"
-    required: false
-    default: "J"
-  strip:
-    description: "Strip components"
-    required: false
-    default: "0"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Unarchive into directory
-      shell: bash
-      run: |
-        mkdir -p ${{ inputs.path }}
-        cd ${{ inputs.path }}
-        curl --no-progress-meter ${{ inputs.url }} | tar -${{ inputs.type }}x --strip-components=${{ inputs.strip }}

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

@@ -65,34 +65,3 @@ runs:
           echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
           echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-    - name: Install Cuda Toolkit 13.1
-      if: ${{ inputs.cuda_version == '13.1' }}
-      shell: pwsh
-      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.2.0.9-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.1.68-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.1.80-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.1.68-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-13.1.78-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_crt-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_cudart-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvcc-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvrtc-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\libcublas-windows-x86_64-13.2.0.9-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\libnvvm-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_nvtx-windows-x86_64-13.1.68-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_profiler_api-windows-x86_64-13.1.80-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\visual_studio_integration-windows-x86_64-13.1.68-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\cuda_cccl-windows-x86_64-13.1.78-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" /E /I /H /Y
-          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

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

@@ -0,0 +1,30 @@
+name: 'Windows - Setup CURL'
+description: 'Composite action, to be reused in other workflow'
+inputs:
+  curl_version:
+    description: 'CURL version'
+    required: false
+    default: '8.6.0_6'
+  architecture:
+    description: 'Architecture of the libcurl to download'
+    required: false
+    default: 'win64'
+outputs:
+  curl_path:
+    description: "Path to the downloaded libcurl"
+    value: ${{ steps.get_libcurl.outputs.curl_path }}
+
+runs:
+  using: "composite"
+  steps:
+    - name: libCURL
+      id: get_libcurl
+      shell: powershell
+      env:
+        CURL_VERSION: ${{ inputs.curl_version }}
+        ARCHITECTURE: ${{ inputs.architecture }}
+      run: |
+        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-${env:ARCHITECTURE}-mingw.zip"
+        mkdir $env:RUNNER_TEMP/libcurl
+        tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
+        echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

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

@@ -1,15 +0,0 @@
-name: "Windows - Setup ROCm"
-description: "Setup ROCm for Windows"
-inputs:
-  version:
-    description: "ROCm version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup ROCm
-      uses: ./.github/actions/install-exe
-      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
-        args: -install

.github/copilot-instructions.md

@@ -0,0 +1,262 @@
+# Copilot Instructions for llama.cpp
+
+## Repository Overview
+
+llama.cpp is a large-scale C/C++ project for efficient LLM (Large Language Model) inference with minimal setup and dependencies. The project enables running language models on diverse hardware with state-of-the-art performance.
+
+**Key Facts:**
+- **Primary language**: C/C++ with Python utility scripts
+- **Size**: ~200k+ lines of code across 1000+ files
+- **Architecture**: Modular design with main library (`libllama`) and 40+ executable tools/examples
+- **Core dependency**: ggml tensor library (vendored in `ggml/` directory)
+- **Backends supported**: CPU (AVX/NEON optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
+- **License**: MIT
+
+## Build Instructions
+
+### Prerequisites
+- CMake 3.14+ (primary build system)
+- C++17 compatible compiler (GCC 13.3+, Clang, MSVC)
+- Optional: ccache for faster compilation
+
+### Basic Build (CPU-only)
+**ALWAYS run these commands in sequence:**
+```bash
+cmake -B build
+cmake --build build --config Release -j $(nproc)
+```
+
+**Build time**: ~10 minutes on 4-core system with ccache enabled, ~25 minutes without ccache.
+
+**Important Notes:**
+- The Makefile is deprecated - always use CMake
+- ccache is automatically detected and used if available
+- Built binaries are placed in `build/bin/`
+- Parallel builds (`-j`) significantly reduce build time
+
+### Backend-Specific Builds
+For CUDA support:
+```bash
+cmake -B build -DGGML_CUDA=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+For Metal (macOS):
+```bash
+cmake -B build -DGGML_METAL=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+**Important Note**: While all backends can be built as long as the correct requirements for that backend are installed, you will not be able to run them without the correct hardware. The only backend that can be run for testing and validation is the CPU backend.
+
+### Debug Builds
+Single-config generators:
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Debug
+cmake --build build
+```
+
+Multi-config generators:
+```bash
+cmake -B build -G "Xcode"
+cmake --build build --config Debug
+```
+
+### Common Build Issues
+- **Issue**: Network tests fail in isolated environments
+  **Solution**: Expected behavior - core functionality tests will still pass
+
+## Testing
+
+### Running Tests
+```bash
+ctest --test-dir build --output-on-failure -j $(nproc)
+```
+
+**Test suite**: 38 tests covering tokenizers, grammar parsing, sampling, backends, and integration
+**Expected failures**: 2-3 tests may fail if network access is unavailable (they download models)
+**Test time**: ~30 seconds for passing tests
+
+### Server Unit Tests
+Run server-specific unit tests after building the server:
+```bash
+# Build the server first
+cmake --build build --target llama-server
+
+# Navigate to server tests and run
+cd tools/server/tests
+source ../../../.venv/bin/activate
+./tests.sh
+```
+**Server test dependencies**: The `.venv` environment includes the required dependencies for server unit tests (pytest, aiohttp, etc.). Tests can be run individually or with various options as documented in `tools/server/tests/README.md`.
+
+### Test Categories
+- Tokenizer tests: Various model tokenizers (BERT, GPT-2, LLaMA, etc.)
+- Grammar tests: GBNF parsing and validation
+- Backend tests: Core ggml operations across different backends
+- Integration tests: End-to-end workflows
+
+### Manual Testing Commands
+```bash
+# Test basic inference
+./build/bin/llama-cli --version
+
+# Test model loading (requires model file)
+./build/bin/llama-cli -m path/to/model.gguf -p "Hello" -n 10
+```
+
+## Code Quality and Linting
+
+### C++ Code Formatting
+**ALWAYS format C++ code before committing:**
+```bash
+git clang-format
+```
+
+Configuration is in `.clang-format` with these key rules:
+- 4-space indentation
+- 120 column limit
+- Braces on same line for functions
+- Pointer alignment: `void * ptr` (middle)
+- Reference alignment: `int & ref` (middle)
+
+### Python Code
+**ALWAYS activate the Python environment in `.venv` and use tools from that environment:**
+```bash
+# Activate virtual environment
+source .venv/bin/activate
+```
+
+Configuration files:
+- `.flake8`: flake8 settings (max-line-length=125, excludes examples/tools)
+- `pyrightconfig.json`: pyright type checking configuration
+
+### Pre-commit Hooks
+Run before committing:
+```bash
+pre-commit run --all-files
+```
+
+## Continuous Integration
+
+### GitHub Actions Workflows
+Key workflows that run on every PR:
+- `.github/workflows/build.yml`: Multi-platform builds
+- `.github/workflows/server.yml`: Server functionality tests
+- `.github/workflows/python-lint.yml`: Python code quality
+- `.github/workflows/python-type-check.yml`: Python type checking
+
+### Local CI Validation
+**Run full CI locally before submitting PRs:**
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```
+
+**CI Runtime**: 30-60 minutes depending on backend configuration
+
+### Triggering CI
+Add `ggml-ci` to commit message to trigger heavy CI workloads on the custom CI infrastructure.
+
+## Project Layout and Architecture
+
+### Core Directories
+- **`src/`**: Main llama library implementation (`llama.cpp`, `llama-*.cpp`)
+- **`include/`**: Public API headers, primarily `include/llama.h`
+- **`ggml/`**: Core tensor library (submodule with custom GGML framework)
+- **`examples/`**: 30+ example applications and tools
+- **`tools/`**: Additional development and utility tools (server benchmarks, tests)
+- **`tests/`**: Comprehensive test suite with CTest integration
+- **`docs/`**: Detailed documentation (build guides, API docs, etc.)
+- **`scripts/`**: Utility scripts for CI, data processing, and automation
+- **`common/`**: Shared utility code used across examples
+
+### Key Files
+- **`CMakeLists.txt`**: Primary build configuration
+- **`include/llama.h`**: Main C API header (~2000 lines)
+- **`src/llama.cpp`**: Core library implementation (~8000 lines)
+- **`CONTRIBUTING.md`**: Coding guidelines and PR requirements
+- **`.clang-format`**: C++ formatting rules
+- **`.pre-commit-config.yaml`**: Git hook configuration
+
+### Built Executables (in `build/bin/`)
+Primary tools:
+- **`llama-cli`**: Main inference tool
+- **`llama-server`**: OpenAI-compatible HTTP server
+- **`llama-quantize`**: Model quantization utility
+- **`llama-perplexity`**: Model evaluation tool
+- **`llama-bench`**: Performance benchmarking
+- **`llama-convert-llama2c-to-ggml`**: Model conversion utilities
+
+### Configuration Files
+- **CMake**: `CMakeLists.txt`, `cmake/` directory
+- **Linting**: `.clang-format`, `.clang-tidy`, `.flake8`
+- **CI**: `.github/workflows/`, `ci/run.sh`
+- **Git**: `.gitignore` (includes build artifacts, models, cache)
+
+### Dependencies
+- **System**: OpenMP, libcurl (for model downloading)
+- **Optional**: CUDA SDK, Metal framework, Vulkan SDK, Intel oneAPI
+- **Bundled**: httplib, json (header-only libraries in vendored form)
+
+## Common Validation Steps
+
+### After Making Changes
+1. **Format code**: `git clang-format`
+2. **Build**: `cmake --build build --config Release`
+3. **Test**: `ctest --test-dir build --output-on-failure`
+4. **Server tests** (if modifying server): `cd tools/server/tests && source ../../../.venv/bin/activate && ./tests.sh`
+5. **Manual validation**: Test relevant tools in `build/bin/`
+
+### Performance Validation
+```bash
+# Benchmark inference performance
+./build/bin/llama-bench -m model.gguf
+
+# Evaluate model perplexity
+./build/bin/llama-perplexity -m model.gguf -f dataset.txt
+```
+
+### Backend Validation
+```bash
+# Test backend operations
+./build/bin/test-backend-ops
+```
+
+## Environment Setup
+
+### Required Tools
+- CMake 3.14+ (install via system package manager)
+- Modern C++ compiler with C++17 support
+- Git (for submodule management)
+- Python 3.9+ with virtual environment (`.venv` is provided)
+
+### Optional but Recommended
+- ccache: `apt install ccache` or `brew install ccache`
+- clang-format 15+: Usually included with LLVM/Clang installation
+- pre-commit: `pip install pre-commit`
+
+### Backend-Specific Requirements
+- **CUDA**: NVIDIA CUDA Toolkit 11.2+
+- **Metal**: Xcode command line tools (macOS only)
+- **Vulkan**: Vulkan SDK
+- **SYCL**: Intel oneAPI toolkit
+
+## Important Guidelines
+
+### Code Changes
+- **Minimal dependencies**: Avoid adding new external dependencies
+- **Cross-platform compatibility**: Test on Linux, macOS, Windows when possible
+- **Performance focus**: This is a performance-critical inference library
+- **API stability**: Changes to `include/llama.h` require careful consideration
+
+### Git Workflow
+- Always create feature branches from `master`
+- **Never** commit build artifacts (`build/`, `.ccache/`, `*.o`, `*.gguf`)
+- Use descriptive commit messages following project conventions
+
+### Trust These Instructions
+Only search for additional information if these instructions are incomplete or found to be incorrect. This document contains validated build and test procedures that work reliably across different environments.
+

.github/labeler.yml

@@ -76,10 +76,6 @@ ggml:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/**
-model:
-    - changed-files:
-        - any-glob-to-any-file:
-            - src/models/**
 nix:
     - changed-files:
         - any-glob-to-any-file:
@@ -89,10 +85,7 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
-jinja parser:
-    - changed-files:
-        - any-glob-to-any-file:
-            - common/jinja/**
+
 Ascend NPU:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-cache.yml

@@ -1,89 +0,0 @@
-name: Build Actions Cache
-
-on:
-  workflow_dispatch: # allows manual triggering
-  schedule:
-    - cron: '0 * * * *'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  ubuntu-24-vulkan-cache:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
-
-  ubuntu-24-spacemit-cache:
-    runs-on: ubuntu-24.04
-
-    env:
-      # Make sure this is in sync with build-linux-cross.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
-
-      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-spacemit
-        with:
-          path: ./spacemit_toolchain
-          version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
-
-  windows-2022-rocm-cache:
-    runs-on: windows-2022
-
-    env:
-      # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-rocm
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: rocm-${{ 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 }}

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

@@ -7,7 +7,7 @@ jobs:
   linux:
     runs-on: ubuntu-24.04
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -20,7 +20,7 @@ jobs:
         run: |
           PREFIX="$(pwd)"/inst
           cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_OPENSSL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_CURL=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

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

@@ -4,55 +4,55 @@ on:
   workflow_call:
 
 jobs:
-  # ubuntu-24-riscv64-cpu-cross:
-  #   runs-on: ubuntu-24.04
+  ubuntu-24-riscv64-cpu-cross:
+    runs-on: ubuntu-24.04
 
-  #   steps:
-  #     - uses: actions/checkout@v6
-  #     - name: Setup Riscv
-  #       run: |
-  #         sudo dpkg --add-architecture riscv64
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo dpkg --add-architecture riscv64
 
-  #         # Add arch-specific repositories for non-amd64 architectures
-  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-  #         EOF
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
 
-  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 gcc-14-riscv64-linux-gnu \
-  #                 g++-14-riscv64-linux-gnu
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu
 
-  #     - name: Build
-  #       run: |
-  #         cmake -B build -DLLAMA_OPENSSL=OFF \
-  #                        -DCMAKE_BUILD_TYPE=Release \
-  #                        -DGGML_OPENMP=OFF \
-  #                        -DLLAMA_BUILD_EXAMPLES=ON \
-  #                        -DLLAMA_BUILD_TOOLS=ON \
-  #                        -DLLAMA_BUILD_TESTS=OFF \
-  #                        -DCMAKE_SYSTEM_NAME=Linux \
-  #                        -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-  #                        -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-  #                        -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-  #         cmake --build build --config Release -j $(nproc)
+          cmake --build build --config Release -j $(nproc)
 
   # ubuntu-24-riscv64-vulkan-cross:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v6
+  #     - uses: actions/checkout@v4
   #     - name: Setup Riscv
   #       run: |
   #         sudo dpkg --add-architecture riscv64
@@ -76,7 +76,7 @@ jobs:
 
   #     - name: Build
   #       run: |
-  #         cmake -B build -DLLAMA_OPENSSL=OFF \
+  #         cmake -B build -DLLAMA_CURL=OFF \
   #                        -DCMAKE_BUILD_TYPE=Release \
   #                        -DGGML_VULKAN=ON \
   #                        -DGGML_OPENMP=OFF \
@@ -99,7 +99,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v6
+  #     - uses: actions/checkout@v4
   #     - name: Setup Arm64
   #       run: |
   #         sudo dpkg --add-architecture arm64
@@ -122,7 +122,7 @@ jobs:
 
   #     - name: Build
   #       run: |
-  #         cmake -B build -DLLAMA_OPENSSL=OFF \
+  #         cmake -B build -DLLAMA_CURL=OFF \
   #                        -DCMAKE_BUILD_TYPE=Release \
   #                        -DGGML_VULKAN=ON \
   #                        -DGGML_OPENMP=OFF \
@@ -141,12 +141,103 @@ jobs:
 
   #         cmake --build build --config Release -j $(nproc)
 
+  ubuntu-24-ppc64el-cpu-cross:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  # ubuntu-24-ppc64el-vulkan-cross:
+  #   runs-on: ubuntu-24.04
+
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup PowerPC64le
+  #       run: |
+  #         sudo dpkg --add-architecture ppc64el
+
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
+
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 glslc \
+  #                 gcc-14-powerpc64le-linux-gnu \
+  #                 g++-14-powerpc64le-linux-gnu \
+  #                 libvulkan-dev:ppc64el
+
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_VULKAN=ON \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+  #                        -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+  #         cmake --build build --config Release -j $(nproc)
+
   debian-13-loongarch64-cpu-cross:
     runs-on: ubuntu-24.04
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v4
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -178,7 +269,7 @@ jobs:
 
       - name: Build
         run: |
-          cmake -B build -DLLAMA_OPENSSL=OFF \
+          cmake -B build -DLLAMA_CURL=OFF \
                          -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -201,7 +292,7 @@ jobs:
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v4
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -235,7 +326,7 @@ jobs:
 
       - name: Build
         run: |
-          cmake -B build -DLLAMA_OPENSSL=OFF \
+          cmake -B build -DLLAMA_CURL=OFF \
                          -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
@@ -253,46 +344,3 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
-
-  ubuntu-24-riscv64-cpu-spacemit-ime-cross:
-    runs-on: ubuntu-24.04
-
-    env:
-      # Make sure this is in sync with build-cache.yml
-      SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: Use SpacemiT Toolchain Cache
-        uses: actions/cache@v5
-        id: cache-toolchain
-        with:
-          path: ./spacemit_toolchain
-          key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
-
-      - name: Setup SpacemiT Toolchain
-        if: steps.cache-toolchain.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-spacemit
-        with:
-          path: ./spacemit_toolchain
-          version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
-
-      - name: Build
-        run: |
-          export RISCV_ROOT_PATH=${PWD}/spacemit_toolchain
-          cmake -B build -DLLAMA_OPENSSL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DGGML_CPU_RISCV64_SPACEMIT=ON \
-                         -DGGML_RVV=ON \
-                         -DGGML_RV_ZFH=ON \
-                         -DGGML_RV_ZICBOP=ON \
-                         -DGGML_RV_ZIHINTPAUSE=ON \
-                         -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
-                         -DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake
-
-          cmake --build build --config Release -j $(nproc)

.github/workflows/build-riscv-native.yml

@@ -0,0 +1,60 @@
+name: Build on RISCV Linux Machine by Cloud-V
+on:
+  pull_request:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  debian-13-riscv64-native: # Bianbu 2.2
+    runs-on: self-hosted
+
+    steps:
+      - name: Install prerequisites
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y libatomic1
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu \
+                  ccache \
+                  cmake
+
+      - name: Setup ccache
+        run: |
+          mkdir -p $HOME/.ccache
+          ccache -M 5G -d $HOME/.ccache
+          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
+          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
+          echo "$GITHUB_WORKSPACE"
+          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
+          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
+          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
+          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
+
+      - name: Build
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_SYSTEM_NAME=Linux \
+            -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+            -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+            -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+            -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/check-vendor.yml

@@ -1,52 +0,0 @@
-name: Check vendor
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      'vendor/**',
-      'scripts/sync_vendor.py'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      'vendor/**',
-      'scripts/sync_vendor.py'
-    ]
-
-jobs:
-  check-vendor:
-    runs-on: ubuntu-slim
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          python-version: '3.x'
-
-      - name: Run vendor sync
-        run: |
-          set -euo pipefail
-          python3 scripts/sync_vendor.py
-
-      - name: Check for changes
-        run: |
-          set -euo pipefail
-          # detect modified or untracked files
-          changed=$(git status --porcelain --untracked-files=all || true)
-          if [ -n "$changed" ]; then
-            echo "Vendor sync modified files:"
-            echo "$changed" | awk '{ print $2 }' | sed '/^$/d'
-            echo "Failing because vendor files mismatch. Please update scripts/sync_vendor.py"
-            exit 1
-          else
-            echo "Vendor files are up-to-date."
-          fi

.github/workflows/close-issue.yml

@@ -10,12 +10,12 @@ permissions:
 
 jobs:
   close-issues:
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
     permissions:
       issues: write
       pull-requests: write
     steps:
-      - uses: actions/stale@v10
+      - uses: actions/stale@v5
         with:
           exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
           days-before-issue-stale: 30

.github/workflows/copilot-setup-steps.yml

@@ -26,7 +26,7 @@ jobs:
     # If you do not check out your code, Copilot will do this for you.
     steps:
       - name: Checkout code
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -38,14 +38,14 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libcurl4-openssl-dev
           # Install git-clang-format script for formatting only changed code
           wget -O /tmp/git-clang-format https://raw.githubusercontent.com/llvm/llvm-project/release/18.x/clang/tools/clang-format/git-clang-format
           sudo cp /tmp/git-clang-format /usr/local/bin/git-clang-format
           sudo chmod +x /usr/local/bin/git-clang-format
 
       - name: Set up Python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: '3.11'
 

.github/workflows/docker.yml

@@ -28,7 +28,7 @@ jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
 
-    runs-on: ${{ matrix.config.runs_on }}
+    runs-on: ubuntu-22.04
     env:
       COMMIT_SHA: ${{ github.sha }}
     strategy:
@@ -39,22 +39,20 @@ jobs:
           # Note: the arm64 images are failing, which prevents the amd64 images from being built
           # https://github.com/ggml-org/llama.cpp/issues/11888
           #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda cuda12", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "12.4.0", ubuntu_version: "22.04" }
-          - { tag: "cuda13", dockerfile: ".devops/cuda-new.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "13.1.0", ubuntu_version: "24.04" }
-          - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04-s390x" }
-          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
+          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
+          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
+          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
     steps:
       - name: Check out the repo
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
-        if: ${{ matrix.config.tag != 's390x' }}
         uses: docker/setup-qemu-action@v3
         with:
           image: tonistiigi/binfmt:qemu-v7.0.0-28
@@ -63,48 +61,46 @@ jobs:
         uses: docker/setup-buildx-action@v3
 
       - name: Log in to Docker Hub
-        uses: docker/login-action@v3
+        uses: docker/login-action@v2
         with:
           registry: ghcr.io
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      - name: Determine source tag name
-        id: srctag
-        uses: ./.github/actions/get-tag-name
-        env:
-          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-
-      - name: Determine image tag name
+      - name: Determine tag name
         id: tag
         shell: bash
         run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
           REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
           REPO_NAME="${{ github.event.repository.name }}"
-          PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
 
+          # determine tag name postfix (build number, commit hash)
+          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
+            TAG_POSTFIX="-b${BUILD_NUMBER}"
+          else
+            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
+            TAG_POSTFIX="-${SAFE_NAME}-${SHORT_HASH}"
+          fi
           # list all tags possible
-          tags="${{ matrix.config.tag }}"
-          for tag in $tags; do
-              if [[ "$tag" == "cpu" ]]; then
-                  TYPE=""
-              else
-                  TYPE="-$tag"
-              fi
-              CACHETAGS="${PREFIX}buildcache${TYPE}"
-              FULLTAGS="${FULLTAGS:+$FULLTAGS,}${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
-              LIGHTTAGS="${LIGHTTAGS:+$LIGHTTAGS,}${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
-              SERVERTAGS="${SERVERTAGS:+$SERVERTAGS,}${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
-          done
-          echo "cache_output_tags=$CACHETAGS" >> $GITHUB_OUTPUT
+          if [[ "${{ matrix.config.tag }}" == "cpu" ]]; then
+              TYPE=""
+          else
+              TYPE="-${{ matrix.config.tag }}"
+          fi
+          PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
+          FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}${TAG_POSTFIX}"
+          LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}${TAG_POSTFIX}"
+          SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}${TAG_POSTFIX}"
           echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
           echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
           echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
-          echo "cache_output_tags=$CACHETAGS"  # print out for debugging
           echo "full_output_tags=$FULLTAGS"  # print out for debugging
           echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
           echo "server_output_tags=$SERVERTAGS"  # print out for debugging
         env:
+          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
       - name: Free Disk Space (Ubuntu)
@@ -136,18 +132,12 @@ jobs:
           file: ${{ matrix.config.dockerfile }}
           target: full
           provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
 
       - name: Build and push Light Docker image (tagged + versioned)
         if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
@@ -161,18 +151,12 @@ jobs:
           file: ${{ matrix.config.dockerfile }}
           target: light
           provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
 
       - name: Build and push Server Docker image (tagged + versioned)
         if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
@@ -186,41 +170,9 @@ jobs:
           file: ${{ matrix.config.dockerfile }}
           target: server
           provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
           # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
           # return to this if the experimental github cache is having issues
           #cache-to: type=local,dest=/tmp/.buildx-cache
           #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
-  create_tag:
-    name: Create and push git tag
-    runs-on: ubuntu-22.04
-    permissions:
-      contents: write
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Determine source tag name
-        id: srctag
-        uses: ./.github/actions/get-tag-name
-        env:
-          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-
-      - name: Create and push git tag
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          git tag ${{ steps.srctag.outputs.name }} || exit 0
-          git push origin ${{ steps.srctag.outputs.name }} || exit 0

.github/workflows/editorconfig.yml

@@ -20,9 +20,9 @@ concurrency:
 
 jobs:
   editorconfig:
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v4
       - uses: editorconfig-checker/action-editorconfig-checker@v2
         with:
           version: v3.0.3

.github/workflows/gguf-publish.yml

@@ -24,9 +24,9 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v4
     - name: Set up Python
-      uses: actions/setup-python@v6
+      uses: actions/setup-python@v5
       with:
         python-version: '3.9.x'
     - name: Install dependencies

.github/workflows/labeler.yml

@@ -7,11 +7,11 @@ jobs:
     permissions:
       contents: read
       pull-requests: write
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v4
       with:
         repository: "ggml-org/llama.cpp"
-    - uses: actions/labeler@v6
+    - uses: actions/labeler@v5
       with:
         configuration-path: '.github/labeler.yml'

.github/workflows/pre-tokenizer-hashes.yml

@@ -12,14 +12,14 @@ on:
 
 jobs:
     pre-tokenizer-hashes:
-        runs-on: ubuntu-slim
+        runs-on: ubuntu-latest
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v6
+          uses: actions/checkout@v4
 
         - name: Set up Python
-          uses: actions/setup-python@v6
+          uses: actions/setup-python@v5
           with:
               python-version: '3.11'
 

.github/workflows/python-check-requirements.yml

@@ -20,13 +20,13 @@ concurrency:
 
 jobs:
   python-check-requirements:
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
     name: check-requirements
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script

.github/workflows/python-lint.yml

@@ -15,13 +15,13 @@ concurrency:
 
 jobs:
   flake8-lint:
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
     name: Lint
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: flake8 Lint

.github/workflows/python-type-check.yml

@@ -24,12 +24,14 @@ jobs:
     name: pyright type-check
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt
+      - name: Install Python dependencies
+        # TODO: use a venv
+        run: pip install -r requirements/requirements-all.txt
       - name: Type-check with Pyright
         uses: jakebailey/pyright-action@v2
         with:

.github/workflows/release.yml

@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -37,6 +37,13 @@ jobs:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
 
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+          brew install curl
+
       - name: Build
         id: cmake_build
         run: |
@@ -45,7 +52,6 @@ jobs:
             -DCMAKE_INSTALL_RPATH='@loader_path' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DGGML_RPC=ON \
@@ -60,21 +66,21 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz
-          name: llama-bin-macos-arm64.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
 
   macOS-x64:
-    runs-on: macos-15-intel
+    runs-on: macos-13
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -84,6 +90,13 @@ jobs:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
 
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+          brew install curl
+
       - name: Build
         id: cmake_build
         run: |
@@ -94,7 +107,6 @@ jobs:
             -DCMAKE_INSTALL_RPATH='@loader_path' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
@@ -108,13 +120,13 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz
-          name: llama-bin-macos-x64.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
 
   ubuntu-22-cpu:
     strategy:
@@ -122,8 +134,6 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 's390x'
-            os: ubuntu-24.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm
@@ -133,21 +143,21 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-cpu-cmake-${{ matrix.build }}
+          key: ubuntu-cpu-cmake
           evict-old-files: 1d
 
       - name: Dependencies
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -170,13 +180,13 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
-          name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
+          name: llama-bin-ubuntu-${{ matrix.build }}.zip
 
   ubuntu-22-vulkan:
     runs-on: ubuntu-22.04
@@ -184,7 +194,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -200,7 +210,7 @@ jobs:
           wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
           sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -223,13 +233,13 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
-          name: llama-bin-ubuntu-vulkan-x64.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
+          name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
     runs-on: windows-2025
@@ -242,7 +252,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -257,28 +267,39 @@ jobs:
         run: |
           choco install ninja
 
+      - name: libCURL
+        id: get_libcurl
+        uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
+
       - name: Build
         shell: cmd
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
             -DGGML_OPENMP=ON ^
+            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release
 
       - name: Pack artifacts
         id: pack_artifacts
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
+          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
           Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
-          7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
+          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
@@ -305,7 +326,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -351,16 +372,16 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -S . -B build ${{ matrix.defines }} -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_BACKEND_DL=ON -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -S . -B build ${{ matrix.defines }} -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_BACKEND_DL=ON -DLLAMA_CURL=OFF
           cmake --build build --config Release --target ${{ matrix.target }}
 
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a -snl llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
+          7z a llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
@@ -370,12 +391,12 @@ jobs:
 
     strategy:
       matrix:
-        cuda: ['12.4', '13.1']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -397,7 +418,6 @@ jobs:
       - name: Build
         id: cmake_build
         shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
         run: |
           call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
@@ -405,18 +425,17 @@ jobs:
             -DGGML_NATIVE=OFF ^
             -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_CURL=OFF
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
+          7z a llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -427,11 +446,10 @@ jobs:
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          robocopy "${{env.CUDA_PATH}}\bin\x64" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -444,14 +462,14 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -474,7 +492,7 @@ jobs:
             -DCMAKE_BUILD_TYPE=Release ^
             -DGGML_BACKEND_DL=ON -DBUILD_SHARED_LIBS=ON ^
             -DGGML_CPU=OFF -DGGML_SYCL=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON
+            -DLLAMA_CURL=OFF
           cmake --build build --target ggml-sycl -j
 
       - name: Build the release package
@@ -487,7 +505,6 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
@@ -496,149 +513,44 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
 
-          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/tcm.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/libhwloc-15.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
-
           echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
+          7z a llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
-  ubuntu-22-rocm:
-    runs-on: ubuntu-22.04
-
-    strategy:
-      matrix:
-        include:
-          - ROCM_VERSION: "7.2"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201"
-            build: 'x64'
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt install -y build-essential git cmake wget
-
-      - name: Setup Legacy ROCm
-        if: matrix.ROCM_VERSION == '7.2'
-        id: legacy_env
-        run: |
-          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
-          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
-            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-
-          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
-          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
-          EOF
-
-          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
-          Package: *
-          Pin: release o=repo.radeon.com
-          Pin-Priority: 600
-          EOF
-
-          sudo apt update
-          sudo apt-get install -y libssl-dev rocm-hip-sdk
-
-      - name: Setup TheRock
-        if: matrix.ROCM_VERSION != '7.2'
-        id: therock_env
-        run: |
-          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
-          mkdir install
-          tar -xf *.tar.gz -C install
-          export ROCM_PATH=$(pwd)/install
-          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
-          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
-          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
-
-      - name: Build with native CMake HIP support
-        id: cmake_build
-        run: |
-          cmake -B build -S . \
-            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_BACKEND_DL=ON \
-            -DGGML_NATIVE=OFF \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DGGML_CPU_ALL_VARIANTS=ON \
-            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
-            -DGGML_HIP=ON \
-            -DHIP_PLATFORM=amd \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
-            ${{ env.CMAKE_ARGS }}
-          cmake --build build --config Release -j $(nproc)
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
-          name: llama-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
-
   windows-hip:
     runs-on: windows-2022
 
     env:
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
 
     strategy:
       matrix:
         include:
           - name: "radeon"
-            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
 
-      - name: Grab rocWMMA package
-        id: grab_rocwmma
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
-          7z x rocwmma.deb
-          7z x data.tar
+          git clone https://github.com/rocm/rocwmma --branch develop --depth 1
 
       - name: Cache ROCm Installation
         id: cache-rocm
-        uses: actions/cache@v5
+        uses: actions/cache@v4
         with:
           path: C:\Program Files\AMD\ROCm
           key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
@@ -655,7 +567,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
           $completed = $proc.WaitForExit(600000)
@@ -689,20 +601,20 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `
             -DGGML_CPU=OFF `
-            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
-            -DLLAMA_BUILD_BORINGSSL=ON
+            -DLLAMA_CURL=OFF
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
@@ -710,10 +622,10 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
+          7z a llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
@@ -723,7 +635,7 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -738,7 +650,7 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_CURL=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -763,106 +675,13 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          # Zip file is required for Swift Package Manager, which does not support tar.gz for binary targets.
-          # For more details, see https://developer.apple.com/documentation/xcode/distributing-binary-frameworks-as-swift-packages
-          zip -r -y llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
-          name: llama-${{ steps.tag.outputs.name }}-xcframework.zip
-
-
-  openEuler-cann:
-    strategy:
-      matrix:
-        include:
-          # 910b with aclgraph (both architectures)
-          - arch: x86
-            chip_type: '910b'
-            build: 'Release'
-            use_acl_graph: 'on'
-          - arch: aarch64
-            chip_type: '910b'
-            build: 'Release'
-            use_acl_graph: 'on'
-          # 310p without aclgraph (both architectures)
-          - arch: x86
-            chip_type: '310p'
-            build: 'Release'
-            use_acl_graph: 'off'
-          - arch: aarch64
-            chip_type: '310p'
-            build: 'Release'
-            use_acl_graph: 'off'
-    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' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-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}"':'"${HOST_GID}"' /workspace/build
-            '
-
-      - name: Determine tag name
-        id: tag
-        uses: ./.github/actions/get-tag-name
-
-      - name: Pack artifacts
-        run: |
-          cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
-          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
+          name: llama-${{ steps.tag.outputs.name }}-xcframework
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -880,18 +699,16 @@ jobs:
       - windows-cuda
       - windows-sycl
       - windows-hip
-      - ubuntu-22-rocm
       - ubuntu-22-cpu
       - ubuntu-22-vulkan
       - macOS-arm64
       - macOS-x64
       - ios-xcode-build
-      - openEuler-cann
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -901,7 +718,7 @@ jobs:
 
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v4
         with:
           path: ./artifact
           merge-multiple: true
@@ -941,7 +758,6 @@ jobs:
 
           echo "Moving other artifacts..."
           mv -v artifact/*.zip release
-          mv -v artifact/*.tar.gz release
 
       - name: Create release
         id: create_release
@@ -950,42 +766,10 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
-          body: |
-            <details open>
-
-            ${{ github.event.head_commit.message }}
-
-            </details>
-
-            **macOS/iOS:**
-            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
-            - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz)
-            - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-xcframework.zip)
-
-            **Linux:**
-            - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)
-            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
-            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
-
-            **Windows:**
-            - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
-            - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)
-            - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
-            - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.1-x64.zip) - [CUDA 13.1 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.1-x64.zip)
-            - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
-            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
-
-            **openEuler:**
-            - [openEuler x86 (310p)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-310p-openEuler-x86.tar.gz)
-            - [openEuler x86 (910b, ACL Graph)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-910b-openEuler-x86-aclgraph.tar.gz)
-            - [openEuler aarch64 (310p)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-310p-openEuler-aarch64.tar.gz)
-            - [openEuler aarch64 (910b, ACL Graph)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-910b-openEuler-aarch64-aclgraph.tar.gz)
 
       - name: Upload release
         id: upload_release
-        uses: actions/github-script@v8
+        uses: actions/github-script@v3
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}
           script: |
@@ -993,9 +777,9 @@ jobs:
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
             for (let file of await fs.readdirSync('./release')) {
-              if (path.extname(file) === '.zip' || file.endsWith('.tar.gz')) {
+              if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
-                await github.rest.repos.uploadReleaseAsset({
+                await github.repos.uploadReleaseAsset({
                   owner: context.repo.owner,
                   repo: context.repo.repo,
                   release_id: release_id,

.github/workflows/server-metal.yml

@@ -1,73 +0,0 @@
-name: Server-Metal
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      sha:
-        description: 'Commit SHA1 to build'
-        required: false
-        type: string
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
-
-env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  server-metal:
-    runs-on: [self-hosted, macOS, ARM64]
-
-    name: server-metal (${{ 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"
-          - build_type: Release
-            extra_args: "GGML_METAL_DEVICES=2"
-            wf_name:    "GPUx2"
-          - build_type: Release
-            extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
-            wf_name:    "GPUx2, 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"

.github/workflows/server-webui.yml

@@ -1,99 +0,0 @@
-# Server WebUI build and tests
-name: Server WebUI
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      sha:
-        description: 'Commit SHA1 to build'
-        required: false
-        type: string
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
-
-env:
-  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-check:
-    name: WebUI Checks
-    runs-on: ubuntu-latest
-    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: Setup Node.js
-        id: node
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22"
-          cache: "npm"
-          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/server/webui
-
-      - name: Run type checking
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run check
-        working-directory: tools/server/webui
-
-      - name: Run linting
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run lint
-        working-directory: tools/server/webui
-
-      - name: Build application
-        if: ${{ always() && steps.setup.conclusion == 'success' }}
-        run: npm run build
-        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/server/webui
-
-      - name: Build Storybook
-        if: ${{ always() && steps.playwright.conclusion == 'success' }}
-        run: npm run build-storybook
-        working-directory: tools/server/webui
-
-      - name: Run Client tests
-        if: ${{ always() && steps.playwright.conclusion == 'success' }}
-        run: npm run test:client
-        working-directory: tools/server/webui
-
-      - name: Run Unit tests
-        if: ${{ always() && steps.playwright.conclusion == 'success' }}
-        run: npm run test:unit
-        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/server/webui
-
-      - name: Run E2E tests
-        if: ${{ always() && steps.playwright.conclusion == 'success' }}
-        run: npm run test:e2e
-        working-directory: tools/server/webui

.github/workflows/server.yml

@@ -36,16 +36,12 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
             sanitizer: ""
-            extra_args: ""
-          - build_type: Release
-            sanitizer: ""
-            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-      fail-fast: false
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     steps:
       - name: Dependencies
@@ -60,51 +56,283 @@ jobs:
             curl \
             wget \
             language-pack-en \
-            libssl-dev
+            libcurl4-openssl-dev
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         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
+      - name: Python setup
+        id: setup_python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Tests dependencies
+        id: test_dependencies
         run: |
-          cmake -B build \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_SCHED_NO_REALLOC=ON \
-            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
-            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+          pip install -r tools/server/tests/requirements.txt
+
+  webui-setup:
+    name: WebUI Setup
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          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@v4
+        with:
+          node-version: "22"
+          cache: "npm"
+          cache-dependency-path: "tools/server/webui/package-lock.json"
+
+      - name: Cache node_modules
+        uses: actions/cache@v4
+        id: cache-node-modules
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Install dependencies
+        if: steps.cache-node-modules.outputs.cache-hit != 'true'
+        run: npm ci
+        working-directory: tools/server/webui
+
+  webui-check:
+    needs: webui-setup
+    name: WebUI Check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          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@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Run type checking
+        run: npm run check
+        working-directory: tools/server/webui
+
+      - name: Run linting
+        run: npm run lint
+        working-directory: tools/server/webui
+
+  webui-build:
+    needs: webui-check
+    name: WebUI Build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          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@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Build application
+        run: npm run build
+        working-directory: tools/server/webui
+
+  webui-tests:
+    needs: webui-build
+    name: Run WebUI tests
+    permissions:
+      contents: read
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Install Playwright browsers
+        run: npx playwright install --with-deps
+        working-directory: tools/server/webui
+
+      - name: Build Storybook
+        run: npm run build-storybook
+        working-directory: tools/server/webui
+
+      - name: Run Client tests
+        run: npm run test:client
+        working-directory: tools/server/webui
+
+      - name: Run Server tests
+        run: npm run test:server
+        working-directory: tools/server/webui
+
+      - name: Run UI tests
+        run: npm run test:ui
+        working-directory: tools/server/webui
+
+      - name: Run E2E tests
+        run: npm run test:e2e
+        working-directory: tools/server/webui
+
+  server-build:
+    needs: [webui-tests]
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        build_type: [RelWithDebInfo]
+        include:
+          - build_type: Release
+            sanitizer: ""
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+
+    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 \
+            libcurl4-openssl-dev
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          pip-install: -r tools/server/tests/requirements.txt
+
+      - name: Tests dependencies
+        id: test_dependencies
+        run: |
+          pip install -r tools/server/tests/requirements.txt
+
+      - name: Setup Node.js for WebUI
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+          cache: "npm"
+          cache-dependency-path: "tools/server/webui/package-lock.json"
+
+      - name: Install WebUI dependencies
+        run: npm ci
+        working-directory: tools/server/webui
+
+      - name: Build WebUI
+        run: npm run build
+        working-directory: tools/server/webui
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+              -DGGML_OPENMP=OFF ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Build (sanitizers)
+        id: cmake_build_sanitizers
+        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Build (sanitizers)
+        id: cmake_build
+        if: ${{ matrix.sanitizer == '' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
+          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) }}
+        if: ${{ matrix.sanitizer == '' }}
+        env:
+          GITHUB_ACTIONS: "true"
         run: |
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
+          ./tests.sh
+
+      - name: Tests (sanitizers)
+        id: server_integration_tests_sanitizers
+        if: ${{ matrix.sanitizer != '' }}
+        run: |
+          cd tools/server/tests
+          LLAMA_SANITIZE=1 ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh
+
 
   server-windows:
     runs-on: windows-2022
@@ -112,23 +340,40 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
+      - name: libCURL
+        id: get_libcurl
+        uses: ./.github/actions/windows-setup-curl
+
       - name: Build
         id: cmake_build
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
+          cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          pip-install: -r tools/server/tests/requirements.txt
+
+      - name: Tests dependencies
+        id: test_dependencies
+        run: |
+          pip install -r tools/server/tests/requirements.txt
+
+      - name: Copy Libcurl
+        id: prepare_libcurl
+        env:
+          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
+        run: |
+          cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
 
       - name: Tests
         id: server_integration_tests

.github/workflows/update-ops-docs.yml

@@ -3,25 +3,23 @@ name: Update Operations Documentation
 on:
     push:
         paths:
-            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
     pull_request:
         paths:
-            - 'docs/ops.md'
             - 'docs/ops/**'
             - 'scripts/create_ops_docs.py'
 
 jobs:
     update-ops-docs:
-        runs-on: ubuntu-slim
+        runs-on: ubuntu-latest
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v6
+          uses: actions/checkout@v4
 
         - name: Set up Python
-          uses: actions/setup-python@v6
+          uses: actions/setup-python@v5
           with:
               python-version: '3.x'
 

.github/workflows/winget.yml

@@ -9,7 +9,6 @@ jobs:
   update:
     name: Update Winget Package
     runs-on: ubuntu-latest
-    if: github.repository_owner == 'ggml-org'
 
     steps:
       - name: Install cargo binstall
@@ -17,28 +16,27 @@ jobs:
 
       - name: Install komac
         run: |
-          cargo binstall komac@2.15.0 -y
+          cargo binstall komac@2.11.2 -y
 
       - name: Find latest release
         id: find_latest_release
-        uses: actions/github-script@v8
+        uses: actions/github-script@v6
         with:
           script: |
             const { data: releases } = await github.rest.repos.listReleases({
               owner: context.repo.owner,
               repo: context.repo.repo,
             });
-            const { tag_name: version, assets: assets } = releases.find(({assets}) => assets.find(asset => asset.name.includes('win-vulkan')));
-            const { browser_download_url: asset_url } = assets.find(asset => asset.name.includes('win-vulkan'));
-            console.log("Latest release:", version);
-            core.setOutput('VERSION', version);
-            core.setOutput('ASSETURL', asset_url);
+            console.log("Latest release:", releases[0].tag_name);
+            return releases[0].tag_name;
 
       - name: Update manifest
+        env:
+          VERSION: ${{ steps.find_latest_release.outputs.result }}
         run: |
           echo "Updating manifest..."
-          komac update --version ${{ steps.find_latest_release.outputs.VERSION }} \
-            --urls "${{ steps.find_latest_release.outputs.ASSETURL }}" \
+          komac update --version ${{ env.VERSION }} \
+            --urls "https://github.com/ggml-org/llama.cpp/releases/download/${{ env.VERSION }}/llama-${{ env.VERSION }}-bin-win-vulkan-x64.zip" \
             --token ${{ secrets.WINGET_GITHUB_TOKEN }} \
             --submit \
             ggml.llamacpp

.gitignore

@@ -20,41 +20,52 @@
 *.so
 *.swp
 *.tmp
-*.DS_Store
 
 # IDE / OS
 
-/.cache/
-/.ccls-cache/
-/.direnv/
-/.envrc
-/.idea/
-/.swiftpm
-/.vs/
-/.vscode/
-/nppBackup
+.cache/
+.ccls-cache/
+.direnv/
+.DS_Store
+.envrc
+.idea/
+.swiftpm
+.vs/
+.vscode/
+nppBackup
 
 
 # Coverage
 
-/gcovr-report/
-/lcov-report/
+gcovr-report/
+lcov-report/
 
 # Build Artifacts
 
-/tags
-/.build/
-/build*
-/release
-/debug
+tags
+.build/
+build*
+release
+debug
+!build-info.cmake
+!build-info.cpp.in
+!build-info.sh
+!build.zig
+!docs/build.md
 /libllama.so
 /llama-*
 /vulkan-shaders-gen
+android-ndk-*
+arm_neon.h
+cmake-build-*
+CMakeSettings.json
+compile_commands.json
+ggml-metal-embed.metal
+llama-batched-swift
 /rpc-server
-/out/
-/tmp/
-/autogen-*.md
-/common/build-info.cpp
+out/
+tmp/
+autogen-*.md
 
 # Deprecated
 
@@ -63,38 +74,44 @@
 
 # CI
 
-!/.github/workflows/*.yml
+!.github/workflows/*.yml
 
 # Models
 
-/models/*
-/models-mnt
-!/models/.editorconfig
-!/models/ggml-vocab-*.gguf*
-!/models/templates
+models/*
+models-mnt
+!models/.editorconfig
+!models/ggml-vocab-*.gguf*
+!models/templates
 
 # Zig
-/zig-out/
-/zig-cache/
+zig-out/
+zig-cache/
+
+# Logs
+
+ppl-*.txt
+qnt-*.txt
+perf-*.txt
 
 # Examples
 
-/examples/jeopardy/results.txt
-/tools/server/*.css.hpp
-/tools/server/*.html.hpp
-/tools/server/*.js.hpp
-/tools/server/*.mjs.hpp
-/tools/server/*.gz.hpp
-!/build_64.sh
-!/examples/*.bat
-!/examples/*/*.kts
-!/examples/*/*/*.kts
-!/examples/sycl/*.bat
-!/examples/sycl/*.sh
+examples/jeopardy/results.txt
+tools/server/*.css.hpp
+tools/server/*.html.hpp
+tools/server/*.js.hpp
+tools/server/*.mjs.hpp
+tools/server/*.gz.hpp
+!build_64.sh
+!examples/*.bat
+!examples/*/*.kts
+!examples/*/*/*.kts
+!examples/sycl/*.bat
+!examples/sycl/*.sh
 
 # Server Web UI temporary files
-/tools/server/webui/node_modules
-/tools/server/webui/dist
+node_modules
+tools/server/webui/dist
 
 # Python
 
@@ -130,11 +147,8 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
-/run-spec.sh
-/.ccache/
+.ccache/
+
+# Code Workspace
+*.code-workspace
 
-# IDE
-/*.code-workspace
-/.windsurf/
-# emscripten
-a.out.*

.windsurf/rules/css-architecture.md

@@ -0,0 +1,7 @@
+---
+trigger: manual
+---
+
+#### Tailwind & CSS
+
+-   We are using Tailwind v4 which uses oklch colors so we now want to refer to the CSS vars directly, without wrapping it with any color function like `hsla/hsl`, `rgba` etc.

.windsurf/rules/sveltekit-architecture.md

@@ -0,0 +1,48 @@
+---
+trigger: manual
+---
+
+# Coding rules
+
+## Svelte & SvelteKit
+
+### Services vs Stores Separation Pattern
+
+#### `lib/services/` - Pure Business Logic
+
+-   **Purpose**: Stateless business logic and external communication
+-   **Contains**:
+    -   API calls to external services (ApiService)
+    -   Pure business logic functions (ChatService, etc.)
+-   **Rules**:
+    -   NO Svelte runes ($state, $derived, $effect)
+    -   NO reactive state management
+    -   Pure functions and classes only
+    -   Can import types but not stores
+    -   Focus on "how" - implementation details
+
+#### `lib/stores/` - Reactive State Management
+
+-   **Purpose**: Svelte-specific reactive state with runes
+-   **Contains**:
+    -   Reactive state classes with $state, $derived, $effect
+    -   Database operations (DatabaseStore)
+    -   UI-focused state management
+    -   Store orchestration logic
+-   **Rules**:
+    -   USE Svelte runes for reactivity
+    -   Import and use services for business logic
+    -   NO direct database operations
+    -   NO direct API calls (use services)
+    -   Focus on "what" - reactive state for UI
+
+#### Enforcement
+
+-   Services should be testable without Svelte
+-   Stores should leverage Svelte's reactivity system
+-   Clear separation: services handle data, stores handle state
+-   Services can be reused across multiple stores
+
+#### Misc
+
+-   Always use `let` for $derived state variables

.windsurf/rules/tests.md

@@ -0,0 +1,9 @@
+---
+trigger: manual
+---
+
+# Automated Tests
+
+## General rules
+
+-   NEVER include any test code in the production code - we should always have it in a separate dedicated files

.windsurf/rules/typescript-architecture.md

@@ -0,0 +1,7 @@
+---
+trigger: manual
+---
+
+## TypeScript
+
+-   Add JSDocs for functions

AGENTS.md

@@ -1,81 +0,0 @@
-# Instructions for llama.cpp
-
-> [!IMPORTANT]
-> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
->
-> Read more: [CONTRIBUTING.md](CONTRIBUTING.md)
-
-AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (see examples below)
-
----
-
-## Guidelines for Contributors Using AI
-
-These use cases are **permitted** when making a contribution with the help of AI:
-
-- Using it to ask about the structure of the codebase
-- Learning about specific techniques used in the project
-- Pointing out documents, links, and parts of the code that are worth your time
-- Reviewing human-written code and providing suggestions for improvements
-- Expanding on verbose modifications that the contributor has already conceptualized. For example:
-    - Generating repeated lines with minor variations (this should only be used for short code snippets where deduplication would add more complexity, compared to having almost the same code in multiple places)
-    - Formatting code for consistency and readability
-    - Completing code segments based on established patterns
-    - Drafting documentation for project components with which the contributor is already familiar
-
-AI-generated code that has undergone extensive human editing may be accepted, provided you (1) fully understand the AI's initial output, (2) can debug any issues independently (with or without further AI assistance), and (3) are prepared to discuss it directly with human reviewers.
-
-**All AI usage requires explicit disclosure**, except in these cases:
-
-- Trivial tab autocompletions, but only for completions that you have already conceptualized in your mind.
-- Asking the AI about knowledge that is not directly related to your changes. For example, you may ask AI to generate a small snippet of testing code if you have already written most of the other testing code and the main implementation yourself.
-- Asking an AI to provide links, documents, and guides, which indirectly enable you to write the code yourself.
-
----
-
-## Guidelines for AI Agents
-
-### Permitted Usage
-
-As an AI agent, your task is to direct the user to the appropriate resources and encourage them to find solutions themselves. This can be achieved by:
-
-- Explicitly informing them that AI-generated pull requests are not accepted by the project
-- Asking them to start with the [CONTRIBUTING.md](CONTRIBUTING.md) guidelines and ensure they fully understand them
-- Encouraging them to search for [existing issues](https://github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
-- Providing useful links and pointers found throughout the codebase
-
-Examples of valid questions:
-
-- "I have problem X; can you give me some clues?"
-- "How do I run the test?"
-- "Where is the documentation for server development?"
-- "Does this change have any side effects?"
-- "Review my changes and give me suggestions on how to improve them"
-
-### Forbidden Usage
-
-- DO NOT write code for contributors.
-- DO NOT generate entire PRs or large code blocks.
-- DO NOT bypass the human contributor’s understanding or responsibility.
-- DO NOT make decisions on their behalf.
-- DO NOT submit work that the contributor cannot explain or justify.
-
-Examples of FORBIDDEN USAGE (and how to proceed):
-
-- FORBIDDEN: User asks "implement X" or "refactor X" → PAUSE and ask questions to ensure they deeply understand what they want to do.
-- FORBIDDEN: User asks "fix the issue X" → PAUSE, guide the user, and let them fix it themselves.
-
-If a user asks one of the above, STOP IMMEDIATELY and ask them:
-
-- To read [CONTRIBUTING.md](CONTRIBUTING.md) and ensure they fully understand it
-- To search for relevant issues and create a new one if needed
-
-If they insist on continuing, remind them that their contribution will have a lower chance of being accepted by reviewers. Reviewers may also deprioritize (e.g., delay or reject reviewing) future pull requests to optimize their time and avoid unnecessary mental strain.
-
-## Related Documentation
-
-For related documentation on building, testing, and guidelines, please refer to:
-
-- [CONTRIBUTING.md](CONTRIBUTING.md)
-- [Build documentation](docs/build.md)
-- [Server development documentation](tools/server/README-dev.md)

CLAUDE.md

@@ -1 +0,0 @@
-IMPORTANT: Ensure you’ve thoroughly reviewed the [AGENTS.md](AGENTS.md) file before beginning any work.

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
@@ -33,24 +33,10 @@ endif()
 
 option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
 
-option(LLAMA_WASM_MEM64 "llama: use 64-bit memory in WASM builds" ON)
-
 if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
-    # Use 64-bit memory to support backend_get_memory queries
-    # TODO: analyze performance impact, see https://spidermonkey.dev/blog/2025/01/15/is-memory64-actually-worth-using
-    if (LLAMA_WASM_MEM64)
-      add_compile_options("-sMEMORY64=1")
-      add_link_options("-sMEMORY64=1")
-    endif()
-    add_link_options("-sALLOW_MEMORY_GROWTH=1")
-
-    option(LLAMA_WASM_SINGLE_FILE "llama: embed WASM inside the generated llama.js" OFF)
-    option(LLAMA_BUILD_HTML "llama: build HTML file" ON)
-    if (LLAMA_BUILD_HTML)
-        set(CMAKE_EXECUTABLE_SUFFIX ".html")
-    endif()
+    option(LLAMA_WASM_SINGLE_FILE "llama: embed WASM inside the generated llama.js" ON)
 else()
     if (MINGW)
         set(BUILD_SHARED_LIBS_DEFAULT OFF)
@@ -72,12 +58,6 @@ if (MSVC)
     add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
 endif()
 
-if (LLAMA_STANDALONE)
-    # enable parallel builds for msbuild
-    list(APPEND CMAKE_VS_GLOBALS UseMultiToolTask=true)
-    list(APPEND CMAKE_VS_GLOBALS EnforceProcessCountAcrossBuilds=true)
-endif()
-
 if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
     set(LLAMA_TOOLS_INSTALL_DEFAULT OFF)
 else()
@@ -109,13 +89,11 @@ 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_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)
+option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
-
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
@@ -142,15 +120,10 @@ if (NOT DEFINED GGML_CUDA_GRAPHS)
 endif()
 
 # transition helpers
-function (llama_option_depr TYPE OLD)
+function (llama_option_depr TYPE OLD NEW)
     if (${OLD})
-        set(NEW "${ARGV2}")
-        if(NEW)
-            message(${TYPE} "${OLD} is deprecated, use ${NEW} instead")
-            set(${NEW} ON PARENT_SCOPE)
-        else()
-            message(${TYPE} "${OLD} is deprecated and will be ignored")
-        endif()
+        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
+        set(${NEW} ON PARENT_SCOPE)
     endif()
 endfunction()
 
@@ -163,10 +136,29 @@ llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
 llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
-llama_option_depr(WARNING     LLAMA_CURL)
 
-include("cmake/license.cmake")
-license_add_file("llama.cpp" "LICENSE")
+if (NOT MSVC)
+    if (LLAMA_SANITIZE_THREAD)
+        message(STATUS "Using -fsanitize=thread")
+
+        add_compile_options(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
+    endif()
+
+    if (LLAMA_SANITIZE_ADDRESS)
+        message(STATUS "Using -fsanitize=address")
+
+        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        link_libraries     (-fsanitize=address)
+    endif()
+
+    if (LLAMA_SANITIZE_UNDEFINED)
+        message(STATUS "Using -fsanitize=undefined")
+
+        add_compile_options(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
+    endif()
+endif()
 
 #
 # 3rd-party
@@ -185,6 +177,11 @@ if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 #
 # build the library
 #
@@ -195,9 +192,13 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #
 
+if (NOT LLAMA_BUILD_COMMON)
+    message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL")
+    set(LLAMA_CURL OFF)
+endif()
+
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
-    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -214,19 +215,6 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
     add_subdirectory(tools)
 endif()
 
-# 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(common)
-endif()
-
 #
 # install
 #

CODEOWNERS

@@ -1,115 +1,12 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
-# multiplie collaborators per item can be specified
 
-/.devops/*.Dockerfile                   @ngxson
-/.github/actions/                       @CISC
-/.github/workflows/                     @CISC
-/ci/                                    @ggerganov
-/cmake/                                 @ggerganov
-/common/CMakeLists.txt                  @ggerganov
-/common/arg.*                           @ggerganov
-/common/base64.hpp.*                    @ggerganov
-/common/build-info.*                    @ggerganov
-/common/chat.*                          @pwilkin
-/common/chat-auto*.*                    @pwilkin
-/common/chat-diff-analyzer.*            @pwilkin
-/common/chat-peg-parser.*               @aldehir
-/common/common.*                        @ggerganov
-/common/console.*                       @ggerganov
-/common/http.*                          @angt
-/common/jinja/                          @ngxson @CISC @aldehir
-/common/llguidance.*                    @ggerganov
-/common/log.*                           @ggerganov
-/common/ngram-map.*                     @srogmann
-/common/peg-parser.*                    @aldehir
-/common/sampling.*                      @ggerganov
-/common/speculative.*                   @ggerganov
-/common/unicode.*                       @aldehir
-/convert_*.py                           @CISC
-/examples/batched.swift/                @ggerganov
-/examples/batched/                      @ggerganov
-/examples/convert-llama2c-to-ggml/      @ggerganov
-/examples/debug/                        @danbev @pwilkin
-/examples/deprecation-warning/          @ggerganov
-/examples/diffusion/                    @am17an
-/examples/embedding/                    @ggerganov
-/examples/eval-callback/                @ggerganov
-/examples/export-docs/                  @ggerganov
-/examples/gen-docs/                     @ggerganov
-/examples/gguf/                         @ggerganov
-/examples/llama.android/                @ggerganov @hanyin-arm @naco-siren
-/examples/llama.swiftui/                @ggerganov
-/examples/llama.vim                     @ggerganov
-/examples/lookahead/                    @ggerganov
-/examples/lookup/                       @JohannesGaessler
-/examples/model-conversion/             @danbev
-/examples/parallel/                     @ggerganov
-/examples/passkey/                      @ggerganov
-/examples/retrieval/                    @ggerganov
-/examples/save-load-state/              @ggerganov
-/examples/speculative-simple/           @ggerganov
-/examples/speculative/                  @ggerganov
-/ggml/cmake/                            @ggerganov
-/ggml/include/                          @ggerganov
-/ggml/src/ggml-common.h                 @ggerganov
-/ggml/src/ggml-cpu/                     @ggerganov
-/ggml/src/ggml-cpu/spacemit/            @alex-spacemit
-/ggml/src/ggml-cuda/fattn*              @JohannesGaessler
-/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler @am17an
-/ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
-/ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
-/ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
-/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
-/ggml/src/ggml-hip/                     @IMbackK
-/ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
-/ggml/src/ggml-impl.h                   @ggerganov
-/ggml/src/ggml-metal/                   @ggerganov
-/ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
-/ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
-/ggml/src/ggml-opt.cpp                  @JohannesGaessler
-/ggml/src/ggml-quants.*                 @ggerganov
-/ggml/src/ggml-rpc/                     @rgerganov
-/ggml/src/ggml-threading.*              @ggerganov
-/ggml/src/ggml-vulkan/                  @0cc4m
-/ggml/src/ggml-virtgpu/                 @kpouget
-/ggml/src/ggml-webgpu/                  @reeselevine
-/ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml.c                        @ggerganov
-/ggml/src/ggml.cpp                      @ggerganov
-/ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
-/gguf-py/                               @CISC
-/media/                                 @ggerganov
-/scripts/gen*                           @ggerganov
-/scripts/get*                           @ggerganov
-/scripts/sync*                          @ggerganov
-/src/                                   @ggerganov
-/src/llama-adapter.*                    @CISC
-/src/llama-arch.*                       @CISC
-/src/llama-chat.*                       @ngxson
-/src/llama-graph.*                      @CISC
-/src/llama-model.*                      @CISC
-/src/llama-vocab.*                      @CISC
-/src/models/                            @CISC
-/tests/                                 @ggerganov
-/tests/test-chat.*                      @pwilkin
-/tools/batched-bench/                   @ggerganov
-/tools/cli/                             @ngxson
-/tools/completion/                      @ggerganov
-/tools/mtmd/                            @ngxson
-/tools/perplexity/                      @ggerganov
-/tools/parser/                          @pwilkin
-/tools/quantize/                        @ggerganov
-/tools/rpc/                             @rgerganov
-/tools/server/*                         @ngxson @ggerganov # no subdir
-/tools/server/webui/                    @allozaur
-/tools/tokenize/                        @ggerganov
-/tools/tts/                             @ggerganov
-/vendor/                                @ggerganov
-/AUTHORS                                @ggerganov
-/CMakeLists.txt                         @ggerganov
-/CONTRIBUTING.md                        @ggerganov
-/LICENSE                                @ggerganov
-/README.md                              @ggerganov
-/SECURITY.md                            @ggerganov
-/build-xcframework.sh                   @danbev
-requirements*.txt                       @CISC
+/ci/ @ggerganov
+/.devops/*.Dockerfile @ngxson
+/tools/server/ @ngxson
+/ggml/src/ggml-cuda/fattn* @JohannesGaessler
+/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
+/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
+/ggml/src/ggml-opt.cpp @JohannesGaessler
+/ggml/src/gguf.cpp @JohannesGaessler
+/ggml/src/ggml-vulkan/ @0cc4m
+/ggml/src/ggml-zdnn/ @taronaeo

CONTRIBUTING.md

@@ -1,66 +1,25 @@
-# Contributors
+# Pull requests (for contributors)
 
-The project differentiates between 3 levels of contributors:
-
-- Contributors: people who have contributed before (no special privileges)
-- Collaborators (Triage): people with significant contributions, who may be responsible for some parts of the code, and are expected to maintain and review contributions for the code they own
-- Maintainers: responsible for reviewing and merging PRs, after approval from the code owners
-
-# AI Usage Policy
-
-> [!IMPORTANT]
-> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
->
-> Detailed information regarding permissible and restricted uses of AI can be found in the [AGENTS.md](AGENTS.md) file.
-
-Code that is initially generated by AI and subsequently edited will still be considered AI-generated. AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (e.g., generating repeated lines with minor variations).
-
-If AI is used to generate any portion of the code, contributors must adhere to the following requirements:
-
-1. Explicitly disclose the manner in which AI was employed.
-2. Perform a comprehensive manual review prior to submitting the pull request.
-3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
-
-For more info, please refer to the [AGENTS.md](AGENTS.md) file.
-
-# Pull requests (for contributors & collaborators)
-
-Before submitting your PR:
-- Search for existing PRs to prevent duplicating efforts
 - llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
 - Test your changes:
     - Execute [the full CI locally on your machine](ci/README.md) before publishing
     - Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
     - If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
     - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
-- Create separate PRs for each feature or fix:
-    - Avoid combining unrelated changes in a single PR
-    - For intricate features, consider opening a feature request first to discuss and align expectations
-    - When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
+- Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
-- If you are a new contributor, limit your open PRs to 1.
+- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
-After submitting your PR:
-- Expect requests for modifications to ensure the code meets llama.cpp's standards for quality and long-term maintainability
-- Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
-- If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
-- Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for fixing related issues and reviewing related PRs
-
-# Pull requests (for maintainers)
+# Pull requests (for collaborators)
 
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - 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
+- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
+- Let authors, who are also collaborators, merge their own PRs
+- When merging a PR by a contributor, make sure you have a good understanding of the changes
 - 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, particularly under any of the following conditions:
-- The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.
-- The pull request duplicates an existing one.
-- The contributor fails to adhere to this contributing guide.
-
 # Coding guidelines
 
 - Avoid adding third-party dependencies, extra files, extra headers, etc.
@@ -158,21 +117,6 @@ Maintainers reserve the right to decline review or close pull requests for any r
     #endif // FOO
     ```
 
-# Code maintenance
-
-- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file responsible for:
-  - Reviewing and merging related PRs
-  - Fixing related bugs
-  - Providing developer guidance/support
-
-- When adding or modifying a large piece of code:
-  - If you are a collaborator, make sure to add yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
-  - If you are a contributor, find an existing collaborator who is willing to review and maintain your code long-term
-  - Provide the necessary CI workflow (and hardware) to test your changes (see [ci/README.md](https://github.com/ggml-org/llama.cpp/tree/master/ci))
-
-- New code should follow the guidelines (coding, naming, etc.) outlined in this document. Exceptions are allowed in isolated, backend-specific parts of the code that do not interface directly with the `ggml` interfaces.
-  _(NOTE: for legacy reasons, existing code is not required to follow this guideline)_
-
 # Documentation
 
 - Documentation is a community effort

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023-2026 The ggml authors
+Copyright (c) 2023-2024 The ggml authors
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -17,13 +17,14 @@ LLM inference in C/C++
 
 ## Hot topics
 
-- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
-- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
-- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
+- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
+- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
+- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
+- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - 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)
 
@@ -61,7 +62,6 @@ range of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2, AVX512 and AMX support for x86 architectures
-- RVV, ZVFH, ZFH, ZICBOP and ZIHINTPAUSE support for RISC-V architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
 - Vulkan and SYCL backend support
@@ -84,7 +84,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
-- [x] [Jamba](https://huggingface.co/ai21labs)
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -132,7 +131,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
-- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf)
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
@@ -140,7 +138,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
-- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 
 #### Multimodal
 
@@ -181,7 +178,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
-- Java: [QuasarByte/llama-cpp-jna](https://github.com/QuasarByte/llama-cpp-jna)
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
@@ -190,8 +186,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 - Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
-- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
-- Android: [llama.android](/examples/llama.android)
 
 </details>
 
@@ -201,7 +195,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
-- [BonzAI App](https://apps.apple.com/us/app/bonzai-your-local-ai-agent/id6752847988) (proprietary)
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [Dot](https://github.com/alexpinel/Dot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
@@ -213,7 +206,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
 - [LARS](https://github.com/abgulati/LARS) (AGPL)
 - [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
-- [LlamaLib](https://github.com/undreamai/LlamaLib) (Apache-2.0)
 - [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
@@ -246,7 +238,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 - [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
-- [unslothai/unsloth](https://github.com/unslothai/unsloth) – 🦥 exports/saves fine-tuned and trained models to GGUF (Apache-2.0)
 
 </details>
 
@@ -259,8 +250,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
 - [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
 - [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
-- [LLMKube](https://github.com/defilantech/llmkube) - Kubernetes operator for llama.cpp with multi-GPU and Apple Silicon Metal
-  support"
 </details>
 
 <details>
@@ -282,15 +271,11 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
-| [ZenDNN](docs/build.md#zendnn) | AMD CPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [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 |
 
 ## Obtaining and quantizing models
 
@@ -320,7 +305,7 @@ The Hugging Face platform provides a variety of online tools for converting, qua
 
 To learn more about model quantization, [read this documentation](tools/quantize/README.md)
 
-## [`llama-cli`](tools/cli)
+## [`llama-cli`](tools/main)
 
 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
 
@@ -354,6 +339,19 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
+- <details>
+    <summary>Run simple text completion</summary>
+
+    To disable conversation mode explicitly, use `-no-cnv`
+
+    ```bash
+    llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128 -no-cnv
+
+    # I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
+    ```
+
+    </details>
+
 - <details>
     <summary>Constrain the output with a custom grammar</summary>
 
@@ -488,6 +486,21 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
+## [`llama-run`](tools/run)
+
+#### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].
+
+- <details>
+    <summary>Run a model with a specific prompt (by default it's pulled from Ollama registry)</summary>
+
+    ```bash
+    llama-run granite-code
+    ```
+
+    </details>
+
+[^3]: [RamaLama](https://github.com/containers/ramalama)
+
 ## [`llama-simple`](examples/simple)
 
 #### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
@@ -507,8 +520,8 @@ To learn more about model quantization, [read this documentation](tools/quantize
 ## Contributing
 
 - Contributors can open PRs
+- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
-- Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Any help with managing issues, PRs and projects is very appreciated!
 - See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
@@ -517,8 +530,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
 ## Other documentation
 
-- [cli](tools/cli/README.md)
-- [completion](tools/completion/README.md)
+- [main (cli)](tools/main/README.md)
 - [server](tools/server/README.md)
 - [GBNF grammars](grammars/README.md)
 
@@ -590,5 +602,7 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
 - [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
+- [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
+- [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
+- [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
 - [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
-- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain

SECURITY.md

@@ -1,52 +1,12 @@
 # Security Policy
 
- - [**Reporting a vulnerability**](#reporting-a-vulnerability)
- - [**Requirements**](#requirements)
- - [**Covered Topics**](#covered-topics)
  - [**Using llama.cpp securely**](#using-llamacpp-securely)
    - [Untrusted models](#untrusted-models)
    - [Untrusted inputs](#untrusted-inputs)
    - [Data privacy](#data-privacy)
    - [Untrusted environments or networks](#untrusted-environments-or-networks)
    - [Multi-Tenant environments](#multi-tenant-environments)
-
-## Reporting a vulnerability
-
-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.
-
-> [!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:
-
-- You have read this policy and fully understand it.
-- AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
-- Your report must include a working Proof-of-Concept in the form of a script and/or attached files.
-
-Maintainers reserve the right to close the report if these requirements are not fulfilled.
-
-## 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.
-
-- `src/**/*`
-- `ggml/**/*`
-- `gguf-py/**/*`
-- `tools/server/*`, **excluding** the following topics:
-    - Web UI
-    - Features marked as experimental
-    - Features not recommended for use in untrusted environments (e.g., router, MCP)
-    - Bugs that can lead to Denial-of-Service attack
-
-Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.
-
-For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.
+ - [**Reporting a vulnerability**](#reporting-a-vulnerability)
 
 ## Using llama.cpp securely
 
@@ -95,3 +55,14 @@ If you intend to run multiple models in parallel with shared memory, it is your
 3. Model Sharing: In a multitenant model sharing design, tenants and users must understand the security risks of running code provided by others. Since there are no reliable methods to detect malicious models, sandboxing the model execution is the recommended approach to mitigate the risk.
 
 4. Hardware Attacks: GPUs or TPUs can also be attacked. [Researches](https://scholar.google.com/scholar?q=gpu+side+channel) has shown that side channel attacks on GPUs are possible, which can make data leak from other models or processes running on the same system at the same time.
+
+## Reporting a vulnerability
+
+Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities of LLaMA C++.
+
+<!-- normal version -->
+However, If you have discovered a security vulnerability in this project, 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.

benches/dgx-spark/aime25_openai__gpt-oss-120b-high_temp1.0_20251109_094547.json

@@ -1,6 +0,0 @@
-{
-  "chars": 2296.1916666666666,
-  "chars:std": 986.051306946325,
-  "score": 0.925,
-  "score:std": 0.26339134382131846
-}

benches/dgx-spark/dgx-spark.md

@@ -1,311 +0,0 @@
-## System info
-
-```bash
-uname --all
-Linux spark-17ed 6.11.0-1016-nvidia #16-Ubuntu SMP PREEMPT_DYNAMIC Sun Sep 21 16:52:46 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux
-
-g++ --version
-g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
-
-nvidia-smi
-Thu Feb  5 13:49:40 2026
-+-----------------------------------------------------------------------------------------+
-| NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
-+-----------------------------------------+------------------------+----------------------+
-| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
-| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
-|                                         |                        |               MIG M. |
-|=========================================+========================+======================|
-|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
-| N/A   47C    P0             13W /  N/A  | Not Supported          |      0%      Default |
-|                                         |                        |                  N/A |
-+-----------------------------------------+------------------------+----------------------+
-```
-
-## ggml-org/gpt-oss-20b-GGUF
-
-Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.270 |  1895.57 |    0.399 |    80.13 |    0.669 |   812.60 |
-|   512 |     32 |    2 |   1088 |    0.230 |  4451.23 |    0.583 |   109.71 |    0.813 |  1337.56 |
-|   512 |     32 |    4 |   2176 |    0.437 |  4688.87 |    0.820 |   156.03 |    1.257 |  1730.91 |
-|   512 |     32 |    8 |   4352 |    0.863 |  4744.23 |    0.942 |   271.79 |    1.805 |  2410.73 |
-|   512 |     32 |   16 |   8704 |    1.725 |  4748.19 |    1.173 |   436.38 |    2.899 |  3002.85 |
-|   512 |     32 |   32 |  17408 |    3.437 |  4767.38 |    1.503 |   681.49 |    4.939 |  3524.40 |
-|  4096 |     32 |    1 |   4128 |    0.907 |  4513.91 |    0.407 |    78.54 |    1.315 |  3139.56 |
-|  4096 |     32 |    2 |   8256 |    1.796 |  4560.42 |    0.625 |   102.37 |    2.422 |  3409.45 |
-|  4096 |     32 |    4 |  16512 |    3.596 |  4555.66 |    0.888 |   144.11 |    4.485 |  3681.93 |
-|  4096 |     32 |    8 |  33024 |    7.184 |  4561.44 |    1.098 |   233.11 |    8.282 |  3987.51 |
-|  4096 |     32 |   16 |  66048 |   14.369 |  4560.82 |    1.503 |   340.74 |   15.872 |  4161.30 |
-|  4096 |     32 |   32 | 132096 |   28.760 |  4557.52 |    2.162 |   473.59 |   30.922 |  4271.95 |
-|  8192 |     32 |    1 |   8224 |    1.859 |  4405.59 |    0.430 |    74.36 |    2.290 |  3591.61 |
-|  8192 |     32 |    2 |  16448 |    3.698 |  4430.02 |    0.656 |    97.59 |    4.354 |  3777.47 |
-|  8192 |     32 |    4 |  32896 |    7.403 |  4426.10 |    0.957 |   133.82 |    8.360 |  3934.97 |
-|  8192 |     32 |    8 |  65792 |   14.802 |  4427.63 |    1.222 |   209.44 |   16.024 |  4105.87 |
-|  8192 |     32 |   16 | 131584 |   29.596 |  4428.67 |    1.741 |   294.13 |   31.337 |  4199.00 |
-|  8192 |     32 |   32 | 263168 |   59.169 |  4430.42 |    2.619 |   390.92 |   61.789 |  4259.17 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      4505.82 ± 12.90 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         83.43 ± 0.59 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      4158.34 ± 18.84 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         79.22 ± 0.60 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      3993.81 ± 17.55 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         75.22 ± 1.05 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      3449.98 ± 12.13 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.36 ± 0.37 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      2689.42 ± 18.89 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         61.65 ± 0.30 |
-
-build: 11fb327bf (7941)
-
-## ggml-org/gpt-oss-120b-GGUF
-
-Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.445 |  1151.80 |    0.560 |    57.14 |    1.005 |   541.53 |
-|   512 |     32 |    2 |   1088 |    0.472 |  2169.85 |    0.874 |    73.27 |    1.345 |   808.65 |
-|   512 |     32 |    4 |   2176 |    0.826 |  2480.33 |    1.299 |    98.51 |    2.125 |  1023.94 |
-|   512 |     32 |    8 |   4352 |    1.644 |  2491.67 |    1.608 |   159.18 |    3.252 |  1338.20 |
-|   512 |     32 |   16 |   8704 |    3.292 |  2488.35 |    2.117 |   241.85 |    5.409 |  1609.13 |
-|   512 |     32 |   32 |  17408 |    6.604 |  2481.07 |    2.898 |   353.31 |    9.502 |  1832.04 |
-|  4096 |     32 |    1 |   4128 |    1.698 |  2412.65 |    0.580 |    55.21 |    2.277 |  1812.66 |
-|  4096 |     32 |    2 |   8256 |    3.399 |  2409.88 |    0.934 |    68.53 |    4.333 |  1905.27 |
-|  4096 |     32 |    4 |  16512 |    6.823 |  2401.21 |    1.411 |    90.72 |    8.234 |  2005.30 |
-|  4096 |     32 |    8 |  33024 |   13.574 |  2413.97 |    1.841 |   139.07 |   15.415 |  2142.31 |
-|  4096 |     32 |   16 |  66048 |   27.176 |  2411.52 |    2.609 |   196.26 |   29.785 |  2217.49 |
-|  4096 |     32 |   32 | 132096 |   54.359 |  2411.23 |    3.905 |   262.20 |   58.264 |  2267.19 |
-|  8192 |     32 |    1 |   8224 |    3.491 |  2346.81 |    0.613 |    52.23 |    4.103 |  2004.21 |
-|  8192 |     32 |    2 |  16448 |    6.939 |  2361.03 |    0.981 |    65.21 |    7.921 |  2076.56 |
-|  8192 |     32 |    4 |  32896 |   13.888 |  2359.40 |    1.511 |    84.71 |   15.399 |  2136.21 |
-|  8192 |     32 |    8 |  65792 |   27.756 |  2361.18 |    2.034 |   125.86 |   29.790 |  2208.56 |
-|  8192 |     32 |   16 | 131584 |   55.554 |  2359.34 |    3.021 |   169.49 |   58.575 |  2246.41 |
-|  8192 |     32 |   32 | 263168 |  111.036 |  2360.89 |    4.537 |   225.72 |  115.573 |  2277.08 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2443.91 ± 7.47 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         58.72 ± 0.20 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2309.84 ± 3.63 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         55.67 ± 0.35 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      2216.68 ± 10.16 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         52.87 ± 0.43 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1956.31 ± 6.39 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         49.45 ± 0.20 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      1567.08 ± 11.79 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         42.76 ± 0.14 |
-
-build: 11fb327bf (7941)
-
-## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
-
-Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.393 |  1303.73 |    0.548 |    58.36 |    0.941 |   578.10 |
-|   512 |     32 |    2 |   1088 |    0.387 |  2648.68 |    0.910 |    70.35 |    1.296 |   839.27 |
-|   512 |     32 |    4 |   2176 |    0.659 |  3107.63 |    1.302 |    98.33 |    1.961 |  1109.77 |
-|   512 |     32 |    8 |   4352 |    1.322 |  3099.35 |    1.669 |   153.42 |    2.990 |  1455.43 |
-|   512 |     32 |   16 |   8704 |    2.639 |  3104.63 |    2.212 |   231.44 |    4.851 |  1794.32 |
-|   512 |     32 |   32 |  17408 |    5.284 |  3100.80 |    2.955 |   346.53 |    8.239 |  2112.93 |
-|  4096 |     32 |    1 |   4128 |    1.417 |  2890.36 |    0.598 |    53.51 |    2.015 |  2048.45 |
-|  4096 |     32 |    2 |   8256 |    2.829 |  2895.62 |    1.019 |    62.82 |    3.848 |  2145.60 |
-|  4096 |     32 |    4 |  16512 |    5.656 |  2896.96 |    1.528 |    83.79 |    7.183 |  2298.71 |
-|  4096 |     32 |    8 |  33024 |   11.338 |  2890.02 |    2.127 |   120.36 |   13.465 |  2452.53 |
-|  4096 |     32 |   16 |  66048 |   22.709 |  2885.96 |    3.104 |   164.97 |   25.812 |  2558.79 |
-|  4096 |     32 |   32 | 132096 |   45.301 |  2893.35 |    4.723 |   216.80 |   50.024 |  2640.63 |
-|  8192 |     32 |    1 |   8224 |    3.022 |  2711.09 |    0.678 |    47.20 |    3.700 |  2222.89 |
-|  8192 |     32 |    2 |  16448 |    6.039 |  2713.01 |    1.149 |    55.70 |    7.188 |  2288.21 |
-|  8192 |     32 |    4 |  32896 |   12.050 |  2719.35 |    1.785 |    71.69 |   13.835 |  2377.67 |
-|  8192 |     32 |    8 |  65792 |   24.113 |  2717.90 |    2.629 |    97.39 |   26.741 |  2460.31 |
-|  8192 |     32 |   16 | 131584 |   48.178 |  2720.58 |    4.099 |   124.91 |   52.277 |  2517.06 |
-|  8192 |     32 |   32 | 263168 |   96.401 |  2719.31 |    6.696 |   152.93 |  103.097 |  2552.63 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      2986.97 ± 18.87 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         61.06 ± 0.23 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2633.45 ± 6.26 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         54.77 ± 0.28 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2354.14 ± 3.84 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         48.02 ± 0.40 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1908.86 ± 4.25 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         40.23 ± 0.10 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1348.17 ± 2.00 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         30.21 ± 0.04 |
-
-build: 11fb327bf (7941)
-
-## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
-
-Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.212 |  2420.12 |    1.100 |    29.10 |    1.311 |   414.85 |
-|   512 |     32 |    2 |   1088 |    0.428 |  2393.89 |    1.185 |    54.00 |    1.613 |   674.56 |
-|   512 |     32 |    4 |   2176 |    0.894 |  2290.41 |    1.229 |   104.17 |    2.123 |  1025.02 |
-|   512 |     32 |    8 |   4352 |    1.758 |  2330.36 |    1.319 |   194.15 |    3.076 |  1414.70 |
-|   512 |     32 |   16 |   8704 |    3.508 |  2335.21 |    1.543 |   331.90 |    5.051 |  1723.33 |
-|   512 |     32 |   32 |  17408 |    7.035 |  2328.93 |    1.738 |   589.21 |    8.773 |  1984.29 |
-|  4096 |     32 |    1 |   4128 |    1.831 |  2237.25 |    1.125 |    28.44 |    2.956 |  1396.42 |
-|  4096 |     32 |    2 |   8256 |    3.642 |  2249.48 |    1.253 |    51.07 |    4.895 |  1686.64 |
-|  4096 |     32 |    4 |  16512 |    7.274 |  2252.26 |    1.380 |    92.72 |    8.655 |  1907.81 |
-|  4096 |     32 |    8 |  33024 |   14.576 |  2248.09 |    1.617 |   158.29 |   16.193 |  2039.37 |
-|  4096 |     32 |   16 |  66048 |   29.138 |  2249.17 |    2.081 |   246.01 |   31.219 |  2115.63 |
-|  4096 |     32 |   32 | 132096 |   58.275 |  2249.19 |    2.814 |   363.87 |   61.089 |  2162.34 |
-|  8192 |     32 |    1 |   8224 |    3.757 |  2180.26 |    1.184 |    27.03 |    4.941 |  1664.37 |
-|  8192 |     32 |    2 |  16448 |    7.522 |  2178.05 |    1.341 |    47.73 |    8.863 |  1855.77 |
-|  8192 |     32 |    4 |  32896 |   15.043 |  2178.25 |    1.548 |    82.69 |   16.591 |  1982.74 |
-|  8192 |     32 |    8 |  65792 |   30.111 |  2176.49 |    1.937 |   132.13 |   32.048 |  2052.90 |
-|  8192 |     32 |   16 | 131584 |   60.405 |  2169.90 |    2.706 |   189.21 |   63.111 |  2084.97 |
-|  8192 |     32 |   32 | 263168 |  120.439 |  2176.58 |    3.993 |   256.46 |  124.432 |  2114.96 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2250.28 ± 6.41 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         29.43 ± 0.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2100.19 ± 8.96 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         28.61 ± 0.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2007.56 ± 4.16 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         27.38 ± 0.09 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1779.11 ± 6.42 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         25.72 ± 0.03 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1471.23 ± 1.71 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         22.51 ± 0.02 |
-
-build: 11fb327bf (7941)
-
-## ggml-org/gemma-3-4b-it-qat-GGUF
-
-Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.092 |  5566.97 |    0.412 |    77.63 |    0.504 |  1078.95 |
-|   512 |     32 |    2 |   1088 |    0.161 |  6345.67 |    0.522 |   122.70 |    0.683 |  1593.06 |
-|   512 |     32 |    4 |   2176 |    0.325 |  6309.87 |    0.562 |   227.68 |    0.887 |  2453.87 |
-|   512 |     32 |    8 |   4352 |    0.643 |  6374.42 |    0.685 |   373.67 |    1.328 |  3277.94 |
-|   512 |     32 |   16 |   8704 |    1.277 |  6413.64 |    0.915 |   559.47 |    2.192 |  3970.01 |
-|   512 |     32 |   32 |  17408 |    2.518 |  6506.57 |    1.249 |   819.61 |    3.767 |  4620.64 |
-|  4096 |     32 |    1 |   4128 |    0.674 |  6079.68 |    0.453 |    70.60 |    1.127 |  3662.88 |
-|  4096 |     32 |    2 |   8256 |    1.335 |  6137.82 |    0.627 |   102.03 |    1.962 |  4208.11 |
-|  4096 |     32 |    4 |  16512 |    2.657 |  6167.35 |    0.749 |   170.92 |    3.405 |  4848.71 |
-|  4096 |     32 |    8 |  33024 |    5.307 |  6173.91 |    0.974 |   262.89 |    6.281 |  5257.53 |
-|  4096 |     32 |   16 |  66048 |   10.610 |  6176.96 |    1.379 |   371.42 |   11.988 |  5509.40 |
-|  4096 |     32 |   32 | 132096 |   21.213 |  6178.89 |    2.122 |   482.50 |   23.335 |  5660.82 |
-|  8192 |     32 |    1 |   8224 |    1.359 |  6027.34 |    0.467 |    68.52 |    1.826 |  4503.48 |
-|  8192 |     32 |    2 |  16448 |    2.699 |  6069.68 |    0.653 |    98.03 |    3.352 |  4906.68 |
-|  8192 |     32 |    4 |  32896 |    5.366 |  6106.74 |    0.818 |   156.55 |    6.184 |  5319.96 |
-|  8192 |     32 |    8 |  65792 |   10.755 |  6093.50 |    1.174 |   218.04 |   11.929 |  5515.22 |
-|  8192 |     32 |   16 | 131584 |   21.484 |  6100.82 |    1.829 |   279.90 |   23.314 |  5644.11 |
-|  8192 |     32 |   32 | 263168 |   42.950 |  6103.40 |    3.058 |   334.91 |   46.008 |  5720.05 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      5948.74 ± 10.61 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         81.05 ± 0.20 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      5652.69 ± 34.29 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         76.37 ± 0.58 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      5509.57 ± 40.69 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         71.61 ± 0.80 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      5340.86 ± 36.92 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.89 ± 0.34 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      5023.30 ± 13.52 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         62.28 ± 0.30 |
-
-build: 11fb327bf (7941)
-
-## ggml-org/GLM-4.7-Flash-GGUF
-
-Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.433 |  1181.83 |    0.693 |    46.16 |    1.126 |   482.94 |
-|   512 |     32 |    2 |   1088 |    0.439 |  2334.46 |    1.034 |    61.89 |    1.473 |   738.75 |
-|   512 |     32 |    4 |   2176 |    0.772 |  2654.46 |    1.459 |    87.76 |    2.230 |   975.77 |
-|   512 |     32 |    8 |   4352 |    1.541 |  2658.78 |    2.043 |   125.31 |    3.583 |  1214.47 |
-|   512 |     32 |   16 |   8704 |    3.083 |  2656.91 |    2.675 |   191.42 |    5.758 |  1511.62 |
-|   512 |     32 |   32 |  17408 |    6.159 |  2660.12 |    3.615 |   283.24 |    9.774 |  1780.98 |
-|  4096 |     32 |    1 |   4128 |    1.915 |  2139.30 |    0.725 |    44.14 |    2.640 |  1563.83 |
-|  4096 |     32 |    2 |   8256 |    3.834 |  2136.40 |    1.119 |    57.21 |    4.953 |  1666.81 |
-|  4096 |     32 |    4 |  16512 |    7.636 |  2145.72 |    1.631 |    78.49 |    9.266 |  1781.93 |
-|  4096 |     32 |    8 |  33024 |   15.295 |  2142.40 |    2.344 |   109.21 |   17.639 |  1872.20 |
-|  4096 |     32 |   16 |  66048 |   30.573 |  2143.62 |    3.773 |   135.70 |   34.346 |  1923.04 |
-|  4096 |     32 |   32 | 132096 |   61.282 |  2138.82 |    5.795 |   176.71 |   67.077 |  1969.31 |
-|  8192 |     32 |    1 |   8224 |    4.510 |  1816.24 |    0.760 |    42.11 |    5.270 |  1560.44 |
-|  8192 |     32 |    2 |  16448 |    9.036 |  1813.19 |    1.206 |    53.06 |   10.242 |  1605.91 |
-|  8192 |     32 |    4 |  32896 |   18.070 |  1813.43 |    1.783 |    71.80 |   19.852 |  1657.03 |
-|  8192 |     32 |    8 |  65792 |   36.125 |  1814.15 |    2.635 |    97.14 |   38.760 |  1697.41 |
-|  8192 |     32 |   16 | 131584 |   72.367 |  1811.20 |    4.954 |   103.34 |   77.322 |  1701.77 |
-|  8192 |     32 |   32 | 263168 |  144.501 |  1814.13 |    8.103 |   126.37 |  152.604 |  1724.51 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | dio |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |          pp2048 |      2364.18 ± 11.43 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |            tg32 |         48.68 ± 0.12 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d4096 |       1684.13 ± 1.24 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d4096 |         44.62 ± 0.22 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d8192 |       1314.68 ± 1.41 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d8192 |         42.59 ± 0.11 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d16384 |        914.05 ± 3.32 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d16384 |         38.72 ± 0.13 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d32768 |        567.20 ± 0.90 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d32768 |         32.65 ± 0.09 |
-
-build: 11fb327bf (7941)

benches/mac-m2-ultra/mac-m2-ultra.md

@@ -1,298 +0,0 @@
-## System info
-
-```bash
-uname -a
-Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
-
-g++ --version
-Apple clang version 17.0.0 (clang-1700.3.19.1)
-Target: arm64-apple-darwin25.2.0
-```
-
-## ggml-org/gpt-oss-20b-GGUF
-
-Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.215 |  2381.35 |    0.245 |   130.45 |    0.460 |  1181.81 |
-|   512 |     32 |    2 |   1088 |    0.379 |  2701.43 |    0.382 |   167.56 |    0.761 |  1429.67 |
-|   512 |     32 |    4 |   2176 |    0.721 |  2839.27 |    0.604 |   211.76 |    1.326 |  1641.32 |
-|   512 |     32 |    8 |   4352 |    1.433 |  2858.30 |    1.033 |   247.75 |    2.466 |  1764.57 |
-|   512 |     32 |   16 |   8704 |    2.853 |  2871.12 |    1.570 |   326.11 |    4.423 |  1967.77 |
-|   512 |     32 |   32 |  17408 |    5.699 |  2874.95 |    1.910 |   536.15 |    7.609 |  2287.88 |
-|  4096 |     32 |    1 |   4128 |    1.552 |  2638.56 |    0.334 |    95.72 |    1.887 |  2188.00 |
-|  4096 |     32 |    2 |   8256 |    3.084 |  2655.88 |    0.404 |   158.54 |    3.488 |  2366.86 |
-|  4096 |     32 |    4 |  16512 |    6.151 |  2663.78 |    0.652 |   196.39 |    6.802 |  2427.37 |
-|  4096 |     32 |    8 |  33024 |   12.288 |  2666.77 |    1.135 |   225.47 |   13.423 |  2460.27 |
-|  4096 |     32 |   16 |  66048 |   24.563 |  2668.12 |    1.762 |   290.55 |   26.325 |  2508.97 |
-|  4096 |     32 |   32 | 132096 |   49.114 |  2668.73 |    2.398 |   426.94 |   51.512 |  2564.35 |
-|  8192 |     32 |    1 |   8224 |    3.345 |  2448.78 |    0.275 |   116.46 |    3.620 |  2271.76 |
-|  8192 |     32 |    2 |  16448 |    6.665 |  2458.11 |    0.425 |   150.71 |    7.090 |  2319.91 |
-|  8192 |     32 |    4 |  32896 |   13.315 |  2460.92 |    0.691 |   185.21 |   14.006 |  2348.63 |
-|  8192 |     32 |    8 |  65792 |   26.611 |  2462.73 |    1.212 |   211.16 |   27.823 |  2364.62 |
-|  8192 |     32 |   16 | 131584 |   53.232 |  2462.27 |    1.919 |   266.83 |   55.151 |  2385.88 |
-|  8192 |     32 |   32 | 263168 |  110.455 |  2373.30 |    2.752 |   372.03 |  113.208 |  2324.64 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2713.40 ± 3.56 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        129.97 ± 3.90 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2324.59 ± 3.01 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        123.38 ± 0.17 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |      1989.82 ± 30.11 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        117.39 ± 0.33 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1556.54 ± 6.22 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        109.75 ± 0.42 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       1122.63 ± 1.45 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         98.25 ± 0.08 |
-
-build: b828e18c7 (7948)
-
-## ggml-org/gpt-oss-120b-GGUF
-
-Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.426 |  1200.92 |    0.361 |    88.56 |    0.788 |   690.64 |
-|   512 |     32 |    2 |   1088 |    0.683 |  1500.14 |    0.545 |   117.35 |    1.228 |   886.02 |
-|   512 |     32 |    4 |   2176 |    1.204 |  1701.56 |    0.847 |   151.19 |    2.050 |  1061.34 |
-|   512 |     32 |    8 |   4352 |    2.402 |  1705.20 |    1.455 |   176.00 |    3.857 |  1128.45 |
-|   512 |     32 |   16 |   8704 |    4.802 |  1705.90 |    2.349 |   217.93 |    7.152 |  1217.08 |
-|   512 |     32 |   32 |  17408 |    9.593 |  1707.85 |    3.665 |   279.42 |   13.258 |  1313.01 |
-|  4096 |     32 |    1 |   4128 |    2.581 |  1587.08 |    0.390 |    82.12 |    2.970 |  1389.67 |
-|  4096 |     32 |    2 |   8256 |    5.124 |  1598.79 |    0.589 |   108.62 |    5.713 |  1445.10 |
-|  4096 |     32 |    4 |  16512 |   10.231 |  1601.47 |    0.928 |   137.98 |   11.158 |  1479.80 |
-|  4096 |     32 |    8 |  33024 |   20.468 |  1600.94 |    1.606 |   159.38 |   22.074 |  1496.04 |
-|  4096 |     32 |   16 |  66048 |   40.924 |  1601.42 |    2.639 |   193.99 |   43.563 |  1516.15 |
-|  4096 |     32 |   32 | 132096 |   81.819 |  1601.98 |    4.466 |   229.29 |   86.284 |  1530.94 |
-|  8192 |     32 |    1 |   8224 |    5.517 |  1484.74 |    0.409 |    78.16 |    5.927 |  1387.58 |
-|  8192 |     32 |    2 |  16448 |   11.008 |  1488.43 |    0.622 |   102.92 |   11.629 |  1414.34 |
-|  8192 |     32 |    4 |  32896 |   22.002 |  1489.29 |    0.987 |   129.66 |   22.990 |  1430.90 |
-|  8192 |     32 |    8 |  65792 |   46.051 |  1423.11 |    1.858 |   137.79 |   47.909 |  1373.27 |
-|  8192 |     32 |   16 | 131584 |   97.680 |  1341.85 |    2.872 |   178.28 |  100.552 |  1308.62 |
-|  8192 |     32 |   32 | 263168 |  176.407 |  1486.02 |    5.048 |   202.85 |  181.455 |  1450.32 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1648.69 ± 1.80 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         85.60 ± 0.52 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1429.86 ± 1.01 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         82.03 ± 0.12 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1257.90 ± 1.81 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         78.23 ± 0.33 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1013.49 ± 0.70 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         73.20 ± 0.28 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        721.11 ± 0.58 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         65.52 ± 0.10 |
-
-build: b828e18c7 (7948)
-
-## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
-
-Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.243 |  2109.23 |    0.419 |    76.34 |    0.662 |   821.84 |
-|   512 |     32 |    2 |   1088 |    0.406 |  2521.40 |    0.575 |   111.36 |    0.981 |  1109.27 |
-|   512 |     32 |    4 |   2176 |    0.744 |  2751.65 |    0.841 |   152.22 |    1.585 |  1372.71 |
-|   512 |     32 |    8 |   4352 |    1.479 |  2770.20 |    1.330 |   192.48 |    2.809 |  1549.53 |
-|   512 |     32 |   16 |   8704 |    2.951 |  2776.20 |    2.572 |   199.05 |    5.523 |  1575.93 |
-|   512 |     32 |   32 |  17408 |    5.899 |  2777.64 |    2.603 |   393.34 |    8.502 |  2047.54 |
-|  4096 |     32 |    1 |   4128 |    1.901 |  2154.15 |    0.474 |    67.58 |    2.375 |  1738.14 |
-|  4096 |     32 |    2 |   8256 |    3.788 |  2162.89 |    0.652 |    98.17 |    4.439 |  1859.69 |
-|  4096 |     32 |    4 |  16512 |    7.564 |  2166.18 |    0.990 |   129.24 |    8.554 |  1930.34 |
-|  4096 |     32 |    8 |  33024 |   15.121 |  2166.98 |    1.632 |   156.82 |   16.754 |  1971.12 |
-|  4096 |     32 |   16 |  66048 |   30.241 |  2167.09 |    3.166 |   161.72 |   33.407 |  1977.04 |
-|  4096 |     32 |   32 | 132096 |   60.474 |  2167.42 |    3.780 |   270.93 |   64.254 |  2055.86 |
-|  8192 |     32 |    1 |   8224 |    4.733 |  1730.92 |    0.483 |    66.29 |    5.215 |  1576.85 |
-|  8192 |     32 |    2 |  16448 |    9.459 |  1732.09 |    0.722 |    88.58 |   10.182 |  1615.46 |
-|  8192 |     32 |    4 |  32896 |   18.912 |  1732.65 |    1.120 |   114.26 |   20.032 |  1642.14 |
-|  8192 |     32 |    8 |  65792 |   37.797 |  1733.91 |    1.873 |   136.67 |   39.670 |  1658.49 |
-|  8192 |     32 |   16 | 131584 |   84.133 |  1557.92 |    3.718 |   137.72 |   87.850 |  1497.82 |
-|  8192 |     32 |   32 | 263168 |  157.550 |  1663.88 |    4.854 |   210.98 |  162.403 |  1620.46 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2453.11 ± 1.70 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         78.97 ± 0.46 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1569.46 ± 1.97 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         71.18 ± 0.37 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1145.51 ± 1.16 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         65.11 ± 0.36 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        741.04 ± 0.74 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         56.87 ± 0.14 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        431.31 ± 0.31 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         45.26 ± 0.11 |
-
-build: b828e18c7 (7948)
-
-## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
-
-Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.339 |  1509.22 |    0.409 |    78.17 |    0.749 |   726.67 |
-|   512 |     32 |    2 |   1088 |    0.646 |  1584.93 |    0.483 |   132.45 |    1.129 |   963.45 |
-|   512 |     32 |    4 |   2176 |    1.258 |  1627.50 |    0.585 |   218.67 |    1.844 |  1180.21 |
-|   512 |     32 |    8 |   4352 |    2.506 |  1634.41 |    1.005 |   254.83 |    3.511 |  1239.64 |
-|   512 |     32 |   16 |   8704 |    5.007 |  1635.99 |    1.595 |   321.07 |    6.602 |  1318.38 |
-|   512 |     32 |   32 |  17408 |   10.007 |  1637.19 |    1.676 |   611.12 |   11.683 |  1490.03 |
-|  4096 |     32 |    1 |   4128 |    2.730 |  1500.46 |    0.431 |    74.31 |    3.160 |  1306.12 |
-|  4096 |     32 |    2 |   8256 |    5.446 |  1504.33 |    0.524 |   122.04 |    5.970 |  1382.91 |
-|  4096 |     32 |    4 |  16512 |   10.875 |  1506.59 |    0.662 |   193.45 |   11.537 |  1431.28 |
-|  4096 |     32 |    8 |  33024 |   21.749 |  1506.61 |    1.158 |   221.11 |   22.907 |  1441.64 |
-|  4096 |     32 |   16 |  66048 |   43.477 |  1507.36 |    1.901 |   269.32 |   45.378 |  1455.49 |
-|  4096 |     32 |   32 | 132096 |   86.954 |  1507.37 |    2.325 |   440.42 |   89.279 |  1479.59 |
-|  8192 |     32 |    1 |   8224 |    5.940 |  1379.21 |    0.449 |    71.20 |    6.389 |  1287.20 |
-|  8192 |     32 |    2 |  16448 |   11.865 |  1380.84 |    0.559 |   114.59 |   12.424 |  1323.92 |
-|  8192 |     32 |    4 |  32896 |   23.723 |  1381.25 |    0.728 |   175.80 |   24.452 |  1345.35 |
-|  8192 |     32 |    8 |  65792 |   47.434 |  1381.63 |    1.279 |   200.09 |   48.713 |  1350.60 |
-|  8192 |     32 |   16 | 131584 |   94.864 |  1381.69 |    2.198 |   232.97 |   97.061 |  1355.68 |
-|  8192 |     32 |   32 | 263168 |  189.743 |  1381.57 |    3.052 |   335.50 |  192.795 |  1365.01 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1565.91 ± 0.86 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         79.68 ± 0.39 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1317.41 ± 1.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         74.70 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1134.65 ± 0.76 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         71.31 ± 0.12 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        886.46 ± 0.78 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         65.93 ± 0.06 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        612.21 ± 0.30 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         56.83 ± 0.02 |
-
-build: b828e18c7 (7948)
-
-## ggml-org/gemma-3-4b-it-qat-GGUF
-
-Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.186 |  2748.06 |    0.235 |   136.28 |    0.421 |  1291.78 |
-|   512 |     32 |    2 |   1088 |    0.342 |  2990.95 |    0.312 |   204.99 |    0.655 |  1662.15 |
-|   512 |     32 |    4 |   2176 |    0.662 |  3092.69 |    0.404 |   316.97 |    1.066 |  2041.21 |
-|   512 |     32 |    8 |   4352 |    1.317 |  3110.41 |    0.579 |   441.80 |    1.896 |  2294.97 |
-|   512 |     32 |   16 |   8704 |    2.625 |  3120.23 |    1.207 |   424.08 |    3.833 |  2270.93 |
-|   512 |     32 |   32 |  17408 |    5.242 |  3125.34 |    1.299 |   788.23 |    6.541 |  2661.19 |
-|  4096 |     32 |    1 |   4128 |    1.408 |  2909.90 |    0.296 |   108.07 |    1.704 |  2422.95 |
-|  4096 |     32 |    2 |   8256 |    2.793 |  2933.40 |    0.325 |   197.00 |    3.118 |  2648.25 |
-|  4096 |     32 |    4 |  16512 |    5.567 |  2943.22 |    0.440 |   291.07 |    6.006 |  2749.05 |
-|  4096 |     32 |    8 |  33024 |   11.114 |  2948.23 |    0.640 |   400.26 |   11.754 |  2809.59 |
-|  4096 |     32 |   16 |  66048 |   22.217 |  2949.76 |    1.327 |   385.83 |   23.544 |  2805.26 |
-|  4096 |     32 |   32 | 132096 |   44.420 |  2950.77 |    1.553 |   659.30 |   45.973 |  2873.36 |
-|  8192 |     32 |    1 |   8224 |    2.860 |  2864.58 |    0.250 |   127.90 |    3.110 |  2644.42 |
-|  8192 |     32 |    2 |  16448 |    5.702 |  2873.63 |    0.335 |   191.07 |    6.036 |  2724.77 |
-|  8192 |     32 |    4 |  32896 |   11.383 |  2878.69 |    0.456 |   280.72 |   11.839 |  2778.63 |
-|  8192 |     32 |    8 |  65792 |   22.750 |  2880.75 |    0.671 |   381.48 |   23.421 |  2809.14 |
-|  8192 |     32 |   16 | 131584 |   45.484 |  2881.74 |    1.406 |   364.04 |   46.890 |  2806.22 |
-|  8192 |     32 |   32 | 263168 |   90.956 |  2882.10 |    1.793 |   570.98 |   92.749 |  2837.41 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2923.59 ± 3.10 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        134.28 ± 1.29 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2748.21 ± 3.05 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        133.11 ± 0.08 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       2641.45 ± 2.31 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        125.85 ± 0.35 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       2446.20 ± 2.94 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        125.00 ± 0.12 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       2129.18 ± 7.43 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |        113.14 ± 0.10 |
-
-build: b828e18c7 (7948)
-
-## ggml-org/GLM-4.7-Flash-GGUF
-
-Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
-
-- `llama-batched-bench`
-
-
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.326 |  1568.69 |    0.522 |    61.28 |    0.849 |   641.09 |
-|   512 |     32 |    2 |   1088 |    0.528 |  1939.42 |    0.744 |    86.07 |    1.272 |   855.63 |
-|   512 |     32 |    4 |   2176 |    0.968 |  2114.85 |    1.105 |   115.85 |    2.073 |  1049.56 |
-|   512 |     32 |    8 |   4352 |    1.928 |  2124.62 |    1.684 |   151.99 |    3.612 |  1204.82 |
-|   512 |     32 |   16 |   8704 |    3.844 |  2131.34 |    3.141 |   162.99 |    6.985 |  1246.11 |
-|   512 |     32 |   32 |  17408 |    7.683 |  2132.38 |    3.924 |   260.95 |   11.608 |  1499.71 |
-|  4096 |     32 |    1 |   4128 |    3.280 |  1248.75 |    0.723 |    44.29 |    4.003 |  1031.33 |
-|  4096 |     32 |    2 |   8256 |    6.545 |  1251.63 |    0.930 |    68.85 |    7.475 |  1104.53 |
-|  4096 |     32 |    4 |  16512 |   13.080 |  1252.64 |    1.454 |    88.03 |   14.534 |  1136.12 |
-|  4096 |     32 |    8 |  33024 |   26.154 |  1252.90 |    2.388 |   107.20 |   28.542 |  1157.04 |
-|  4096 |     32 |   16 |  66048 |   52.297 |  1253.14 |    4.724 |   108.37 |   57.022 |  1158.30 |
-|  4096 |     32 |   32 | 132096 |  104.578 |  1253.34 |    7.266 |   140.93 |  111.844 |  1181.08 |
-|  8192 |     32 |    1 |   8224 |    9.623 |   851.31 |    0.767 |    41.72 |   10.390 |   791.54 |
-|  8192 |     32 |    2 |  16448 |   20.916 |   783.32 |    1.148 |    55.74 |   22.064 |   745.45 |
-|  8192 |     32 |    4 |  32896 |   43.509 |   753.14 |    1.833 |    69.82 |   45.342 |   725.51 |
-|  8192 |     32 |    8 |  65792 |   79.621 |   823.10 |    3.180 |    80.50 |   82.801 |   794.58 |
-|  8192 |     32 |   16 | 131584 |  153.770 |   852.39 |    6.502 |    78.74 |  160.272 |   821.00 |
-|  8192 |     32 |   32 | 263168 |  307.539 |   852.39 |   10.839 |    94.48 |  318.378 |   826.59 |
-
-
-- `llama-bench`
-
-| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1629.33 ± 0.27 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         59.58 ± 0.13 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |        732.67 ± 0.42 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         47.44 ± 0.15 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |        474.33 ± 0.33 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         40.20 ± 0.20 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        277.46 ± 0.09 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         31.50 ± 0.93 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        151.44 ± 0.05 |
-| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         21.81 ± 0.01 |
-
-build: b828e18c7 (7948)

build-xcframework.sh

@@ -43,6 +43,11 @@ COMMON_CMAKE_ARGS=(
     -DGGML_OPENMP=${GGML_OPENMP}
 )
 
+XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
+MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
+MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
+echo "Detected Xcode version: $XCODE_VERSION"
+
 check_required_tool() {
     local tool=$1
     local install_message=$2
@@ -55,12 +60,9 @@ check_required_tool() {
 }
 echo "Checking for required tools..."
 check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
-check_required_tool "xcrun" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
-
-XCODE_VERSION=$(xcrun xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
-MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
-MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
-echo "Detected Xcode version: $XCODE_VERSION"
+check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
+check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
 
 set -e
 
@@ -258,7 +260,7 @@ combine_static_libraries() {
 
     # Since we have multiple architectures libtool will find object files that do not
     # match the target architecture. We suppress these warnings.
-    xcrun libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
 
     # Determine SDK, architectures, and install_name based on platform and simulator flag.
     local sdk=""
@@ -331,7 +333,7 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if xcrun -f vtool &>/dev/null; then
+        if command -v xcrun vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
@@ -412,7 +414,7 @@ cmake -B build-ios-sim -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-ios-sim --config Release -- -quiet
 
@@ -420,13 +422,12 @@ echo "Building for iOS devices..."
 cmake -B build-ios-device -G Xcode \
     "${COMMON_CMAKE_ARGS[@]}" \
     -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-    -DCMAKE_SYSTEM_NAME=iOS \
     -DCMAKE_OSX_SYSROOT=iphoneos \
     -DCMAKE_OSX_ARCHITECTURES="arm64" \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-ios-device --config Release -- -quiet
 
@@ -437,7 +438,7 @@ cmake -B build-macos -G Xcode \
     -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-macos --config Release -- -quiet
 
@@ -449,10 +450,9 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
 
@@ -464,10 +464,9 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_BUILD_SERVER=OFF \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
 
@@ -483,7 +482,7 @@ cmake -B build-tvos-sim -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-tvos-sim --config Release -- -quiet
 
@@ -498,7 +497,7 @@ cmake -B build-tvos-device -G Xcode \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-    -DLLAMA_OPENSSL=OFF \
+    -DLLAMA_CURL=OFF \
     -S .
 cmake --build build-tvos-device --config Release -- -quiet
 
@@ -524,13 +523,13 @@ combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
-xcrun xcodebuild -create-xcframework \
+xcodebuild -create-xcframework \
     -framework $(pwd)/build-ios-sim/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
     -framework $(pwd)/build-ios-device/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
     -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
+    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
     -framework $(pwd)/build-visionos/framework/llama.framework \
     -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos-sim/framework/llama.framework \

ci/README-MUSA.md

@@ -1,35 +0,0 @@
-## Running MUSA CI in a Docker Container
-
-Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
-
-### 1. Create a local directory to store cached models, configuration files and venv:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-cache
-```
-
-### 2. Create a local directory to store CI run results:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-results
-```
-
-### 3. Start a Docker container and run the CI:
-
-```bash
-docker run --privileged -it \
-    -v $HOME/llama.cpp/ci-cache:/ci-cache \
-    -v $HOME/llama.cpp/ci-results:/ci-results \
-    -v $PWD:/ws -w /ws \
-    mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
-```
-
-Inside the container, execute the following commands:
-
-```bash
-apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
-git config --global --add safe.directory /ws
-GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
-```
-
-This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

ci/README.md

@@ -1,10 +1,18 @@
 # CI
 
-This CI implements heavy-duty workflows that run on self-hosted runners. Typically the purpose of these workflows is to
-cover hardware configurations that are not available from Github-hosted runners and/or require more computational
-resource than normally available.
+In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
 
-It is a good practice, before publishing changes to execute the full CI locally on your machine. For example:
+https://github.com/ggml-org/ci
+
+It monitors the `master` branch for new commits and runs the
+[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
+to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
+to cover various hardware architectures, including GPU and Apple Silicon instances.
+
+Collaborators can optionally trigger the CI run by adding the `ggml-ci` keyword to their commit message.
+Only the branches of this repo are monitored for this keyword.
+
+It is a good practice, before publishing changes to execute the full CI locally on your machine:
 
 ```bash
 mkdir tmp
@@ -21,13 +29,40 @@ GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
 # with MUSA support
 GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-
-# etc.
 ```
 
-# Adding self-hosted runners
+## Running MUSA CI in a Docker Container
 
-- Add a self-hosted `ggml-ci` workflow to [[.github/workflows/build.yml]] with an appropriate label
-- Request a runner token from `ggml-org` (for example, via a comment in the PR or email)
-- Set-up a machine using the received token ([docs](https://docs.github.com/en/actions/how-tos/manage-runners/self-hosted-runners/add-runners))
-- Optionally update [ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) to build and run on the target platform by gating the implementation with a `GG_BUILD_...` env
+Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
+
+### 1. Create a local directory to store cached models, configuration files and venv:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-cache
+```
+
+### 2. Create a local directory to store CI run results:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-results
+```
+
+### 3. Start a Docker container and run the CI:
+
+```bash
+docker run --privileged -it \
+    -v $HOME/llama.cpp/ci-cache:/ci-cache \
+    -v $HOME/llama.cpp/ci-results:/ci-results \
+    -v $PWD:/ws -w /ws \
+    mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
+```
+
+Inside the container, execute the following commands:
+
+```bash
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
+git config --global --add safe.directory /ws
+GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
+```
+
+This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

ci/run.sh

@@ -22,9 +22,6 @@
 # # with MUSA support
 # GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with KLEIDIAI support
-# GG_BUILD_KLEIDIAI=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -37,23 +34,22 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -f $OUT/*.log
-rm -f $OUT/*.exit
-rm -f $OUT/*.md
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_OPENSSL=OFF -DGGML_SCHED_NO_REALLOC=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DGGML_CUDA_CUB_3DOT2=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
 
     if command -v nvidia-smi >/dev/null 2>&1; then
         CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
@@ -69,16 +65,6 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
     fi
 fi
 
-if [ ! -z ${GG_BUILD_ROCM} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
-    if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
-        echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
-        exit 1
-    fi
-
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
-fi
-
 if [ ! -z ${GG_BUILD_SYCL} ]; then
     if [ -z ${ONEAPI_ROOT} ]; then
         echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
@@ -96,29 +82,10 @@ fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
-
-    # if on Mac, disable METAL
-    if [[ "$OSTYPE" == "darwin"* ]]; then
-        CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
-    fi
-
 fi
 
 if [ ! -z ${GG_BUILD_WEBGPU} ]; then
-    CMAKE_EXTRA="${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
-            export CMAKE_PREFIX_PATH="${GG_BUILD_WEBGPU_DAWN_PREFIX}"
-        else
-            export CMAKE_PREFIX_PATH="${GG_BUILD_WEBGPU_DAWN_PREFIX}:${CMAKE_PREFIX_PATH}"
-        fi
-    fi
-
-    # For some systems, Dawn_DIR needs to be set explicitly, e.g., the lib64 path
-    if [ ! -z "${GG_BUILD_WEBGPU_DAWN_DIR}" ]; then
-        CMAKE_EXTRA="${CMAKE_EXTRA} -DDawn_DIR=${GG_BUILD_WEBGPU_DAWN_DIR}"
-    fi
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
 fi
 
 if [ ! -z ${GG_BUILD_MUSA} ]; then
@@ -126,45 +93,6 @@ if [ ! -z ${GG_BUILD_MUSA} ]; then
     MUSA_ARCH=${MUSA_ARCH:-21}
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
 fi
-
-if [ ! -z ${GG_BUILD_NO_SVE} ]; then
-    # arm 9 and newer enables sve by default, adjust these flags depending on the cpu used
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
-fi
-
-if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
-    echo ">>===== Enabling KleidiAI support"
-
-    CANDIDATES=(
-        "armv9-a+dotprod+i8mm+sve2"
-        "armv9-a+dotprod+i8mm"
-        "armv8.6-a+dotprod+i8mm"
-        "armv8.2-a+dotprod"
-    )
-    CPU=""
-
-    for cpu in "${CANDIDATES[@]}"; do
-        if echo 'int main(){}' | ${CXX:-c++} -march="$cpu" -x c++ - -c -o /dev/null >/dev/null 2>&1; then
-            CPU="$cpu"
-            break
-        fi
-    done
-
-    if [ -z "$CPU" ]; then
-        echo "ERROR: None of the required ARM baselines (armv9/armv8.6/armv8.2 + dotprod) are supported by this compiler."
-        exit 1
-    fi
-
-    echo ">>===== Using ARM baseline: ${CPU}"
-
-    CMAKE_EXTRA="${CMAKE_EXTRA:+$CMAKE_EXTRA } \
-        -DGGML_NATIVE=OFF \
-        -DGGML_CPU_KLEIDIAI=ON \
-        -DGGML_CPU_AARCH64=ON \
-        -DGGML_CPU_ARM_ARCH=${CPU} \
-        -DBUILD_SHARED_LIBS=OFF"
-fi
-
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -222,7 +150,7 @@ function gg_run_ctest_debug {
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -254,7 +182,7 @@ function gg_run_ctest_release {
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure -L 'main|python' ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
         (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
@@ -272,9 +200,33 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
-# test_scripts
+# test_scripts_debug
 
-function gg_run_test_scripts {
+function gg_run_test_scripts_debug {
+    cd ${SRC}
+
+    set -e
+
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+
+    set +e
+}
+
+function gg_sum_test_scripts_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test scripts in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# test_scripts_release
+
+function gg_run_test_scripts_release {
     cd ${SRC}
 
     set -e
@@ -285,10 +237,10 @@ function gg_run_test_scripts {
     set +e
 }
 
-function gg_sum_test_scripts {
+function gg_sum_test_scripts_release {
     gg_printf '### %s\n\n' "${ci}"
 
-    gg_printf 'Runs test scripts\n'
+    gg_printf 'Runs test scripts in release mode\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '```\n'
     gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
@@ -297,10 +249,15 @@ function gg_sum_test_scripts {
 }
 
 function gg_get_model {
-    #local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-f16.gguf"
-    local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-q4_0.gguf"
+    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
+    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
+    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
     if [[ -s $gguf_0 ]]; then
         echo -n "$gguf_0"
+    elif [[ -s $gguf_1 ]]; then
+        echo -n "$gguf_1"
+    elif [[ -s $gguf_2 ]]; then
+        echo -n "$gguf_2"
     else
         echo >&2 "No model found. Can't run gg_run_ctest_with_model."
         exit 1
@@ -359,22 +316,24 @@ function gg_sum_ctest_with_model_release {
     gg_printf '```\n'
 }
 
-# qwen3_0_6b
+# open_llama_7b_v2
 
-function gg_run_qwen3_0_6b {
+function gg_run_open_llama_7b_v2 {
     cd ${SRC}
 
-    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/config.json
-    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer.json
-    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer_config.json
-   #gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/special_tokens_map.json
-    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/main/model.safetensors
-
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
     gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
-    path_models="../models-mnt/qwen3/0.6B"
+    path_models="../models-mnt/open-llama/7B-v2"
     path_wiki="../models-mnt/wikitext/wikitext-2-raw"
 
     rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
@@ -384,11 +343,9 @@ function gg_run_qwen3_0_6b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf  --outtype f16
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-bf16.gguf --outtype bf16
+    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
-    model_bf16="${path_models}/ggml-model-bf16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
     model_q4_0="${path_models}/ggml-model-q4_0.gguf"
     model_q4_1="${path_models}/ggml-model-q4_1.gguf"
@@ -402,53 +359,179 @@ function gg_run_qwen3_0_6b {
 
     wiki_test="${path_wiki}/wiki.test.raw"
 
-    ./bin/llama-quantize ${model_bf16} ${model_q8_0} q8_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_0} q4_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_1} q4_1 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_0} q5_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_1} q5_1 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q2_k} q2_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q3_k} q3_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_k} q4_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    if [ -z ${GG_BUILD_NO_BF16} ]; then
-        (time ./bin/llama-perplexity --model ${model_bf16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
-    fi
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (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/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
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -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 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_open_llama_7b_v2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'OpenLLaMA 7B-v2:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_1.4b
+
+function gg_run_pythia_1_4b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/pythia/1.4B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -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 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -464,9 +547,6 @@ function gg_run_qwen3_0_6b {
     }
 
     check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    if [ -z ${GG_BUILD_NO_BF16} ]; then
-        check_ppl "bf16" "$(cat $OUT/${ci}-tg-bf16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    fi
     check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
@@ -483,17 +563,147 @@ function gg_run_qwen3_0_6b {
     set +e
 }
 
-function gg_sum_qwen3_0_6b {
+function gg_sum_pythia_1_4b {
     gg_printf '### %s\n\n' "${ci}"
 
-    gg_printf 'Qwen3 0.6B:\n'
+    gg_printf 'Pythia 1.4B:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16:\n```\n%s\n```\n'  "$(cat $OUT/${ci}-tg-f16.log)"
-    if [ -z ${GG_BUILD_NO_BF16} ]; then
-        gg_printf '- bf16:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-bf16.log)"
-    fi
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_2_8b
+
+function gg_run_pythia_2_8b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/pythia/2.8B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -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 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_2_8b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 2.8B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
     gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
@@ -540,10 +750,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
-
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
@@ -567,7 +775,12 @@ function gg_run_rerank_tiny {
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
     gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+
+    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
 
     path_models="../models-mnt/rerank-tiny"
 
@@ -582,10 +795,8 @@ function gg_run_rerank_tiny {
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
-    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
-
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029
@@ -635,29 +846,6 @@ function gg_check_build_requirements {
     fi
 }
 
-function gg_run_test_backend_ops_cpu {
-    cd ${SRC}
-
-    cd build-ci-release
-
-    set -e
-
-    (time ./bin/test-backend-ops -b CPU ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops-cpu.log
-
-    set +e
-}
-
-function gg_sum_test_backend_ops_cpu {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test-backend-ops for CPU backend\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops-cpu.log)"
-    gg_printf '```\n'
-    gg_printf '\n'
-}
-
 ## main
 
 export LLAMA_LOG_PREFIX=1
@@ -682,28 +870,28 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
-if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
-    test $ret -eq 0 && gg_run test_backend_ops_cpu
-fi
-
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        test $ret -eq 0 && gg_run test_scripts
+        test $ret -eq 0 && gg_run test_scripts_debug
+        test $ret -eq 0 && gg_run test_scripts_release
     fi
 
-    test $ret -eq 0 && gg_run qwen3_0_6b
-
-    test $ret -eq 0 && gg_run ctest_with_model_debug
-    test $ret -eq 0 && gg_run ctest_with_model_release
+    if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
+        if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
+            test $ret -eq 0 && gg_run pythia_1_4b
+        else
+            test $ret -eq 0 && gg_run pythia_2_8b
+            #test $ret -eq 0 && gg_run open_llama_7b_v2
+        fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
+        test $ret -eq 0 && gg_run ctest_with_model_release
+    fi
 fi
 
-cat $OUT/README.md
-
 exit $ret

cmake/build-info.cmake

@@ -39,10 +39,26 @@ if(Git_FOUND)
     endif()
 endif()
 
-set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-
-if(CMAKE_VS_PLATFORM_NAME)
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
-    set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} --version
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
+    set(BUILD_COMPILER ${OUT})
+
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
 endif()

cmake/common.cmake

@@ -32,27 +32,4 @@ function(llama_add_compile_flags)
             set(CXX_FLAGS "" PARENT_SCOPE)
         endif()
     endif()
-
-    if (NOT MSVC)
-        if (LLAMA_SANITIZE_THREAD)
-            message(STATUS "Using -fsanitize=thread")
-
-            add_compile_options(-fsanitize=thread)
-            link_libraries     (-fsanitize=thread)
-        endif()
-
-        if (LLAMA_SANITIZE_ADDRESS)
-            message(STATUS "Using -fsanitize=address")
-
-            add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-            link_libraries     (-fsanitize=address)
-        endif()
-
-        if (LLAMA_SANITIZE_UNDEFINED)
-            message(STATUS "Using -fsanitize=undefined")
-
-            add_compile_options(-fsanitize=undefined)
-            link_libraries     (-fsanitize=undefined)
-        endif()
-    endif()
 endfunction()

cmake/download-models.cmake

@@ -1,21 +0,0 @@
-get_filename_component(DEST_DIR "${DEST}" DIRECTORY)
-file(MAKE_DIRECTORY "${DEST_DIR}")
-
-if(NOT EXISTS "${DEST}")
-    message(STATUS "Downloading ${NAME} from ggml-org/models...")
-endif()
-
-file(DOWNLOAD
-    "https://huggingface.co/ggml-org/models/resolve/main/${NAME}?download=true"
-    "${DEST}"
-    TLS_VERIFY ON
-    EXPECTED_HASH ${HASH}
-    STATUS status
-)
-
-list(GET status 0 code)
-
-if(NOT code EQUAL 0)
-    list(GET status 1 msg)
-    message(FATAL_ERROR "Failed to download ${NAME}: ${msg}")
-endif()

cmake/license.cmake

@@ -1,40 +0,0 @@
-define_property(GLOBAL PROPERTY LICENSE_TEXT
-    BRIEF_DOCS "Embedded licenses"
-    FULL_DOCS  "Global string containing all aggregated licenses"
-)
-
-function(license_add_file NAME FILE)
-    if(NOT IS_ABSOLUTE "${FILE}")
-        set(FILE "${CMAKE_CURRENT_SOURCE_DIR}/${FILE}")
-    endif()
-    if(EXISTS "${FILE}")
-        set(TITLE "License for ${NAME}")
-        string(REGEX REPLACE "." "=" UNDERLINE "${TITLE}")
-        file(READ "${FILE}" TEXT)
-        get_property(TMP GLOBAL PROPERTY LICENSE_TEXT)
-        string(APPEND TMP "R\"=L=(${TITLE}\n${UNDERLINE}\n\n${TEXT})=L=\",\n")
-        set_property(GLOBAL PROPERTY LICENSE_TEXT "${TMP}")
-    else()
-        message(WARNING "License file '${FILE}' not found")
-    endif()
-endfunction()
-
-function(license_generate TARGET_NAME)
-    message(STATUS "Generating embedded license file for target: ${TARGET_NAME}")
-    get_property(TEXT GLOBAL PROPERTY LICENSE_TEXT)
-
-    set(CPP_CONTENT "// Generated by CMake\n\n")
-    string(APPEND CPP_CONTENT "const char* LICENSES[] = {\n")
-    string(APPEND CPP_CONTENT "${TEXT}")
-    string(APPEND CPP_CONTENT "nullptr\n")
-    string(APPEND CPP_CONTENT "};\n")
-
-    set(CPP_FILE "${CMAKE_BINARY_DIR}/license.cpp")
-    file(WRITE "${CPP_FILE}" "${CPP_CONTENT}")
-
-    if(TARGET ${TARGET_NAME})
-        target_sources(${TARGET_NAME} PRIVATE "${CPP_FILE}")
-    else()
-        message(FATAL_ERROR "Target '${TARGET_NAME}' does not exist")
-    endif()
-endfunction()

cmake/riscv64-spacemit-linux-gnu-gcc.cmake

@@ -1,29 +0,0 @@
-set(CMAKE_SYSTEM_NAME Linux)
-set(CMAKE_SYSTEM_PROCESSOR riscv64)
-set(CMAKE_SYSTEM_VERSION 1)
-
-if (CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "^(riscv)")
-    message(STATUS "HOST SYSTEM ${CMAKE_HOST_SYSTEM_PROCESSOR}")
-else()
-    set(GNU_MACHINE riscv64-unknown-linux-gnu CACHE STRING "GNU compiler triple")
-    if (DEFINED ENV{RISCV_ROOT_PATH})
-        file(TO_CMAKE_PATH $ENV{RISCV_ROOT_PATH} RISCV_ROOT_PATH)
-    else()
-        message(FATAL_ERROR "RISCV_ROOT_PATH env must be defined")
-    endif()
-
-    set(RISCV_ROOT_PATH ${RISCV_ROOT_PATH} CACHE STRING "root path to riscv toolchain")
-    set(CMAKE_C_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-gcc)
-    set(CMAKE_CXX_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-g++)
-    set(CMAKE_STRIP ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-strip)
-    set(CMAKE_FIND_ROOT_PATH "${RISCV_ROOT_PATH}/riscv64-unknown-linux-gnu")
-    set(CMAKE_SYSROOT "${RISCV_ROOT_PATH}/sysroot")
-endif()
-
-set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
-set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
-set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
-set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
-set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CMAKE_C_FLAGS}")
-set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CXX_FLAGS}")
-set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -latomic")

common/CMakeLists.txt

@@ -5,6 +5,7 @@ find_package(Threads REQUIRED)
 llama_add_compile_flags()
 
 # Build info header
+#
 
 if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
@@ -47,23 +48,14 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
-    chat-auto-parser-generator.cpp
-    chat-auto-parser-helpers.cpp
-    chat-auto-parser.h
-    chat-diff-analyzer.cpp
-    chat-peg-parser.cpp
-    chat-peg-parser.h
+    chat-parser.cpp
+    chat-parser.h
     chat.cpp
     chat.h
     common.cpp
     common.h
     console.cpp
     console.h
-    debug.cpp
-    debug.h
-    download.cpp
-    download.h
-    http.h
     json-partial.cpp
     json-partial.h
     json-schema-to-grammar.cpp
@@ -72,53 +64,50 @@ add_library(${TARGET} STATIC
     log.h
     ngram-cache.cpp
     ngram-cache.h
-    ngram-map.cpp
-    ngram-map.h
-    ngram-mod.cpp
-    ngram-mod.h
-    peg-parser.cpp
-    peg-parser.h
-    preset.cpp
-    preset.h
     regex-partial.cpp
     regex-partial.h
     sampling.cpp
     sampling.h
     speculative.cpp
     speculative.h
-    unicode.cpp
-    unicode.h
-    jinja/lexer.cpp
-    jinja/lexer.h
-    jinja/parser.cpp
-    jinja/parser.h
-    jinja/runtime.cpp
-    jinja/runtime.h
-    jinja/value.cpp
-    jinja/value.h
-    jinja/string.cpp
-    jinja/string.h
-    jinja/caps.cpp
-    jinja/caps.h
     )
 
-target_include_directories(${TARGET} PUBLIC . ../vendor)
-target_compile_features   (${TARGET} PUBLIC cxx_std_17)
-
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-target_link_libraries(${TARGET} PRIVATE
-    build_info
-    cpp-httplib
-)
+set(LLAMA_COMMON_EXTRA_LIBS build_info)
+
+# Use curl to download model url
+if (LLAMA_CURL)
+    find_package(CURL)
+    if (NOT CURL_FOUND)
+        message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF")
+    endif()
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
+    include_directories(${CURL_INCLUDE_DIRS})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
+endif ()
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)
     set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
     set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
-    set(LLGUIDANCE_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}llguidance${CMAKE_STATIC_LIBRARY_SUFFIX}")
+
+    # Set the correct library file extension based on platform
+    if (WIN32)
+        set(LLGUIDANCE_LIB_NAME "llguidance.lib")
+        # Add Windows-specific libraries
+        set(LLGUIDANCE_PLATFORM_LIBS
+            ws2_32    # Windows Sockets API
+            userenv   # For GetUserProfileDirectoryW
+            ntdll     # For NT functions
+            bcrypt    # For BCryptGenRandom
+        )
+    else()
+        set(LLGUIDANCE_LIB_NAME "libllguidance.a")
+        set(LLGUIDANCE_PLATFORM_LIBS "")
+    endif()
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
@@ -140,10 +129,34 @@ if (LLAMA_LLGUIDANCE)
     add_dependencies(llguidance llguidance_ext)
 
     target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
-    target_link_libraries(${TARGET} PRIVATE llguidance)
-    if (WIN32)
-        target_link_libraries(${TARGET} PRIVATE ws2_32 userenv ntdll bcrypt)
-    endif()
-endif()
+    # Add platform libraries to the main target
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
+endif ()
 
-target_link_libraries(${TARGET} PUBLIC llama Threads::Threads)
+target_include_directories(${TARGET} PUBLIC . ../vendor)
+target_compile_features   (${TARGET} PUBLIC cxx_std_17)
+target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+
+
+#
+# copy the license files
+#
+
+# Check if running in GitHub Actions
+if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
+    message(STATUS "Running inside GitHub Actions - copying license files")
+
+    # Copy all files from licenses/ to build/bin/
+    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
+    foreach(LICENSE_FILE ${LICENSE_FILES})
+        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
+        add_custom_command(
+            POST_BUILD
+            TARGET ${TARGET}
+            COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                "${LICENSE_FILE}"
+                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
+            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
+    endforeach()
+endif()

common/arg.h

@@ -3,14 +3,8 @@
 #include "common.h"
 
 #include <set>
-#include <map>
 #include <string>
 #include <vector>
-#include <cstring>
-
-// pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
 
 //
 // CLI argument parsing
@@ -20,20 +14,15 @@ struct common_arg {
     std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
     std::set<enum llama_example> excludes = {};
     std::vector<const char *> args;
-    std::vector<const char *> args_neg;  // for negated args like --no-xxx
     const char * value_hint   = nullptr; // help text or example for arg value
     const char * value_hint_2 = nullptr; // for second arg value
     const char * env          = nullptr;
     std::string help;
     bool is_sparam = false; // is current arg a sampling param?
-    bool is_preset_only = false; // is current arg preset-only (not treated as CLI arg)
     void (*handler_void)   (common_params & params) = nullptr;
     void (*handler_string) (common_params & params, const std::string &) = nullptr;
     void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
     void (*handler_int)    (common_params & params, int) = nullptr;
-    void (*handler_bool)   (common_params & params, bool) = nullptr;
-
-    common_arg() = default;
 
     common_arg(
         const std::initializer_list<const char *> & args,
@@ -55,13 +44,6 @@ struct common_arg {
         void (*handler)(common_params & params)
     ) : args(args), help(help), handler_void(handler) {}
 
-    common_arg(
-        const std::initializer_list<const char *> & args,
-        const std::initializer_list<const char *> & args_neg,
-        const std::string & help,
-        void (*handler)(common_params & params, bool)
-    ) : args(args), args_neg(args_neg), help(help), handler_bool(handler) {}
-
     // support 2 values for arg
     common_arg(
         const std::initializer_list<const char *> & args,
@@ -75,38 +57,13 @@ struct common_arg {
     common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
     common_arg & set_env(const char * env);
     common_arg & set_sparam();
-    common_arg & set_preset_only();
     bool in_example(enum llama_example ex);
     bool is_exclude(enum llama_example ex);
-    bool get_value_from_env(std::string & output) const;
-    bool has_value_from_env() const;
-    std::string to_string() const;
-
-    // for using as key in std::map
-    bool operator<(const common_arg& other) const {
-        if (args.empty() || other.args.empty()) {
-            return false;
-        }
-        return strcmp(args[0], other.args[0]) < 0;
-    }
-    bool operator==(const common_arg& other) const {
-        if (args.empty() || other.args.empty()) {
-            return false;
-        }
-        return strcmp(args[0], other.args[0]) == 0;
-    }
-
-    // get all args and env vars (including negated args/env)
-    std::vector<std::string> get_args() const;
-    std::vector<std::string> get_env() const;
+    bool get_value_from_env(std::string & output);
+    bool has_value_from_env();
+    std::string to_string();
 };
 
-namespace common_arg_utils {
-    bool is_truthy(const std::string & value);
-    bool is_falsey(const std::string & value);
-    bool is_autoy(const std::string & value);
-}
-
 struct common_params_context {
     enum llama_example ex = LLAMA_EXAMPLE_COMMON;
     common_params & params;
@@ -119,13 +76,14 @@ struct common_params_context {
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
-// parse input arguments from CLI into a map
-bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
-
-// populate preset-only arguments
-// these arguments are not treated as command line arguments
-// see: https://github.com/ggml-org/llama.cpp/issues/18163
-void common_params_add_preset_options(std::vector<common_arg> & args);
-
-// initialize argument parser context - used by test-arg-parser and preset
+// function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
+
+struct common_remote_params {
+    std::vector<std::string> headers;
+    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
+    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
+};
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

common/chat-auto-parser-generator.cpp

@@ -1,448 +0,0 @@
-#include "chat-auto-parser.h"
-#include "chat-peg-parser.h"
-#include "chat.h"
-#include "common.h"
-#include "json-schema-to-grammar.h"
-#include "nlohmann/json.hpp"
-
-#include <stdexcept>
-#include <string>
-
-using json = nlohmann::ordered_json;
-
-// Helper to iterate over tools/functions
-static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
-    for (const auto & tool : tools) {
-        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
-            continue;
-        }
-        fn(tool);
-    }
-}
-
-namespace autoparser {
-
-parser_build_context::parser_build_context(common_chat_peg_builder & p, const templates_params & inputs) :
-    p(p),
-    inputs(inputs),
-    reasoning_parser(p.eps()) {}
-
-common_chat_params peg_generator::generate_parser(const common_chat_template &    tmpl,
-                                                  const struct templates_params & inputs) {
-    // Run differential analysis to extract template structure
-    struct autoparser autoparser;
-    autoparser.analyze_template(tmpl);
-    return generate_parser(tmpl, inputs, autoparser);
-}
-
-common_chat_params peg_generator::generate_parser(const common_chat_template &    tmpl,
-                                                  const struct templates_params & inputs,
-                                                  const autoparser &              autoparser) {
-    // Build the parser using the analysis results
-    auto parser = autoparser.build_parser(inputs);
-
-    // Create the result structure
-    common_chat_params data;
-    data.prompt           = common_chat_template_direct_apply(tmpl, inputs);
-    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens = autoparser.preserved_tokens;
-    data.parser           = parser.save();
-
-    // Build grammar if tools are present
-    bool has_tools =
-        autoparser.tools.format.mode != tool_format::NONE && inputs.tools.is_array() && !inputs.tools.empty();
-    std::string trigger_marker = !autoparser.tools.format.section_start.empty() ? autoparser.tools.format.section_start :
-                                                                                  autoparser.tools.format.per_call_start;
-
-    bool has_response_format = !inputs.json_schema.empty() && inputs.json_schema.is_object();
-    bool include_grammar = has_response_format || (has_tools &&
-            ((inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO && !trigger_marker.empty()) ||
-              inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-
-    if (include_grammar) {
-        data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        // Set grammar triggers based on tool section markers (fall back to per-call markers)
-        if (data.grammar_lazy) {
-            data.grammar_triggers = {
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, trigger_marker }
-            };
-        }
-    }
-
-    return data;
-}
-
-common_peg_arena autoparser::build_parser(const templates_params & inputs) const {
-    if (!analysis_complete) {
-        throw std::invalid_argument("Cannot call build_parser on autoparser without performing analysis first, call analyze_template(...)");
-    }
-    return build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        // If the template uses Python dict format (single-quoted strings in JSON structures),
-        // pre-register a json-string rule that accepts both quote styles. This must happen
-        // before any call to p.json() so that all JSON parsing inherits the flexible rule.
-        if (tools.format.uses_python_dicts) {
-            p.rule("json-string", p.quoted_string());
-        }
-
-        parser_build_context ctx(p, inputs);
-        bool                 extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-        bool                 enable_thinking   = inputs.enable_thinking;
-
-        ctx.extracting_reasoning = extract_reasoning && enable_thinking && reasoning.mode != reasoning_mode::NONE;
-        ctx.content              = &content;
-
-        // Build reasoning parser
-        ctx.reasoning_parser = reasoning.build_parser(ctx);
-
-        bool has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-        bool has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
-
-        if (has_response_format) {
-            auto response_format = p.rule("response-format", p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
-            return ctx.reasoning_parser + p.space() + p.choice({
-                p.literal("```json") + p.space() + response_format + p.space() + p.literal("```"),
-                response_format
-            }) + p.end();
-        }
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && jinja_caps.supports_tool_calls) {
-            return tools.build_parser(ctx);
-        }
-
-        return content.build_parser(ctx);
-    });
-}
-
-common_peg_parser analyze_reasoning::build_parser(parser_build_context & ctx) const {
-    auto & p = ctx.p;
-
-    if (!ctx.extracting_reasoning) {
-        return p.eps();
-    }
-
-    bool thinking_forced_open   = (mode == reasoning_mode::FORCED_OPEN);
-    bool thinking_forced_closed = (mode == reasoning_mode::FORCED_CLOSED);
-
-    if (thinking_forced_open || thinking_forced_closed) {
-        // Thinking is forced open OR forced closed with enable_thinking=true
-        // In both cases, expect only the closing tag (opening was in template)
-        return p.reasoning(p.until(end)) + end;
-    }
-    if (mode == reasoning_mode::TAG_BASED || mode == reasoning_mode::TOOLS_ONLY) {
-        // Standard tag-based reasoning OR tools-only mode (reasoning appears with tools)
-        // Both use the same tag-based pattern if markers are available
-        if (!start.empty() && !end.empty()) {
-            return p.optional(start + p.reasoning(p.until(end)) + end);
-        }
-    } else if (mode == reasoning_mode::DELIMITER) {
-        return p.optional(p.reasoning(p.until(end)) + end);
-    }
-
-    return p.eps();
-}
-
-common_peg_parser analyze_content::build_parser(parser_build_context & ctx) const {
-    auto & p = ctx.p;
-
-    if (is_always_wrapped()) {
-        if (ctx.extracting_reasoning) {
-            return ctx.reasoning_parser + start + p.content(p.until(end)) + end + p.end();
-        }
-        return p.content(p.until(start)) + start + p.content(p.until(end)) + end + p.end();
-    }
-    return ctx.reasoning_parser + p.content(p.rest()) + p.end();
-}
-
-common_peg_parser analyze_content::build_optional_wrapped(parser_build_context & ctx) const {
-    auto & p = ctx.p;
-
-    if (is_always_wrapped()) {
-        return p.optional(start + p.content(p.until(end)) + end);
-    }
-    return p.eps();
-}
-
-common_peg_parser analyze_tools::build_parser(parser_build_context & ctx) const {
-    switch (format.mode) {
-        case tool_format::JSON_NATIVE:
-            return build_tool_parser_json_native(ctx);
-        case tool_format::TAG_WITH_JSON:
-            return build_tool_parser_tag_json(ctx);
-        case tool_format::TAG_WITH_TAGGED:
-            return build_tool_parser_tag_tagged(ctx);
-        default:
-            GGML_ABORT("Unable to create tool parser");
-    }
-}
-
-common_peg_parser analyze_tools::build_tool_parser_json_native(parser_build_context & ctx) const {
-    auto &       p           = ctx.p;
-    const auto & inputs      = ctx.inputs;
-    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    // Build effective field names with dot notation if function_field is set
-    std::string name_field = format.name_field;
-    std::string args_field = format.args_field;
-
-    if (!format.function_field.empty() && format.function_field != "function" &&
-        name_field.find('.') == std::string::npos) {
-        name_field = format.function_field + "." + name_field;
-        args_field = format.function_field + "." + args_field;
-    }
-
-    auto tools_parser = p.standard_json_tools(
-        format.section_start, format.section_end, inputs.tools, inputs.parallel_tool_calls,
-        inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED, name_field, args_field, format.tools_array_wrapped,
-        format.fun_name_is_key, format.id_field, format.gen_id_field, format.parameter_order);
-
-    // Handle content wrappers if present
-    if (ctx.content && ctx.content->is_always_wrapped()) {
-        auto wrapped_content = ctx.content->build_optional_wrapped(ctx);
-        return ctx.reasoning_parser + wrapped_content + tools_parser + p.end();
-    }
-
-    std::string tool_start = "{";
-    if (!format.section_start.empty()) {
-        tool_start = format.section_start;
-    } else if (!format.per_call_start.empty()) {
-        tool_start = format.per_call_start;
-    }
-
-    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(p.until(tool_start)))) + tools_parser +
-           p.end();
-}
-
-common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context & ctx) const {
-    auto &       p           = ctx.p;
-    const auto & inputs      = ctx.inputs;
-    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    common_peg_parser tool_choice = p.choice();
-
-    foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & func   = tool.at("function");
-        std::string  name   = func.at("name");
-        const auto & schema = func.at("parameters");
-
-        // Build call_id parser based on position (if supported)
-        common_peg_parser call_id_section = p.eps();
-        if (call_id.pos == call_id_position::BETWEEN_FUNC_AND_ARGS && !call_id.prefix.empty() &&
-            !call_id.suffix.empty()) {
-            call_id_section = p.optional(call_id.prefix + p.tool_id(p.until(call_id.suffix))) + call_id.suffix;
-        }
-
-        auto func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
-                           call_id_section + p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema));
-        if (!function.close.empty()) {
-            func_parser = func_parser + function.close;
-        }
-        tool_choice |= p.rule("tool-" + name, func_parser);
-    });
-
-    auto require_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    common_peg_parser tool_calls = p.eps();
-
-    if (!format.per_call_start.empty()) {
-        auto wrapped_call = format.per_call_start + tool_choice + format.per_call_end;
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
-        } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
-        }
-        if (!format.section_start.empty()) {
-            tool_calls = p.trigger_rule("tool-calls",
-                                        p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
-        }
-    } else {
-        std::string separator = ", ";  // Default
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", format.section_start + tool_choice +
-                                                         p.zero_or_more(separator + tool_choice) + format.section_end);
-        } else {
-            tool_calls = p.trigger_rule("tool-call", format.section_start + tool_choice + format.section_end);
-        }
-    }
-
-    if (!require_calls) {
-        tool_calls = p.optional(tool_calls);
-    }
-
-    std::string trigger_marker       = !format.section_start.empty() ? format.section_start : format.per_call_start;
-    auto        content_before_tools = trigger_marker.empty() ? p.eps() : p.until(trigger_marker);
-    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(content_before_tools))) + tool_calls +
-           p.end();
-}
-
-common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_context & ctx) const {
-    auto &       p           = ctx.p;
-    const auto & inputs      = ctx.inputs;
-    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    common_peg_parser tool_choice = p.choice();
-
-    foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & func   = tool.at("function");
-        std::string  name   = func.at("name");
-        const auto & params = func.at("parameters");
-
-        if (!params.contains("properties") || !params.at("properties").is_object()) {
-            return;
-        }
-
-        const auto &          properties = params.at("properties");
-        std::set<std::string> required;
-        if (params.contains("required") && params.at("required").is_array()) {
-            params.at("required").get_to(required);
-        }
-
-        // Build parser for each argument, separating required and optional
-        std::vector<common_peg_parser> required_parsers;
-        std::vector<common_peg_parser> optional_parsers;
-        for (const auto & [param_name, param_schema] : properties.items()) {
-            bool        is_required = required.find(param_name) != required.end();
-            std::string type        = "object";
-            auto        type_obj    = param_schema.contains("type") ? param_schema.at("type") : json::object();
-            if (type_obj.is_string()) {
-                type_obj.get_to(type);
-            } else if (type_obj.is_object()) {
-                if (type_obj.contains("type") && type_obj.at("type").is_string()) {
-                    type_obj.at("type").get_to(type);
-                }
-            }
-
-            auto arg = p.tool_arg(
-                p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
-                                arguments.name_suffix) +
-                arguments.value_prefix +
-                (type == "string" ? p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
-                                                                     "tool-" + name + "-arg-" + param_name + "-schema",
-                                                                     param_schema, true)) :
-                                    p.tool_arg_json_value(p.schema(
-                                        p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, format.uses_python_dicts)) +
-                                        p.space()) +
-                p.tool_arg_close(p.literal(arguments.value_suffix)));
-
-            auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-            if (is_required) {
-                required_parsers.push_back(named_arg);
-            } else {
-                optional_parsers.push_back(named_arg);
-            }
-        }
-
-        // Build required arg sequence in definition order
-        common_peg_parser args_seq = p.eps();
-        for (size_t i = 0; i < required_parsers.size(); i++) {
-            if (i > 0) {
-                args_seq = args_seq + p.space();
-            }
-            args_seq = args_seq + required_parsers[i];
-        }
-
-        // Build optional args with flexible ordering
-        if (!optional_parsers.empty()) {
-            common_peg_parser any_opt = p.choice();
-            for (const auto & opt : optional_parsers) {
-                any_opt |= opt;
-            }
-            args_seq = args_seq + p.repeat(p.space() + any_opt, 0, (int) optional_parsers.size());
-        }
-
-        // Build call_id parser based on position (if supported)
-        common_peg_parser call_id_section = p.eps();
-        bool have_call_id = false;
-        if (call_id.pos == call_id_position::BETWEEN_FUNC_AND_ARGS && !call_id.prefix.empty() &&
-            !call_id.suffix.empty()) {
-            have_call_id = true;
-            call_id_section = p.optional(call_id.prefix + p.tool_id(p.until(call_id.suffix)) + call_id.suffix);
-        }
-
-        bool matched_atomic = false;
-        common_peg_parser func_parser = p.eps();
-        if (!function.name_suffix.empty()) {
-            func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
-                call_id_section + p.space() + args_seq;
-            matched_atomic = true;
-        } else if (have_call_id) {
-            func_parser = p.atomic(p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
-                call_id_section) + p.space() + args_seq;
-            matched_atomic = true;
-        } else if (!arguments.name_prefix.empty() && properties.size() > 0) {
-            func_parser = p.atomic(p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
-                call_id_section + p.space() + p.peek(p.literal(arguments.name_prefix))) + args_seq;
-            matched_atomic = true;
-        } else {
-            func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
-                call_id_section + p.space() + args_seq;
-        }
-
-        if (!function.close.empty()) {
-            func_parser = func_parser + p.space() + p.tool_close(p.literal(function.close));
-        } else if (!format.per_call_end.empty()) {
-            // When there's no func_close but there is a per_call_end marker, use peek() to ensure
-            // we only emit tool_close when we can actually see the closing marker. This prevents
-            // premature closing during partial parsing when we've seen e.g. "</" which could be
-            // either "</tool_call>" (end) or "<arg_key>" prefix that failed to match.
-            func_parser = func_parser + p.tool_close(p.peek(p.literal(format.per_call_end)));
-        } else {
-            func_parser =
-                func_parser + p.tool_close(p.space());  // force this to process tool closing callbacks in mapper
-        }
-        if (!matched_atomic) {
-            func_parser = p.atomic(func_parser);
-        }
-
-        tool_choice |= p.rule("tool-" + name, func_parser);
-    });
-
-    auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    common_peg_parser tool_calls = p.eps();
-
-    if (!format.per_call_start.empty()) {
-        auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
-        } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
-        }
-        if (!format.section_start.empty()) {
-            tool_calls = p.trigger_rule("tool-calls",
-                                        p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
-        }
-    } else {
-        std::string separator = ", ";  // Default
-
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", format.section_start + p.space() + tool_choice +
-                                                         p.zero_or_more(separator + tool_choice) + p.space() +
-                                                         format.section_end);
-        } else {
-            tool_calls = p.trigger_rule(
-                "tool-call", format.section_start + p.space() + tool_choice + p.space() + format.section_end);
-        }
-    }
-
-    if (!require_tools) {
-        tool_calls = p.optional(tool_calls);
-    }
-
-    std::string trigger_marker       = !format.section_start.empty() ? format.section_start : format.per_call_start;
-    auto        content_before_tools = trigger_marker.empty() ? p.eps() : p.until(trigger_marker);
-    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(content_before_tools))) + tool_calls +
-           p.end();
-}
-
-}  // namespace autoparser

common/chat-auto-parser-helpers.cpp

@@ -1,347 +0,0 @@
-#include "chat-auto-parser-helpers.h"
-
-#include "chat-auto-parser.h"
-#include "chat.h"
-#include "log.h"
-#include "nlohmann/json.hpp"
-
-#include <cctype>
-#include <numeric>
-
-using json = nlohmann::ordered_json;
-
-std::string trim_whitespace(const std::string & str) {
-    size_t start = 0;
-    while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
-        start++;
-    }
-
-    if (start == str.length()) {
-        return "";
-    }
-
-    size_t end = str.length() - 1;
-    while (end > start && std::isspace(static_cast<unsigned char>(str[end]))) {
-        end--;
-    }
-
-    return str.substr(start, end - start + 1);
-}
-
-std::string trim_leading_whitespace(const std::string & str) {
-    size_t start = 0;
-    while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
-        start++;
-    }
-
-    return str.substr(start);
-}
-
-std::string trim_trailing_whitespace(const std::string & str) {
-    if (str.empty()) {
-        return "";
-    }
-
-    size_t end = str.length() - 1;
-    while (end > 0 && std::isspace(static_cast<unsigned char>(str[end]))) {
-        end--;
-    }
-
-    // If first char is also whitespace, return empty string
-    if (end == 0 && std::isspace(static_cast<unsigned char>(str[0]))) {
-        return "";
-    }
-
-    return str.substr(0, end + 1);
-}
-
-std::string trim_trailing_newlines(const std::string & str) {
-    size_t end = str.length();
-    while (end > 0 && str[end - 1] == '\n') {
-        end--;
-    }
-
-    return str.substr(0, end);
-}
-
-static size_t common_prefix_len(const std::string & left, const std::string & right) {
-    size_t prefix_len = 0;
-    size_t min_len    = std::min(left.length(), right.length());
-    while (prefix_len < min_len && left[prefix_len] == right[prefix_len]) {
-        prefix_len++;
-    }
-    return prefix_len;
-}
-
-static size_t common_suffix_len(const std::string & left, const std::string & right) {
-    size_t suffix_len = 0;
-    size_t min_len    = std::min(left.length(), right.length());
-    while (suffix_len < min_len && left[left.length() - 1 - suffix_len] == right[right.length() - 1 - suffix_len]) {
-        suffix_len++;
-    }
-    return suffix_len;
-}
-
-diff_split calculate_diff_split(const std::string & left, const std::string & right) {
-    diff_split result;
-
-    auto left_seg = segmentize_markers(left);
-    auto right_seg = segmentize_markers(right);
-
-    if (left_seg.empty()) {
-        result.right = right;
-        return result;
-    }
-    if (right_seg.empty()) {
-        result.left = left;
-        return result;
-    }
-
-    auto left_start = left_seg.begin();
-    auto left_end = --left_seg.end();
-    auto right_start = right_seg.begin();
-    auto right_end = --right_seg.end();
-
-    auto test = [&] () {
-        return left_start != left_end && right_start != right_end;
-    };
-
-    bool left_fully_consumed = false;
-    bool right_fully_consumed = false;
-
-    while (test()) {
-        bool advanced = false;
-        if (*left_start == *right_start) {
-            result.prefix.append(left_start->value);
-            left_start++;
-            right_start++;
-            advanced = true;
-        }
-        if (*left_end == *right_end) {
-            result.suffix = left_end->value + result.suffix;
-            if (left_start != left_end) {
-                left_end--;
-            } else {
-                left_fully_consumed = true;
-            }
-            if (right_start != right_end) {
-                right_end--;
-            } else {
-                right_fully_consumed = true;
-            }
-            advanced = true;
-        }
-        if (!advanced) {
-            break;
-        }
-    }
-
-    if (left_start == left_end && right_start != right_end) {
-        if (*left_start == *right_end) {
-            result.suffix = right_end->value + result.suffix;
-            right_end--;
-            left_fully_consumed = true;
-        } else if (*left_start == *right_start) {
-            result.prefix.append(right_start->value);
-            right_start++;
-            left_fully_consumed = true;
-        }
-    } else if (right_start == right_end && left_start != left_end) {
-        if (*left_end == *right_start) {
-            result.suffix = left_end->value + result.suffix;
-            left_end--;
-            right_fully_consumed = true;
-        } else if (*left_start == *right_start) {
-            result.prefix.append(left_start->value);
-            left_start++;
-            right_fully_consumed = true;
-        }
-    } else if (left_start == left_end && right_start == right_end && *left_start == *right_start && left_start->type == segment_type::MARKER) {
-        result.prefix.append(right_start->value);
-        left_fully_consumed = true;
-        right_fully_consumed = true;
-    }
-
-    auto eat_segment = [](std::string str, const segment & seg) -> std::string { return std::move(str) + seg.value; };
-
-    bool can_have_text_suffix = left_end->type == segment_type::TEXT && right_end->type == segment_type::TEXT;
-    bool can_have_text_prefix = right_start->type == segment_type::TEXT && left_start->type == segment_type::TEXT;
-
-    std::string remainder_left = std::accumulate(left_start, left_fully_consumed ? left_end : ++left_end, std::string(), eat_segment);
-    std::string remainder_right = std::accumulate(right_start, right_fully_consumed ? right_end : ++right_end, std::string(), eat_segment);
-
-    size_t suffix_len = can_have_text_suffix ? common_suffix_len(remainder_left, remainder_right) : 0;
-    // avoid overlaps between prefix and suffix
-    size_t prefix_len = can_have_text_prefix ? common_prefix_len(remainder_left.substr(0, remainder_left.size() - suffix_len),
-        remainder_right.substr(0, remainder_right.size() - suffix_len)) : 0;
-
-    result.prefix.append(remainder_left.substr(0, prefix_len));
-    result.suffix = remainder_left.substr(remainder_left.length() - suffix_len, suffix_len) + result.suffix;
-    result.left = remainder_left.substr(prefix_len, remainder_left.length() - prefix_len - suffix_len);
-    result.right = remainder_right.substr(prefix_len, remainder_right.length() - prefix_len - suffix_len);
-
-    if (result.left == "" && result.right == "") {
-        // degenerate case, no diff
-        result.prefix = left;
-        result.suffix = "";
-        // pick prefix = all as representation
-    }
-    return result;
-}
-
-// Returns the prefix of `full` up until the first occurrence of the common prefix of `left` and `right`
-std::string until_common_prefix(const std::string & full, const std::string & left, const std::string & right) {
-    // Find the common prefix of left and right
-    size_t common_prefix_len = 0;
-    size_t min_len           = std::min(left.length(), right.length());
-    while (common_prefix_len < min_len && left[common_prefix_len] == right[common_prefix_len]) {
-        common_prefix_len++;
-    }
-
-    // If there's no common prefix, return empty string
-    if (common_prefix_len == 0) {
-        return "";
-    }
-
-    // Find the common prefix in the full string
-    std::string common_prefix = left.substr(0, common_prefix_len);
-    size_t      pos           = full.find(common_prefix);
-
-    // If not found, return empty string
-    if (pos == std::string::npos) {
-        return "";
-    }
-
-    // Return everything before the common prefix
-    return full.substr(0, pos);
-}
-
-// Returns the suffix of `full` after the last occurrence of the common suffix of `left` and `right`
-std::string after_common_suffix(const std::string & full, const std::string & left, const std::string & right) {
-    // Find the common suffix of left and right (compare from the end)
-    size_t common_suffix_len = 0;
-    size_t min_len           = std::min(left.length(), right.length());
-    while (common_suffix_len < min_len &&
-           left[left.length() - 1 - common_suffix_len] == right[right.length() - 1 - common_suffix_len]) {
-        common_suffix_len++;
-    }
-
-    // If there's no common suffix, return empty string
-    if (common_suffix_len == 0) {
-        return "";
-    }
-
-    // Extract the common suffix
-    std::string common_suffix = left.substr(left.length() - common_suffix_len);
-
-    // Find the last occurrence of the common suffix in the full string
-    size_t pos = full.rfind(common_suffix);
-
-    // If not found, return empty string
-    if (pos == std::string::npos) {
-        return "";
-    }
-
-    // Return everything after the common suffix
-    return full.substr(pos + common_suffix_len);
-}
-
-// TODO: segmentize will treat a JSON array inside tags as a tag: <calls>[{ "fun": { ... } }]</calls> will be three markers
-// not too worried about that because it hasn't turned out as a problem anywhere, but noting here in case it will
-// Might have to put some restrictions on tag contents as well (like "no { }")
-std::vector<segment> segmentize_markers(const std::string & text) {
-    std::vector<segment> retval;
-    bool in_marker = false;
-    char marker_opener = '\0';
-
-    auto is_marker_opener = [](char c) -> bool { return c == '<' || c == '['; };
-    auto is_marker_closer = [](char op, char c) -> bool { return (op == '<' && c == '>') || (op == '[' && c == ']'); };
-
-    size_t last_border = 0;
-
-    for (size_t cur_pos = 0; cur_pos < text.length(); cur_pos++) {
-        if (!in_marker && is_marker_opener(text[cur_pos])) {
-            if (last_border < cur_pos) {
-                retval.push_back(segment(segment_type::TEXT, text.substr(last_border, cur_pos - last_border)));
-            }
-            last_border = cur_pos;
-            in_marker = true;
-            marker_opener = text[cur_pos];
-        } else if (in_marker && is_marker_closer(marker_opener, text[cur_pos])) {
-            // no need to check because last_border will always be smaller
-                retval.push_back(segment(segment_type::MARKER, text.substr(last_border, cur_pos - last_border + 1)));
-            last_border = cur_pos + 1;
-            in_marker = false;
-            marker_opener = '\0';
-        }
-    }
-    if (last_border < text.length()) {
-            retval.push_back(segment(segment_type::TEXT, text.substr(last_border)));
-    }
-    return retval;
-}
-
-std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments) {
-    std::vector<segment> result;
-    for (const auto & seg : segments) {
-        if (!trim_whitespace(seg.value).empty()) {
-            result.push_back(seg);
-        }
-    }
-    return result;
-}
-
-namespace autoparser {
-
-std::string apply_template(const common_chat_template & tmpl, const template_params & params) {
-    templates_params tmpl_params;
-    tmpl_params.messages              = params.messages;
-    tmpl_params.tools                 = params.tools;
-    tmpl_params.add_generation_prompt = params.add_generation_prompt;
-    tmpl_params.enable_thinking       = params.enable_thinking;
-
-    if (params.extra_context) {
-        tmpl_params.extra_context = *params.extra_context;
-    }
-    tmpl_params.extra_context["enable_thinking"] = params.enable_thinking;
-
-    try {
-        return common_chat_template_direct_apply(tmpl, tmpl_params);
-    } catch (const std::exception & e) {
-        LOG_DBG("Template application failed: %s\n", e.what());
-        return "";
-    }
-}
-
-std::optional<compare_variants_result> compare_variants(
-    const common_chat_template &                   tmpl,
-    const template_params &                        params_A,
-    const std::function<void(template_params &)> & params_modifier) {
-    // Create variant B by copying A
-    template_params params_B = params_A;
-
-    // Apply modifier to create variant B
-    if (params_modifier) {
-        params_modifier(params_B);
-    }
-
-    // Apply template to both variants
-    std::string output_A = apply_template(tmpl, params_A);
-    std::string output_B = apply_template(tmpl, params_B);
-
-    // Check for template application failures
-    if (output_A.empty() || output_B.empty()) {
-        return std::nullopt;
-    }
-
-    // Calculate diff and return result with both outputs
-    compare_variants_result result;
-    result.diff     = calculate_diff_split(output_A, output_B);
-    result.output_A = output_A;
-    result.output_B = output_B;
-
-    return result;
-}
-
-}  // namespace autoparser
-

common/chat-auto-parser-helpers.h

@@ -1,73 +0,0 @@
-#pragma once
-
-#include "chat-auto-parser.h"
-#include <functional>
-#include <optional>
-#include <string>
-
-std::string trim_whitespace(const std::string & str);
-std::string trim_leading_whitespace(const std::string & str);
-std::string trim_trailing_whitespace(const std::string & str);
-std::string trim_trailing_newlines(const std::string & str);
-
-// calculate a diff split (longest common prefix, longest common suffix excluding prefix,
-// mismatched part on the left, mismatched part on the right) between two strings
-// account for markers - align prefix and suffix endings so that they end on markers
-// * eg.:
-// calculate_diff_split("<html><body><div></div></body></html>", "<html><body><p>Something</p></body><html>") ->
-//  { "prefix": "<html><body>" (not: "<html><body><"), "suffix": "</body></html>", "left": "<div></div>", "right": "<p>Something</p>" }
-// calculate_diff_split("<html><body>Something</body></html>", "<html><body></body><html>") ->
-//  { "prefix": "<html><body>", "suffix": "</body></html>", "left": "Something", "right": "" }
-diff_split calculate_diff_split(const std::string & left, const std::string & right);
-
-// Returns the prefix of `full` up until the first occurrence of the common prefix of `left` and `right`
-// Returns empty string if there's no common prefix
-// * eg.:
-// until_common_prefix("really want a FUNCTION call", "FUNCTION alpha", "FUNCTION beta") -> "really want a "
-// until_common_prefix("<tool_call>", "<something>", "<something_else>") -> ""
-// until_common_prefix("some text", "1234", "abcd") -> ""
-// until_common_prefix("one arg two args three args four", "argument alpha", "argument beta") -> "one ""
-std::string until_common_prefix(const std::string & full, const std::string & left, const std::string & right);
-
-// Returns the suffix of `full` after the last occurrence of the common suffix of `left` and `right`
-// Returns empty string if there's no common suffix
-// Mirror function of `until_common_prefix`
-// * eg.:
-// after_common_suffix("really want a FUNCTION call", "first FUNCTION", "second FUNCTION") -> " call"
-// after_common_suffix("one arg two-args three args four", "alpha-args", "beta-args") -> " three args four"
-std::string after_common_suffix(const std::string & full, const std::string & left, const std::string & right);
-
-// Segmentize text into markers and non-marker fragments
-// * eg.:
-// segmentize_markers("<html><head><title>The site title</title><body><div>Here's some <b>content</b></div></body></html>" ->
-//  [ (MARKER, "<html>"), (MARKER, "<head>"), (MARKER, "<title>"), (TEXT, "The site title"), (MARKER, "</title>"),
-//    (MARKER, "<body>"), (MARKER, "<div>"), (TEXT, "Here's some "), (MARKER, "<b>"), (TEXT, "content"), (MARKER, "</b>"),
-//    (MARKER, "</div>"), (MARKER, "</body>"), (MARKER, "</html>")
-//  ]
-// segmentize_markers("<|tool_call|>[args]{ are here }[/args]<|tool_call_end|>") ->
-//  [ (MARKER, "<|tool_call|>"), (MARKER, "[args]"), (TEXT, "{ are here }"), (MARKER, "[/args]"), (MARKER, "<|tool_call_end|>") ]
-std::vector<segment> segmentize_markers(const std::string & text);
-
-// Prune whitespace-only segments from a vector of segments
-// * eg.:
-// segmentize_markers("<tool_call>\n<function=foo>\n<arg=bar>\n   \n</arg>\n</function>\n</tool_call>") ->
-//  X = [ (MARKER, "<tool_call>"), (TEXT, "\n"), (MARKER, "<function=foo>"), (TEXT, "\n"), (MARKER, "<arg=bar>"), (TEXT, "\n   \n"),
-//        (MARKER, "</arg>"), (TEXT, "\n"), (MARKER, "</function>"), (TEXT, "\n"), (MARKER, "</tool_call>") ]
-// prune_whitespace_segments(X) -> [ (MARKER, "<tool_call>"), (MARKER, "<function=foo>"), (MARKER, "<arg=bar>"), (MARKER, "</arg>"),
-//                                   (MARKER, "</function>"), (MARKER, "</tool_call>") ]
-std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments);
-
-namespace autoparser {
-
-// Apply a template with the given parameters, returning the rendered string (empty on failure)
-std::string apply_template(const common_chat_template & tmpl, const template_params & params);
-
-// Factorized differential comparison function
-// Takes base params and a single modifier lambda to create variant B
-// Returns compare_variants_result containing diff and both outputs, or std::nullopt on failure
-std::optional<compare_variants_result> compare_variants(
-    const common_chat_template &                   tmpl,
-    const template_params &                        params_A,
-    const std::function<void(template_params &)> & params_modifier);
-
-}  // namespace autoparser

common/chat-auto-parser.h

@@ -1,433 +0,0 @@
-#pragma once
-
-#include "chat.h"
-#include "common.h"
-#include "jinja/caps.h"
-#include "peg-parser.h"
-
-#include <chrono>
-#include <optional>
-#include <string>
-#include <utility>
-#include <vector>
-
-using json = nlohmann::ordered_json;
-
-class common_chat_peg_builder;
-
-// ============================================================================
-// Parameters for template application (low-level, used by diff analysis)
-// ============================================================================
-struct template_params {
-    json                messages;
-    json                tools;
-    bool                add_generation_prompt = false;
-    bool                enable_thinking       = true;
-    std::optional<json> extra_context         = std::nullopt;
-};
-
-struct diff_split {
-    std::string prefix;
-    std::string suffix;
-    std::string left;
-    std::string right;
-
-    bool operator==(struct diff_split & other) const {
-        return prefix == other.prefix && suffix == other.suffix && left == other.left && right == other.right;
-    }
-};
-
-// Result of compare_variants containing diff and original outputs
-struct compare_variants_result {
-    diff_split  diff;
-    std::string output_A;
-    std::string output_B;
-};
-
-namespace autoparser {
-
-// ============================================================================
-// High-level params for parser generation
-// ============================================================================
-
-struct templates_params {
-    json                                  messages;
-    json                                  tools;
-    common_chat_tool_choice               tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
-    json                                  json_schema;
-    bool                                  parallel_tool_calls = true;
-    common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_AUTO;
-    bool                                  stream              = true;
-    std::string                           grammar;
-    bool                                  add_generation_prompt = false;
-    bool                                  enable_thinking       = true;
-    std::chrono::system_clock::time_point now                   = std::chrono::system_clock::now();
-    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
-};
-
-// ============================================================================
-// Analysis Result Enums
-// ============================================================================
-
-// Reasoning handling mode (derived from R1-R3 comparisons)
-enum class reasoning_mode {
-    NONE,           // No reasoning markers detected
-    TAG_BASED,      // Standard tag-based: <think>...</think>
-    DELIMITER,      // Delimiter-based: [BEGIN FINAL RESPONSE] (reasoning ends at delimiter)
-    FORCED_OPEN,    // Template ends with open reasoning tag (empty start, non-empty end)
-    FORCED_CLOSED,  // Template ends with open reasoning tag on enabled thinking but
-                    // with both opened and closed tag for disabled thinking
-    TOOLS_ONLY      // Only reason on tool calls, not on normal content
-};
-
-inline std::ostream & operator<<(std::ostream & os, const reasoning_mode & mode) {
-    switch (mode) {
-        case reasoning_mode::NONE:
-            return os << "NONE";
-        case reasoning_mode::TAG_BASED:
-            return os << "TAG_BASED";
-        case reasoning_mode::DELIMITER:
-            return os << "DELIMITER";
-        case reasoning_mode::FORCED_OPEN:
-            return os << "FORCED_OPEN";
-        case reasoning_mode::FORCED_CLOSED:
-            return os << "FORCED_CLOSED";
-        case reasoning_mode::TOOLS_ONLY:
-            return os << "TOOLS_ONLY";
-        default:
-            return os << "UNKNOWN";
-    }
-}
-
-// Content wrapping mode (derived from C1 comparison)
-enum class content_mode {
-    PLAIN,                   // No content markers
-    ALWAYS_WRAPPED,          // Content always wrapped with markers
-    WRAPPED_WITH_REASONING,  // Content wrapped only when reasoning present
-};
-
-inline std::ostream & operator<<(std::ostream & os, const content_mode & mode) {
-    switch (mode) {
-        case content_mode::PLAIN:
-            return os << "PLAIN";
-        case content_mode::ALWAYS_WRAPPED:
-            return os << "ALWAYS_WRAPPED";
-        case content_mode::WRAPPED_WITH_REASONING:
-            return os << "WRAPPED_WITH_REASONING";
-        default:
-            return os << "UNKNOWN";
-    }
-}
-
-// Call ID position in tool calls (for non-JSON formats)
-enum class call_id_position {
-    NONE,                   // No call ID support detected
-    PRE_FUNC_NAME,          // Call ID before function name: [CALL_ID]id[FUNC]name{args}
-    BETWEEN_FUNC_AND_ARGS,  // Call ID between function and args: [FUNC]name[CALL_ID]id{args}
-    POST_ARGS,              // Call ID after arguments: [FUNC]name{args}[CALL_ID]id
-};
-
-inline std::ostream & operator<<(std::ostream & os, const call_id_position & pos) {
-    switch (pos) {
-        case call_id_position::NONE:
-            return os << "NONE";
-        case call_id_position::PRE_FUNC_NAME:
-            return os << "PRE_FUNC_NAME";
-        case call_id_position::BETWEEN_FUNC_AND_ARGS:
-            return os << "BETWEEN_FUNC_AND_ARGS";
-        case call_id_position::POST_ARGS:
-            return os << "POST_ARGS";
-        default:
-            return os << "UNKNOWN";
-    }
-}
-
-// Tool call format classification (derived from T1-T5, A1-A3 comparisons)
-enum class tool_format {
-    NONE,             // No tool support detected
-    JSON_NATIVE,      // Pure JSON: {"name": "X", "arguments": {...}}
-    TAG_WITH_JSON,    // Tag-based with JSON args: <function=X>{...}</function>
-    TAG_WITH_TAGGED,  // Tag-based with tagged args: <param=key>value</param>
-};
-
-inline std::ostream & operator<<(std::ostream & os, const tool_format & format) {
-    switch (format) {
-        case tool_format::NONE:
-            return os << "NONE";
-        case tool_format::JSON_NATIVE:
-            return os << "JSON_NATIVE";
-        case tool_format::TAG_WITH_JSON:
-            return os << "TAG_WITH_JSON";
-        case tool_format::TAG_WITH_TAGGED:
-            return os << "TAG_WITH_TAGGED";
-        default:
-            return os << "UNKNOWN";
-    }
-}
-
-// ============================================================================
-// Sub-structs for tool analysis
-// ============================================================================
-
-struct tool_format_analysis {
-    tool_format mode = tool_format::NONE;
-
-    std::string section_start;   // e.g., "<tool_call>", "[TOOL_CALLS]", ""
-    std::string section_end;     // e.g., "</tool_call>", ""
-    std::string per_call_start;  // e.g., "<|tool_call_begin|>", "" (for multi-call templates)
-    std::string per_call_end;    // e.g., "<|tool_call_end|>", ""
-
-    bool fun_name_is_key = false;       // In JSON format function name is JSON key, i.e. { "<funname>": { ... arguments ... } }
-    bool tools_array_wrapped = false;   // Tool calls wrapped in JSON array [...]
-    bool uses_python_dicts = false;     // Tool call args use Python dict format (single-quoted strings)
-
-    std::string              function_field = "function";
-    std::string              name_field     = "name";
-    std::string              args_field     = "arguments";
-    std::string              id_field;
-    std::string              gen_id_field;
-    std::vector<std::string> parameter_order;
-};
-
-struct tool_function_analysis {
-    std::string name_prefix;  // e.g., "<function=", "\"name\": \"", "functions."
-    std::string name_suffix;  // e.g., ">", "\"", ":0"
-    std::string close;        // e.g., "</function>", "" (for tag-based)
-};
-
-struct tool_arguments_analysis {
-    std::string start;          // e.g., "<|tool_call_argument_begin|>", "<args>"
-    std::string end;            // e.g., "<|tool_call_argument_end|>", "</args>"
-    std::string name_prefix;   // e.g., "<param=", "<arg_key>", "\""
-    std::string name_suffix;   // e.g., ">", "</arg_key>", "\":"
-    std::string value_prefix;  // e.g., "", "<arg_value>", ""
-    std::string value_suffix;  // e.g., "</param>", "</arg_value>", ""
-    std::string separator;     // e.g., "", "\n", ","
-};
-
-struct tool_id_analysis {
-    call_id_position pos = call_id_position::NONE;
-
-    std::string prefix;  // e.g., "[CALL_ID]" (marker before call ID value)
-    std::string suffix;  // e.g., "" (marker after call ID value, before next section)
-};
-
-// ============================================================================
-// Parser build context (shared interface for build_parser methods)
-// ============================================================================
-
-struct analyze_content;
-
-struct parser_build_context {
-    common_chat_peg_builder & p;
-    const templates_params &          inputs;
-    common_peg_parser                 reasoning_parser;
-    bool                              extracting_reasoning = false;
-    const analyze_content *           content              = nullptr;
-
-    parser_build_context(common_chat_peg_builder & p, const templates_params & inputs);
-};
-
-// ============================================================================
-// Base class for analyzers with parser building
-// ============================================================================
-
-struct analyze_base {
-    virtual ~analyze_base() = default;
-    virtual common_peg_parser build_parser(parser_build_context & ctx) const = 0;
-
-  protected:
-    const common_chat_template * tmpl = nullptr;
-
-    analyze_base() = default;
-    explicit analyze_base(const common_chat_template & tmpl) : tmpl(&tmpl) {}
-};
-
-// ============================================================================
-// Reasoning analyzer
-// ============================================================================
-
-struct analyze_reasoning : analyze_base {
-    reasoning_mode mode = reasoning_mode::NONE;
-
-    std::string start;  // e.g., "<think>", "[THINK]", "<|START_THINKING|>", ""
-    std::string end;    // e.g., "</think>", "[BEGIN FINAL RESPONSE]", "<|END_THINKING|>"
-
-    analyze_reasoning() = default;
-    analyze_reasoning(const common_chat_template & tmpl, bool supports_tools);
-
-    common_peg_parser build_parser(parser_build_context & ctx) const override;
-
-  private:
-    // Look for reasoning markers in rendered content
-    void compare_reasoning_presence();
-
-    // Compare generation prompt with enable_thinking=true vs false
-    void compare_thinking_enabled();
-
-    // Check if reasoning is always possible or only in tool calls
-    void compare_reasoning_scope();
-};
-
-// ============================================================================
-// Content analyzer
-// ============================================================================
-
-struct analyze_content : analyze_base {
-    content_mode mode = content_mode::PLAIN;
-
-    std::string start;  // e.g., "<response>", ">>>all\n", ""
-    std::string end;    // e.g., "</response>", ""
-
-    bool requires_nonnull_content = false;
-
-    analyze_content() = default;
-    analyze_content(const common_chat_template & tmpl, const analyze_reasoning & reasoning);
-
-    common_peg_parser build_parser(parser_build_context & ctx) const override;
-
-    bool is_always_wrapped() const;
-    common_peg_parser build_optional_wrapped(parser_build_context & ctx) const;
-};
-
-// ============================================================================
-// Tool analyzer
-// ============================================================================
-
-struct analyze_tools : analyze_base {
-    tool_format_analysis    format;
-    tool_function_analysis  function;
-    tool_arguments_analysis arguments;
-    tool_id_analysis        call_id;
-
-    analyze_tools() = default;
-    analyze_tools(const common_chat_template & tmpl,
-                  const jinja::caps &          caps,
-                  const analyze_reasoning &    reasoning);
-
-    common_peg_parser build_parser(parser_build_context & ctx) const override;
-
-  private:
-    // Extract tool calling 'haystack' for further analysis and delegate further analysis based on format
-    void analyze_tool_calls(const analyze_reasoning & reasoning);
-
-    // Analyze format based on position of function and argument name in needle
-    void analyze_tool_call_format(const std::string &       haystack,
-                                  const std::string &       fun_name_needle,
-                                  const std::string &       arg_name_needle,
-                                  const analyze_reasoning & reasoning);
-
-    // Analyze specifics of JSON native format (entire tool call is a JSON object)
-    void analyze_tool_call_format_json_native(const std::string & clean_haystack,
-                                              const std::string & fun_name_needle,
-                                              const std::string & arg_name_needle);
-
-    // Analyze specifics of non-JSON native format (tags for function name or for function name and arguments)
-    void analyze_tool_call_format_non_json(const std::string & clean_haystack,
-                                           const std::string & fun_name_needle);
-
-    // Check for and extract specific per-call markers for non-native-JSON templates with parallel call support
-    void check_per_call_markers();
-
-    // Extract function name markers
-    void extract_function_markers();
-
-    // Delegates to separate functions for: separator analysis, argument name analysis, argument value analysis
-    void analyze_arguments();
-
-    // Extract argument name markers
-    void extract_argument_name_markers();
-
-    // Extract argument value markers
-    void extract_argument_value_markers();
-
-    // Extract argument separator, if specified (eg. <arg=foo>...</arg><sep><arg=bar>...</arg>)
-    void extract_argument_separator();
-
-    // Extract argument wrapper markers, if present (eg. '<args><arg=foo>...</arg><arg=bar>...</arg></args>')
-    void extract_args_markers();
-
-    // Extract call ID markers, if present
-    void extract_call_id_markers();
-
-    // Per-format tool parser builders
-    common_peg_parser build_tool_parser_json_native(parser_build_context & ctx) const;
-    common_peg_parser build_tool_parser_tag_json(parser_build_context & ctx) const;
-    common_peg_parser build_tool_parser_tag_tagged(parser_build_context & ctx) const;
-};
-
-// ============================================================================
-// Main autoparser class
-// ============================================================================
-
-struct autoparser {
-    jinja::caps          jinja_caps;
-    analyze_reasoning    reasoning;
-    analyze_content      content;
-    analyze_tools        tools;
-    bool                 analysis_complete = false;
-
-    // Preserved tokens for tokenizer (union of all non-empty markers)
-    std::vector<std::string> preserved_tokens;
-
-    autoparser() = default;
-
-    // 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 templates_params & inputs) const;
-
-  private:
-    // Collect tokens from entire analysis to preserve
-    void collect_preserved_tokens();
-};
-
-// ============================================================================
-// Parser generator
-// ============================================================================
-
-class peg_generator {
-  public:
-    static common_chat_params generate_parser(const common_chat_template &    tmpl,
-                                              const struct templates_params & inputs);
-
-    static common_chat_params generate_parser(const common_chat_template &    tmpl,
-                                              const struct templates_params & inputs,
-                                              const autoparser &              autoparser);
-};
-
-}  // namespace autoparser
-
-enum segment_type { TEXT, MARKER };
-
-inline std::ostream & operator<<(std::ostream & os, const segment_type & type) {
-    switch (type) {
-        case segment_type::TEXT:
-            return os << "TEXT";
-        case segment_type::MARKER:
-            return os << "MARKER";
-        default:
-            return os << "UNKNOWN";
-    }
-}
-
-struct segment {
-    segment_type type;
-    std::string  value;
-
-    segment(segment_type type, std::string value) : type(type), value(std::move(value)) {}
-
-    bool operator==(const segment & other) const {
-        return type == other.type && value == other.value;
-    }
-
-    bool operator!=(const segment & other) const {
-        return !(*this == other);
-    }
-};

common/chat-parser.cpp

@@ -0,0 +1,393 @@
+#include "chat-parser.h"
+#include "common.h"
+#include "log.h"
+#include "regex-partial.h"
+
+#include <optional>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+    : input_(input), is_partial_(is_partial), syntax_(syntax)
+{
+    result_.role = "assistant";
+
+    while (true) {
+        std::string id = std::to_string(std::rand());
+        if (input.find(id) == std::string::npos) {
+            healing_marker_ = id;
+            break;
+        }
+    }
+}
+
+std::string common_chat_msg_parser::str(const common_string_range & rng) const {
+    GGML_ASSERT(rng.begin <= rng.end);
+    return input_.substr(rng.begin, rng.end - rng.begin);
+}
+
+void common_chat_msg_parser::add_content(const std::string &content) {
+    result_.content += content;
+}
+
+void common_chat_msg_parser::add_reasoning_content(const std::string &reasoning_content) {
+    result_.reasoning_content += reasoning_content;
+}
+
+bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::string & id, const std::string & arguments) {
+    if (name.empty()) {
+        return false;
+    }
+
+    common_chat_tool_call tool_call;
+    tool_call.name = name;
+    tool_call.arguments = arguments;
+    tool_call.id = id;
+
+    // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
+    result_.tool_calls.emplace_back(tool_call);
+
+    return true;
+}
+bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
+    std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
+    std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
+    std::string arguments = "";
+    if (tool_call.contains("arguments")) {
+        if (tool_call.at("arguments").is_object()) {
+            arguments = tool_call.at("arguments").dump();
+        } else {
+            arguments = tool_call.at("arguments");
+        }
+    }
+
+    return add_tool_call(name, id, arguments);
+}
+
+bool common_chat_msg_parser::add_tool_calls(const json & arr) {
+    for (const auto & item : arr) {
+        if (!add_tool_call(item)) {
+            return false;
+        }
+    }
+    return true;
+}
+void common_chat_msg_parser::finish() {
+    if (!is_partial_ && pos_ != input_.size()) {
+        throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));
+    }
+}
+
+bool common_chat_msg_parser::consume_spaces() {
+    const auto length = input_.size();
+    auto consumed = false;
+    while (pos_ < length && std::isspace(input_[pos_])) {
+        ++pos_;
+        consumed = true;
+    }
+    return consumed;
+}
+
+bool common_chat_msg_parser::try_consume_literal(const std::string & literal) {
+    auto pos = pos_;
+    for (auto i = 0u; i < literal.size(); ++i) {
+        if (pos >= input_.size()) {
+            return false;
+        }
+        if (input_[pos] != literal[i]) {
+            return false;
+        }
+        ++pos;
+    }
+    pos_ = pos;
+    return true;
+}
+
+std::optional<common_chat_msg_parser::find_regex_result>  common_chat_msg_parser::try_find_literal(const std::string & literal) {
+    auto idx = input_.find(literal, pos_);
+    if (idx != std::string::npos) {
+        find_regex_result res;
+        res.prelude = input_.substr(pos_, idx - pos_);
+        auto end = idx + literal.size();
+        res.groups.emplace_back(common_string_range{idx, end});
+        move_to(end);
+        return res;
+    }
+    if (is_partial_) {
+        idx = string_find_partial_stop(input_, literal);
+        if (idx != std::string::npos && idx >= pos_) {
+            find_regex_result res;
+            res.prelude = input_.substr(pos_, idx - pos_);
+            auto end = input_.size();
+            res.groups.emplace_back(common_string_range{idx, end});
+            move_to(end);
+            return res;
+        }
+    }
+    return std::nullopt;
+}
+
+void common_chat_msg_parser::consume_literal(const std::string & literal) {
+    if (!try_consume_literal(literal)) {
+        throw common_chat_msg_partial_exception(literal);
+    }
+}
+
+bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
+    auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
+        auto stripped_reasoning = string_strip(reasoning);
+        if (stripped_reasoning.empty()) {
+            return;
+        }
+        if (syntax_.reasoning_in_content) {
+            add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "<think>" : start_think);
+            add_content(stripped_reasoning);
+            if (closed) {
+                add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
+            }
+        } else {
+            add_reasoning_content(stripped_reasoning);
+        }
+    };
+    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
+        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
+            if (auto res = try_find_literal(end_think)) {
+                handle_reasoning(res->prelude, /* closed */ true);
+                consume_spaces();
+                return true;
+            }
+            auto rest = consume_rest();
+            if (!rest.empty()) {
+                handle_reasoning(rest, /* closed */ !is_partial());
+            }
+            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
+            // if (!syntax_.thinking_forced_open) {
+            //     throw common_chat_msg_partial_exception(end_think);
+            // }
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string common_chat_msg_parser::consume_rest() {
+    auto rest = input_.substr(pos_);
+    pos_ = input_.size();
+    return rest;
+}
+
+// Tries to find the regex, consumes it (pos right after it) and gives the prelude (right before it) and the groups to the callback.
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from, bool add_prelude_to_content) {
+    auto m = regex.search(input_, from == std::string::npos ? pos_ : from);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
+    pos_ = m.groups[0].end;
+
+    if (add_prelude_to_content) {
+        add_content(prelude);
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    return find_regex_result{prelude, m.groups};
+}
+
+common_chat_msg_parser::find_regex_result common_chat_msg_parser::consume_regex(const common_regex & regex) {
+    if (auto result = try_consume_regex(regex)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception(regex.str());
+}
+
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_consume_regex(const common_regex & regex) {
+    auto m = regex.search(input_, pos_);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    if (m.groups[0].begin != pos_) {
+        // Didn't match at the current position.
+        return std::nullopt;
+    }
+    pos_ = m.groups[0].end;
+
+    return find_regex_result {
+        /* .prelude = */ "",
+        m.groups,
+    };
+}
+
+std::optional<common_json> common_chat_msg_parser::try_consume_json() {
+    auto it = input_.cbegin() + pos_;
+    const auto end = input_.cend();
+    common_json result;
+    if (!common_json_parse(it, end, healing_marker_, result)) {
+        return std::nullopt;
+    }
+    pos_ = std::distance(input_.cbegin(), it);
+    if (result.healing_marker.marker.empty()) {
+        // No healing marker, just return the parsed json
+        return result;
+    }
+    if (!is_partial()) {
+        throw common_chat_msg_partial_exception("JSON");
+    }
+    return result;
+}
+
+common_json common_chat_msg_parser::consume_json() {
+    if (auto result = try_consume_json()) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+common_chat_msg_parser::consume_json_result common_chat_msg_parser::consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    if (auto result = try_consume_json_with_dumped_args(args_paths, content_paths)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parser::try_consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    auto partial = try_consume_json();
+    if (!partial) {
+        return std::nullopt;
+    }
+    auto is_arguments_path = [&](const std::vector<std::string> & path) {
+        return std::find(args_paths.begin(), args_paths.end(), path) != args_paths.end();
+    };
+    auto is_content_path = [&](const std::vector<std::string> & path) {
+        return std::find(content_paths.begin(), content_paths.end(), path) != content_paths.end();
+    };
+
+    if (partial->healing_marker.marker.empty()) {
+        if (args_paths.empty()) {
+            // No arguments to dump, and JSON was parsed fully.
+            return consume_json_result {
+                partial->json,
+                /* .is_partial = */ false,
+            };
+        }
+        if (is_arguments_path({})) {
+            // Entire JSON is the arguments and was parsed fully.
+            return consume_json_result {
+                partial->json.dump(),
+                /* .is_partial = */ false,
+            };
+        }
+    }
+
+    LOG_DBG("Parsed partial JSON: %s (json_healing_marker: %s)\n", partial->json.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+
+    auto found_healing_marker = false;
+    std::vector<std::string> path;
+    std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
+        if (is_arguments_path(path)) {
+            auto arguments = j.dump();
+            if (is_partial() && !partial->healing_marker.marker.empty()) {
+                auto idx = arguments.find(partial->healing_marker.json_dump_marker);
+                if (idx != std::string::npos) {
+                    arguments.resize(idx);
+                    found_healing_marker = true;
+                }
+                if (arguments == "\"") {
+                    // This happens because of completing `:"$magic` after `"arguments"`
+                    arguments = "";
+                }
+            }
+            return arguments;
+        }
+        if (is_content_path(path)) {
+            if (!j.is_string()) {
+                throw std::runtime_error("Content path must be a string");
+            }
+            std::string str = j;
+            auto idx = str.find(partial->healing_marker.marker); // not using json_dump_marker as we're inside a string
+            if (idx != std::string::npos) {
+                str.resize(idx);
+                found_healing_marker = true;
+            }
+            return str;
+        }
+        if (j.is_object()) {
+            auto obj = json::object();
+            for (const auto & p : j.items()) {
+                const auto & key = p.key();
+                const auto & value = p.value();
+                const std::string key_str = key; // NOLINT
+                auto idx = key_str.find(healing_marker_);
+                if (idx != std::string::npos) {
+                    found_healing_marker = true;
+                    break;
+                }
+                path.push_back(key_str);
+                if (value.is_string()) {
+                    const std::string value_str = value;
+                    if (value_str.find(healing_marker_) != std::string::npos) {
+                        found_healing_marker = true;
+                        if (is_content_path(path)) {
+                            if (partial->healing_marker.marker == partial->healing_marker.json_dump_marker) {
+                                // The healing occurred inside the string: good. Otherwise we just ditch the entire key/value pair.
+                                obj[key] = remove_unsupported_healings_and_dump_args(value);
+                            }
+                        }
+                        break;
+                    }
+                    obj[key] = value;
+                } else {
+                    obj[key] = remove_unsupported_healings_and_dump_args(value);
+                }
+                path.pop_back();
+            }
+            return obj;
+        }
+        if (j.is_array()) {
+            auto arr = json::array();
+            for (const auto & value : j) {
+                if (value.is_string()) {
+                    std::string str = value;
+                    auto idx = str.find(healing_marker_);
+                    if (idx != std::string::npos) {
+                        // Don't heal array values that aren't in the arguments.
+                        found_healing_marker = true;
+                        break;
+                    }
+                }
+                arr.push_back(remove_unsupported_healings_and_dump_args(value));
+            }
+            return arr;
+        }
+        return j;
+    };
+
+    auto cleaned = remove_unsupported_healings_and_dump_args(partial->json);
+    LOG_DBG("Cleaned up JSON %s to %s (json_healing_marker : '%s')\n", partial->json.dump().c_str(), cleaned.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+    return consume_json_result {
+        cleaned,
+        /* .is_partial = */ found_healing_marker,
+    };
+}
+
+void common_chat_msg_parser::clear_tools() {
+    result_.tool_calls.clear();
+}

common/chat-parser.h

@@ -0,0 +1,120 @@
+#pragma once
+
+#include "chat.h"
+#include "json-partial.h"
+#include "regex-partial.h"
+
+#include <nlohmann/json.hpp>
+
+#include <optional>
+#include <string>
+#include <vector>
+
+class common_chat_msg_partial_exception : public std::runtime_error {
+  public:
+    common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
+};
+
+class common_chat_msg_parser {
+    std::string input_;
+    bool is_partial_;
+    common_chat_syntax syntax_;
+    std::string healing_marker_;
+
+    size_t pos_ = 0;
+    common_chat_msg result_;
+
+  public:
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    const std::string & input() const { return input_; }
+    size_t pos() const { return pos_; }
+    const std::string & healing_marker() const { return healing_marker_; }
+    const bool & is_partial() const { return is_partial_; }
+    const common_chat_msg & result() const { return result_; }
+    const common_chat_syntax & syntax() const { return syntax_; }
+
+    void move_to(size_t pos) {
+        if (pos > input_.size()) {
+            throw std::runtime_error("Invalid position!");
+        }
+        pos_ = pos;
+    }
+    void move_back(size_t n) {
+        if (pos_ < n) {
+            throw std::runtime_error("Can't move back that far!");
+        }
+        pos_ -= n;
+    }
+
+    // Get the substring of the input at the given range
+    std::string str(const common_string_range & rng) const;
+
+    // Appends to the result.content field
+    void add_content(const std::string & content);
+
+    // Appends to the result.reasoning_content field
+    void add_reasoning_content(const std::string & reasoning_content);
+
+    // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
+    bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
+
+    // Adds a tool call using the "name", "id" and "arguments" fields of the json object
+    bool add_tool_call(const nlohmann::ordered_json & tool_call);
+
+    // Adds an array of tool calls using their "name", "id" and "arguments" fields.
+    bool add_tool_calls(const nlohmann::ordered_json & arr);
+
+    void finish();
+
+    bool consume_spaces();
+
+    void consume_literal(const std::string & literal);
+
+    bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
+
+    std::string consume_rest();
+
+    struct find_regex_result {
+        std::string prelude;
+        std::vector<common_string_range> groups;
+    };
+
+    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
+
+    bool try_consume_literal(const std::string & literal);
+
+    std::optional<find_regex_result> try_find_literal(const std::string & literal);
+
+    find_regex_result consume_regex(const common_regex & regex);
+
+    std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
+
+    std::optional<common_json> try_consume_json();
+    common_json consume_json();
+
+    struct consume_json_result {
+        nlohmann::ordered_json value;
+        bool is_partial;
+    };
+
+    /*
+        Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
+
+        By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
+        e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
+
+        But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
+        - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
+        - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
+    */
+    consume_json_result consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+    std::optional<consume_json_result> try_consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+
+    void clear_tools();
+};

common/chat-peg-parser.cpp

@@ -1,839 +0,0 @@
-#include "chat-peg-parser.h"
-
-#include "chat-auto-parser.h"
-#include "ggml.h"
-#include "peg-parser.h"
-
-#include <nlohmann/json.hpp>
-
-using json = nlohmann::ordered_json;
-
-static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
-    int count = 0;
-    while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.back()))) {
-        if (max != -1 && count >= max) {
-            break;
-        }
-        sv.remove_suffix(1);
-        count++;
-    }
-    return sv;
-}
-
-static std::string_view trim_leading_space(std::string_view sv, int max = -1) {
-    int count = 0;
-    while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.front()))) {
-        if (max != -1 && count >= max) {
-            break;
-        }
-        sv.remove_prefix(1);
-        count++;
-    }
-    return sv;
-}
-
-static std::string_view trim(std::string_view sv) {
-    return trim_trailing_space(trim_leading_space(sv, 1));
-}
-
-// Count the number of unclosed '{' braces in a JSON-like string,
-// properly skipping braces inside quoted strings.
-static int json_brace_depth(const std::string & s) {
-    int  depth     = 0;
-    bool in_string = false;
-    bool escaped   = false;
-    for (char c : s) {
-        if (escaped) {
-            escaped = false;
-            continue;
-        }
-        if (c == '\\' && in_string) {
-            escaped = true;
-            continue;
-        }
-        if (c == '"') {
-            in_string = !in_string;
-            continue;
-        }
-        if (!in_string) {
-            if (c == '{') {
-                depth++;
-            } else if (c == '}') {
-                depth--;
-            }
-        }
-    }
-    return depth;
-}
-
-// JSON-escape a string and return the inner content (without surrounding quotes).
-static std::string escape_json_string_inner(const std::string & s) {
-    std::string escaped = json(s).dump();
-    if (escaped.size() >= 2 && escaped.front() == '"' && escaped.back() == '"') {
-        return escaped.substr(1, escaped.size() - 2);
-    }
-    return escaped;
-}
-
-// Convert Python-style single-quoted strings to JSON double-quoted strings
-// Only converts outer string delimiters, properly handling escape sequences:
-// - {'key': 'value'} -> {"key": "value"}
-// - {'code': 'print(\'hello\')'} -> {"code": "print('hello')"}
-// - {'msg': 'He said "hi"'} -> {"msg": "He said \"hi\""}
-static std::string normalize_quotes_to_json(const std::string & input) {
-    std::string result;
-    result.reserve(input.size() + 16);  // May need extra space for escaping
-
-    bool in_single_quoted = false;
-    bool in_double_quoted = false;
-
-    for (size_t i = 0; i < input.size(); ++i) {
-        char c = input[i];
-
-        // Handle escape sequences
-        if (c == '\\' && i + 1 < input.size()) {
-            char next = input[i + 1];
-
-            if (in_single_quoted) {
-                // Inside a single-quoted string being converted to double quotes
-                if (next == '\'') {
-                    // \' -> ' (escaped single quote becomes unescaped in double-quoted string)
-                    result += '\'';
-                    ++i;
-                    continue;
-                }
-                if (next == '"') {
-                    // \" stays as \" (already escaped, works in double-quoted string)
-                    result += "\\\"";
-                    ++i;
-                    continue;
-                }
-                // Other escapes (\n, \\, etc.): pass through both characters
-                result += c;
-                result += next;
-                ++i;
-                continue;
-            }
-
-            if (in_double_quoted) {
-                // Inside a double-quoted string - pass through escape sequences as-is
-                result += c;
-                result += next;
-                ++i;
-                continue;
-            }
-
-            // Outside any string - just pass through the backslash
-            result += c;
-            continue;
-        }
-
-        // Handle quote characters
-        if (c == '"') {
-            if (in_single_quoted) {
-                // Unescaped double quote inside single-quoted string -> must escape for JSON
-                result += "\\\"";
-            } else {
-                // Double quote as string delimiter or outside strings
-                in_double_quoted = !in_double_quoted;
-                result += c;
-            }
-        } else if (c == '\'') {
-            if (in_double_quoted) {
-                // Single quote inside double-quoted string -> pass through
-                result += c;
-            } else if (in_single_quoted) {
-                // Closing single quote -> convert to double quote
-                in_single_quoted = false;
-                result += '"';
-            } else {
-                // Opening single quote -> convert to double quote
-                in_single_quoted = true;
-                result += '"';
-            }
-        } else {
-            result += c;
-        }
-    }
-
-    return result;
-}
-
-void tag_based_peg_mapper::from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result) {
-    arena.visit(result, [this](const common_peg_ast_node & node) {
-        if (!node.tag.empty()) {
-            tags[node.tag] = std::string(node.text);
-        }
-    });
-}
-
-tagged_parse_result tagged_peg_parser::parse_and_extract(const std::string & input, common_peg_parse_flags extra_flags) const {
-    common_peg_parse_context ctx(input, flags | extra_flags);
-    auto parse_result = arena.parse(ctx);
-
-    tag_based_peg_mapper mapper;
-    mapper.from_ast(ctx.ast, parse_result);
-
-    return { std::move(parse_result), std::move(mapper.tags) };
-}
-
-tagged_parse_result tagged_peg_parser::parse_anywhere_and_extract(const std::string & input) const {
-    if (input.empty()) {
-        return parse_and_extract(input);
-    }
-    for (size_t i = 0; i < input.size(); i++) {
-        common_peg_parse_context ctx(input, flags);
-        auto parse_result = arena.parse(ctx, i);
-        if (parse_result.success() || i == input.size() - 1) {
-            tag_based_peg_mapper mapper;
-            mapper.from_ast(ctx.ast, parse_result);
-            return { std::move(parse_result), std::move(mapper.tags) };
-        }
-    }
-    GGML_ABORT("Should not happen");
-}
-
-tagged_peg_parser build_tagged_peg_parser(
-    const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {
-    common_peg_parser_builder builder;
-    builder.set_root(fn(builder));
-    return { builder.build() };
-}
-
-common_peg_parser common_chat_peg_builder::tag_with_safe_content(const std::string &       tag_name,
-                                                                 const std::string &       marker,
-                                                                 const common_peg_parser & p) {
-    if (marker.empty()) {
-        return zero_or_more(choice({ p, rule(tag_name, content(any())) }));
-    }
-    auto content_chunk = rule(tag_name, content(negate(literal(marker)) + any() + until(marker)));
-    return zero_or_more(choice({ p, content_chunk }));
-}
-
-std::string & common_chat_peg_mapper::args_target() {
-    return (current_tool && !current_tool->name.empty()) ? current_tool->arguments : args_buffer;
-}
-
-void common_chat_peg_mapper::from_ast(const common_peg_ast_arena &    arena,
-                                      const common_peg_parse_result & parse_result_arg) {
-    arena.visit(parse_result_arg, [this](const common_peg_ast_node & node) { map(node); });
-    // Flush any pending tool call that was started but never got a name
-    // This happens during partial parsing when the tool call is incomplete
-    if (pending_tool_call.has_value() && !pending_tool_call->name.empty()) {
-        if (!args_buffer.empty()) {
-            pending_tool_call->arguments = args_buffer;
-        }
-        if (closing_quote_pending && !pending_tool_call->arguments.empty()) {
-            pending_tool_call->arguments += "\"";
-        }
-        result.tool_calls.push_back(pending_tool_call.value());
-        pending_tool_call.reset();
-    }
-}
-
-void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
-    // Handle reasoning/content tags
-    bool is_reasoning = node.tag == common_chat_peg_builder::REASONING;
-    bool is_content   = node.tag == common_chat_peg_builder::CONTENT;
-
-    if (is_reasoning) { // GPT OSS can have more than 1 reasoning block, so concatenate here
-        result.reasoning_content += std::string(node.text);
-    }
-
-    if (is_content) {
-        // Concatenate content from multiple content nodes (e.g., when reasoning markers
-        // are preserved before content markers in reasoning_format=NONE mode)
-        result.content += std::string(node.text);
-    }
-
-    // Handle tool-related tags (supporting both JSON and tagged formats)
-    bool is_tool_open  = node.tag == common_chat_peg_builder::TOOL_OPEN;
-    bool is_tool_close = node.tag == common_chat_peg_builder::TOOL_CLOSE;
-    bool is_tool_name  = node.tag == common_chat_peg_builder::TOOL_NAME;
-    bool is_tool_id    = node.tag == common_chat_peg_builder::TOOL_ID;
-    bool is_tool_args  = node.tag == common_chat_peg_builder::TOOL_ARGS;
-    bool is_arg_open   = node.tag == common_chat_peg_builder::TOOL_ARG_OPEN;
-    bool is_arg_close  = node.tag == common_chat_peg_builder::TOOL_ARG_CLOSE;
-    bool is_arg_name         = node.tag == common_chat_peg_builder::TOOL_ARG_NAME;
-    bool is_arg_value        = node.tag == common_chat_peg_builder::TOOL_ARG_VALUE;
-    bool is_arg_string_value = node.tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE;
-
-    if (is_tool_open) {
-        pending_tool_call     = common_chat_tool_call();
-        current_tool          = &pending_tool_call.value();
-        arg_count             = 0;
-        args_buffer.clear();
-        closing_quote_pending = false;
-    }
-
-    if (is_tool_id && current_tool) {
-        auto text = trim_trailing_space(node.text);
-        if (text.size() >= 2 && text.front() == '"' && text.back() == '"') {
-            text = text.substr(1, text.size() - 2);
-        }
-        current_tool->id = std::string(text);
-    }
-
-    if (is_tool_name && current_tool) {
-        current_tool->name = std::string(trim_trailing_space(node.text));
-        // Now that we have the name, populate the arguments from the buffer
-        if (!args_buffer.empty()) {
-            current_tool->arguments = args_buffer;
-            args_buffer.clear();
-        } else if (current_tool->arguments.empty()) {
-            current_tool->arguments = "{";
-        }
-        // Add the tool call to results so streaming can see it
-        if (pending_tool_call.has_value()) {
-            result.tool_calls.push_back(pending_tool_call.value());
-            pending_tool_call.reset();
-            current_tool = &result.tool_calls.back();
-        }
-    }
-
-    if (is_tool_args && current_tool) {
-        // For JSON format: arguments come as a complete JSON object
-        // For tagged format: built up from individual arg_name/arg_value nodes
-        auto text = trim_trailing_space(node.text);
-        if (!text.empty() && text.front() == '{') {
-            args_target() = std::string(text);
-        }
-    }
-
-    if (is_arg_open) {
-        closing_quote_pending = false;
-    }
-
-    if (is_arg_name && current_tool) {
-        std::string arg_entry;
-        if (arg_count > 0) {
-            arg_entry = ",";
-        }
-        arg_entry += json(trim(node.text)).dump() + ":";
-        ++arg_count;
-
-        auto & target = args_target();
-        if (target.empty()) {
-            target = "{";
-        }
-        target += arg_entry;
-    }
-
-    if ((is_arg_value || is_arg_string_value) && current_tool) {
-        std::string value_content = std::string(trim_trailing_space(trim_leading_space(node.text, 1), 1));
-
-        std::string value_to_add;
-        if (value_content.empty() && is_arg_string_value) {
-            // Empty string value - arg_close will add the closing quote
-            value_to_add          = "\"";
-            closing_quote_pending = true;
-        } else if (!value_content.empty() && is_arg_string_value) {
-            // Schema declares this as string type - always treat as literal string value
-            if (!closing_quote_pending) {
-                value_to_add          = "\"";
-                closing_quote_pending = true;
-            }
-            value_to_add += escape_json_string_inner(value_content);
-        } else if (!value_content.empty()) {
-            // For potential containers, normalize Python-style single quotes to JSON double quotes
-            bool is_potential_container = value_content[0] == '[' || value_content[0] == '{';
-            if (is_potential_container) {
-                value_content = normalize_quotes_to_json(value_content);
-            }
-
-            // Try to parse as JSON value (number, bool, null, object, array)
-            try {
-                json parsed = 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;
-    }
-
-    if (is_arg_close && current_tool) {
-        if (closing_quote_pending) {
-            args_target() += "\"";
-            closing_quote_pending = false;
-        }
-    }
-
-    if (is_tool_close && current_tool) {
-        // Flush buffer to arguments if tool name was never seen
-        if (current_tool->name.empty() && !args_buffer.empty()) {
-            current_tool->arguments = args_buffer;
-            args_buffer.clear();
-        }
-        // Close any pending string quote
-        if (closing_quote_pending) {
-            current_tool->arguments += "\"";
-            closing_quote_pending = false;
-        }
-        // Close any unclosed braces (accounts for nested objects)
-        for (int d = json_brace_depth(current_tool->arguments); d > 0; d--) {
-            current_tool->arguments += "}";
-        }
-        // Add tool call to results if named; otherwise discard
-        if (pending_tool_call.has_value()) {
-            if (!current_tool->name.empty()) {
-                result.tool_calls.push_back(pending_tool_call.value());
-            }
-            pending_tool_call.reset();
-        }
-    }
-}
-
-common_peg_parser common_chat_peg_builder::standard_constructed_tools(
-    const std::map<std::string, std::string> & markers,
-    const nlohmann::json &                     tools,
-    bool                                       parallel_tool_calls,
-    bool                                       force_tool_calls) {
-    if (!tools.is_array() || tools.empty()) {
-        return eps();
-    }
-
-    // Extract markers with defaults
-    auto get_marker = [&markers](const std::string & key, const std::string & default_val = "") -> std::string {
-        auto it = markers.find(key);
-        return it != markers.end() ? it->second : default_val;
-    };
-
-    std::string section_start    = get_marker("tool_call_start_marker", "<tool_call>");
-    std::string section_end      = get_marker("tool_call_end_marker", "</tool_call>");
-    std::string func_opener      = get_marker("function_opener", "<function=");
-    std::string func_name_suffix = get_marker("function_name_suffix", ">");
-    std::string func_closer      = get_marker("function_closer", "</function>");
-    std::string param_key_prefix = get_marker("parameter_key_prefix", "<param=");
-    std::string param_key_suffix = get_marker("parameter_key_suffix", ">");
-    std::string param_closer     = get_marker("parameter_closer", "</param>");
-
-    // Build tool choices for tagged format
-    auto tool_choices = choice();
-
-    for (const auto & tool_def : tools) {
-        if (!tool_def.contains("function")) {
-            continue;
-        }
-        const auto &   function = tool_def.at("function");
-        std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
-
-        // Build argument parsers
-        auto args = eps();
-        if (params.contains("properties") && !params["properties"].empty()) {
-            auto arg_choice = choice();
-            for (const auto & el : params["properties"].items()) {
-                const std::string & prop_name = el.key();
-
-                auto arg_name_parser =
-                    choice({ literal(prop_name), literal("\"" + prop_name + "\""), literal("'" + prop_name + "'") });
-
-                auto arg_rule = tool_arg(tool_arg_open(literal(param_key_prefix)) + tool_arg_name(arg_name_parser) +
-                                         literal(param_key_suffix) + tool_arg_value(until(param_closer)) +
-                                         tool_arg_close(literal(param_closer)));
-                arg_choice |= arg_rule;
-            }
-            args = zero_or_more(arg_choice + space());
-        }
-
-        // Build function parser: <function=name>args</function>
-        auto tool_parser = tool(tool_open(literal(func_opener) + tool_name(literal(name)) + literal(func_name_suffix)) +
-                                space() + tool_args(args) + space() + tool_close(literal(func_closer)));
-
-        tool_choices |= rule("tool-" + name, tool_parser);
-    }
-
-    // Build the section with markers
-    auto section =
-        parallel_tool_calls ?
-            trigger_rule("tool-call", literal(section_start) + space() + one_or_more(tool_choices + space()) +
-                                          literal(section_end)) :
-            trigger_rule("tool-call", literal(section_start) + space() + tool_choices + space() + literal(section_end));
-
-    return force_tool_calls ? section : optional(section);
-}
-
-// Python-style tool calls: name(arg1="value1", arg2=123)
-// Used only by LFM2 for now, so we don't merge it into autoparser
-common_peg_parser common_chat_peg_builder::python_style_tool_calls(
-    const nlohmann::json & tools,
-    bool                   parallel_tool_calls) {
-    if (!tools.is_array() || tools.empty()) {
-        return eps();
-    }
-
-    auto tool_choices = choice();
-
-    for (const auto & tool_def : tools) {
-        if (!tool_def.contains("function")) {
-            continue;
-        }
-        const auto &   function = tool_def.at("function");
-        std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
-
-        auto args = eps();
-        if (params.contains("properties") && !params["properties"].empty()) {
-            auto arg_choice = choice();
-            for (const auto & el : params["properties"].items()) {
-                const std::string & prop_name = el.key();
-                const auto & prop_def = el.value();
-                bool is_string_type = (prop_def.contains("type") && prop_def["type"] == "string");
-
-                auto arg_name_parser = literal(prop_name);
-
-                common_peg_parser arg_value_parser = eps();
-                auto string_value_parser = choice({
-                    literal("\"") + tool_arg_string_value(string_content('"')) + literal("\""),
-                    literal("'") + tool_arg_string_value(string_content('\'')) + literal("'")
-                });
-
-                if (is_string_type) {
-                    arg_value_parser = string_value_parser;
-                } else {
-                    arg_value_parser = tool_arg_value(python_value());
-                }
-
-                // Full argument: name="value" or name=value
-                auto arg_rule = tool_arg(
-                    tool_arg_open(eps()) +
-                    tool_arg_name(arg_name_parser) +
-                    literal("=") +
-                    arg_value_parser +
-                    tool_arg_close(eps())
-                );
-                arg_choice |= arg_rule;
-            }
-
-            args = arg_choice + zero_or_more("," + space() + arg_choice);
-        }
-
-        auto tool_parser = tool(tool_open(tool_name(literal(name)) + literal("(")) +
-            space() + tool_args(args) + space() + tool_close(literal(")"))
-        );
-
-        tool_choices |= rule("tool-" + name, tool_parser);
-    }
-
-    if (parallel_tool_calls) {
-        return "[" + space() + tool_choices + zero_or_more("," + space() + tool_choices) + space() + "]";
-    }
-    return "[" + space() + tool_choices + space() + "]";
-}
-
-// Helper: Parse dot notation key into prefix and field name
-static std::pair<std::string, std::string> parse_key_spec(const std::string & key) {
-    auto dot_pos = key.find('.');
-    if (dot_pos == std::string::npos) {
-        return {"", key};  // Top-level field
-    }
-    return {key.substr(0, dot_pos), key.substr(dot_pos + 1)};
-}
-
-// Mode 1: function_is_key — parse {"function_name": {...}}
-common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
-    const nlohmann::json & tools,
-    const std::string &    args_key,
-    const std::string &    effective_args_key,
-    const std::string &    call_id_key,
-    const std::string &    gen_call_id_key) {
-
-    auto tool_choices = choice();
-
-    for (const auto & tool_def : tools) {
-        if (!tool_def.contains("function")) {
-            continue;
-        }
-        const auto &   function = tool_def.at("function");
-        std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
-
-        // Build inner object fields
-        std::vector<common_peg_parser> inner_fields;
-
-        if (!call_id_key.empty()) {
-            auto id_parser = atomic(
-                literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
-                literal("\"") + tool_id(string_content('"')) + literal("\"")
-            );
-            inner_fields.push_back(optional(id_parser + space() + optional(literal(",") + space())));
-        }
-
-        if (!gen_call_id_key.empty()) {
-            auto gen_id_parser = atomic(
-                literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
-                choice({
-                    literal("\"") + tool_id(string_content('"')) + literal("\""),
-                    tool_id(json_number())
-                })
-            );
-            inner_fields.push_back(optional(gen_id_parser + space() + optional(literal(",") + space())));
-        }
-
-        // Arguments — either wrapped in args_key or parsed directly
-        common_peg_parser args_parser = eps();
-        if (args_key.empty()) {
-            args_parser = tool_args(schema(json(), "tool-" + name + "-schema", params));
-        } else {
-            args_parser = literal("\"" + effective_args_key + "\"") + space() + literal(":") + space() +
-                          tool_args(schema(json(), "tool-" + name + "-schema", params));
-        }
-        inner_fields.push_back(args_parser);
-
-        // Build inner object parser
-        common_peg_parser inner_object = eps();
-        if (args_key.empty() && inner_fields.size() == 1) {
-            inner_object = inner_fields[0];
-        } else {
-            inner_object = literal("{") + space();
-            for (size_t i = 0; i < inner_fields.size(); i++) {
-                inner_object = inner_object + inner_fields[i];
-                if (i < inner_fields.size() - 1) {
-                    inner_object = inner_object + space();
-                }
-            }
-            inner_object = inner_object + space() + literal("}");
-        }
-
-        auto tool_parser = tool(
-            tool_open(literal("{")) + space() +
-            literal("\"") + tool_name(literal(name)) + literal("\"") +
-            space() + literal(":") + space() +
-            inner_object +
-            space() + tool_close(literal("}"))
-        );
-
-        tool_choices |= rule("tool-" + name, tool_parser);
-    }
-
-    return tool_choices;
-}
-
-// Mode 2: Nested keys (dot notation like "function.name")
-common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
-    const nlohmann::json & tools,
-    const std::string &    effective_name_key,
-    const std::string &    effective_args_key,
-    const std::string &    call_id_key,
-    const std::string &    gen_call_id_key) {
-
-    auto tool_choices = choice();
-
-    auto name_spec = parse_key_spec(effective_name_key);
-    auto args_spec = parse_key_spec(effective_args_key);
-
-    std::string nested_prefix     = !name_spec.first.empty() ? name_spec.first  : args_spec.first;
-    std::string nested_name_field = !name_spec.first.empty() ? name_spec.second  : effective_name_key;
-    std::string nested_args_field = !args_spec.first.empty() ? args_spec.second  : effective_args_key;
-
-    for (const auto & tool_def : tools) {
-        if (!tool_def.contains("function")) {
-            continue;
-        }
-        const auto &   function = tool_def.at("function");
-        std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
-
-        auto nested_name = literal("\"" + nested_name_field + "\"") + space() + literal(":") + space() +
-                          literal("\"") + tool_name(literal(name)) + literal("\"");
-        auto nested_args = literal("\"" + nested_args_field + "\"") + space() + literal(":") + space() +
-                          tool_args(schema(json(), "tool-" + name + "-schema", params));
-
-        auto nested_object = literal("{") + space() +
-                            nested_name + space() + literal(",") + space() +
-                            nested_args +
-                            space() + literal("}");
-
-        // Format: { id?, "function": {...} }
-        auto tool_parser_body = tool_open(literal("{")) + space();
-
-        if (!call_id_key.empty()) {
-            auto id_spec = parse_key_spec(call_id_key);
-            if (id_spec.first.empty()) {
-                auto id_parser = atomic(
-                    literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
-                    literal("\"") + tool_id(string_content('"')) + literal("\"")
-                );
-                tool_parser_body = tool_parser_body + optional(id_parser + space() + literal(",") + space());
-            }
-        }
-
-        if (!gen_call_id_key.empty()) {
-            auto gen_id_spec = parse_key_spec(gen_call_id_key);
-            if (gen_id_spec.first.empty()) {
-                auto gen_id_parser = atomic(
-                    literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
-                    choice({
-                        literal("\"") + tool_id(string_content('"')) + literal("\""),
-                        tool_id(json_number())
-                    })
-                );
-                tool_parser_body = tool_parser_body + optional(gen_id_parser + space() + literal(",") + space());
-            }
-        }
-
-        auto nested_field = literal("\"" + nested_prefix + "\"") + space() + literal(":") + space() + nested_object;
-        tool_parser_body = tool_parser_body + nested_field + space() + tool_close(literal("}"));
-
-        tool_choices |= rule("tool-" + name, tool(tool_parser_body));
-    }
-
-    return tool_choices;
-}
-
-// Mode 3: Flat keys with optional ID fields and parameter ordering
-common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
-    const nlohmann::json &           tools,
-    const std::string &              effective_name_key,
-    const std::string &              effective_args_key,
-    const std::string &              call_id_key,
-    const std::string &              gen_call_id_key,
-    const std::vector<std::string> & parameters_order) {
-
-    auto tool_choices    = choice();
-    auto name_key_parser = literal("\"" + effective_name_key + "\"");
-    auto args_key_parser = literal("\"" + effective_args_key + "\"");
-
-    for (const auto & tool_def : tools) {
-        if (!tool_def.contains("function")) {
-            continue;
-        }
-        const auto &   function = tool_def.at("function");
-        std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
-
-        auto tool_name_ = name_key_parser + space() + literal(":") + space() +
-                         literal("\"") + tool_name(literal(name)) + literal("\"");
-        auto tool_args_ = args_key_parser + space() + literal(":") + space() +
-                         tool_args(schema(json(), "tool-" + name + "-schema", params));
-
-        // Build ID parsers if keys are provided
-        common_peg_parser id_parser = eps();
-        if (!call_id_key.empty()) {
-            id_parser = atomic(
-                literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
-                choice({
-                    literal("\"") + tool_id(string_content('"')) + literal("\""),
-                    tool_id(json_number())
-                })
-            );
-        }
-
-        common_peg_parser gen_id_parser = eps();
-        if (!gen_call_id_key.empty()) {
-            gen_id_parser = atomic(
-                literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
-                choice({
-                    literal("\"") + tool_id(string_content('"')) + literal("\""),
-                    tool_id(json_number())
-                })
-            );
-        }
-
-        // Create (parser, key) pairs for all fields, then sort by parameters_order
-        std::vector<std::pair<common_peg_parser, std::string>> parser_pairs;
-        parser_pairs.emplace_back(tool_name_, effective_name_key);
-        parser_pairs.emplace_back(tool_args_, effective_args_key);
-        if (!call_id_key.empty()) {
-            parser_pairs.emplace_back(optional(id_parser), call_id_key);
-        }
-        if (!gen_call_id_key.empty()) {
-            parser_pairs.emplace_back(optional(gen_id_parser), gen_call_id_key);
-        }
-
-        std::sort(parser_pairs.begin(), parser_pairs.end(),
-            [&parameters_order](const auto & a, const auto & b) {
-                auto pos_a = std::find(parameters_order.begin(), parameters_order.end(), a.second);
-                auto pos_b = std::find(parameters_order.begin(), parameters_order.end(), b.second);
-                size_t idx_a = (pos_a == parameters_order.end()) ? parameters_order.size() : std::distance(parameters_order.begin(), pos_a);
-                size_t idx_b = (pos_b == parameters_order.end()) ? parameters_order.size() : std::distance(parameters_order.begin(), pos_b);
-                return idx_a < idx_b;
-            });
-
-        auto ordered_body = tool_open(literal("{")) + space();
-        for (size_t i = 0; i < parser_pairs.size(); i++) {
-            ordered_body = ordered_body + parser_pairs[i].first;
-            if (i < parser_pairs.size() - 1) {
-                ordered_body = ordered_body + space() + literal(",") + space();
-            }
-        }
-        ordered_body = ordered_body + space() + tool_close(literal("}"));
-
-        tool_choices |= rule("tool-" + name, tool(ordered_body));
-    }
-
-    return tool_choices;
-}
-
-common_peg_parser common_chat_peg_builder::standard_json_tools(
-                                                       const std::string &              section_start,
-                                                       const std::string &              section_end,
-                                                       const nlohmann::json &           tools,
-                                                       bool                             parallel_tool_calls,
-                                                       bool                             force_tool_calls,
-                                                       const std::string &              name_key,
-                                                       const std::string &              args_key,
-                                                       bool                             array_wrapped,
-                                                       bool                             function_is_key,
-                                                       const std::string &              call_id_key,
-                                                       const std::string &              gen_call_id_key,
-                                                       const std::vector<std::string> & parameters_order) {
-    if (!tools.is_array() || tools.empty()) {
-        return eps();
-    }
-
-    std::string effective_name_key = name_key.empty() ? "name" : name_key;
-    std::string effective_args_key = args_key.empty() ? "arguments" : args_key;
-
-    // Dispatch to the appropriate builder based on the JSON layout mode
-    common_peg_parser tool_choices = eps();
-    if (function_is_key) {
-        tool_choices = build_json_tools_function_is_key(tools, args_key, effective_args_key, call_id_key, gen_call_id_key);
-    } else {
-        auto name_spec = parse_key_spec(effective_name_key);
-        auto args_spec = parse_key_spec(effective_args_key);
-        if (!name_spec.first.empty() || !args_spec.first.empty()) {
-            tool_choices = build_json_tools_nested_keys(tools, effective_name_key, effective_args_key, call_id_key, gen_call_id_key);
-        } else {
-            tool_choices = build_json_tools_flat_keys(tools, effective_name_key, effective_args_key, call_id_key, gen_call_id_key, parameters_order);
-        }
-    }
-
-    // Build the section with markers
-    auto tool_calls = tool_choices;
-    if (parallel_tool_calls) {
-        tool_calls = tool_calls + zero_or_more(space() + literal(",") + space() + tool_choices);
-    }
-
-    if (array_wrapped) {
-        tool_calls = literal("[") + space() + tool_calls + space() + literal("]");
-    }
-
-    auto section =
-        trigger_rule("tool-call", literal(section_start) + space() + tool_calls + space() + literal(section_end));
-
-    return force_tool_calls ? section : optional(section);
-}

common/chat-peg-parser.h

@@ -1,181 +0,0 @@
-#pragma once
-
-#include "chat.h"
-#include "peg-parser.h"
-
-#include <map>
-#include <optional>
-#include <vector>
-
-class common_chat_peg_mapper {
-  public:
-    common_chat_msg & result;
-
-    common_chat_peg_mapper(common_chat_msg & msg) : result(msg) {}
-
-    virtual ~common_chat_peg_mapper() = default;
-
-    virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
-    virtual void map(const common_peg_ast_node & node);
-    private:
-      // Tool call handling state
-      std::optional<common_chat_tool_call> pending_tool_call;  // Tool call waiting for name
-      common_chat_tool_call *              current_tool          = nullptr;
-      int                                  arg_count             = 0;
-      bool                                 closing_quote_pending = false;
-      std::string                          args_buffer;  // Buffer to delay arguments until tool name is known
-
-      // Returns a reference to the active argument destination string.
-      // Before tool_name is known, writes go to args_buffer; after, to current_tool->arguments.
-      std::string & args_target();
-};
-
-struct content_structure;
-struct tool_call_structure;
-
-class common_chat_peg_builder : public common_peg_parser_builder {
-  public:
-    // Tag constants (from former common_chat_peg_base_builder)
-    static constexpr const char * REASONING_BLOCK = "reasoning-block";
-    static constexpr const char * REASONING       = "reasoning";
-    static constexpr const char * CONTENT         = "content";
-
-    // Tag constants
-    static constexpr const char * TOOL           = "tool";
-    static constexpr const char * TOOL_OPEN      = "tool-open";
-    static constexpr const char * TOOL_CLOSE     = "tool-close";
-    static constexpr const char * TOOL_ID        = "tool-id";
-    static constexpr const char * TOOL_NAME      = "tool-name";
-    static constexpr const char * TOOL_ARGS      = "tool-args";
-    static constexpr const char * TOOL_ARG       = "tool-arg";
-    static constexpr const char * TOOL_ARG_OPEN  = "tool-arg-open";
-    static constexpr const char * TOOL_ARG_CLOSE = "tool-arg-close";
-    static constexpr const char * TOOL_ARG_NAME         = "tool-arg-name";
-    static constexpr const char * TOOL_ARG_VALUE        = "tool-arg-value";
-    static constexpr const char * TOOL_ARG_STRING_VALUE = "tool-arg-string-value";  // For schema-declared string types
-
-    // Low-level tag methods (from former common_chat_peg_base_builder)
-    common_peg_parser reasoning_block(const common_peg_parser & p) { return tag(REASONING_BLOCK, p); }
-
-    common_peg_parser reasoning(const common_peg_parser & p) { return tag(REASONING, p); }
-
-    common_peg_parser content(const common_peg_parser & p) { return tag(CONTENT, p); }
-
-    common_peg_parser tag_with_safe_content(const std::string &       tag_name,
-                        const std::string &       marker,
-                        const common_peg_parser & p);
-
-    // Low-level tag methods
-    common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
-    common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
-    common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
-    common_peg_parser tool_id(const common_peg_parser & p) { return atomic(tag(TOOL_ID, p)); }
-    common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
-    common_peg_parser tool_args(const common_peg_parser & p) { return tag(TOOL_ARGS, p); }
-    common_peg_parser tool_arg(const common_peg_parser & p) { return tag(TOOL_ARG, p); }
-    common_peg_parser tool_arg_open(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_OPEN, p)); }
-    common_peg_parser tool_arg_close(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_CLOSE, p)); }
-    common_peg_parser tool_arg_name(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_NAME, p)); }
-    common_peg_parser tool_arg_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
-
-    // 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 atomic(tag(TOOL_ARG_VALUE, p)); }
-
-    // Legacy-compatible helper for building standard JSON tool calls
-    // Used by tests and manual parsers
-    // name_key/args_key: JSON key names for function name and arguments
-    //   Empty or "name"/"arguments" will accept both common variations
-    //   Supports dot notation for nested objects (e.g., "function.name")
-    // array_wrapped: if true, tool calls are wrapped in JSON array [...]
-    // function_is_key: if true, function name is the JSON key (e.g., {"func_name": {...}})
-    // call_id_key: JSON key for string call ID (e.g., "id")
-    // gen_call_id_key: JSON key for generated integer call ID (e.g., "tool_call_id")
-    // parameters_order: order in which JSON fields should be parsed
-    common_peg_parser standard_json_tools(const std::string &              section_start,
-                                          const std::string &              section_end,
-                                          const nlohmann::json &           tools,
-                                          bool                             parallel_tool_calls,
-                                          bool                             force_tool_calls,
-                                          const std::string &              name_key = "",
-                                          const std::string &              args_key = "",
-                                          bool                             array_wrapped = false,
-                                          bool                             function_is_key = false,
-                                          const std::string &              call_id_key = "",
-                                          const std::string &              gen_call_id_key = "",
-                                          const std::vector<std::string> & parameters_order = {});
-
-    // Legacy-compatible helper for building XML/tagged style tool calls
-    // Used by tests and manual parsers
-    common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const nlohmann::json &                     tools,
-                                                 bool                                       parallel_tool_calls,
-                                                 bool                                       force_tool_calls);
-
-    // Helper for Python-style function call format: name(arg1="value1", arg2=123)
-    // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const nlohmann::json & tools,
-                                              bool                   parallel_tool_calls);
-
-  private:
-    // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const nlohmann::json & tools,
-                                                       const std::string &    args_key,
-                                                       const std::string &    effective_args_key,
-                                                       const std::string &    call_id_key,
-                                                       const std::string &    gen_call_id_key);
-
-    common_peg_parser build_json_tools_nested_keys(const nlohmann::json & tools,
-                                                   const std::string &    effective_name_key,
-                                                   const std::string &    effective_args_key,
-                                                   const std::string &    call_id_key,
-                                                   const std::string &    gen_call_id_key);
-
-    common_peg_parser build_json_tools_flat_keys(const nlohmann::json &           tools,
-                                                 const std::string &              effective_name_key,
-                                                 const std::string &              effective_args_key,
-                                                 const std::string &              call_id_key,
-                                                 const std::string &              gen_call_id_key,
-                                                 const std::vector<std::string> & parameters_order);
-};
-
-inline common_peg_arena build_chat_peg_parser(
-  const std::function<common_peg_parser(common_chat_peg_builder & builder)> & fn) {
-  common_chat_peg_builder builder;
-  builder.set_root(fn(builder));
-  return builder.build();
-}
-
-class tag_based_peg_mapper {
-  public:
-    std::map<std::string, std::string> tags;
-
-    void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
-};
-
-struct tagged_parse_result {
-    common_peg_parse_result              result;
-    std::map<std::string, std::string> tags;
-};
-
-struct tagged_peg_parser {
-    common_peg_arena arena;
-    common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_NONE;
-
-    tagged_peg_parser & withDebug() {
-      flags |= COMMON_PEG_PARSE_FLAG_DEBUG;
-      return *this;
-    }
-
-    tagged_peg_parser & withoutDebug() {
-      flags = flags & ~COMMON_PEG_PARSE_FLAG_DEBUG;
-      return *this;
-    }
-
-    tagged_parse_result parse_and_extract(const std::string & input, common_peg_parse_flags extra_flags = COMMON_PEG_PARSE_FLAG_NONE) const;
-    tagged_parse_result parse_anywhere_and_extract(const std::string & input) const;
-};
-
-tagged_peg_parser build_tagged_peg_parser(
-    const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn);
-

common/chat.h

@@ -3,30 +3,14 @@
 #pragma once
 
 #include "common.h"
-#include "jinja/parser.h"
-#include "nlohmann/json_fwd.hpp"
-#include "peg-parser.h"
-#include "jinja/runtime.h"
-#include "jinja/caps.h"
-#include "nlohmann/json.hpp"
-
-#include <chrono>
 #include <functional>
-#include <map>
+#include <chrono>
 #include <string>
 #include <vector>
-
-using chat_template_caps = jinja::caps;
-using json = nlohmann::ordered_json;
-
-#include <nlohmann/json_fwd.hpp>
+#include <map>
 
 struct common_chat_templates;
 
-namespace autoparser {
-struct templates_params;
-}  // namespace autoparser
-
 struct common_chat_tool_call {
     std::string name;
     std::string arguments;
@@ -41,95 +25,26 @@ struct common_chat_msg_content_part {
     std::string type;
     std::string text;
 
-    // TODO @ngxson : no known chat templates support reasoning_content in content parts yet
-    //                this can be useful for models with interleaved thinking (like Kimi-K2)
-    //                if you see any templates explicitly support this, please ping me
-    // std::string reasoning_content;
-
     bool operator==(const common_chat_msg_content_part & other) const {
         return type == other.type && text == other.text;
     }
 };
 
-struct common_chat_template {
-    jinja::program prog;
-    std::string bos_tok;
-    std::string eos_tok;
-    std::string src;
-    chat_template_caps caps;
-
-    common_chat_template(const std::string & src, const std::string & bos_token, const std::string & eos_token) {
-        jinja::lexer lexer;
-        auto lexer_res = lexer.tokenize(src);
-        this->prog = jinja::parse_from_tokens(lexer_res);
-
-        this->src = lexer_res.source;
-        this->bos_tok = bos_token;
-        this->eos_tok = eos_token;
-
-        this->caps = jinja::caps_get(prog);
-        // LOG_INF("%s: caps:\n%s\n", __func__, this->caps.to_string().c_str());
-    }
-
-    const std::string & source() const { return src; }
-    const std::string & bos_token() const { return bos_tok; }
-    const std::string & eos_token() const { return eos_tok; }
-
-    // TODO: this is ugly, refactor it somehow
-    json add_system(const json & messages, const std::string & system_prompt) const {
-        GGML_ASSERT(messages.is_array());
-        auto msgs_copy = messages;
-        if (!caps.supports_system_role) {
-            if (msgs_copy.empty()) {
-                msgs_copy.insert(msgs_copy.begin(), json{
-                    {"role", "user"},
-                    {"content", system_prompt}
-                });
-            } else {
-                auto & first_msg = msgs_copy[0];
-                if (!first_msg.contains("content")) {
-                    first_msg["content"] = "";
-                }
-                first_msg["content"] = system_prompt + "\n\n"
-                    + first_msg["content"].get<std::string>();
-            }
-        } else {
-            if (msgs_copy.empty() || msgs_copy[0].at("role") != "system") {
-                msgs_copy.insert(msgs_copy.begin(), json{
-                    {"role", "system"},
-                    {"content", system_prompt}
-                });
-            } else if (msgs_copy[0].at("role") == "system") {
-                msgs_copy[0]["content"] = system_prompt;
-            }
-        }
-        return msgs_copy;
-    }
-
-    chat_template_caps original_caps() const {
-        return caps;
-    }
-
-};
-
 struct common_chat_msg {
-    std::string                               role;
-    std::string                               content;
-    std::vector<common_chat_msg_content_part> content_parts;
-    std::vector<common_chat_tool_call>        tool_calls;
-    std::string                               reasoning_content;
-    std::string                               tool_name;
-    std::string                               tool_call_id;
+    std::string role;
+    std::string content;
+    std::vector<common_chat_msg_content_part> content_parts = {};
+    std::vector<common_chat_tool_call> tool_calls = {};
+    std::string reasoning_content;
+    std::string tool_name;
+    std::string tool_call_id;
 
-    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
+    template <class T> T to_json_oaicompat() const;
 
     bool empty() const {
-        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() &&
-               tool_name.empty() && tool_call_id.empty();
+        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
     }
-
-    void set_tool_call_ids(std::vector<std::string> &           ids_cache,
-                           const std::function<std::string()> & gen_tool_call_id) {
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
         for (auto i = 0u; i < tool_calls.size(); i++) {
             if (ids_cache.size() <= i) {
                 auto id = tool_calls[i].id;
@@ -141,28 +56,32 @@ struct common_chat_msg {
             tool_calls[i].id = ids_cache[i];
         }
     }
-
     bool operator==(const common_chat_msg & other) const {
-        return role == other.role && content == other.content && content_parts == other.content_parts &&
-               tool_calls == other.tool_calls && reasoning_content == other.reasoning_content &&
-               tool_name == other.tool_name && tool_call_id == other.tool_call_id;
+        return role == other.role
+            && content == other.content
+            && content_parts == other.content_parts
+            && tool_calls == other.tool_calls
+            && reasoning_content == other.reasoning_content
+            && tool_name == other.tool_name
+            && tool_call_id == other.tool_call_id;
+    }
+    bool operator!=(const common_chat_msg & other) const {
+        return !(*this == other);
     }
-
-    bool operator!=(const common_chat_msg & other) const { return !(*this == other); }
 };
 
 struct common_chat_msg_diff {
-    std::string           reasoning_content_delta;
-    std::string           content_delta;
-    size_t                tool_call_index = std::string::npos;
+    std::string reasoning_content_delta;
+    std::string content_delta;
+    size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;
 
-    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & msg_prv,
-                                                           const common_chat_msg & msg_new);
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg);
 
     bool operator==(const common_chat_msg_diff & other) const {
-        return content_delta == other.content_delta && tool_call_index == other.tool_call_index &&
-               tool_call_delta == other.tool_call_delta;
+        return content_delta == other.content_delta
+        && tool_call_index == other.tool_call_index
+        && tool_call_delta == other.tool_call_delta;
     }
 };
 
@@ -180,61 +99,61 @@ enum common_chat_tool_choice {
 
 enum common_chat_format {
     COMMON_CHAT_FORMAT_CONTENT_ONLY,
+    COMMON_CHAT_FORMAT_GENERIC,
+    COMMON_CHAT_FORMAT_MISTRAL_NEMO,
+    COMMON_CHAT_FORMAT_LLAMA_3_X,
+    COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
+    COMMON_CHAT_FORMAT_DEEPSEEK_R1,
+    COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
+    COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
+    COMMON_CHAT_FORMAT_HERMES_2_PRO,
+    COMMON_CHAT_FORMAT_COMMAND_R7B,
+    COMMON_CHAT_FORMAT_GRANITE,
+    COMMON_CHAT_FORMAT_GPT_OSS,
+    COMMON_CHAT_FORMAT_SEED_OSS,
+    COMMON_CHAT_FORMAT_NEMOTRON_V2,
 
-    // These are intended to be parsed by the PEG parser
-    COMMON_CHAT_FORMAT_PEG_SIMPLE,
-    COMMON_CHAT_FORMAT_PEG_NATIVE,
-
-    COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
+    COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
 
 struct common_chat_templates_inputs {
-    std::vector<common_chat_msg>          messages;
-    std::string                           grammar;
-    std::string                           json_schema;
-    bool                                  add_generation_prompt = true;
-    bool                                  use_jinja             = true;
+    std::vector<common_chat_msg> messages;
+    std::string grammar;
+    std::string json_schema;
+    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;
-    bool                                  parallel_tool_calls = false;
-    common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool enable_thinking"
-    bool                                  enable_thinking     = true;
-    std::chrono::system_clock::time_point now                 = std::chrono::system_clock::now();
-    std::map<std::string, std::string>    chat_template_kwargs;
-    bool                                  add_bos = false;
-    bool                                  add_eos = false;
+    std::vector<common_chat_tool> tools;
+    common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
+    bool parallel_tool_calls = false;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    bool enable_thinking = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    std::map<std::string, std::string> chat_template_kwargs;
+    bool add_bos = false;
+    bool add_eos = false;
 };
 
 struct common_chat_params {
     common_chat_format                  format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     std::string                         prompt;
     std::string                         grammar;
-    bool                                grammar_lazy         = false;
+    bool                                grammar_lazy = false;
     bool                                thinking_forced_open = false;
-    bool                                supports_thinking    = false;
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
-    std::string                         parser;
 };
 
-// per-message parsing syntax
-// should be derived from common_chat_params
-struct common_chat_parser_params {
-    common_chat_format      format               = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    common_reasoning_format reasoning_format     = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool parse_reasoning"
+struct common_chat_syntax {
+    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
     // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
-    bool                    reasoning_in_content = false;
-    bool                    thinking_forced_open = false;
-    bool                    parse_tool_calls     = true;
-    bool                    debug                = false;  // Enable debug output for PEG parser
-    common_peg_arena        parser               = {};
-    common_chat_parser_params() = default;
-    common_chat_parser_params(const common_chat_params & chat_params) {
-        format               = chat_params.format;
-        thinking_forced_open = chat_params.thinking_forced_open;
-    }
+    bool                     reasoning_in_content  = false;
+    bool                     thinking_forced_open  = false;
+    bool                     parse_tool_calls      = true;
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
@@ -242,64 +161,55 @@ bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
 
 void common_chat_templates_free(struct common_chat_templates * tmpls);
 
-struct common_chat_templates_deleter {
-    void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); }
-};
+struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } };
 
 typedef std::unique_ptr<struct common_chat_templates, common_chat_templates_deleter> common_chat_templates_ptr;
 
-common_chat_templates_ptr common_chat_templates_init(const struct llama_model * model,
-                                                     const std::string &        chat_template_override,
-                                                     const std::string &        bos_token_override = "",
-                                                     const std::string &        eos_token_override = "");
+common_chat_templates_ptr common_chat_templates_init(
+                                    const struct llama_model * model,
+                                           const std::string & chat_template_override,
+                                           const std::string & bos_token_override = "",
+                                           const std::string & eos_token_override = "");
 
 bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
-std::string  common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant = "");
+const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
 
-struct common_chat_params common_chat_templates_apply(const struct common_chat_templates *        tmpls,
-                                                      const struct common_chat_templates_inputs & inputs);
+
+struct common_chat_params      common_chat_templates_apply(
+    const struct common_chat_templates * tmpls,
+    const struct common_chat_templates_inputs & inputs);
 
 // Format single message, while taking into account the position of that message in chat history
-std::string common_chat_format_single(const struct common_chat_templates * tmpls,
-                                      const std::vector<common_chat_msg> & past_msg,
-                                      const common_chat_msg &              new_msg,
-                                      bool                                 add_ass,
-                                      bool                                 use_jinja);
+std::string common_chat_format_single(
+        const struct common_chat_templates * tmpls,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
+        bool add_ass,
+        bool use_jinja);
 
 // Returns an example of formatted chat
-std::string common_chat_format_example(const struct common_chat_templates *       tmpls,
-                                       bool                                       use_jinja,
-                                       const std::map<std::string, std::string> & chat_template_kwargs);
+std::string common_chat_format_example(
+    const struct common_chat_templates * tmpls,
+    bool use_jinja,
+    const std::map<std::string, std::string> & chat_template_kwargs);
 
-const char *            common_chat_format_name(common_chat_format format);
-common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & params);
-common_chat_msg           common_chat_peg_parse(const common_peg_arena & src_parser, const std::string & input, bool is_partial, const common_chat_parser_params & params);
-
-// used by arg and server
-const char *            common_reasoning_format_name(common_reasoning_format format);
-common_reasoning_format common_reasoning_format_from_name(const std::string & format);
+const char*               common_chat_format_name(common_chat_format format);
+const char*               common_reasoning_format_name(common_reasoning_format format);
+common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
+template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-// 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);
+// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
+template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
-nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
-
-nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
-
-// get template caps, useful for reporting to server /props endpoint
-std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
-
-std::string common_chat_template_direct_apply(
-    const common_chat_template & tmpl,
-    const autoparser::templates_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);
+template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);

common/common.cpp

@@ -1,17 +1,19 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "ggml.h"
 #include "gguf.h"
 
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "sampling.h"
-#include "unicode.h"
 
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
-#include <chrono>
+#include <codecvt>
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
@@ -23,6 +25,7 @@
 #include <sstream>
 #include <string>
 #include <thread>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -47,23 +50,10 @@
 #include <unistd.h>
 #endif
 
-#if defined(__linux__)
-#include <sys/types.h>
-#include <pwd.h>
-#endif
-
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
-
-common_time_meas::~common_time_meas() {
-    if (t_start_us >= 0) {
-        t_acc += ggml_time_us() - t_start_us;
-    }
-}
-
 //
 // CPU utils
 //
@@ -247,7 +237,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
         case GGML_SCHED_PRIO_REALTIME: p = -20; break;
     }
 
-    if (setpriority(PRIO_PROCESS, 0, p) != 0) {
+    if (!setpriority(PRIO_PROCESS, 0, p)) {
         LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
@@ -359,7 +349,11 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set(common_log_default_callback, NULL);
+    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
+        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+            common_log_add(common_log_main(), level, "%s", text);
+        }
+    }, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -452,6 +446,34 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}
+
+bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
+    bool has_suffix = string_ends_with(str, suffix);
+    if (has_suffix) {
+        str = str.substr(0, str.size() - suffix.size());
+    }
+    return has_suffix;
+}
+
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
+    if (!str.empty() && !stop.empty()) {
+        const char text_last_char = str.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
+                const auto current_partial = stop.substr(0, char_index + 1);
+                if (string_ends_with(str, current_partial)) {
+                    return str.size() - char_index - 1;
+                }
+            }
+        }
+    }
+
+    return std::string::npos;
+}
+
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
     return std::regex_replace(s, special_chars, "\\$&");
@@ -662,7 +684,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
+bool fs_validate_filename(const std::string & filename) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -674,28 +696,45 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
         return false;
     }
 
-    size_t offset = 0;
-    while (offset < filename.size()) {
-        utf8_parse_result result = common_parse_utf8_codepoint(filename, offset);
+    std::u32string filename_utf32;
+    try {
+#if defined(__clang__)
+        // disable C++17 deprecation warning for std::codecvt_utf8
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+#endif
 
-        if (result.status != utf8_parse_result::SUCCESS) {
+        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
+#endif
+
+        filename_utf32 = converter.from_bytes(filename);
+
+        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
+        // or invalid encodings were encountered. Reject such attempts
+        std::string filename_reencoded = converter.to_bytes(filename_utf32);
+        if (filename_reencoded != filename) {
             return false;
         }
-        uint32_t c = result.codepoint;
+    } catch (const std::exception &) {
+        return false;
+    }
 
-        if ((result.bytes_consumed == 2 && c < 0x80) ||
-            (result.bytes_consumed == 3 && c < 0x800) ||
-            (result.bytes_consumed == 4 && c < 0x10000)) {
-            return false;
-        }
-
-        // Check for forbidden codepoints:
-        // - Control characters
-        // - Unicode equivalents of illegal characters
-        // - UTF-16 surrogate pairs
-        // - UTF-8 replacement character
-        // - Byte order mark (BOM)
-        // - Illegal characters: / \ : * ? " < > |
+    // Check for forbidden codepoints:
+    // - Control characters
+    // - Unicode equivalents of illegal characters
+    // - UTF-16 surrogate pairs
+    // - UTF-8 replacement character
+    // - Byte order mark (BOM)
+    // - Illegal characters: / \ : * ? " < > |
+    for (char32_t c : filename_utf32) {
         if (c <= 0x1F // Control characters (C0)
             || c == 0x7F // Control characters (DEL)
             || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
@@ -703,18 +742,12 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             || c == 0x2215 // Division Slash (forward slash equivalent)
             || c == 0x2216 // Set Minus (backslash equivalent)
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
-            || c > 0x10FFFF // Max Unicode limit
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == ':' || c == '*' // Illegal characters
+            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
-        if (!allow_subdirs && (c == '/' || c == '\\')) {
-            // Subdirectories not allowed, reject path separators
-            return false;
-        }
-        offset += result.bytes_consumed;
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -739,29 +772,11 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
 #include <iostream>
 
 
-#ifdef _WIN32
-static std::wstring utf8_to_wstring(const std::string & str) {
-    if (str.empty()) {
-        return std::wstring();
-    }
-
-    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
-
-    if (size <= 0) {
-        return std::wstring();
-    }
-
-    std::wstring wstr(size, 0);
-    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
-
-    return wstr;
-}
-#endif
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
-    std::wstring wpath = utf8_to_wstring(path);
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -834,11 +849,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
-bool fs_is_directory(const std::string & path) {
-    std::filesystem::path dir(path);
-    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
-}
-
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -851,31 +861,16 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
-        defined(__OpenBSD__) || defined(__NetBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else if (std::getenv("HOME")) {
-            cache_directory = std::getenv("HOME") + std::string("/.cache/");
         } else {
-#if defined(__linux__)
-            /* no $HOME is defined, fallback to getpwuid */
-            struct passwd *pw = getpwuid(getuid());
-            if ((!pw) || (!pw->pw_dir)) {
-                throw std::runtime_error("Failed to find $HOME directory");
-            }
-
-            cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
-#else /* defined(__linux__) */
-            throw std::runtime_error("Failed to find $HOME directory");
-#endif /* defined(__linux__) */
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
         }
 #elif defined(__APPLE__)
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#elif defined(__EMSCRIPTEN__)
-        GGML_ABORT("not implemented on this platform");
 #else
 #  error Unknown architecture
 #endif
@@ -895,293 +890,34 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
-    std::vector<common_file_info> files;
-    if (path.empty()) return files;
-
-    std::filesystem::path dir(path);
-    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
-        return files;
-    }
-
-    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
-        try {
-            // Only include regular files (skip directories)
-            const auto & p = entry.path();
-            if (std::filesystem::is_regular_file(p)) {
-                common_file_info info;
-                info.path   = p.string();
-                info.name   = p.filename().string();
-                info.is_dir = false;
-                try {
-                    info.size = static_cast<size_t>(std::filesystem::file_size(p));
-                } catch (const std::filesystem::filesystem_error &) {
-                    info.size = 0;
-                }
-                files.push_back(std::move(info));
-            } else if (include_directories && std::filesystem::is_directory(p)) {
-                common_file_info info;
-                info.path   = p.string();
-                info.name   = p.filename().string();
-                info.size   = 0; // Directories have no size
-                info.is_dir = true;
-                files.push_back(std::move(info));
-            }
-        } catch (const std::filesystem::filesystem_error &) {
-            // skip entries we cannot inspect
-            continue;
-        }
-    }
-
-    return files;
-}
-
-//
-// TTY utils
-//
-
-bool tty_can_use_colors() {
-    // Check NO_COLOR environment variable (https://no-color.org/)
-    if (const char * no_color = std::getenv("NO_COLOR")) {
-        if (no_color[0] != '\0') {
-            return false;
-        }
-    }
-
-    // Check TERM environment variable
-    if (const char * term = std::getenv("TERM")) {
-        if (std::strcmp(term, "dumb") == 0) {
-            return false;
-        }
-    }
-
-    // Check if stdout and stderr are connected to a terminal
-    // We check both because log messages can go to either
-    bool stdout_is_tty = isatty(fileno(stdout));
-    bool stderr_is_tty = isatty(fileno(stderr));
-
-    return stdout_is_tty || stderr_is_tty;
-}
 
 //
 // Model utils
 //
 
-// TODO: move to common/sampling
-static void common_init_sampler_from_model(
-    const llama_model * model,
-    common_params_sampling & sparams) {
-
-    const uint64_t config = sparams.user_sampling_config;
-
-    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
-        if (config & user_config) {
-            return;
-        }
-
-        char buf[64] = {0};
-        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
-            char * end = nullptr;
-            int32_t v = strtol(buf, &end, 10);
-            if (end && end != buf) {
-                dst = v;
-            }
-        }
-    };
-
-    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
-        if (config & user_config) {
-            return;
-        }
-
-        char buf[128] = {0};
-        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
-            char * end = nullptr;
-            float v = strtof(buf, &end);
-            if (end && end != buf) {
-                dst = v;
-            }
-        }
-    };
-
-    // Sampling sequence
-    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
-        char buf[512] = {0};
-        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
-            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
-            if (!sampler_names.empty()) {
-                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
-            }
-        }
-    }
-
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
-}
-
-struct common_init_result::impl {
-    impl() = default;
-    ~impl() = default;
-
-    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
-
-    llama_model_ptr   model;
-    llama_context_ptr context;
-
-    std::vector<llama_adapter_lora_ptr> lora;
-
-    std::vector<common_sampler_ptr> samplers;
-    std::vector<llama_sampler_seq_config> samplers_seq_config;
-};
-
-common_init_result::common_init_result(common_params & params) :
-    pimpl(new impl{}) {
+struct common_init_result common_init_from_params(common_params & params) {
+    common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
-    auto cparams = common_context_params_to_llama(params);
-
-    if (params.fit_params) {
-        LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
-        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split,
-            params.tensor_buft_overrides.data(),
-            params.fit_params_target.data(),
-            params.fit_params_min_ctx,
-            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);
     if (model == NULL) {
-        return;
+        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
+        return iparams;
     }
 
-    pimpl->model.reset(model);
-
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    // load and optionally apply lora adapters (must be loaded before context creation)
-    for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
-        if (lora == nullptr) {
-            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
-            pimpl->model.reset(model);
-            return;
-        }
-
-        char buf[1024];
-        la.ptr = lora.get();
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
-        la.task_name = buf;
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
-        la.prompt_prefix = buf;
-        pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
-    }
-
-    // updates params.sampling
-    // TODO: fix naming
-    common_init_sampler_from_model(model, params.sampling);
-
-    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sampling.ignore_eos = false;
-    }
-
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
-    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
-    }
-
-    //if (params.sampling.penalty_last_n == -1) {
-    //    LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    //if (params.sampling.dry_penalty_last_n == -1) {
-    //    LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    // init the backend samplers as part of the context creation
-    pimpl->samplers.resize(cparams.n_seq_max);
-    pimpl->samplers_seq_config.resize(cparams.n_seq_max);
-
-    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
-        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
-        pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
-    }
-
-    if (params.sampling.backend_sampling) {
-        cparams.samplers   = pimpl->samplers_seq_config.data();
-        cparams.n_samplers = pimpl->samplers_seq_config.size();
-    }
+    auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
-        return;
+        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
+        llama_model_free(model);
+        return iparams;
     }
 
-    pimpl->context.reset(lctx);
-}
-
-llama_model * common_init_result::model() {
-    return pimpl->model.get();
-}
-
-llama_context * common_init_result::context() {
-    return pimpl->context.get();
-}
-
-common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
-    return pimpl->samplers[seq_id].get();
-}
-
-void common_init_result::reset_samplers() {
-    for (int i = 0; i < (int) pimpl->samplers.size(); ++i) {
-        llama_sampler_reset(common_sampler_get(pimpl->samplers[i].get()));
-    }
-}
-
-std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
-    return pimpl->lora;
-}
-
-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) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
-        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());
-        return res;
-    }
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
@@ -1193,10 +929,13 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
 
-        int err = llama_set_adapter_cvec(
+        int err = llama_apply_adapter_cvec(
                 lctx,
                 cvec.data.data(),
                 cvec.data.size(),
@@ -1204,7 +943,10 @@ common_init_result_ptr common_init_from_params(common_params & params) {
                 params.control_vector_layer_start,
                 params.control_vector_layer_end);
         if (err) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
     }
 
@@ -1218,24 +960,79 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
         bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
         bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-        bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;
 
-        if (!has_eos && !has_sep && !has_rerank_prompt) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
             ok = false;
         } else if (!has_eos) {
             LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
         }
 
         if (!ok) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
     }
 
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
+            LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
+            llama_free(lctx);
+            llama_model_free(model);
+            return iparams;
+        }
+
+        char buf[1024];
+        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
+        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+    }
+
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
+        params.sampling.ignore_eos = false;
+    }
+
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
+        }
+    }
+
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
+    if (params.sampling.penalty_last_n == -1) {
+        LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+        params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    }
+
+    if (params.sampling.dry_penalty_last_n == -1) {
+        LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+        params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    }
+
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
@@ -1272,15 +1069,13 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
-
-        // reset samplers to reset RNG state after warmup to the seeded state
-        res->reset_samplers();
     }
 
-    return res;
-}
+    iparams.model.reset(model);
+    iparams.context.reset(lctx);
 
-common_init_result::~common_init_result() = default;
+    return iparams;
+}
 
 std::string get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
@@ -1290,23 +1085,18 @@ std::string get_model_endpoint() {
     std::string model_endpoint = "https://huggingface.co/";
     if (endpoint_env) {
         model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+        if (model_endpoint.back() != '/') model_endpoint += '/';
     }
     return model_endpoint;
 }
 
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    std::vector<llama_adapter_lora *> loras;
-    std::vector<float> scales;
-
-    for (auto & la: lora) {
-        loras.push_back(la.ptr);
-        scales.push_back(la.scale);
+    llama_clear_adapter_lora(ctx);
+    for (auto & la : lora) {
+        if (la.scale != 0.0f) {
+            llama_set_adapter_lora(ctx, la.ptr, la.scale);
+        }
     }
-
-    llama_set_adapters_lora(ctx, loras.data(), loras.size(), scales.data());
 }
 
 struct llama_model_params common_model_params_to_llama(common_params & params) {
@@ -1316,16 +1106,17 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.devices = params.devices.data();
     }
 
-    mparams.n_gpu_layers    = params.n_gpu_layers;
+    if (params.n_gpu_layers != -1) {
+        mparams.n_gpu_layers = params.n_gpu_layers;
+    }
+
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
-    mparams.use_direct_io   = params.use_direct_io;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;
-    mparams.no_host         = params.no_host;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
@@ -1426,6 +1217,66 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
+//
+// Token utils
+//
+
+size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+
+    return i;
+}
+
+size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
+    // check for empty sequences
+    if (a.empty() || b.empty()) {
+        return 0;
+    }
+
+    // get the lengths of the input sequences
+    size_t a_len = a.size();
+    size_t b_len = b.size();
+
+    // initialize the maximum length of the longest common subsequence (LCS)
+    size_t max_length = 0;
+
+    // use two rows instead of a 2D matrix to optimize space
+    std::vector<size_t> prev_row(b_len + 1, 0);
+    std::vector<size_t> curr_row(b_len + 1, 0);
+
+    // iterate through the elements of a
+    for (size_t i = 1; i <= a_len; i++) {
+        // iterate through the elements of b
+        for (size_t j = 1; j <= b_len; j++) {
+            // if elements at the current positions match
+            if (a[i - 1] == b[j - 1]) {
+                // if it's the first element of either sequences, set LCS length to 1
+                if (i == 1 || j == 1) {
+                    curr_row[j] = 1;
+                } else {
+                    // increment LCS length by 1 compared to the previous element
+                    curr_row[j] = prev_row[j - 1] + 1;
+                }
+
+                // update max_length if necessary
+                if (curr_row[j] > max_length) {
+                    max_length = curr_row[j];
+                }
+            } else {
+                // reset LCS length if elements don't match
+                curr_row[j] = 0;
+            }
+        }
+
+        // update the previous row for the next iteration
+        prev_row = curr_row;
+    }
+
+    // return the maximum length of the LCS
+    return max_length;
+}
+
 //
 // Vocab utils
 //
@@ -1760,65 +1611,3 @@ float lr_opt::get_lr(float epoch) const {
     LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
     return r;
 }
-
-bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos) {
-    llama_batch batch = llama_batch_get_one(&last_token, 1);
-    batch.pos = &pos;
-    if (llama_decode(ctx, batch)) {
-        LOG_ERR("%s: failed to replay last token\n", __func__);
-        return false;
-    }
-    return true;
-}
-
-bool common_prompt_batch_decode(
-              struct llama_context * ctx,
-    const std::vector<llama_token> & tokens,
-                               int & n_past,
-                               int   n_batch,
-                  std::string_view   state_path,
-                              bool   save_state) {
-    const int n_eval = tokens.size();
-    if (n_eval == 0) {
-        return true;
-    }
-
-    if (save_state && n_eval > 1) {
-        const int n_tokens_before_last = n_eval - 1;
-
-        GGML_ASSERT(n_eval <= n_batch);
-
-        // Decode all but the last token so we can save the memory state before decoding the last token.
-        // This is done so we can restore the session state later and replay the last token.
-        // Memory implementations in recurrent/hybrid models don't support removing tokens from their
-        // memory, so we can't just remove the last token from the memory and replay the last token which
-        // is the reason for this logic.
-        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_tokens_before_last))) {
-            LOG_ERR("%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_tokens_before_last;
-
-        llama_state_save_file(ctx, state_path.data(), tokens.data(), n_tokens_before_last);
-        LOG_INF("saved session before last token to %s, n_tokens = %d\n", state_path.data(), n_tokens_before_last);
-
-        llama_token last_token = tokens.back();
-        llama_batch batch = llama_batch_get_one(&last_token, 1);
-        int32_t pos = n_past;
-        batch.pos = &pos;
-
-        if (llama_decode(ctx, batch)) {
-            LOG_ERR("%s : failed to eval last token\n", __func__);
-            return false;
-        }
-        n_past++;
-    } else {
-        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_eval))) {
-            LOG_ERR("%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_eval;
-    }
-
-    return true;
-}

common/common.h

@@ -2,19 +2,17 @@
 
 #pragma once
 
-#include "ggml-opt.h"
-#include "llama-cpp.h"
-
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
+#include <sstream>
+#include <cmath>
 
-#if defined(_WIN32) && !defined(_WIN32_WINNT)
-#define _WIN32_WINNT 0x0A00
-#endif
+#include "ggml-opt.h"
+#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -30,14 +28,7 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-struct common_time_meas {
-    common_time_meas(int64_t & t_acc, bool disable = false);
-    ~common_time_meas();
-
-    const int64_t t_start_us;
-
-    int64_t & t_acc;
-};
+#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
 struct common_adapter_lora_info {
     std::string path;
@@ -57,8 +48,6 @@ extern const char * LLAMA_COMMIT;
 extern const char * LLAMA_COMPILER;
 extern const char * LLAMA_BUILD_TARGET;
 
-const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
-
 struct common_control_vector_load_info;
 
 //
@@ -82,12 +71,9 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
-    LLAMA_EXAMPLE_BATCHED,
-    LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_COMPLETION,
-    LLAMA_EXAMPLE_CLI,
+    LLAMA_EXAMPLE_MAIN,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -103,8 +89,6 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
-    LLAMA_EXAMPLE_FIT_PARAMS,
-    LLAMA_EXAMPLE_RESULTS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -122,7 +106,6 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_INFILL      = 9,
     COMMON_SAMPLER_TYPE_PENALTIES   = 10,
     COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
-    COMMON_SAMPLER_TYPE_ADAPTIVE_P  = 12,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -150,71 +133,41 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
-enum common_params_sampling_config : uint64_t {
-    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
-    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
-    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
-    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
-    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
-    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
-    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
-    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
-};
-
-enum common_speculative_type {
-    COMMON_SPECULATIVE_TYPE_NONE,          // no speculative decoding
-    COMMON_SPECULATIVE_TYPE_DRAFT,         // draft model
-    COMMON_SPECULATIVE_TYPE_EAGLE3,        // eagle draft model
-    COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding
-    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
-    COMMON_SPECULATIVE_TYPE_NGRAM_MOD,
-    COMMON_SPECULATIVE_TYPE_NGRAM_CACHE,   // self-speculative decoding with 3-level n-gram cache
-    COMMON_SPECULATIVE_TYPE_COUNT          // number of types, unknown type
-};
-
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev             = 64;     // number of previous tokens to remember
-    int32_t n_probs            = 0;      // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep           = 0;      // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k              = 40;     // <= 0 to use vocab size
-    float   top_p              = 0.95f;  // 1.0 = disabled
-    float   min_p              = 0.05f;  // 0.0 = disabled
-    float   xtc_probability    = 0.00f;  // 0.0 = disabled
-    float   xtc_threshold      = 0.10f;  // > 0.5 disables XTC
-    float   typ_p              = 1.00f;  // typical_p, 1.0 = disabled
-    float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range     = 0.00f;  // 0.0 = disabled
-    float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat     = 1.00f;  // 1.0 = disabled
-    float   penalty_freq       = 0.00f;  // 0.0 = disabled
-    float   penalty_present    = 0.00f;  // 0.0 = disabled
-    float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
-    float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
-    int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
-    float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
-    float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
-    int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   top_n_sigma        = -1.00f; // -1.0 = disabled
-    float   mirostat_tau       = 5.00f;  // target entropy
-    float   mirostat_eta       = 0.10f;  // learning rate
+    int32_t n_prev             = 64;    // number of previous tokens to remember
+    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k              = 40;    // <= 0 to use vocab size
+    float   top_p              = 0.95f; // 1.0 = disabled
+    float   min_p              = 0.05f; // 0.0 = disabled
+    float   xtc_probability    = 0.00f; // 0.0 = disabled
+    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
+    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
+    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range     = 0.00f; // 0.0 = disabled
+    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat     = 1.00f; // 1.0 = disabled
+    float   penalty_freq       = 0.00f; // 0.0 = disabled
+    float   penalty_present    = 0.00f; // 0.0 = disabled
+    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
+    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
+    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
+    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   top_n_sigma        = -1.00f;// -1.0 = disabled
+    float   mirostat_tau       = 5.00f; // target entropy
+    float   mirostat_eta       = 0.10f; // learning rate
     bool    ignore_eos         = false;
-    bool    no_perf            = false;  // disable performance metrics
+    bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
-    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
-
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
+
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
@@ -235,12 +188,6 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
-    bool backend_sampling = false;
-
-    bool has_logit_bias() const {
-        return !logit_bias.empty();
-    }
-
     // print the parameters into a string
     std::string print() const;
 };
@@ -251,42 +198,19 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
-    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
-struct common_ngram_mod;
-
 struct common_params_speculative {
-    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
+    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    // general-purpose speculative decoding parameters
-
-    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   = 0.75f; // minimum speculative decoding probability (greedy)
-
-    // ngram-based speculative decoding
-
-    uint16_t ngram_size_n     = 12; // ngram size for lookup
-    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
-
-    std::shared_ptr<common_ngram_mod> ngram_mod;
-
-    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding          // NOLINT
-
-    // draft-model speculative decoding
-
-    struct common_params_model mparams_dft;
-
-    llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts
-
-    llama_context_params cparams_dft; // these are the parameters for the draft llama_context
-
-    int32_t n_ctx        = 0;  // draft context size
-    int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    int32_t n_ctx        =     0; // draft context size
+    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
+    int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    float   p_split      =  0.1f; // speculative decoding split probability
+    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
+    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@@ -294,14 +218,7 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
-
-    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
-
-    bool has_dft() const {
-        return !mparams_dft.path.empty() || !mparams_dft.hf_repo.empty();
-    }
+    struct common_params_model model;
 };
 
 struct common_params_vocoder {
@@ -327,7 +244,6 @@ struct common_params_diffusion {
 };
 
 // reasoning API response format (not to be confused as chat template's reasoning format)
-// only used by server
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
@@ -358,8 +274,8 @@ struct lr_opt {
 struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 
 struct common_params {
-    int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
-    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
+    int32_t n_predict             =    -1; // new tokens to predict
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -380,14 +296,9 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
-    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
-    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
-    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
-
-    // margin per device in bytes for fitting parameters to free memory:
-    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
+    int32_t n_gpu_layers      = -1;  // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu          = 0;   // the GPU that is used for scratch and small tensors
+    float   tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -411,8 +322,7 @@ struct common_params {
 
     struct common_params_model model;
 
-    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 model_alias          = ""; // model alias                                                   // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT
@@ -420,13 +330,10 @@ struct common_params {
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
     std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
-    // llama-debug specific options
-    std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
-    bool        save_logits       = false;  // whether to save logits to files                              // NOLINT
-    std::vector<std::string> tensor_filter; // filter tensor names for debug output (regex)                 // NOLINT
-
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -437,7 +344,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
+    int32_t verbosity                  = 0;
     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;
@@ -457,8 +364,6 @@ struct common_params {
 
     bool   kl_divergence    = false; // compute KL divergence
 
-    bool check             = false; // check rather than generate results for llama-results
-
     bool usage             = false; // print usage
     bool completion        = false; // print source-able completion script
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -473,14 +378,12 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
-    bool show_timings      = true;  // show timing information on CLI
-    bool ctx_shift         = false; // context shift on infinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool use_mmap          = true;  // enable mmap to use filesystem cache
-    bool use_direct_io     = false; // read from disk without buffering
+    bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
@@ -489,7 +392,6 @@ struct common_params {
     bool check_tensors     = false; // validate tensor data
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
     bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
-    bool no_host           = false; // bypass host buffer allowing extra buffers to be used
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -503,8 +405,6 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
-    int image_min_tokens = -1;
-    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -519,26 +419,22 @@ struct common_params {
     std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
-    int32_t port                = 8080;          // server listens on this network port
-    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
-    int32_t n_ctx_checkpoints   = 32;     // max number of context checkpoints per slot
-    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.
+    int32_t port              = 8080;         // server listens on this network port
+    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
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = true;                                                                                  // NOLINT
+    bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
     int reasoning_budget = -1;
-    bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
-    int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
 
@@ -547,32 +443,20 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
-    // webui configs
-    bool webui = true;
-    bool webui_mcp_proxy = false;
-    std::string webui_config_json;
-
     // "advanced" endpoints are disabled by default for better security
+    bool webui            = true;
     bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
-    // router server configs
-    std::string models_dir    = ""; // directory containing models for the router server
-    std::string models_preset = ""; // directory containing model presets for the router server
-    int models_max = 4;             // maximum number of models to load simultaneously
-    bool models_autoload = true;    // automatically load models when requested via the router server
-
     bool log_json = false;
 
     std::string slot_save_path;
-    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
-    bool is_pp_shared   = false;
-    bool is_tg_separate = false;
+    bool is_pp_shared = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -676,55 +560,30 @@ static std::vector<T> string_split(const std::string & str, char delim) {
 }
 
 template<>
-inline std::vector<std::string> string_split<std::string>(const std::string & str, char delim)
+std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
 {
     std::vector<std::string> parts;
     size_t begin_pos = 0;
-    size_t delim_pos = str.find(delim);
-    while (delim_pos != std::string::npos) {
-        std::string part = str.substr(begin_pos, delim_pos - begin_pos);
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(begin_pos, separator_pos - begin_pos);
         parts.emplace_back(part);
-        begin_pos = delim_pos + 1;
-        delim_pos = str.find(delim, begin_pos);
+        begin_pos = separator_pos + 1;
+        separator_pos = input.find(separator, begin_pos);
     }
-    parts.emplace_back(str.substr(begin_pos));
+    parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
     return parts;
 }
 
-// remove when moving to c++20
-inline bool string_starts_with(std::string_view str, std::string_view prefix) {
-    return str.size() >= prefix.size() &&
-           str.compare(0, prefix.size(), prefix) == 0;
+static bool string_starts_with(const std::string & str,
+                               const std::string & prefix) {  // While we wait for C++20's std::string::starts_with...
+    return str.rfind(prefix, 0) == 0;
 }
 
-// remove when moving to c++20
-inline bool string_ends_with(std::string_view str, std::string_view suffix) {
-    return str.size() >= suffix.size() &&
-           str.compare(str.size() - suffix.size(), suffix.size(), suffix) == 0;
-}
-
-inline bool string_remove_suffix(std::string & str, std::string_view suffix) {
-    if (string_ends_with(str, suffix)) {
-        str.resize(str.size() - suffix.size());
-        return true;
-    }
-    return false;
-}
-
-inline size_t string_find_partial_stop(std::string_view str, std::string_view stop) {
-    if (!str.empty() && !stop.empty()) {
-        const size_t max_len = std::min(str.size(), stop.size());
-        const char last_char = str.back();
-        for (size_t len = max_len; len > 0; --len) {
-            if (stop[len - 1] == last_char) {
-                if (string_ends_with(str, stop.substr(0, len))) {
-                    return str.size() - len;
-                }
-            }
-        }
-    }
-    return std::string::npos;
-}
+// While we wait for C++20's std::string::ends_with...
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+bool string_remove_suffix(std::string & str, const std::string_view & suffix);
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -738,55 +597,25 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
+bool fs_validate_filename(const std::string & filename);
 bool fs_create_directory_with_parents(const std::string & path);
-bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
-struct common_file_info {
-    std::string path;
-    std::string name;
-    size_t      size = 0; // in bytes
-    bool        is_dir = false;
-};
-std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
-
-//
-// TTY utils
-//
-
-// Auto-detect if colors can be enabled based on terminal and environment
-bool tty_can_use_colors();
-
 //
 // Model utils
 //
 
-struct common_sampler;
-
-// note: defines the model, context, samplers, ets. lifetimes
+// note: defines object's lifetime
 struct common_init_result {
-    common_init_result(common_params & params);
-    ~common_init_result();
+    llama_model_ptr   model;
+    llama_context_ptr context;
 
-    llama_model * model();
-    llama_context * context();
-
-    common_sampler * sampler(llama_seq_id seq_id);
-    void reset_samplers();
-
-    std::vector<llama_adapter_lora_ptr> & lora();
-
-private:
-    struct impl;
-    std::unique_ptr<impl> pimpl;
+    std::vector<llama_adapter_lora_ptr> lora;
 };
 
-using common_init_result_ptr = std::unique_ptr<common_init_result>;
-
-common_init_result_ptr common_init_from_params(common_params & params);
+struct common_init_result     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);
@@ -810,22 +639,15 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
-// decodes a single batch of tokens for a prompt and manages session tokens
 //
-// Note: We save state before the last token so that we can replay it to ensure
-// compatibility with all memory types. Recurrent/hybrid models cannot remove
-// tokens from memory, so this approach works across all model architectures.
-bool common_prompt_batch_decode(
-              struct llama_context * ctx,
-    const std::vector<llama_token> & embd,
-                               int & n_past,
-                               int   n_batch,
-                  std::string_view   state_path,
-                              bool   save_state);
+// Token utils
+//
 
-// replays the last token after loading state to regenerate logits
-// used after loading session state to ensure the sampling context has valid logits
-bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos);
+// longest common prefix
+size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
+
+// longet common subsequence
+size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
 
 //
 // Vocab utils
@@ -874,7 +696,7 @@ std::string common_detokenize(
 // Embedding utils
 //
 
-// TODO: replace embd_norm with an enum
+// TODO: repace embd_norm with an enum
 void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
 
 float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
@@ -916,13 +738,13 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 // MoE utils
 //
 
-const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_exps";
 
-inline std::string llm_ffn_exps_block_regex(int idx) {
+static std::string llm_ffn_exps_block_regex(int idx) {
     return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
 }
 
-inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
     return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
 }
 

common/console.cpp

@@ -1,16 +1,6 @@
 #include "console.h"
-#include "log.h"
 #include <vector>
 #include <iostream>
-#include <cassert>
-#include <cstddef>
-#include <cctype>
-#include <cwctype>
-#include <cstdint>
-#include <condition_variable>
-#include <mutex>
-#include <thread>
-#include <stdarg.h>
 
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
@@ -40,48 +30,28 @@
 #define ANSI_COLOR_BLUE    "\x1b[34m"
 #define ANSI_COLOR_MAGENTA "\x1b[35m"
 #define ANSI_COLOR_CYAN    "\x1b[36m"
-#define ANSI_COLOR_GRAY    "\x1b[90m"
 #define ANSI_COLOR_RESET   "\x1b[0m"
 #define ANSI_BOLD          "\x1b[1m"
 
 namespace console {
 
-#if defined (_WIN32)
-    namespace {
-        // Use private-use unicode values to represent special keys that are not reported
-        // as characters (e.g. arrows on Windows). These values should never clash with
-        // real input and let the rest of the code handle navigation uniformly.
-        static constexpr char32_t KEY_ARROW_LEFT       = 0xE000;
-        static constexpr char32_t KEY_ARROW_RIGHT      = 0xE001;
-        static constexpr char32_t KEY_ARROW_UP         = 0xE002;
-        static constexpr char32_t KEY_ARROW_DOWN       = 0xE003;
-        static constexpr char32_t KEY_HOME             = 0xE004;
-        static constexpr char32_t KEY_END              = 0xE005;
-        static constexpr char32_t KEY_CTRL_ARROW_LEFT  = 0xE006;
-        static constexpr char32_t KEY_CTRL_ARROW_RIGHT = 0xE007;
-        static constexpr char32_t KEY_DELETE           = 0xE008;
-    }
-
     //
     // Console state
     //
-#endif
 
-    static bool         advanced_display = false;
-    static bool         simple_io        = true;
-    static display_type current_display  = DISPLAY_TYPE_RESET;
+    static bool      advanced_display = false;
+    static bool      simple_io        = true;
+    static display_t current_display  = reset;
 
-    static FILE*        out              = stdout;
+    static FILE*     out              = stdout;
 
 #if defined (_WIN32)
-    static void*        hConsole;
+    static void*     hConsole;
 #else
-    static FILE*        tty              = nullptr;
-    static termios      initial_state;
+    static FILE*     tty              = nullptr;
+    static termios   initial_state;
 #endif
 
-    static completion_callback completion_cb = nullptr;
-
     //
     // Init and cleanup
     //
@@ -150,7 +120,7 @@ namespace console {
 
     void cleanup() {
         // Reset console display
-        set_display(DISPLAY_TYPE_RESET);
+        set_display(reset);
 
 #if !defined(_WIN32)
         // Restore settings on POSIX systems
@@ -170,26 +140,20 @@ namespace console {
     //
 
     // Keep track of current display and only emit ANSI code if it changes
-    void set_display(display_type display) {
+    void set_display(display_t display) {
         if (advanced_display && current_display != display) {
-            common_log_flush(common_log_main());
+            fflush(stdout);
             switch(display) {
-                case DISPLAY_TYPE_RESET:
+                case reset:
                     fprintf(out, ANSI_COLOR_RESET);
                     break;
-                case DISPLAY_TYPE_INFO:
-                    fprintf(out, ANSI_COLOR_MAGENTA);
-                    break;
-                case DISPLAY_TYPE_PROMPT:
+                case prompt:
                     fprintf(out, ANSI_COLOR_YELLOW);
                     break;
-                case DISPLAY_TYPE_REASONING:
-                    fprintf(out, ANSI_COLOR_GRAY);
-                    break;
-                case DISPLAY_TYPE_USER_INPUT:
+                case user_input:
                     fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
                     break;
-                case DISPLAY_TYPE_ERROR:
+                case error:
                     fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
             }
             current_display = display;
@@ -212,18 +176,7 @@ namespace console {
             if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
                 wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
                 if (wc == 0) {
-                    const DWORD ctrl_mask = LEFT_CTRL_PRESSED | RIGHT_CTRL_PRESSED;
-                    const bool ctrl_pressed = (record.Event.KeyEvent.dwControlKeyState & ctrl_mask) != 0;
-                    switch (record.Event.KeyEvent.wVirtualKeyCode) {
-                        case VK_LEFT:   return ctrl_pressed ? KEY_CTRL_ARROW_LEFT  : KEY_ARROW_LEFT;
-                        case VK_RIGHT:  return ctrl_pressed ? KEY_CTRL_ARROW_RIGHT : KEY_ARROW_RIGHT;
-                        case VK_UP:     return KEY_ARROW_UP;
-                        case VK_DOWN:   return KEY_ARROW_DOWN;
-                        case VK_HOME:   return KEY_HOME;
-                        case VK_END:    return KEY_END;
-                        case VK_DELETE: return KEY_DELETE;
-                        default:        continue;
-                    }
+                    continue;
                 }
 
                 if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
@@ -362,52 +315,6 @@ namespace console {
 #endif
     }
 
-    static char32_t decode_utf8(const std::string & input, size_t pos, size_t & advance) {
-        unsigned char c = static_cast<unsigned char>(input[pos]);
-        if ((c & 0x80u) == 0u) {
-            advance = 1;
-            return c;
-        }
-        if ((c & 0xE0u) == 0xC0u && pos + 1 < input.size()) {
-            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
-            if ((c1 & 0xC0u) != 0x80u) {
-                advance = 1;
-                return 0xFFFD;
-            }
-            advance = 2;
-            return ((c & 0x1Fu) << 6) | (static_cast<unsigned char>(input[pos + 1]) & 0x3Fu);
-        }
-        if ((c & 0xF0u) == 0xE0u && pos + 2 < input.size()) {
-            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
-            unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
-            if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u) {
-                advance = 1;
-                return 0xFFFD;
-            }
-            advance = 3;
-            return ((c & 0x0Fu) << 12) |
-                   ((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 6) |
-                   (static_cast<unsigned char>(input[pos + 2]) & 0x3Fu);
-        }
-        if ((c & 0xF8u) == 0xF0u && pos + 3 < input.size()) {
-            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
-            unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
-            unsigned char c3 = static_cast<unsigned char>(input[pos + 3]);
-            if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u || (c3 & 0xC0u) != 0x80u) {
-                advance = 1;
-                return 0xFFFD;
-            }
-            advance = 4;
-            return ((c & 0x07u) << 18) |
-                   ((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 12) |
-                   ((static_cast<unsigned char>(input[pos + 2]) & 0x3Fu) << 6) |
-                   (static_cast<unsigned char>(input[pos + 3]) & 0x3Fu);
-        }
-
-        advance = 1;
-        return 0xFFFD; // replacement character for invalid input
-    }
-
     static void append_utf8(char32_t ch, std::string & out) {
         if (ch <= 0x7F) {
             out.push_back(static_cast<unsigned char>(ch));
@@ -429,327 +336,22 @@ namespace console {
     }
 
     // Helper function to remove the last UTF-8 character from a string
-    static size_t prev_utf8_char_pos(const std::string & line, size_t pos) {
-        if (pos == 0) return 0;
-        pos--;
-        while (pos > 0 && (line[pos] & 0xC0) == 0x80) {
-            pos--;
-        }
-        return pos;
-    }
-
-    static size_t next_utf8_char_pos(const std::string & line, size_t pos) {
-        if (pos >= line.length()) return line.length();
-        pos++;
-        while (pos < line.length() && (line[pos] & 0xC0) == 0x80) {
-            pos++;
-        }
-        return pos;
-    }
-
-    static void move_cursor(int delta);
-    static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
-    static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
-    static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths);
-    static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
-
-    static void delete_at_cursor(std::string & line, std::vector<int> & widths, size_t & char_pos, size_t & byte_pos) {
-        if (char_pos >= widths.size()) {
+    static void pop_back_utf8_char(std::string & line) {
+        if (line.empty()) {
             return;
         }
 
-        size_t next_pos = next_utf8_char_pos(line, byte_pos);
-        int w = widths[char_pos];
-        size_t char_len = next_pos - byte_pos;
+        size_t pos = line.length() - 1;
 
-        line.erase(byte_pos, char_len);
-        widths.erase(widths.begin() + char_pos);
-
-        size_t p = byte_pos;
-        int tail_width = 0;
-        for (size_t i = char_pos; i < widths.size(); ++i) {
-            size_t following = next_utf8_char_pos(line, p);
-            put_codepoint(line.c_str() + p, following - p, widths[i]);
-            tail_width += widths[i];
-            p = following;
+        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
+        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
+            if ((line[pos] & 0xC0) != 0x80) {
+                break; // Found the start of the character
+            }
         }
-
-        for (int i = 0; i < w; ++i) {
-            fputc(' ', out);
-        }
-
-        move_cursor(-(tail_width + w));
+        line.erase(pos);
     }
 
-    static void clear_current_line(const std::vector<int> & widths) {
-        int total_width = 0;
-        for (int w : widths) {
-            total_width += (w > 0 ? w : 1);
-        }
-
-        if (total_width > 0) {
-            std::string spaces(total_width, ' ');
-            fwrite(spaces.c_str(), 1, total_width, out);
-            move_cursor(-total_width);
-        }
-    }
-
-    static void set_line_contents(std::string new_line, std::string & line, std::vector<int> & widths, size_t & char_pos,
-                                  size_t & byte_pos, int cursor_byte_pos = -1) {
-        move_to_line_start(char_pos, byte_pos, widths);
-        clear_current_line(widths);
-
-        line = std::move(new_line);
-        widths.clear();
-        byte_pos = 0;
-        char_pos = 0;
-
-        size_t idx = 0;
-        int back_width = 0;
-        while (idx < line.size()) {
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, idx, advance);
-            int expected_width = estimateWidth(cp);
-            int real_width = put_codepoint(line.c_str() + idx, advance, expected_width);
-            if (real_width < 0) real_width = 0;
-            widths.push_back(real_width);
-            idx += advance;
-            if (cursor_byte_pos >= 0 && static_cast<size_t>(cursor_byte_pos) < idx) {
-                back_width += real_width;
-            } else {
-                ++char_pos;
-                byte_pos = idx;
-            }
-        }
-        if (cursor_byte_pos >= 0) {
-            move_cursor(-back_width);
-        }
-    }
-
-    static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths) {
-        int back_width = 0;
-        for (size_t i = 0; i < char_pos; ++i) {
-            back_width += widths[i];
-        }
-        move_cursor(-back_width);
-        char_pos = 0;
-        byte_pos = 0;
-    }
-
-    static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
-        int forward_width = 0;
-        for (size_t i = char_pos; i < widths.size(); ++i) {
-            forward_width += widths[i];
-        }
-        move_cursor(forward_width);
-        char_pos = widths.size();
-        byte_pos = line.length();
-    }
-
-    static bool has_ctrl_modifier(const std::string & params) {
-        size_t start = 0;
-        while (start < params.size()) {
-            size_t end = params.find(';', start);
-            size_t len = (end == std::string::npos) ? params.size() - start : end - start;
-            if (len > 0) {
-                int value = 0;
-                for (size_t i = 0; i < len; ++i) {
-                    char ch = params[start + i];
-                    if (!std::isdigit(static_cast<unsigned char>(ch))) {
-                        value = -1;
-                        break;
-                    }
-                    value = value * 10 + (ch - '0');
-                }
-                if (value == 5) {
-                    return true;
-                }
-            }
-
-            if (end == std::string::npos) {
-                break;
-            }
-            start = end + 1;
-        }
-        return false;
-    }
-
-    static bool is_space_codepoint(char32_t cp) {
-        return std::iswspace(static_cast<wint_t>(cp)) != 0;
-    }
-
-    static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
-        if (char_pos == 0) {
-            return;
-        }
-
-        size_t new_char_pos = char_pos;
-        size_t new_byte_pos = byte_pos;
-        int move_width = 0;
-
-        while (new_char_pos > 0) {
-            size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, prev_byte, advance);
-            if (!is_space_codepoint(cp)) {
-                break;
-            }
-            move_width += widths[new_char_pos - 1];
-            new_char_pos--;
-            new_byte_pos = prev_byte;
-        }
-
-        while (new_char_pos > 0) {
-            size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, prev_byte, advance);
-            if (is_space_codepoint(cp)) {
-                break;
-            }
-            move_width += widths[new_char_pos - 1];
-            new_char_pos--;
-            new_byte_pos = prev_byte;
-        }
-
-        move_cursor(-move_width);
-        char_pos = new_char_pos;
-        byte_pos = new_byte_pos;
-    }
-
-    static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
-        if (char_pos >= widths.size()) {
-            return;
-        }
-
-        size_t new_char_pos = char_pos;
-        size_t new_byte_pos = byte_pos;
-        int move_width = 0;
-
-        while (new_char_pos < widths.size()) {
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, new_byte_pos, advance);
-            if (!is_space_codepoint(cp)) {
-                break;
-            }
-            move_width += widths[new_char_pos];
-            new_char_pos++;
-            new_byte_pos += advance;
-        }
-
-        while (new_char_pos < widths.size()) {
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, new_byte_pos, advance);
-            if (is_space_codepoint(cp)) {
-                break;
-            }
-            move_width += widths[new_char_pos];
-            new_char_pos++;
-            new_byte_pos += advance;
-        }
-
-        while (new_char_pos < widths.size()) {
-            size_t advance = 0;
-            char32_t cp = decode_utf8(line, new_byte_pos, advance);
-            if (!is_space_codepoint(cp)) {
-                break;
-            }
-            move_width += widths[new_char_pos];
-            new_char_pos++;
-            new_byte_pos += advance;
-        }
-
-        move_cursor(move_width);
-        char_pos = new_char_pos;
-        byte_pos = new_byte_pos;
-    }
-
-    static void move_cursor(int delta) {
-        if (delta == 0) return;
-#if defined(_WIN32)
-        if (hConsole != NULL) {
-            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
-            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
-            COORD newCursorPosition = bufferInfo.dwCursorPosition;
-            int width = bufferInfo.dwSize.X;
-            int newX = newCursorPosition.X + delta;
-            int newY = newCursorPosition.Y;
-
-            while (newX >= width) {
-                newX -= width;
-                newY++;
-            }
-            while (newX < 0) {
-                newX += width;
-                newY--;
-            }
-
-            newCursorPosition.X = newX;
-            newCursorPosition.Y = newY;
-            SetConsoleCursorPosition(hConsole, newCursorPosition);
-        }
-#else
-        if (delta < 0) {
-            for (int i = 0; i < -delta; i++) fprintf(out, "\b");
-        } else {
-            for (int i = 0; i < delta; i++) fprintf(out, "\033[C");
-        }
-#endif
-    }
-
-    struct history_t {
-        std::vector<std::string> entries;
-        size_t viewing_idx = SIZE_MAX;
-        std::string backup_line; // current line before viewing history
-        void add(const std::string & line) {
-            if (line.empty()) {
-                return;
-            }
-            // avoid duplicates with the last entry
-            if (entries.empty() || entries.back() != line) {
-                entries.push_back(line);
-            }
-            // also clear viewing state
-            end_viewing();
-        }
-        bool prev(std::string & cur_line) {
-            if (entries.empty()) {
-                return false;
-            }
-            if (viewing_idx == SIZE_MAX) {
-                return false;
-            }
-            if (viewing_idx > 0) {
-                viewing_idx--;
-            }
-            cur_line = entries[viewing_idx];
-            return true;
-        }
-        bool next(std::string & cur_line) {
-            if (entries.empty() || viewing_idx == SIZE_MAX) {
-                return false;
-            }
-            viewing_idx++;
-            if (viewing_idx >= entries.size()) {
-                cur_line = backup_line;
-                end_viewing();
-            } else {
-                cur_line = entries[viewing_idx];
-            }
-            return true;
-        }
-        void begin_viewing(const std::string & line) {
-            backup_line = line;
-            viewing_idx = entries.size();
-        }
-        void end_viewing() {
-            viewing_idx = SIZE_MAX;
-            backup_line.clear();
-        }
-        bool is_viewing() const {
-            return viewing_idx != SIZE_MAX;
-        }
-    } history;
-
     static bool readline_advanced(std::string & line, bool multiline_input) {
         if (out != stdout) {
             fflush(stdout);
@@ -760,33 +362,8 @@ namespace console {
         bool is_special_char = false;
         bool end_of_stream = false;
 
-        size_t byte_pos = 0; // current byte index
-        size_t char_pos = 0; // current character index (one char can be multiple bytes)
-
         char32_t input_char;
         while (true) {
-            assert(char_pos <= byte_pos);
-            assert(char_pos <= widths.size());
-            auto history_prev = [&]() {
-                if (!history.is_viewing()) {
-                    history.begin_viewing(line);
-                }
-                std::string new_line;
-                if (!history.prev(new_line)) {
-                    return;
-                }
-                set_line_contents(new_line, line, widths, char_pos, byte_pos);
-            };
-            auto history_next = [&]() {
-                if (history.is_viewing()) {
-                    std::string new_line;
-                    if (!history.next(new_line)) {
-                        return;
-                    }
-                    set_line_contents(new_line, line, widths, char_pos, byte_pos);
-                }
-            };
-
             fflush(out); // Ensure all output is displayed before waiting for input
             input_char = getchar32();
 
@@ -794,97 +371,20 @@ namespace console {
                 break;
             }
 
-            if (completion_cb && input_char == '\t') {
-                auto candidates = completion_cb(line, byte_pos);
-
-                if (!candidates.empty()) {
-                    if (candidates.size() > 1 || candidates[0].first != line) {
-                        // TODO?: Display all candidates
-                        set_line_contents(candidates[0].first, line, widths, char_pos, byte_pos, candidates[0].second);
-                    } else {
-                        // TODO: Move cursor to new byte_pos
-                    }
-                    continue;
-                }
-            }
-
-            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D */) {
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
                 end_of_stream = true;
                 break;
             }
 
             if (is_special_char) {
+                set_display(user_input);
                 replace_last(line.back());
                 is_special_char = false;
             }
 
             if (input_char == '\033') { // Escape sequence
                 char32_t code = getchar32();
-                if (code == '[') {
-                    std::string params;
-                    while (true) {
-                        code = getchar32();
-                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~' || code == (char32_t) WEOF) {
-                            break;
-                        }
-                        params.push_back(static_cast<char>(code));
-                    }
-
-                    const bool ctrl_modifier = has_ctrl_modifier(params);
-
-                    if (code == 'D') { // left
-                        if (ctrl_modifier) {
-                            move_word_left(char_pos, byte_pos, widths, line);
-                        } else if (char_pos > 0) {
-                            int w = widths[char_pos - 1];
-                            move_cursor(-w);
-                            char_pos--;
-                            byte_pos = prev_utf8_char_pos(line, byte_pos);
-                        }
-                    } else if (code == 'C') { // right
-                        if (ctrl_modifier) {
-                            move_word_right(char_pos, byte_pos, widths, line);
-                        } else if (char_pos < widths.size()) {
-                            int w = widths[char_pos];
-                            move_cursor(w);
-                            char_pos++;
-                            byte_pos = next_utf8_char_pos(line, byte_pos);
-                        }
-                    } else if (code == 'H') { // home
-                        move_to_line_start(char_pos, byte_pos, widths);
-                    } else if (code == 'F') { // end
-                        move_to_line_end(char_pos, byte_pos, widths, line);
-                    } else if (code == 'A' || code == 'B') {
-                        // up/down
-                        if (code == 'A') {
-                            history_prev();
-                            is_special_char = false;
-                        } else if (code == 'B') {
-                            history_next();
-                            is_special_char = false;
-                        }
-                    } else if ((code == '~' || (code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z')) && !params.empty()) {
-                        std::string digits;
-                        for (char ch : params) {
-                            if (ch == ';') {
-                                break;
-                            }
-                            if (std::isdigit(static_cast<unsigned char>(ch))) {
-                                digits.push_back(ch);
-                            }
-                        }
-
-                        if (code == '~') {
-                            if (digits == "1" || digits == "7") { // home
-                                move_to_line_start(char_pos, byte_pos, widths);
-                            } else if (digits == "4" || digits == "8") { // end
-                                move_to_line_end(char_pos, byte_pos, widths, line);
-                            } else if (digits == "3") { // delete
-                                delete_at_cursor(line, widths, char_pos, byte_pos);
-                            }
-                        }
-                    }
-                } else if (code == 0x1B) {
+                if (code == '[' || code == 0x1B) {
                     // Discard the rest of the escape sequence
                     while ((code = getchar32()) != (char32_t) WEOF) {
                         if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
@@ -892,110 +392,32 @@ namespace console {
                         }
                     }
                 }
-#if defined(_WIN32)
-            } else if (input_char == KEY_ARROW_LEFT) {
-                if (char_pos > 0) {
-                    int w = widths[char_pos - 1];
-                    move_cursor(-w);
-                    char_pos--;
-                    byte_pos = prev_utf8_char_pos(line, byte_pos);
-                }
-            } else if (input_char == KEY_ARROW_RIGHT) {
-                if (char_pos < widths.size()) {
-                    int w = widths[char_pos];
-                    move_cursor(w);
-                    char_pos++;
-                    byte_pos = next_utf8_char_pos(line, byte_pos);
-                }
-            } else if (input_char == KEY_CTRL_ARROW_LEFT) {
-                move_word_left(char_pos, byte_pos, widths, line);
-            } else if (input_char == KEY_CTRL_ARROW_RIGHT) {
-                move_word_right(char_pos, byte_pos, widths, line);
-            } else if (input_char == KEY_HOME) {
-                move_to_line_start(char_pos, byte_pos, widths);
-            } else if (input_char == KEY_END) {
-                move_to_line_end(char_pos, byte_pos, widths, line);
-            } else if (input_char == KEY_DELETE) {
-                delete_at_cursor(line, widths, char_pos, byte_pos);
-            } else if (input_char == KEY_ARROW_UP || input_char == KEY_ARROW_DOWN) {
-                if (input_char == KEY_ARROW_UP) {
-                    history_prev();
-                    is_special_char = false;
-                } else if (input_char == KEY_ARROW_DOWN) {
-                    history_next();
-                    is_special_char = false;
-                }
-#endif
             } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
-                if (char_pos > 0) {
-                    int w = widths[char_pos - 1];
-                    move_cursor(-w);
-                    char_pos--;
-                    size_t prev_pos = prev_utf8_char_pos(line, byte_pos);
-                    size_t char_len = byte_pos - prev_pos;
-                    byte_pos = prev_pos;
-
-                    // remove the character
-                    line.erase(byte_pos, char_len);
-                    widths.erase(widths.begin() + char_pos);
-
-                    // redraw tail
-                    size_t p = byte_pos;
-                    int tail_width = 0;
-                    for (size_t i = char_pos; i < widths.size(); ++i) {
-                        size_t next_p = next_utf8_char_pos(line, p);
-                        put_codepoint(line.c_str() + p, next_p - p, widths[i]);
-                        tail_width += widths[i];
-                        p = next_p;
-                    }
-
-                    // clear display
-                    for (int i = 0; i < w; ++i) {
-                        fputc(' ', out);
-                    }
-                    move_cursor(-(tail_width + w));
+                if (!widths.empty()) {
+                    int count;
+                    do {
+                        count = widths.back();
+                        widths.pop_back();
+                        // Move cursor back, print space, and move cursor back again
+                        for (int i = 0; i < count; i++) {
+                            replace_last(' ');
+                            pop_cursor();
+                        }
+                        pop_back_utf8_char(line);
+                    } while (count == 0 && !widths.empty());
                 }
             } else {
-                // insert character
-                std::string new_char_str;
-                append_utf8(input_char, new_char_str);
-                int w = estimateWidth(input_char);
-
-                if (char_pos == widths.size()) {
-                    // insert at the end
-                    line += new_char_str;
-                    int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
-                    if (real_w < 0) real_w = 0;
-                    widths.push_back(real_w);
-                    byte_pos += new_char_str.length();
-                    char_pos++;
-                } else {
-                    // insert in middle
-                    line.insert(byte_pos, new_char_str);
-
-                    int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
-                    if (real_w < 0) real_w = 0;
-
-                    widths.insert(widths.begin() + char_pos, real_w);
-
-                    // print the tail
-                    size_t p = byte_pos + new_char_str.length();
-                    int tail_width = 0;
-                    for (size_t i = char_pos + 1; i < widths.size(); ++i) {
-                        size_t next_p = next_utf8_char_pos(line, p);
-                        put_codepoint(line.c_str() + p, next_p - p, widths[i]);
-                        tail_width += widths[i];
-                        p = next_p;
-                    }
-
-                    move_cursor(-tail_width);
-
-                    byte_pos += new_char_str.length();
-                    char_pos++;
+                int offset = line.length();
+                append_utf8(input_char, line);
+                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
+                if (width < 0) {
+                    width = 0;
                 }
+                widths.push_back(width);
             }
 
             if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
+                set_display(prompt);
                 replace_last(line.back());
                 is_special_char = true;
             }
@@ -1029,15 +451,6 @@ namespace console {
             }
         }
 
-        if (!end_of_stream && !line.empty()) {
-            // remove the trailing newline for history storage
-            if (!line.empty() && line.back() == '\n') {
-                line.pop_back();
-            }
-            // TODO: maybe support multiline history entries?
-            history.add(line);
-        }
-
         fflush(out);
         return has_more;
     }
@@ -1080,86 +493,12 @@ namespace console {
     }
 
     bool readline(std::string & line, bool multiline_input) {
+        set_display(user_input);
+
         if (simple_io) {
             return readline_simple(line, multiline_input);
         }
         return readline_advanced(line, multiline_input);
     }
 
-    void set_completion_callback(completion_callback cb) {
-        completion_cb = cb;
-    }
-
-    namespace spinner {
-        static const char LOADING_CHARS[] = {'|', '/', '-', '\\'};
-        static std::condition_variable cv_stop;
-        static std::thread th;
-        static size_t frame = 0; // only modified by one thread
-        static bool running = false;
-        static std::mutex mtx;
-        static auto wait_time = std::chrono::milliseconds(100);
-        static void draw_next_frame() {
-            // don't need lock because only one thread modifies running
-            frame = (frame + 1) % sizeof(LOADING_CHARS);
-            replace_last(LOADING_CHARS[frame]);
-            fflush(out);
-        }
-        void start() {
-            std::unique_lock<std::mutex> lock(mtx);
-            if (simple_io || running) {
-                return;
-            }
-            common_log_flush(common_log_main());
-            fprintf(out, "%c", LOADING_CHARS[0]);
-            fflush(out);
-            frame = 1;
-            running = true;
-            th = std::thread([]() {
-                std::unique_lock<std::mutex> lock(mtx);
-                while (true) {
-                    if (cv_stop.wait_for(lock, wait_time, []{ return !running; })) {
-                        break;
-                    }
-                    draw_next_frame();
-                }
-            });
-        }
-        void stop() {
-            {
-                std::unique_lock<std::mutex> lock(mtx);
-                if (simple_io || !running) {
-                    return;
-                }
-                running = false;
-                cv_stop.notify_all();
-            }
-            if (th.joinable()) {
-                th.join();
-            }
-            replace_last(' ');
-            pop_cursor();
-            fflush(out);
-        }
-    }
-
-    void log(const char * fmt, ...) {
-        va_list args;
-        va_start(args, fmt);
-        vfprintf(out, fmt, args);
-        va_end(args);
-    }
-
-    void error(const char * fmt, ...) {
-        va_list args;
-        va_start(args, fmt);
-        display_type cur = current_display;
-        set_display(DISPLAY_TYPE_ERROR);
-        vfprintf(out, fmt, args);
-        set_display(cur); // restore previous color
-        va_end(args);
-    }
-
-    void flush() {
-        fflush(out);
-    }
 }