cont [no ci]

ggml-ci

.devops/cpu.Dockerfile

@@ -1,92 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-
-FROM ubuntu:$UBUNTU_VERSION AS build
-
-ARG TARGETARCH
-
-ARG GGML_CPU_ARM_ARCH=armv8-a
-
-RUN apt-get update && \
-    apt-get install -y build-essential git cmake libcurl4-openssl-dev
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
-    elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
-    else \
-        echo "Unsupported architecture"; \
-        exit 1; \
-    fi && \
-    cmake --build build -j $(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
-
-RUN mkdir -p /app/full \
-    && cp build/bin/* /app/full \
-    && cp *.py /app/full \
-    && cp -r gguf-py /app/full \
-    && cp -r requirements /app/full \
-    && cp requirements.txt /app/full \
-    && cp .devops/tools.sh /app/full/tools.sh
-
-## Base image
-FROM ubuntu:$UBUNTU_VERSION 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 \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -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
-
-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

@@ -1,94 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.6.0
-# 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 ${BASE_CUDA_DEV_CONTAINER} AS build
-
-# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=default
-
-RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-    export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release -j$(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /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 \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -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
-
-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/full-cuda.Dockerfile

@@ -0,0 +1,33 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=12.6.0
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release -j$(nproc) && \
+    cp build/bin/* .
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-musa.Dockerfile

@@ -0,0 +1,33 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release -j$(nproc) && \
+    cp build/bin/* .
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-rocm.Dockerfile

@@ -0,0 +1,50 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH="\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102"
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV GGML_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
+RUN make -j$(nproc)
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full.Dockerfile

@@ -0,0 +1,38 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build -j $(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+WORKDIR /app
+
+RUN apt-get update && \
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
+
+COPY requirements.txt   /app/requirements.txt
+COPY requirements       /app/requirements
+COPY .devops/tools.sh   /app/tools.sh
+
+RUN pip install --upgrade pip setuptools wheel && \
+    pip install -r /app/requirements.txt
+
+COPY --from=build /app/build/bin/ /app/
+COPY --from=build /app/lib/ /app/
+COPY --from=build /app/convert_hf_to_gguf.py /app/
+COPY --from=build /app/gguf-py /app/gguf-py
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT ["/app/tools.sh"]

.devops/intel.Dockerfile

@@ -1,91 +0,0 @@
-ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
-
-## Build Image
-
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
-
-ARG GGML_SYCL_F16=OFF
-RUN apt-get update && \
-    apt-get install -y git libcurl4-openssl-dev
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
-        echo "GGML_SYCL_F16 is set" \
-        && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
-    fi && \
-    echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
-    cmake --build build --config Release -j$(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /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
-
-FROM intel/oneapi-basekit:$ONEAPI_VERSION 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
-
-### Full
-FROM base AS full
-
-COPY --from=build /app/lib/ /app
-COPY --from=build /app/full /app
-
-WORKDIR /app
-
-RUN apt-get update \
-    && apt-get install -y \
-    git \
-    python3 \
-    python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -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/lib/ /app
-COPY --from=build /app/full/llama-cli /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/lib/ /app
-COPY --from=build /app/full/llama-server /app
-
-WORKDIR /app
-
-HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
-
-ENTRYPOINT [ "/app/llama-server" ]
-

.devops/llama-cli-cuda.Dockerfile

@@ -0,0 +1,38 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=12.6.0
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/lib/ /
+COPY --from=build /app/build/bin/llama-cli /
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-intel.Dockerfile

@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+
+ARG GGML_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
+        echo "GGML_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
+    fi && \
+    echo "Building with static libs" && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
+    ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
+    cmake --build build --config Release --target llama-cli
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-musa.Dockerfile

@@ -0,0 +1,38 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/lib/ /
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-rocm.Dockerfile

@@ -0,0 +1,45 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH="\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102"
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV GGML_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make -j$(nproc) llama-cli
+
+ENTRYPOINT [ "/app/llama-cli" ]

.devops/llama-cli-vulkan.Dockerfile

@@ -0,0 +1,27 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget libgomp1
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 && \
+    cmake --build build --config Release --target llama-cli
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/llama-cli /llama-cli && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build -j $(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+WORKDIR /app
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/build/bin/llama-cli /app/
+COPY --from=build /app/lib/ /app/
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/app/llama-cli" ]

.devops/llama-server-cuda.Dockerfile

@@ -0,0 +1,43 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=12.6.0
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/lib/ /
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-intel.Dockerfile

@@ -0,0 +1,34 @@
+ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+
+ARG GGML_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
+        echo "GGML_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
+    fi && \
+    echo "Building with dynamic libs" && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
+    cmake --build build --config Release --target llama-server
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev curl
+
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-musa.Dockerfile

@@ -0,0 +1,43 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/lib/ /
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-rocm.Dockerfile

@@ -0,0 +1,54 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH="\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102"
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV GGML_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev curl
+
+RUN make -j$(nproc) llama-server
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/app/llama-server" ]

.devops/llama-server-vulkan.Dockerfile

@@ -0,0 +1,31 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK and cURL
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+    cmake --build build --config Release --target llama-server
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/llama-server /llama-server && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server.Dockerfile

@@ -0,0 +1,33 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build -j $(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+WORKDIR /app
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/build/bin/llama-server /app/
+COPY --from=build /app/lib/ /app/
+
+ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/app/llama-server" ]

.devops/musa.Dockerfile

@@ -1,108 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-# This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc3.1.0
-# Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_MUSA_DEV_CONTAINER} AS build
-
-# MUSA architecture to build for (defaults to all supported archs)
-ARG MUSA_DOCKER_ARCH=default
-
-RUN apt-get update && \
-    apt-get install -y \
-    build-essential \
-    cmake \
-    python3 \
-    python3-pip \
-    git \
-    libcurl4-openssl-dev \
-    libgomp1
-
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
-
-RUN pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt
-
-WORKDIR /app
-
-COPY . .
-
-# Use the default MUSA archs if not specified
-RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release -j$(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /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_MUSA_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 \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -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
-
-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/rocm.Dockerfile

@@ -1,113 +0,0 @@
-ARG UBUNTU_VERSION=24.04
-
-# This needs to generally match the container host's environment.
-ARG ROCM_VERSION=6.3
-ARG AMDGPU_VERSION=6.3
-
-# Target the CUDA build image
-ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
-
-### Build image
-FROM ${BASE_ROCM_DEV_CONTAINER} AS build
-
-# Unless otherwise specified, we make a fat build.
-# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
-# This is mostly tied to rocBLAS supported archs.
-# gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
-# gfx906 is deprecated
-#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
-
-#ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
-ARG ROCM_DOCKER_ARCH=gfx1100
-
-# Set nvcc architectured
-ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
-# Enable ROCm
-# ENV CC=/opt/rocm/llvm/bin/clang
-# ENV CXX=/opt/rocm/llvm/bin/clang++
-
-RUN apt-get update \
-    && apt-get install -y \
-    build-essential \
-    cmake \
-    git \
-    libcurl4-openssl-dev \
-    curl \
-    libgomp1
-
-WORKDIR /app
-
-COPY . .
-
-RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
-    && cmake --build build --config Release -j$(nproc)
-
-RUN mkdir -p /app/lib \
-    && find build -name "*.so" -exec cp {} /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_ROCM_DEV_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-pip \
-    python3 \
-    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
-
-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/tools.sh

@@ -13,13 +13,9 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     exec ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     exec ./llama-cli "$@"
-elif [[ "$arg1" == '--bench' || "$arg1" == '-b' ]]; then
-    exec ./llama-bench "$@"
-elif [[ "$arg1" == '--perplexity' || "$arg1" == '-p' ]]; then
-    exec ./llama-perplexity "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
-    for i in $(ls $1/$2/ggml-model-f16.bin*); do
+    for i in `ls $1/$2/ggml-model-f16.bin*`; do
         if [ -f "${i/f16/q4_0}" ]; then
             echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
         else
@@ -34,10 +30,6 @@ else
     echo "Available commands: "
     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."
-    echo "              ex: -m model.gguf -f file.txt"
     echo "  --convert (-c): Convert a llama model into ggml"
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"

.devops/vulkan.Dockerfile

@@ -1,89 +0,0 @@
-ARG UBUNTU_VERSION=24.04
-
-FROM ubuntu:$UBUNTU_VERSION AS build
-
-# Install build tools
-RUN apt update && apt install -y git build-essential cmake wget
-
-# Install Vulkan SDK and cURL
-RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
-    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list https://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
-    apt update -y && \
-    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
-
-# Build it
-WORKDIR /app
-
-COPY . .
-
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
-    cmake --build build --config Release -j$(nproc)
-
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /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 ubuntu:$UBUNTU_VERSION AS base
-
-RUN apt-get update \
-    && apt-get install -y libgomp1 curl libvulkan-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 --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
-
-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" ]

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

@@ -65,22 +65,12 @@ body:
         If possible, please do a git bisect and identify the exact commit that introduced the bug.
     validations:
       required: false
-  - type: textarea
-    id: command
-    attributes:
-      label: Compile command
-      description: >
-        Please provide the exact command you used to compile llama.cpp. For example: `cmake -B ...`.
-        This will be automatically formatted into code, so no need for backticks.
-      render: shell
-    validations:
-      required: true
   - type: textarea
     id: logs
     attributes:
       label: Relevant log output
       description: >
-          Please copy and paste any relevant log output, including any generated text.
+          Please copy and paste any relevant log output, including the command that you entered and any generated text.
           This will be automatically formatted into code, so no need for backticks.
       render: shell
     validations:

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

@@ -52,16 +52,6 @@ body:
         - Other (Please specify in the next section)
     validations:
       required: false
-  - type: textarea
-    id: command
-    attributes:
-      label: Command line
-      description: >
-        Please provide the exact commands you entered, if applicable. For example: `llama-server -m ... -c ...`, `llama-cli -m ...`, etc.
-        This will be automatically formatted into code, so no need for backticks.
-      render: shell
-    validations:
-      required: false
   - type: textarea
     id: info
     attributes:
@@ -84,7 +74,7 @@ body:
     attributes:
       label: Relevant log output
       description: >
-          If applicable, please copy and paste any relevant log output, including any generated text.
+          If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
           This will be automatically formatted into code, so no need for backticks.
       render: shell
     validations:

.github/workflows/build.yml

@@ -56,12 +56,12 @@ jobs:
           mkdir build
           cd build
           cmake .. \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_CURL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -88,7 +88,6 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
-          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -121,11 +120,11 @@ jobs:
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_CURL=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -152,7 +151,6 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
-          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -162,8 +160,8 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
           name: llama-bin-macos-x64.zip
 
-  ubuntu-cpu-cmake:
-    runs-on: ubuntu-22.04
+  ubuntu-latest-cmake:
+    runs-on: ubuntu-latest
 
     steps:
       - name: Clone
@@ -183,10 +181,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_CURL=ON \
-            -DGGML_RPC=ON
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -224,7 +219,6 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
-          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -242,7 +236,7 @@ jobs:
     strategy:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug]
+        build_type: [Debug, Release]
 
     steps:
       - name: Clone
@@ -261,10 +255,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Build (no OpenMP)
@@ -273,11 +264,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-            -DGGML_OPENMP=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DGGML_OPENMP=OFF
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -307,8 +294,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. \
-            -DGGML_RPC=ON
+          cmake -DGGML_RPC=ON ..
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -331,24 +317,16 @@ jobs:
           wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
           sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk
+          sudo apt-get install -y build-essential vulkan-sdk
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. \
-            -DGGML_VULKAN=ON
+          cmake -DGGML_VULKAN=ON ..
           cmake --build . --config Release -j $(nproc)
 
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 1800
-
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.0.2
@@ -367,18 +345,13 @@ jobs:
       - name: Build with native CMake HIP support
         id: cmake_build
         run: |
-          cmake -B build -S . \
-            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DGGML_HIP=ON
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
       - name: Build with legacy HIP support
         id: cmake_build_legacy_hip
         run: |
-          cmake -B build2 -S . \
-            -DCMAKE_C_COMPILER=hipcc \
-            -DCMAKE_CXX_COMPILER=hipcc \
-            -DGGML_HIP=ON
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIP=ON
           cmake --build build2 --config Release -j $(nproc)
 
   ubuntu-22-cmake-musa:
@@ -399,8 +372,7 @@ jobs:
       - name: Build with native CMake MUSA support
         id: cmake_build
         run: |
-          cmake -B build -S . \
-            -DGGML_MUSA=ON
+          cmake -B build -S . -DGGML_MUSA=ON
           cmake --build build --config Release -j $(nproc)
 
   ubuntu-22-cmake-sycl:
@@ -441,10 +413,7 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake .. \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
   ubuntu-22-cmake-sycl-fp16:
@@ -485,13 +454,42 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake .. \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx \
-            -DGGML_SYCL_F16=ON
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON ..
           cmake --build . --config Release -j $(nproc)
 
+  # TODO: build with GGML_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
+  #       would be great if we fix these
+  macOS-latest-cmake:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -614,7 +612,6 @@ jobs:
           msystem: ${{matrix.sys}}
           install: >-
             base-devel
-            git
             mingw-w64-${{matrix.env}}-toolchain
             mingw-w64-${{matrix.env}}-cmake
             mingw-w64-${{matrix.env}}-openblas
@@ -648,23 +645,23 @@ jobs:
       matrix:
         include:
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64-opencl-adreno'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
@@ -795,7 +792,6 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -823,13 +819,7 @@ jobs:
 
         - name: Build with CMake
           run: |
-            cmake -S . -B build -G Ninja \
-              -DCMAKE_BUILD_TYPE=Release \
-              -DCMAKE_CUDA_ARCHITECTURES=89-real \
-              -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
-              -DLLAMA_FATAL_WARNINGS=ON \
-              -DGGML_NATIVE=OFF \
-              -DGGML_CUDA=ON
+            cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
             cmake --build build
 
   windows-2019-cmake-cuda:
@@ -918,11 +908,7 @@ jobs:
         shell: cmd
         run: |
           call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
-          cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DGGML_NATIVE=OFF ^
-            -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1075,12 +1061,7 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DGGML_HIP=ON `
-            -DGGML_RPC=ON
+          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" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
   windows-latest-cmake-hip-release:
@@ -1118,13 +1099,7 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . `
-            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_BUILD_TYPE=Release `
-            -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
-            -DGGML_HIP=ON `
-            -DGGML_RPC=ON
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
@@ -1218,7 +1193,8 @@ jobs:
     runs-on: ubuntu-latest
 
     needs:
-      - ubuntu-cpu-cmake
+      - ubuntu-latest-cmake
+      - macOS-latest-cmake
       - windows-latest-cmake
       - windows-2019-cmake-cuda
       - windows-latest-cmake-hip-release
@@ -1257,7 +1233,7 @@ jobs:
 
       - name: Create release
         id: create_release
-        uses: ggml-org/action-create-release@v1
+        uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
@@ -1477,37 +1453,3 @@ jobs:
 #          popd
 #          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
 #          make
-
-  openEuler-latest-cmake-cann:
-    if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
-    defaults:
-      run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
-    strategy:
-      matrix:
-        cann:
-          - '8.0.rc3.beta1-910b-openeuler22.03-py3.10'
-        device:
-          - 'ascend910b3'
-        build:
-          - 'Release'
-    container: ascendai/cann:${{ matrix.cann }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        run: |
-          yum update -y
-          yum install -y git gcc gcc-c++ make cmake
-
-      - name: Build
-        run: |
-          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=${{ matrix.build }} \
-              -DGGML_CANN=on \
-              -DSOC_TYPE=${{ matrix.device }}
-          cmake --build build -j $(nproc)

.github/workflows/docker.yml

@@ -28,21 +28,27 @@ jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
 
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-latest
     env:
       COMMIT_SHA: ${{ github.sha }}
     strategy:
-      fail-fast: false
       matrix:
         config:
-          # Multi-stage build
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
+          - { tag: "light", dockerfile: ".devops/llama-cli.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/llama-server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-musa", dockerfile: ".devops/llama-cli-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-musa", dockerfile: ".devops/llama-server-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-musa", dockerfile: ".devops/full-musa.Dockerfile", platforms: "linux/amd64" }
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
-          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true }
+          #- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          #- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          #- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
@@ -50,10 +56,10 @@ jobs:
           fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v2
 
       - name: Log in to Docker Hub
         uses: docker/login-action@v2
@@ -73,34 +79,26 @@ jobs:
 
           # determine tag name postfix (build number, commit hash)
           if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
-            TAG_POSTFIX="-b${BUILD_NUMBER}"
+            TAG_POSTFIX="b${BUILD_NUMBER}"
           else
             SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
-            TAG_POSTFIX="-${SAFE_NAME}-${SHORT_HASH}"
+            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
           fi
+
           # list all tags possible
-          if [[ "${{ 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" >> $GITHUB_OUTPUT
-          echo "full_output_tags=$FULLTAGS"  # print out for debugging
-          echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
-          echo "server_output_tags=$SERVERTAGS"  # print out for debugging
+          TAGS=""
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
+
+          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
+          echo "output_tags=$TAGS"  # print out for debugging
         env:
           GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
+      # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
-        if: ${{ matrix.config.free_disk_space == true }}
-        uses: ggml-org/free-disk-space@v1.3.1
+        uses: jlumbroso/free-disk-space@main
         with:
           # this might remove tools that are actually needed,
           # if set to "true" but frees about 6 GB
@@ -115,59 +113,13 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Build and push Full Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.full == true }}
+      - name: Build and push Docker image (tagged + versioned)
+        if: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
         uses: docker/build-push-action@v6
         with:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
           # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.full_output_tags }}
+          tags: ${{ steps.tag.outputs.output_tags }}
           file: ${{ matrix.config.dockerfile }}
-          target: full
-          provenance: false
-          # using github experimental cache
-          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
-
-      - 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 }}
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.light_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: light
-          provenance: false
-          # using github experimental cache
-          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
-
-      - 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 }}
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.server_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: server
-          provenance: false
-          # using github experimental cache
-          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

.github/workflows/editorconfig.yml

@@ -23,7 +23,5 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - uses: editorconfig-checker/action-editorconfig-checker@v2
-        with:
-          version: v3.0.3
+      - uses: editorconfig-checker/action-editorconfig-checker@main
       - run: editorconfig-checker

.github/workflows/server.yml

@@ -112,9 +112,9 @@ jobs:
               -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' }}
+      - name: Build
+        id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
         run: |
           cmake -B build \
               -DGGML_NATIVE=OFF \
@@ -124,31 +124,12 @@ jobs:
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
-      - name: Build (sanitizers)
-        id: cmake_build
-        if: ${{ matrix.sanitizer == '' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DLLAMA_CURL=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
       - name: Tests
         id: server_integration_tests
-        if: ${{ matrix.sanitizer == '' }}
         run: |
           cd examples/server/tests
           ./tests.sh
 
-      - name: Tests (sanitizers)
-        id: server_integration_tests_sanitizers
-        if: ${{ matrix.sanitizer != '' }}
-        run: |
-          cd examples/server/tests
-          LLAMA_SANITIZE=1 ./tests.sh
-
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}

.gitignore

@@ -18,7 +18,6 @@
 *.metallib
 *.o
 *.so
-*.swp
 *.tmp
 
 # IDE / OS

CMakeLists.txt

@@ -16,7 +16,6 @@ endif()
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
 if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
@@ -50,8 +49,6 @@ endif()
 if (MSVC)
     add_compile_options("$<$<COMPILE_LANGUAGE:C>:/utf-8>")
     add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/utf-8>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/bigobj>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
 endif()
 
 #
@@ -86,8 +83,11 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
 # override ggml options
-set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
-set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
+set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
+set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
+set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
+set(GGML_ALL_WARNINGS       ${LLAMA_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
@@ -117,62 +117,16 @@ 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)
 
-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
+# build the library
 #
 
 if (NOT TARGET ggml)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
-
-#
-# build the library
-#
-
 add_subdirectory(src)
 
-#
-# utils, programs, examples and tests
-#
-
-if (LLAMA_BUILD_COMMON)
-    add_subdirectory(common)
-endif()
-
-if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    include(CTest)
-    add_subdirectory(tests)
-endif()
-
-if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
-    add_subdirectory(examples)
-    add_subdirectory(pocs)
-endif()
-
 #
 # install
 #
@@ -188,14 +142,27 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
+# At the moment some compile definitions are placed within the ggml/src
+# directory but not exported on the `ggml` target. This could be improved by
+# determining _precisely_ which defines are necessary for the llama-config
+# package.
+#
+set(GGML_TRANSIENT_DEFINES)
+get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
+get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
+if (GGML_DIR_DEFINES)
+    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_DIR_DEFINES})
+endif()
+get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
+if (GGML_TARGET_DEFINES)
+    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
+endif()
+get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
+# all public headers
 set(LLAMA_PUBLIC_HEADERS
     ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
     ${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
-
-set_target_properties(llama
-    PROPERTIES
-        PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
-
+set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(
@@ -233,3 +200,21 @@ configure_file(cmake/llama.pc.in
 
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
+
+#
+# utils, programs, examples and tests
+#
+
+if (LLAMA_BUILD_COMMON)
+    add_subdirectory(common)
+endif()
+
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+    include(CTest)
+    add_subdirectory(tests)
+endif()
+
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+    add_subdirectory(pocs)
+endif()

CODEOWNERS

@@ -1,11 +1,5 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
 
 /ci/ @ggerganov
-/.devops/*.Dockerfile @ngxson
+/.devops/ @ngxson
 /examples/server/ @ngxson
-/ggml/src/ggml-cuda/fattn* @JohannesGaessler
-/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
-/ggml/src/ggml-cuda/mmv.* @JohannesGaessler
-/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
-/ggml/src/ggml-opt.cpp @JohannesGaessler
-/ggml/src/gguf.cpp @JohannesGaessler

CONTRIBUTING.md

@@ -1,10 +1,10 @@
 # Pull requests (for contributors)
 
 - 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`
+  - 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`
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
@@ -20,104 +20,14 @@
 - Avoid adding third-party dependencies, extra files, extra headers, etc.
 - Always consider cross-compatibility with other operating systems and architectures
 - Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
-- Vertical alignment makes things more readable and easier to batch edit
+- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
-- Use sized integer types such as `int32_t` in the public API, e.g. `size_t` may also be appropriate for allocation sizes or byte offsets
-- Declare structs with `struct foo {}` instead of `typedef struct foo {} foo`
-    - In C++ code omit optional `struct` and `enum` keyword whenever they are not necessary
-    ```cpp
-    // OK
-    llama_context * ctx;
-    const llama_rope_type rope_type;
-
-    // not OK
-    struct llama_context * ctx;
-    const enum llama_rope_type rope_type;
-    ```
-
-    _(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline.)_
-
-- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code
-- For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines)
+- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
 - Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
 
 ![matmul](media/matmul.png)
 
-# Naming guidelines
-
-- Use `snake_case` for function, variable and type names
-- Naming usually optimizes for longest common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
-
-    ```cpp
-    // not OK
-    int small_number;
-    int big_number;
-
-    // OK
-    int number_small;
-    int number_big;
-    ```
-
-- Enum values are always in upper case and prefixed with the enum name
-
-    ```cpp
-    enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_NONE = 0,
-        LLAMA_VOCAB_TYPE_SPM  = 1,
-        LLAMA_VOCAB_TYPE_BPE  = 2,
-        LLAMA_VOCAB_TYPE_WPM  = 3,
-        LLAMA_VOCAB_TYPE_UGM  = 4,
-        LLAMA_VOCAB_TYPE_RWKV = 5,
-    };
-    ```
-
-- The general naming pattern is `<class>_<method>`, with `<method>` being `<action>_<noun>`
-
-    ```cpp
-    llama_model_init();           // class: "llama_model",         method: "init"
-    llama_sampler_chain_remove(); // class: "llama_sampler_chain", method: "remove"
-    llama_sampler_get_seed();     // class: "llama_sampler",       method: "get_seed"
-    llama_set_embeddings();       // class: "llama_context",       method: "set_embeddings"
-    llama_n_threads();            // class: "llama_context",       method: "n_threads"
-    llama_adapter_lora_free();    // class: "llama_adapter_lora",  method: "free"
-    ```
-
-    - The `get` `<action>` can be omitted
-    - The `<noun>` can be omitted if not necessary
-    - The `_context` suffix of the `<class>` is optional. Use it to disambiguate symbols when needed
-    - Use `init`/`free` for constructor/destructor `<action>`
-
-- Use the `_t` suffix when a type is supposed to be opaque to the user - it's not relevant to them if it is a struct or anything else
-
-    ```cpp
-    typedef struct llama_context * llama_context_t;
-
-    enum llama_pooling_type llama_pooling_type(const llama_context_t ctx);
-    ```
-
-    _(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline)_
-
-- C/C++ filenames are all lowercase with dashes. Headers use the `.h` extension. Source files use the `.c` or `.cpp` extension
-- Python filenames are all lowercase with underscores
-
-- _(TODO: abbreviations usage)_
-
-# Preprocessor directives
-
-- _(TODO: add guidelines with examples and apply them to the codebase)_
-
-    ```cpp
-    #ifdef FOO
-    #endif // FOO
-    ```
-
-# Documentation
-
-- Documentation is a community effort
-- When you need to look into the source code to figure out how to use an API consider adding a short summary to the header file for future reference
-- When you notice incorrect or outdated documentation, please update it
-
 # Resources
 
 The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:

Makefile

@@ -1361,9 +1361,7 @@ llama-server: \
 	examples/server/httplib.h \
 	examples/server/index.html.hpp \
 	examples/server/loading.html.hpp \
-	common/chat-template.hpp \
 	common/json.hpp \
-	common/minja.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)

README.md

@@ -16,10 +16,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggerganov/llama.cpp/pull/11427
-- **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/ggerganov/llama.cpp/discussions/10123
+- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
@@ -72,7 +69,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
 - [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
 - [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
-- [x] [PhiMoE](https://github.com/ggerganov/llama.cpp/pull/11003)
 - [x] [GPT-2](https://huggingface.co/gpt2)
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
@@ -102,8 +98,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [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-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)
 
 #### Multimodal
 
@@ -206,8 +200,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
-- [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
-- [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
 
 </details>
 
@@ -228,7 +220,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
 | [MUSA](docs/build.md#musa) | Moore Threads MTT GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
-| [HIP](docs/build.md#hip) | AMD GPU |
+| [hipBLAS](docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 
@@ -249,8 +241,6 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from Hugging Face by using this CLI argument: `-hf <user>/<model>[:quant]`
-
 After downloading a model, use the CLI tools to run it locally - see below.
 
 `llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
@@ -269,12 +259,21 @@ To learn more about model quantization, [read this documentation](examples/quant
 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
 
 - <details open>
-    <summary>Run in conversation mode</summary>
-
-    Models with a built-in chat template will automatically activate conversation mode. If this doesn't occur, you can manually enable it by adding `-cnv` and specifying a suitable chat template with `--chat-template NAME`
+    <summary>Run simple text completion</summary>
 
     ```bash
-    llama-cli -m model.gguf
+    llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128
+
+    # 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>Run in conversation mode</summary>
+
+    ```bash
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv
 
     # > hi, who are you?
     # Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
@@ -286,28 +285,17 @@ To learn more about model quantization, [read this documentation](examples/quant
     </details>
 
 - <details>
-    <summary>Run in conversation mode with custom chat template</summary>
+    <summary>Run with custom chat template</summary>
 
     ```bash
-    # use the "chatml" template (use -h to see the list of supported templates)
-    llama-cli -m model.gguf -cnv --chat-template chatml
+    # use the "chatml" template
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
 
     # use a custom template
-    llama-cli -m model.gguf -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
     ```
 
-    </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.
-    ```
+    [Supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
 
     </details>
 
@@ -422,10 +410,10 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
-[^1]: [examples/perplexity/README.md](./examples/perplexity/README.md)
+[^1]: [examples/perplexity/README.md](examples/perplexity/README.md)
 [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
 
-## [`llama-bench`](examples/llama-bench)
+## [`llama-bench`](example/bench)
 
 #### Benchmark the performance of the inference for various parameters.
 
@@ -459,7 +447,7 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
-[^3]: [RamaLama](https://github.com/containers/ramalama)
+[^3]: [https://github.com/containers/ramalama](RamaLama)
 
 ## [`llama-simple`](examples/simple)
 

ci/run.sh

@@ -326,17 +326,17 @@ function gg_run_open_llama_7b_v2 {
     ./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-cli --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 --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 --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 --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 --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 --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 --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 --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 --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 --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 --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
@@ -460,17 +460,17 @@ function gg_run_pythia_1_4b {
     ./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-cli --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 --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 --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 --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 --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 --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 --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 --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 --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 --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 --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
@@ -591,17 +591,17 @@ function gg_run_pythia_2_8b {
     ./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-cli --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 --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 --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 --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 --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 --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 --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 --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 --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 --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 --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

cmake/build-info.cmake

@@ -44,7 +44,7 @@ if(MSVC)
     set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
 else()
     execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
         OUTPUT_VARIABLE OUT
         OUTPUT_STRIP_TRAILING_WHITESPACE
     )

cmake/llama-config.cmake.in

@@ -3,13 +3,159 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
 
+set(GGML_STATIC @GGML_STATIC@)
+set(GGML_NATIVE @GGML_NATIVE@)
+set(GGML_LTO    @GGML_LTO@)
+set(GGML_CCACHE @GGML_CCACHE@)
+set(GGML_AVX    @GGML_AVX@)
+set(GGML_AVX2   @GGML_AVX2@)
+set(GGML_AVX512 @GGML_AVX512@)
+set(GGML_AVX512_VBMI @GGML_AVX512_VBMI@)
+set(GGML_AVX512_VNNI @GGML_AVX512_VNNI@)
+set(GGML_AVX512_BF16 @GGML_AVX512_BF16@)
+set(GGML_AMX_TILE @GGML_AMX_TILE@)
+set(GGML_AMX_INT8 @GGML_AMX_INT8@)
+set(GGML_AMX_BF16 @GGML_AMX_BF16@)
+set(GGML_FMA  @GGML_FMA@)
+set(GGML_LASX @GGML_LASX@)
+set(GGML_LSX  @GGML_LSX@)
+set(GGML_RVV  @GGML_RVV@)
+set(GGML_SVE  @GGML_SVE@)
+
+set(GGML_ACCELERATE @GGML_ACCELERATE@)
+set(GGML_OPENMP  @GGML_OPENMP@)
+set(GGML_CPU_HBM @GGML_CPU_HBM@)
+set(GGML_BLAS_VENDOR @GGML_BLAS_VENDOR@)
+
+set(GGML_CUDA_FORCE_MMQ    @GGML_CUDA_FORCE_MMQ@)
+set(GGML_CUDA_FORCE_CUBLAS @GGML_CUDA_FORCE_CUBLAS@)
+set(GGML_CUDA_F16          @GGML_CUDA_F16@)
+set(GGML_CUDA_PEER_MAX_BATCH_SIZE @GGML_CUDA_PEER_MAX_BATCH_SIZE@)
+set(GGML_CUDA_NO_PEER_COPY  @GGML_CUDA_NO_PEER_COPY@)
+set(GGML_CUDA_NO_VMM        @GGML_CUDA_NO_VMM@)
+set(GGML_CUDA_FA_ALL_QUANTS @GGML_CUDA_FA_ALL_QUANTS@)
+set(GGML_CUDA_GRAPHS        @GGML_CUDA_GRAPHS@)
+
+set(GGML_HIP_UMA @GGML_HIP_UMA@)
+
+set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
+set(GGML_VULKAN_DEBUG         @GGML_VULKAN_DEBUG@)
+set(GGML_VULKAN_MEMORY_DEBUG  @GGML_VULKAN_MEMORY_DEBUG@)
+set(GGML_VULKAN_SHADER_DEBUG_INFO @GGML_VULKAN_SHADER_DEBUG_INFO@)
+set(GGML_VULKAN_PERF      @GGML_VULKAN_PERF@)
+set(GGML_VULKAN_VALIDATE  @GGML_VULKAN_VALIDATE@)
+set(GGML_VULKAN_RUN_TESTS @GGML_VULKAN_RUN_TESTS@)
+
+set(GGML_METAL_USE_BF16 @GGML_METAL_USE_BF16@)
+set(GGML_METAL_NDEBUG   @GGML_METAL_NDEBUG@)
+set(GGML_METAL_SHADER_DEBUG  @GGML_METAL_SHADER_DEBUG@)
+set(GGML_METAL_EMBED_LIBRARY @GGML_METAL_EMBED_LIBRARY@)
+set(GGML_METAL_MACOSX_VERSION_MIN @GGML_METAL_MACOSX_VERSION_MIN@)
+set(GGML_METAL_STD @GGML_METAL_STD@)
+
+set(GGML_SYCL_F16    @GGML_SYCL_F16@)
+set(GGML_SYCL_TARGET @GGML_SYCL_TARGET@)
+set(GGML_SYCL_DEVICE_ARCH @GGML_SYCL_DEVICE_ARCH@)
+
+
 @PACKAGE_INIT@
 
 set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
 set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
 set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
 
-find_package(ggml REQUIRED HINTS ${LLAMA_LIB_DIR}/cmake)
+find_package(Threads REQUIRED)
+
+set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
+set(_llama_link_deps "")
+set(_llama_link_opts "")
+foreach(_ggml_lib ggml ggml-base)
+    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
+    find_library(${_ggml_lib_var} ${_ggml_lib}
+        REQUIRED
+        HINTS ${LLAMA_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+    list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
+    message(STATUS "Found ${${_ggml_lib_var}}")
+endforeach()
+
+foreach(backend amx blas cann cpu cuda hip kompute metal musa rpc sycl vulkan)
+    string(TOUPPER "GGML_${backend}" backend_id)
+    set(_ggml_lib "ggml-${backend}")
+    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
+
+    find_library(${_ggml_lib_var} ${_ggml_lib}
+        HINTS ${LLAMA_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+    if(${_ggml_lib_var})
+        list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
+        set(${backend_id} ON)
+        message(STATUS "Found backend ${${_ggml_lib_var}}")
+    else()
+        set(${backend_id} OFF)
+    endif()
+endforeach()
+
+if (NOT LLAMA_SHARED_LIB)
+    if (APPLE AND GGML_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+        list(APPEND _llama_link_deps ${ACCELERATE_FRAMEWORK})
+    endif()
+
+    if (GGML_OPENMP)
+        find_package(OpenMP REQUIRED)
+        list(APPEND _llama_link_deps OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+    endif()
+
+    if (GGML_CPU_HBM)
+        find_library(memkind memkind REQUIRED)
+        list(APPEND _llama_link_deps memkind)
+    endif()
+
+    if (GGML_BLAS)
+        find_package(BLAS REQUIRED)
+        list(APPEND _llama_link_deps ${BLAS_LIBRARIES})
+        list(APPEND _llama_link_opts ${BLAS_LINKER_FLAGS})
+    endif()
+
+    if (GGML_CUDA)
+        find_package(CUDAToolkit REQUIRED)
+    endif()
+
+    if (GGML_METAL)
+        find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+        find_library(METAL_FRAMEWORK    Metal REQUIRED)
+        find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+        list(APPEND _llama_link_deps ${FOUNDATION_LIBRARY}
+                                     ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    endif()
+
+    if (GGML_VULKAN)
+        find_package(Vulkan REQUIRED)
+        list(APPEND _llama_link_deps Vulkan::Vulkan)
+    endif()
+
+    if (GGML_HIP)
+        find_package(hip     REQUIRED)
+        find_package(hipblas REQUIRED)
+        find_package(rocblas REQUIRED)
+        list(APPEND _llama_link_deps hip::host roc::rocblas roc::hipblas)
+    endif()
+
+    if (GGML_SYCL)
+        find_package(DNNL)
+        if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+            list(APPEND _llama_link_deps DNNL::dnnl)
+        endif()
+        if (WIN32)
+            find_package(IntelSYCL REQUIRED)
+            find_package(MKL       REQUIRED)
+            list(APPEND _llama_link_deps IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+        endif()
+    endif()
+endif()
 
 find_library(llama_LIBRARY llama
     REQUIRED
@@ -21,10 +167,12 @@ add_library(llama UNKNOWN IMPORTED)
 set_target_properties(llama
     PROPERTIES
         INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
-        INTERFACE_LINK_LIBRARIES "ggml::ggml;ggml::ggml-base;"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        INTERFACE_LINK_OPTIONS   "${_llama_link_opts}"
+        INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
         IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
         IMPORTED_LOCATION "${llama_LIBRARY}"
-        INTERFACE_COMPILE_FEATURES c_std_90
-        POSITION_INDEPENDENT_CODE ON)
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
 
 check_required_components(Llama)

common/CMakeLists.txt

@@ -56,7 +56,6 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
-    chat-template.hpp
     common.cpp
     common.h
     console.cpp
@@ -65,7 +64,6 @@ add_library(${TARGET} STATIC
     json.hpp
     log.cpp
     log.h
-    minja.hpp
     ngram-cache.cpp
     ngram-cache.h
     sampling.cpp

common/arg.cpp

@@ -22,11 +22,6 @@ common_arg & common_arg::set_examples(std::initializer_list<enum llama_example>
     return *this;
 }
 
-common_arg & common_arg::set_excludes(std::initializer_list<enum llama_example> excludes) {
-    this->excludes = std::move(excludes);
-    return *this;
-}
-
 common_arg & common_arg::set_env(const char * env) {
     help = help + "\n(env: " + env + ")";
     this->env = env;
@@ -42,10 +37,6 @@ bool common_arg::in_example(enum llama_example ex) {
     return examples.find(ex) != examples.end();
 }
 
-bool common_arg::is_exclude(enum llama_example ex) {
-    return excludes.find(ex) != excludes.end();
-}
-
 bool common_arg::get_value_from_env(std::string & output) {
     if (env == nullptr) return false;
     char * value = std::getenv(env);
@@ -128,43 +119,29 @@ std::string common_arg::to_string() {
 // utils
 //
 
-static void common_params_handle_model_default(
-        std::string & model,
-        const std::string & model_url,
-        std::string & hf_repo,
-        std::string & hf_file,
-        const std::string & hf_token,
-        const std::string & model_default) {
-    if (!hf_repo.empty()) {
+static void common_params_handle_model_default(common_params & params) {
+    if (!params.hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
-        if (hf_file.empty()) {
-            if (model.empty()) {
-                auto auto_detected = common_get_hf_file(hf_repo, hf_token);
-                if (auto_detected.first.empty() || auto_detected.second.empty()) {
-                    exit(1); // built without CURL, error message already printed
-                }
-                hf_repo = auto_detected.first;
-                hf_file = auto_detected.second;
-            } else {
-                hf_file = model;
+        if (params.hf_file.empty()) {
+            if (params.model.empty()) {
+                throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
             }
-        }
-        // make sure model path is present (for caching purposes)
-        if (model.empty()) {
+            params.hf_file = params.model;
+        } else if (params.model.empty()) {
             // this is to avoid different repo having same file name, or same file name in different subdirs
-            std::string filename = hf_repo + "_" + hf_file;
+            std::string filename = params.hf_repo + "_" + params.hf_file;
             // to make sure we don't have any slashes in the filename
             string_replace_all(filename, "/", "_");
-            model = fs_get_cache_file(filename);
+            params.model = fs_get_cache_file(filename);
         }
-    } else if (!model_url.empty()) {
-        if (model.empty()) {
-            auto f = string_split<std::string>(model_url, '#').front();
+    } else if (!params.model_url.empty()) {
+        if (params.model.empty()) {
+            auto f = string_split<std::string>(params.model_url, '#').front();
             f = string_split<std::string>(f, '?').front();
-            model = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            params.model = fs_get_cache_file(string_split<std::string>(f, '/').back());
         }
-    } else if (model.empty()) {
-        model = model_default;
+    } else if (params.model.empty()) {
+        params.model = DEFAULT_MODEL_PATH;
     }
 }
 
@@ -299,10 +276,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // TODO: refactor model params in a common struct
-    common_params_handle_model_default(params.model,             params.model_url,             params.hf_repo,             params.hf_file,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
-    common_params_handle_model_default(params.vocoder.model,     params.vocoder.model_url,     params.vocoder.hf_repo,     params.vocoder.hf_file,     params.hf_token, "");
+    common_params_handle_model_default(params);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -325,14 +299,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
     }
 
-    if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
-        throw std::runtime_error(string_format(
-            "error: the supplied chat template is not supported: %s%s\n",
-            params.chat_template.c_str(),
-            params.use_jinja ? "" : "\nnote: llama.cpp was started without --jinja, we only support commonly used templates"
-        ));
-    }
-
     return true;
 }
 
@@ -386,30 +352,6 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
-static void add_rpc_devices(std::string servers) {
-    auto rpc_servers = string_split<std::string>(servers, ',');
-    if (rpc_servers.empty()) {
-        throw std::invalid_argument("no RPC servers specified");
-    }
-    ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
-    if (!rpc_reg) {
-        throw std::invalid_argument("failed to find RPC backend");
-    }
-    typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
-    ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
-    if (!ggml_backend_rpc_add_device_fn) {
-        throw std::invalid_argument("failed to find RPC device add function");
-    }
-    for (const auto & server : rpc_servers) {
-        ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
-        if (dev) {
-            ggml_backend_device_register(dev);
-        } else {
-            throw std::invalid_argument("failed to register RPC device");
-        }
-    }
-}
-
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
@@ -472,7 +414,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
      * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example
      */
     auto add_opt = [&](common_arg arg) {
-        if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
+        if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) {
             ctx_arg.options.push_back(std::move(arg));
         }
     };
@@ -678,7 +620,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.ctx_shift = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--chunks"}, "N",
         string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -701,7 +643,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.prompt = value;
         }
-    ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    ));
     add_opt(common_arg(
         {"--no-perf"},
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@@ -725,7 +667,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.prompt.pop_back();
             }
         }
-    ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    ));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
         "an input file (repeat to specify multiple files)",
@@ -752,7 +694,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.prompt = ss.str();
             fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str());
         }
-    ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    ));
     add_opt(common_arg(
         {"-e", "--escape"},
         string_format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"),
@@ -811,19 +753,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-cnv", "--conversation"},
-        "run in conversation mode:\n"
-        "- does not print special tokens and suffix/prefix\n"
-        "- interactive mode is also enabled\n"
-        "(default: auto enabled if chat template is available)",
+        string_format(
+            "run in conversation mode:\n"
+            "- does not print special tokens and suffix/prefix\n"
+            "- interactive mode is also enabled\n"
+            "(default: %s)",
+            params.conversation ? "true" : "false"
+        ),
         [](common_params & params) {
-            params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
-        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
-    add_opt(common_arg(
-        {"-no-cnv", "--no-conversation"},
-        "force disable conversation mode (default: false)",
-        [](common_params & params) {
-            params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
+            params.conversation = true;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
@@ -877,7 +815,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -904,7 +842,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--sampling-seq", "--sampler-seq"}, "SEQUENCE",
+        {"--sampling-seq"}, "SEQUENCE",
         string_format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()),
         [](common_params & params, const std::string & value) {
             params.sampling.samplers = common_sampler_types_from_chars(value);
@@ -917,6 +855,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.ignore_eos = true;
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"--penalize-nl"},
+        string_format("penalize newline tokens (default: %s)", params.sampling.penalize_nl ? "true" : "false"),
+        [](common_params & params) {
+            params.sampling.penalize_nl = true;
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--temp"}, "N",
         string_format("temperature (default: %.1f)", (double)params.sampling.temp),
@@ -971,9 +916,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--repeat-last-n"}, "N",
         string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
         [](common_params & params, int value) {
-            if (value < -1) {
-                throw std::runtime_error(string_format("error: invalid repeat-last-n = %d\n", value));
-            }
             params.sampling.penalty_last_n = value;
             params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
         }
@@ -1028,9 +970,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--dry-penalty-last-n"}, "N",
         string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
         [](common_params & params, int value) {
-            if (value < -1) {
-                throw std::runtime_error(string_format("error: invalid dry-penalty-last-n = %d\n", value));
-            }
             params.sampling.dry_penalty_last_n = value;
         }
     ).set_sparam());
@@ -1419,8 +1358,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             {"--rpc"}, "SERVERS",
             "comma separated list of RPC servers",
             [](common_params & params, const std::string & value) {
-                add_rpc_devices(value);
-                GGML_UNUSED(params);
+                params.rpc_servers = value;
             }
         ).set_env("LLAMA_ARG_RPC"));
     }
@@ -1569,7 +1507,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & value) {
-            params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
+            params.lora_adapters.push_back({ std::string(value), 1.0 });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@@ -1577,7 +1515,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora-scaled"}, "FNAME", "SCALE",
         "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & fname, const std::string & scale) {
-            params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
+            params.lora_adapters.push_back({ fname, std::stof(scale) });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@@ -1631,42 +1569,19 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
-        {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
-        "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
-        "example: unsloth/phi-4-GGUF:q4_k_m\n"
-        "(default: unused)",
+        {"-hfr", "--hf-repo"}, "REPO",
+        "Hugging Face model repository (default: unused)",
         [](common_params & params, const std::string & value) {
             params.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO"));
-    add_opt(common_arg(
-        {"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
-        "Same as --hf-repo, but for the draft model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.speculative.hf_repo = value;
-        }
-    ).set_env("LLAMA_ARG_HFD_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
-        "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
+        "Hugging Face model file (default: unused)",
         [](common_params & params, const std::string & value) {
             params.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE"));
-    add_opt(common_arg(
-        {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
-        "Hugging Face model repository for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.hf_repo = value;
-        }
-    ).set_env("LLAMA_ARG_HF_REPO_V"));
-    add_opt(common_arg(
-        {"-hffv", "--hf-file-v"}, "FILE",
-        "Hugging Face model file for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.hf_file = value;
-        }
-    ).set_env("LLAMA_ARG_HF_FILE_V"));
     add_opt(common_arg(
         {"-hft", "--hf-token"}, "TOKEN",
         "Hugging Face access token (default: value from HF_TOKEN environment variable)",
@@ -1955,44 +1870,24 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--jinja"},
-        "use jinja template for chat (default: disabled)",
-        [](common_params & params) {
-            params.use_jinja = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
             "set custom jinja chat template (default: template taken from model's metadata)\n"
             "if suffix/prefix are specified, template will be disabled\n"
-            "only commonly used templates are accepted (unless --jinja is set before this flag):\n"
             "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
         ),
         [](common_params & params, const std::string & value) {
+            if (!common_chat_verify_template(value)) {
+                throw std::runtime_error(string_format(
+                    "error: the supplied chat template is not supported: %s\n"
+                    "note: llama.cpp does not use jinja parser, we only support commonly used templates\n",
+                    value.c_str()
+                ));
+            }
             params.chat_template = value;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
-    add_opt(common_arg(
-        {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
-        string_format(
-            "set custom jinja chat template file (default: template taken from model's metadata)\n"
-            "if suffix/prefix are specified, template will be disabled\n"
-            "only commonly used templates are accepted (unless --jinja is set before this flag):\n"
-            "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
-        ),
-        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
-        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
     add_opt(common_arg(
         {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
         string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),
@@ -2284,32 +2179,5 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
 
-    add_opt(common_arg(
-        {"-mv", "--model-vocoder"}, "FNAME",
-        "vocoder model for audio generation (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.model = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
-     add_opt(common_arg(
-        {"--tts-use-guide-tokens"},
-        "Use guide tokens to improve TTS word recall",
-        [](common_params & params) {
-            params.vocoder.use_guide_tokens = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
-
-    // model-specific
-    add_opt(common_arg(
-        {"--tts-oute-default"},
-        string_format("use default OuteTTS models (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf";
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS}));
-
     return ctx_arg;
 }

common/arg.h

@@ -12,7 +12,6 @@
 
 struct common_arg {
     std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
-    std::set<enum llama_example> excludes = {};
     std::vector<const char *> args;
     const char * value_hint   = nullptr; // help text or example for arg value
     const char * value_hint_2 = nullptr; // for second arg value
@@ -54,11 +53,9 @@ struct common_arg {
     ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
 
     common_arg & set_examples(std::initializer_list<enum llama_example> examples);
-    common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
     common_arg & set_env(const char * env);
     common_arg & set_sparam();
     bool in_example(enum llama_example ex);
-    bool is_exclude(enum llama_example ex);
     bool get_value_from_env(std::string & output);
     bool has_value_from_env();
     std::string to_string();

common/chat-template.hpp

@@ -1,366 +0,0 @@
-/*
-    Copyright 2024 Google LLC
-
-    Use of this source code is governed by an MIT-style
-    license that can be found in the LICENSE file or at
-    https://opensource.org/licenses/MIT.
-*/
-// SPDX-License-Identifier: MIT
-#pragma once
-
-#include "minja.hpp"
-#include <json.hpp>
-#include <string>
-#include <vector>
-
-using json = nlohmann::ordered_json;
-
-namespace minja {
-
-struct chat_template_caps {
-    bool supports_tools = false;
-    bool supports_tool_calls = false;
-    bool supports_tool_responses = false;
-    bool supports_system_role = false;
-    bool supports_parallel_tool_calls = false;
-    bool supports_tool_call_id = false;
-    // meta-llama/Llama-3.1-8B-Instruct expects arguments to be an object.
-    // Most other templates (and OpenAI's API) expect the arguments object to be stringified.
-    bool requires_object_arguments = false;
-    // CohereForAI/c4ai-command-r-plus simple variant
-    bool requires_non_null_content = false;
-    // MiniMaxAI/MiniMax-Text-01 special
-    bool requires_typed_content = false;
-};
-
-class chat_template {
-
-  private:
-    chat_template_caps caps_;
-    std::string source_;
-    std::string bos_token_;
-    std::string eos_token_;
-    std::shared_ptr<minja::TemplateNode> template_root_;
-
-    std::string try_raw_render(
-        const nlohmann::ordered_json & messages,
-        const nlohmann::ordered_json & tools,
-        bool add_generation_prompt,
-        const nlohmann::ordered_json & extra_context = nlohmann::ordered_json()) const
-    {
-        try {
-            auto prompt = apply(messages, tools, add_generation_prompt, extra_context, /* adjust_inputs= */ false);
-            // fprintf(stderr, "try_raw_render: %s\n", prompt.c_str());
-            return prompt;
-        } catch (const std::exception & e) {
-            // fprintf(stderr, "try_raw_render error: %s\n", e.what());
-            return "";
-        }
-    }
-
-  public:
-
-    chat_template(const std::string & source, const std::string & bos_token, const std::string & eos_token)
-        : source_(source), bos_token_(bos_token), eos_token_(eos_token)
-    {
-        template_root_ = minja::Parser::parse(source_, {
-            /* .trim_blocks = */ true,
-            /* .lstrip_blocks = */ true,
-            /* .keep_trailing_newline = */ false,
-        });
-
-        auto contains = [](const std::string & haystack, const std::string & needle) {
-            return haystack.find(needle) != std::string::npos;
-        };
-
-        const std::string user_needle = "<User Needle>";
-        const std::string sys_needle = "<System Needle>";
-        const json dummy_str_user_msg = {{"role", "user"}, {"content", user_needle}};
-        const json dummy_typed_user_msg = {{"role", "user"}, {"content", json::array({{{"type", "text"}, {"text", user_needle}}})}};
-
-        caps_.requires_typed_content =
-            !contains(try_raw_render(json::array({dummy_str_user_msg}), {}, false), user_needle)
-            && contains(try_raw_render(json::array({dummy_typed_user_msg}), {}, false), user_needle);
-
-        const auto dummy_user_msg = caps_.requires_typed_content
-            ? dummy_typed_user_msg
-            : dummy_str_user_msg;
-        const json needle_system_msg = {
-            {"role", "system"},
-            {"content", caps_.requires_typed_content ? json::array({{{"type", "text"}, {"text", sys_needle}}}) : json(sys_needle)},
-        };
-
-        caps_.supports_system_role = contains(try_raw_render({needle_system_msg, dummy_user_msg,}, {}, false), sys_needle);
-
-        auto out = try_raw_render(json::array({
-            dummy_user_msg
-        }), json::array({
-            {
-                {"name", "some_tool"},
-                {"type", "function"},
-                {"function", {
-                    {"name", "some_tool"},
-                    {"description", "Some tool."},
-                    {"parameters", {
-                        {"type", "object"},
-                        {"properties", {
-                            {"arg", {
-                                {"type", "string"},
-                                {"description", "Some argument."},
-                            }},
-                        }},
-                        {"required", json::array({ "arg" })},
-                    }},
-                }},
-            },
-        }), false);
-        caps_.supports_tools = contains(out, "some_tool");
-
-        auto make_tool_calls_msg = [&](const json & tool_calls) {
-            return json {
-                {"role", "assistant"},
-                {"content", nullptr},
-                {"tool_calls", tool_calls},
-            };
-        };
-        auto make_tool_call = [](const std::string & tool_name, const json & arguments) {
-            return json {
-                {"id", "call_1___"},
-                {"type", "function"},
-                {"function", {
-                    {"arguments", arguments},
-                    {"name", tool_name},
-                }},
-            };
-        };
-        const json dummy_args_obj {{"argument_needle", "print('Hello, World!')"}};
-
-        // Note: the arguments are rendered in both cases, but may be double-escaped, which we don't want.
-        out = try_raw_render(json::array({
-            dummy_user_msg,
-            make_tool_calls_msg(json::array({make_tool_call("ipython", dummy_args_obj.dump())})),
-        }), {}, false);
-        auto tool_call_renders_str_arguments = contains(out, "\"argument_needle\":") || contains(out, "'argument_needle':");
-        out = try_raw_render(json::array({
-            dummy_user_msg,
-            make_tool_calls_msg(json::array({make_tool_call("ipython", dummy_args_obj)})),
-        }), {}, false);
-        auto tool_call_renders_obj_arguments = contains(out, "\"argument_needle\":") || contains(out, "'argument_needle':");
-
-        caps_.supports_tool_calls = tool_call_renders_str_arguments || tool_call_renders_obj_arguments;
-        caps_.requires_object_arguments = !tool_call_renders_str_arguments && tool_call_renders_obj_arguments;
-        auto out_empty = try_raw_render(json::array({dummy_user_msg, {{"role", "assistant"}, {"content", ""}}}), {}, false);
-        auto out_null = try_raw_render(json::array({dummy_user_msg, {{"role", "assistant"}, {"content", nullptr}}}), {}, false);
-        caps_.requires_non_null_content = contains(out_empty, user_needle) && !contains(out_null, user_needle);
-
-        if (caps_.supports_tool_calls) {
-            auto dummy_args = caps_.requires_object_arguments ? dummy_args_obj : json(dummy_args_obj.dump());
-            auto tc1 = make_tool_call("test_tool1", dummy_args);
-            auto tc2 = make_tool_call("test_tool2", dummy_args);
-            auto out = try_raw_render(json::array({
-                dummy_user_msg,
-                make_tool_calls_msg(json::array({tc1, tc2})),
-            }), {}, false);
-            caps_.supports_parallel_tool_calls = contains(out, "test_tool1") && contains(out, "test_tool2");
-
-            out = try_raw_render(json::array({
-                dummy_user_msg,
-                make_tool_calls_msg(json::array({tc1})),
-                {
-                    {"role", "tool"},
-                    {"name", "test_tool1"},
-                    {"content", "Some response!"},
-                    {"tool_call_id", "call_911_"},
-                }
-            }), {}, false);
-            caps_.supports_tool_responses = contains(out, "Some response!");
-            caps_.supports_tool_call_id = contains(out, "call_911_");
-        }
-    }
-
-    const std::string & source() const { return source_; }
-    const std::string & bos_token() const { return bos_token_; }
-    const std::string & eos_token() const { return eos_token_; }
-    const chat_template_caps & original_caps() const { return caps_; }
-
-    std::string apply(
-        const nlohmann::ordered_json & messages,
-        const nlohmann::ordered_json & tools,
-        bool add_generation_prompt,
-        const nlohmann::ordered_json & extra_context = nlohmann::ordered_json(),
-        bool adjust_inputs = true) const
-    {
-        json actual_messages;
-
-        auto needs_adjustments = adjust_inputs && (false
-            || !caps_.supports_system_role
-            || !caps_.supports_tools
-            || !caps_.supports_tool_responses
-            || !caps_.supports_tool_calls
-            || caps_.requires_object_arguments
-            || caps_.requires_typed_content
-        );
-        if (needs_adjustments) {
-            actual_messages = json::array();
-
-            auto add_message = [&](const json & msg) {
-                if (caps_.requires_typed_content && msg.contains("content") && !msg.at("content").is_null() && msg.at("content").is_string()) {
-                    actual_messages.push_back({
-                        {"role", msg.at("role")},
-                        {"content", {{
-                            {"type", "text"},
-                            {"text", msg.at("content")},
-                        }}},
-                    });
-                } else {
-                    actual_messages.push_back(msg);
-                }
-            };
-
-            std::string pending_system;
-            auto flush_sys = [&]() {
-                if (!pending_system.empty()) {
-                    add_message({
-                        {"role", "user"},
-                        {"content", pending_system},
-                    });
-                    pending_system.clear();
-                }
-            };
-            auto needs_tools_in_system = !tools.is_null() && tools.size() > 0 && !caps_.supports_tools;
-
-            for (const auto & message_ : needs_tools_in_system ? add_system(messages, "Available tools: " + tools.dump(2)) : messages) {
-                auto message = message_;
-                if (!message.contains("role") || !message.contains("content")) {
-                    throw std::runtime_error("message must have 'role' and 'content' fields: " + message.dump());
-                }
-                std::string role = message.at("role");
-
-                if (message.contains("tool_calls")) {
-                    if (caps_.requires_object_arguments || !caps_.supports_tool_calls) {
-                        for (auto & tool_call : message.at("tool_calls")) {
-                            if (tool_call["type"] == "function") {
-                                auto & function = tool_call.at("function");
-                                auto & arguments = function.at("arguments");
-                                if (arguments.is_string()) {
-                                    try {
-                                        arguments = json::parse(arguments.get<std::string>());
-                                    } catch (const std::exception & ecvt) {
-                                        fprintf(stderr, "Failed to parse arguments: %s\n", ecvt.what());
-                                    }
-                                }
-                            }
-                        }
-                    }
-                    if (!caps_.supports_tool_calls) {
-                        auto content = message.at("content");
-                        auto tool_calls = json::array();
-                        for (const auto & tool_call : message.at("tool_calls")) {
-                            if (tool_call.at("type") != "function") {
-                                continue;
-                            }
-                            const auto & function = tool_call.at("function");
-                            auto tc = json {
-                                {"name", function.at("name")},
-                                {"arguments", function.at("arguments")},
-                            };
-                            if (tool_call.contains("id")) {
-                                tc["id"] = tool_call["id"];
-                            }
-                            tool_calls.push_back(tc);
-                        }
-                        auto obj = json {
-                            {"tool_calls", tool_calls},
-                        };
-                        if (!content.is_null() && content != "") {
-                            obj["content"] = content;
-                        }
-                        message["content"] = obj.dump(2);
-                        message.erase("tool_calls");
-                    }
-                }
-                if (!caps_.supports_tool_responses && role == "tool") {
-                    message["role"] = "user";
-                    auto obj = json {
-                        {"tool_response", {
-                            {"tool", message.at("name")},
-                            {"content", message.at("content")},
-                        }},
-                    };
-                    if (message.contains("tool_call_id")) {
-                        obj["tool_response"]["tool_call_id"] = message.at("tool_call_id");
-                    }
-                    message["content"] = obj.dump(2);
-                    message.erase("name");
-                }
-
-                if (!message["content"].is_null() && !caps_.supports_system_role) {
-                    std::string content = message.at("content");
-                    if (role == "system") {
-                        if (!pending_system.empty()) pending_system += "\n";
-                        pending_system += content;
-                        continue;
-                    } else {
-                        if (role == "user") {
-                            if (!pending_system.empty()) {
-                                message["content"] = pending_system + (content.empty() ? "" : "\n" + content);
-                                pending_system.clear();
-                            }
-                        } else {
-                            flush_sys();
-                        }
-                    }
-                }
-                add_message(message);
-            }
-            if (!caps_.supports_system_role) {
-                flush_sys();
-            }
-        } else {
-            actual_messages = messages;
-        }
-
-        auto context = minja::Context::make(json({
-            {"messages", actual_messages},
-            {"add_generation_prompt", add_generation_prompt},
-            {"bos_token", bos_token_},
-            {"eos_token", eos_token_},
-        }));
-
-        if (!tools.is_null()) {
-            auto tools_val = minja::Value(tools);
-            context->set("tools", tools_val);
-        }
-        if (!extra_context.is_null()) {
-            for (auto & kv : extra_context.items()) {
-                minja::Value val(kv.value());
-                context->set(kv.key(), val);
-            }
-        }
-
-        auto ret = template_root_->render(context);
-        // fprintf(stderr, "actual_messages: %s\n", actual_messages.dump(2).c_str());
-        // fprintf(stderr, "apply: %s\n\n", ret.c_str());
-        return ret;
-    }
-
-    static nlohmann::ordered_json add_system(const nlohmann::ordered_json & messages, const std::string & system_prompt) {
-        json messages_with_system = messages;
-
-        if (messages_with_system.size() > 0 && messages_with_system[0].at("role") == "system") {
-            std::string existing_system = messages_with_system.at(0).at("content");
-            messages_with_system[0] = json {
-                {"role", "system"},
-                {"content", existing_system + "\n" + system_prompt},
-            };
-        } else {
-            messages_with_system.insert(messages_with_system.begin(), json {
-                {"role", "system"},
-                {"content", system_prompt},
-            });
-        }
-        return messages_with_system;
-    }
-};
-
-}  // namespace minja

common/common.cpp

@@ -2,9 +2,6 @@
 #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
 #endif
 
-#include "ggml.h"
-#include "gguf.h"
-
 #include "common.h"
 #include "log.h"
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
@@ -12,7 +9,6 @@
 #include "json.hpp"
 #include "json-schema-to-grammar.h"
 #include "llama.h"
-#include "chat-template.hpp"
 
 #include <algorithm>
 #include <cinttypes>
@@ -22,7 +18,6 @@
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
-#include <filesystem>
 #include <fstream>
 #include <iostream>
 #include <iterator>
@@ -67,29 +62,11 @@
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
+#define PATH_MAX MAX_PATH
 #else
 #include <sys/syslimits.h>
 #endif
 #define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
-//
-// CURL utils
-//
-
-using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
-
-// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
-struct curl_slist_ptr {
-    struct curl_slist * ptr = nullptr;
-    ~curl_slist_ptr() {
-        if (ptr) {
-            curl_slist_free_all(ptr);
-        }
-    }
-};
 #endif // LLAMA_USE_CURL
 
 using json = nlohmann::ordered_json;
@@ -484,48 +461,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
-std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
-    std::ostringstream result;
-    for (size_t i = 0; i < values.size(); ++i) {
-        if (i > 0) {
-            result << separator;
-        }
-        result << values[i];
-    }
-    return result.str();
-}
-
-std::vector<std::string> string_split(const std::string & str, const std::string & delimiter) {
-    std::vector<std::string> parts;
-    size_t start = 0;
-    size_t end = str.find(delimiter);
-
-    while (end != std::string::npos) {
-        parts.push_back(str.substr(start, end - start));
-        start = end + delimiter.length();
-        end = str.find(delimiter, start);
-    }
-
-    parts.push_back(str.substr(start));
-
-    return parts;
-}
-
-std::string string_repeat(const std::string & str, size_t n) {
-    if (n == 0) {
-        return "";
-    }
-
-    std::string result;
-    result.reserve(str.length() * n);
-
-    for (size_t i = 0; i < n; ++i) {
-        result += str;
-    }
-
-    return result;
-}
-
 std::string string_from(bool value) {
     return value ? "true" : "false";
 }
@@ -908,7 +843,7 @@ struct common_init_result common_init_from_params(common_params & params) {
     } else if (!params.model_url.empty()) {
         model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
     } else {
-        model = llama_model_load_from_file(params.model.c_str(), mparams);
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
 
     if (model == NULL) {
@@ -916,28 +851,26 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     if (params.reranking) {
         bool ok = true;
 
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
+        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
+        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
             ok = false;
         }
 
         if (!ok) {
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
@@ -945,40 +878,40 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     auto cparams = common_context_params_to_llama(params);
 
-    llama_context * lctx = llama_init_from_model(model, cparams);
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
-        llama_model_free(model);
+        llama_free_model(model);
         return iparams;
     }
 
     if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
-        params.ctx_shift = false;
+        LOG_ERR("%s: KV cache shifting is not supported for this model (--no-context-shift to disable)'\n", __func__);
+        llama_free_model(model);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
-        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_model_n_layer(model);
+        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
 
-        int err = llama_apply_adapter_cvec(
-                lctx,
-                cvec.data.data(),
-                cvec.data.size(),
-                cvec.n_embd,
-                params.control_vector_layer_start,
-                params.control_vector_layer_end);
+        int err = llama_control_vector_apply(lctx,
+                                             cvec.data.data(),
+                                             cvec.data.size(),
+                                             cvec.n_embd,
+                                             params.control_vector_layer_start,
+                                             params.control_vector_layer_end);
         if (err) {
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
@@ -986,54 +919,33 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     // 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) {
+        common_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
             return iparams;
         }
-
-        la.ptr = lora.get();
-        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
     }
-
     if (!params.lora_init_without_apply) {
-        common_set_adapter_lora(lctx, params.lora_adapters);
+        common_lora_adapters_apply(lctx, iparams.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__);
+    if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
     }
 
-    if (params.sampling.ignore_eos) {
-        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.push_back({i, -INFINITY});
-            }
-        }
-    }
-
-    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__);
 
         std::vector<llama_token> tmp;
-        llama_token bos = llama_vocab_bos(vocab);
-        llama_token eos = llama_vocab_eos(vocab);
-
+        llama_token bos = llama_token_bos(model);
+        llama_token eos = llama_token_eos(model);
         // some models (e.g. T5) don't have a BOS token
         if (bos != LLAMA_TOKEN_NULL) {
             tmp.push_back(bos);
@@ -1048,7 +960,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_encoder(model)) {
             llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-            if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            if (decoder_start_token_id == -1) {
                 decoder_start_token_id = bos;
             }
             tmp.clear();
@@ -1062,17 +974,17 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_perf_context_reset(lctx);
     }
 
-    iparams.model.reset(model);
-    iparams.context.reset(lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
 
     return iparams;
 }
 
-void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    llama_clear_adapter_lora(ctx);
-    for (auto & la : lora) {
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
         if (la.scale != 0.0f) {
-            llama_set_adapter_lora(ctx, la.ptr, la.scale);
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
         }
     }
 }
@@ -1086,6 +998,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+    mparams.rpc_servers     = params.rpc_servers.c_str();
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
@@ -1163,7 +1076,7 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
     int remaining_attempts = max_attempts;
 
     while (remaining_attempts > 0) {
@@ -1187,9 +1100,9 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
 }
 
 static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
+
     // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
+    std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
     if (!curl) {
         LOG_ERR("%s: error initializing libcurl\n", __func__);
         return false;
@@ -1203,9 +1116,11 @@ static bool common_download_file(const std::string & url, const std::string & pa
 
     // Check if hf-token or bearer-token was specified
     if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+      std::string auth_header = "Authorization: Bearer ";
+      auth_header += hf_token.c_str();
+      struct curl_slist *http_headers = NULL;
+      http_headers = curl_slist_append(http_headers, auth_header.c_str());
+      curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
     }
 
 #if defined(_WIN32)
@@ -1215,7 +1130,8 @@ static bool common_download_file(const std::string & url, const std::string & pa
 #endif
 
     // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
+    struct stat model_file_info;
+    auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
@@ -1257,13 +1173,11 @@ static bool common_download_file(const std::string & url, const std::string & pa
         std::string etag;
         std::string last_modified;
     };
-
     common_load_model_from_url_headers headers;
-
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
+            common_load_model_from_url_headers *headers = (common_load_model_from_url_headers *) userdata;
 
             static std::regex header_regex("([^:]+): (.*)\r\n");
             static std::regex etag_regex("ETag", std::regex_constants::icase);
@@ -1475,7 +1389,7 @@ struct llama_model * common_load_model_from_url(
         }
     }
 
-    return llama_model_load_from_file(local_path.c_str(), params);
+    return llama_load_model_from_file(local_path.c_str(), params);
 }
 
 struct llama_model * common_load_model_from_hf(
@@ -1501,80 +1415,6 @@ struct llama_model * common_load_model_from_hf(
     return common_load_model_from_url(model_url, local_path, hf_token, params);
 }
 
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
-
-    // fetch model info from Hugging Face Hub API
-    json model_info;
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
-        model_info = json::parse(res_str);
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
-    } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
-    }
-
-    // check response
-    if (!model_info.contains("ggufFile")) {
-        throw std::runtime_error("error: model does not have ggufFile");
-    }
-    json & gguf_file = model_info.at("ggufFile");
-    if (!gguf_file.contains("rfilename")) {
-        throw std::runtime_error("error: ggufFile does not have rfilename");
-    }
-
-    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
-}
-
 #else
 
 struct llama_model * common_load_model_from_url(
@@ -1596,11 +1436,6 @@ struct llama_model * common_load_model_from_hf(
     return nullptr;
 }
 
-std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return std::make_pair("", "");
-}
-
 #endif // LLAMA_USE_CURL
 
 //
@@ -1699,23 +1534,21 @@ std::vector<llama_token> common_tokenize(
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_tokenize(vocab, text, add_special, parse_special);
+    return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
 }
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_vocab * vocab,
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -1724,18 +1557,12 @@ std::vector<llama_token> common_tokenize(
 }
 
 std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_token_to_piece(vocab, token, special);
-}
-
-std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
     std::string piece;
     piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
-    const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+    const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
     if (n_chars < 0) {
         piece.resize(-n_chars);
-        int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
         GGML_ASSERT(check == -n_chars);
     }
     else {
@@ -1745,19 +1572,13 @@ std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token
     return piece;
 }
 
-std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_detokenize(vocab, tokens, special);
-}
-
-std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
+std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
     std::string text;
     text.resize(std::max(text.capacity(), tokens.size()));
-    int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
     if (n_chars < 0) {
         text.resize(-n_chars);
-        n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
         GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
     }
 
@@ -1771,75 +1592,63 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
 // Chat template utils
 //
 
-bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
-    if (use_jinja) {
-        try {
-            auto chat_template = minja::chat_template(tmpl, "<s>", "</s>");
-            chat_template.apply({{
-                {"role", "user"},
-                {"content", "test"},
-            }}, json(), true);
-            return true;
-        } catch (const std::exception & e) {
-            LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what());
-            return false;
-        }
-    }
+bool common_chat_verify_template(const std::string & tmpl) {
     llama_chat_message chat[] = {{"user", "test"}};
-    const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
     return res >= 0;
 }
 
-std::string common_chat_apply_template(
-        const common_chat_template & tmpl,
+std::string common_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
         const std::vector<common_chat_msg> & msgs,
-        bool add_ass,
-        bool use_jinja) {
-    if (use_jinja) {
-        auto messages = json::array();
-        for (const auto & msg : msgs) {
-            messages.push_back({{"role", msg.role}, {"content", msg.content}});
-        }
-        return tmpl.apply(messages, /* tools= */ json(), add_ass);
-    }
-
+        bool add_ass) {
     int alloc_size = 0;
+    bool fallback = false; // indicate if we must fallback to default chatml
     std::vector<llama_chat_message> chat;
-    for (const auto & msg : msgs) {
+    for (auto & msg : msgs) {
         chat.push_back({msg.role.c_str(), msg.content.c_str()});
         alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
     }
 
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
     std::vector<char> buf(alloc_size);
 
     // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
 
     // error: chat template is not supported
     if (res < 0) {
-        // if the custom "tmpl" is not supported, we throw an error
-        // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-        throw std::runtime_error("this custom template is not supported");
+        if (ptr_tmpl != nullptr) {
+            // if the custom "tmpl" is not supported, we throw an error
+            // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
+            throw std::runtime_error("this custom template is not supported");
+        } else {
+            // If the built-in template is not supported, we default to chatml
+            res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+            fallback = true;
+        }
     }
 
     // if it turns out that our buffer is too small, we resize it
     if ((size_t) res > buf.size()) {
         buf.resize(res);
-        res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+        res = llama_chat_apply_template(
+            fallback ? nullptr : model,
+            fallback ? "chatml" : ptr_tmpl,
+            chat.data(), chat.size(), add_ass, buf.data(), buf.size());
     }
 
     std::string formatted_chat(buf.data(), res);
     return formatted_chat;
 }
 
-std::string common_chat_format_single(
-        const common_chat_template & tmpl,
+std::string common_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
         const std::vector<common_chat_msg> & past_msg,
         const common_chat_msg & new_msg,
-        bool add_ass,
-        bool use_jinja) {
+        bool add_ass) {
     std::ostringstream ss;
-    auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(tmpl, past_msg, false, use_jinja);
+    auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
     std::vector<common_chat_msg> chat_new(past_msg);
     // if the past_msg ends with a newline, we must preserve it in the formatted version
     if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
@@ -1847,74 +1656,21 @@ std::string common_chat_format_single(
     };
     // format chat with new_msg
     chat_new.push_back(new_msg);
-    auto fmt_new_msg = common_chat_apply_template(tmpl, chat_new, add_ass, use_jinja);
+    auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
     // get the diff part
     ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
     return ss.str();
 }
 
-std::string common_chat_format_example(const common_chat_template & tmpl, bool use_jinja) {
+std::string common_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl) {
     std::vector<common_chat_msg> msgs = {
         {"system",    "You are a helpful assistant"},
         {"user",      "Hello"},
         {"assistant", "Hi there"},
         {"user",      "How are you?"},
     };
-    return common_chat_apply_template(tmpl, msgs, true, use_jinja);
-}
-
-common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
-{
-    auto vocab = llama_model_get_vocab(model);
-    std::string default_template_src = chat_template_override;
-    std::string template_tool_use_src = chat_template_override;
-    bool has_explicit_template = !chat_template_override.empty();
-    if (chat_template_override.empty()) {
-        auto str = llama_model_chat_template(model, /* name */ nullptr);
-        if (str) {
-            default_template_src = str;
-            has_explicit_template = true;
-        }
-        str = llama_model_chat_template(model, /* name */ "tool_use");
-        if (str) {
-            template_tool_use_src = str;
-            has_explicit_template = true;
-        }
-    }
-    if (default_template_src.empty() || default_template_src == "chatml") {
-        if (!template_tool_use_src.empty()) {
-            default_template_src = template_tool_use_src;
-        } else {
-            default_template_src = R"(
-                {%- for message in messages -%}
-                    {{- "<|im_start|>" + message.role + "\n" + message.content + "<|im_end|>\n" -}}
-                {%- endfor -%}
-                {%- if add_generation_prompt -%}
-                    {{- "<|im_start|>assistant\n" -}}
-                {%- endif -%}
-            )";
-        }
-    }
-    const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
-        if (token == LLAMA_TOKEN_NULL) {
-            if (default_template_src.find(jinja_variable_name) != std::string::npos
-                || template_tool_use_src.find(jinja_variable_name) != std::string::npos) {
-                LOG_WRN("%s: warning: vocab does not have a %s token, jinja template won't work as intended.\n", __func__, name);
-            }
-            return std::string();
-        } else {
-            return common_token_to_piece(vocab, token, true);
-        }
-    };
-    auto token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
-    auto token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
-    return {
-        has_explicit_template,
-        std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos),
-        template_tool_use_src.empty()
-            ? nullptr
-            : std::make_unique<minja::chat_template>(template_tool_use_src, token_bos, token_eos)
-    };
+    return common_chat_apply_template(model, tmpl, msgs, true);
 }
 
 //
@@ -2005,9 +1761,7 @@ void common_embd_normalize(const float * inp, float * out, int n, int embd_norm)
             break;
         case 0: // max absolute
             for (int i = 0; i < n; i++) {
-                if (sum < std::abs(inp[i])) {
-                    sum = std::abs(inp[i]);
-                }
+                if (sum < std::abs(inp[i])) sum = std::abs(inp[i]);
             }
             sum /= 32760.0; // make an int16 range
             break;

common/common.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "llama-cpp.h"
+#include "llama.h"
 
 #include <string>
 #include <vector>
@@ -24,11 +24,13 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct common_adapter_lora_info {
+struct common_lora_adapter_info {
     std::string path;
     float scale;
+};
 
-    struct llama_adapter_lora * ptr;
+struct common_lora_adapter_container : common_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
 };
 
 using llama_tokens = std::vector<llama_token>;
@@ -78,7 +80,6 @@ enum llama_example {
     LLAMA_EXAMPLE_LLAVA,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
-    LLAMA_EXAMPLE_TTS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -94,7 +95,6 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
     COMMON_SAMPLER_TYPE_XTC         = 8,
     COMMON_SAMPLER_TYPE_INFILL      = 9,
-    COMMON_SAMPLER_TYPE_PENALTIES   = 10,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -103,12 +103,6 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
-enum common_conversation_mode {
-    COMMON_CONVERSATION_MODE_DISABLED = 0,
-    COMMON_CONVERSATION_MODE_ENABLED  = 1,
-    COMMON_CONVERSATION_MODE_AUTO     = 2,
-};
-
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -136,6 +130,7 @@ struct common_params_sampling {
     int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau       = 5.00f; // target entropy
     float   mirostat_eta       = 0.10f; // learning rate
+    bool    penalize_nl        = false; // consider newlines as a repeatable token
     bool    ignore_eos         = false;
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
@@ -144,7 +139,6 @@ struct common_params_sampling {
 
 
     std::vector<enum common_sampler_type> samplers = {
-        COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
         COMMON_SAMPLER_TYPE_TOP_K,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
@@ -164,7 +158,6 @@ struct common_params_sampling {
 
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
-
     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        =     5; // minimum number of draft tokens to use for speculative decoding
@@ -175,21 +168,7 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model = "";     // draft model for speculative decoding                      // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
-};
-
-struct common_params_vocoder {
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model     = ""; // model path                                                // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
-
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
+    std::string model = ""; // draft model for speculative decoding                          // NOLINT
 };
 
 struct common_params {
@@ -214,13 +193,11 @@ struct common_params {
     float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // 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 - 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
+    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 - 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
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
@@ -234,9 +211,8 @@ struct common_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct common_params_sampling    sampling;
+    struct common_params_sampling sampling;
     struct common_params_speculative speculative;
-    struct common_params_vocoder     vocoder;
 
     std::string model                = ""; // model path                                                    // NOLINT
     std::string model_alias          = ""; // model alias                                                   // NOLINT
@@ -252,13 +228,14 @@ struct common_params {
     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
+    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // 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;
 
-    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
-    std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -286,6 +263,7 @@ struct common_params {
     bool special           = false; // enable special token output
     bool interactive       = false; // interactive mode
     bool interactive_first = false; // wait for user input immediately
+    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
@@ -311,8 +289,6 @@ struct common_params {
     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
 
-    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
-
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
     std::vector<std::string> image; // path to image file(s)
@@ -334,7 +310,6 @@ struct common_params {
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
@@ -429,10 +404,6 @@ std::string string_format(const char * fmt, ...);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
-std::string string_join(const std::vector<std::string> & values, const std::string & separator);
-std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
-std::string string_repeat(const std::string & str, size_t n);
-
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
 
 template<class T>
@@ -471,11 +442,6 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool string_ends_with(const std::string & str,
-                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
-
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
@@ -498,12 +464,10 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// note: defines object's lifetime
 struct common_init_result {
-    llama_model_ptr   model;
-    llama_context_ptr context;
-
-    std::vector<llama_adapter_lora_ptr> lora;
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<common_lora_adapter_container> lora_adapters;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -517,7 +481,6 @@ struct llama_model * common_load_model_from_url(
     const std::string & local_path,
     const std::string & hf_token,
     const struct llama_model_params & params);
-
 struct llama_model * common_load_model_from_hf(
     const std::string & repo,
     const std::string & remote_path,
@@ -525,12 +488,8 @@ struct llama_model * common_load_model_from_hf(
     const std::string & hf_token,
     const struct llama_model_params & params);
 
-std::pair<std::string, std::string> common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & hf_token);
-
 // clear LoRA adapters from context, then apply new list of adapters
-void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
 
 //
 // Batch utils
@@ -568,7 +527,7 @@ std::vector<llama_token> common_tokenize(
                         bool   parse_special = false);
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_vocab * vocab,
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
@@ -580,21 +539,11 @@ std::string common_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
-std::string common_token_to_piece(
-          const struct llama_vocab * vocab,
-                       llama_token   token,
-                       bool          special = true);
-
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
 // optionally renders special/control tokens
 std::string common_detokenize(
-            const struct llama_context * ctx,
-        const std::vector<llama_token> & tokens,
-                                  bool   special = true);
-
-std::string common_detokenize(
-              const struct llama_vocab * vocab,
+                         llama_context * ctx,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
@@ -609,42 +558,26 @@ struct common_chat_msg {
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
-
-namespace minja {
-    class chat_template;
-}
-
-typedef minja::chat_template common_chat_template;
-
-struct common_chat_templates {
-    bool has_explicit_template; // Model had builtin template or template overridde was specified.
-    std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
-    std::unique_ptr<common_chat_template> template_tool_use;
-};
+bool common_chat_verify_template(const std::string & tmpl);
 
 // CPP wrapper for llama_chat_apply_template
 // If the built-in template is not supported, we default to chatml
 // If the custom "tmpl" is not supported, we throw an error
-std::string common_chat_apply_template(
-        const common_chat_template & tmpl,
+std::string common_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
         const std::vector<common_chat_msg> & chat,
-        bool add_ass,
-        bool use_jinja);
+        bool add_ass);
 
 // Format single message, while taking into account the position of that message in chat history
-std::string common_chat_format_single(
-        const common_chat_template & tmpl,
+std::string common_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
         const std::vector<common_chat_msg> & past_msg,
         const common_chat_msg & new_msg,
-        bool add_ass,
-        bool use_jinja);
+        bool add_ass);
 
 // Returns an example of formatted chat
-std::string common_chat_format_example(
-    const common_chat_template & tmpl, bool use_jinja);
-
-common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override);
+std::string common_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl);
 
 //
 // KV cache utils
@@ -660,8 +593,7 @@ void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_si
 // Embedding utils
 //
 
-// TODO: repace embd_norm with an enum
-void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
+void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
 
 float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 
@@ -690,10 +622,6 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 // Split utils
 //
 
-namespace {
-
-const char * const LLM_KV_SPLIT_NO            = "split.no";
-const char * const LLM_KV_SPLIT_COUNT         = "split.count";
-const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
-
-}
+static const char * const LLM_KV_SPLIT_NO            = "split.no";
+static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

common/json-schema-to-grammar.cpp

@@ -1,6 +1,4 @@
 #include "json-schema-to-grammar.h"
-#include "common.h"
-
 #include <algorithm>
 #include <fstream>
 #include <map>
@@ -13,6 +11,11 @@
 
 using json = nlohmann::ordered_json;
 
+template <typename Iterator>
+static std::string join(Iterator begin, Iterator end, const std::string & separator);
+
+static std::string repeat(const std::string & str, size_t n);
+
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
@@ -125,8 +128,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
                 if (sub_len > 0) {
                     auto from_sub = from.substr(i + 1);
                     auto to_sub = to.substr(i + 1);
-                    auto sub_zeros = string_repeat("0", sub_len);
-                    auto sub_nines = string_repeat("9", sub_len);
+                    auto sub_zeros = repeat("0", sub_len);
+                    auto sub_nines = repeat("9", sub_len);
 
                     auto to_reached = false;
                     out << "(";
@@ -185,8 +188,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         auto max_digits = max_s.length();
 
         for (auto digits = min_digits; digits < max_digits; digits++) {
-            uniform_range(min_s, string_repeat("9", digits));
-            min_s = "1" + string_repeat("0", digits);
+            uniform_range(min_s, repeat("9", digits));
+            min_s = "1" + repeat("0", digits);
             out << " | ";
         }
         uniform_range(min_s, max_s);
@@ -315,6 +318,49 @@ std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
 std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
+template <typename Iterator>
+std::string join(Iterator begin, Iterator end, const std::string & separator) {
+    std::ostringstream result;
+    if (begin != end) {
+        result << *begin;
+        for (Iterator it = begin + 1; it != end; ++it) {
+            result << separator << *it;
+        }
+    }
+    return result.str();
+}
+
+static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        tokens.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    tokens.push_back(str.substr(start));
+
+    return tokens;
+}
+
+static std::string repeat(const std::string & str, size_t n) {
+    if (n == 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(str.length() * n);
+
+    for (size_t i = 0; i < n; ++i) {
+        result += str;
+    }
+
+    return result;
+}
+
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
     std::smatch match;
     std::string result;
@@ -343,7 +389,6 @@ static std::string format_literal(const std::string & literal) {
 
 class SchemaConverter {
 private:
-    friend std::string build_grammar(const std::function<void(const llama_grammar_builder &)> & cb);
     std::function<json(const std::string &)> _fetch_json;
     bool _dotall;
     std::map<std::string, std::string> _rules;
@@ -373,7 +418,7 @@ private:
         for (size_t i = 0; i < alt_schemas.size(); i++) {
             rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
         }
-        return string_join(rules, " | ");
+        return join(rules.begin(), rules.end(), " | ");
     }
 
     std::string _visit_pattern(const std::string & pattern, const std::string & name) {
@@ -436,7 +481,7 @@ private:
                 for (const auto & item : ret) {
                     results.push_back(to_rule(item));
                 }
-                return std::make_pair(string_join(results, " "), false);
+                return std::make_pair(join(results.begin(), results.end(), " "), false);
             };
 
             while (i < length) {
@@ -494,7 +539,7 @@ private:
                     }
                     curly_brackets += '}';
                     i++;
-                    auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
+                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                     int min_times = 0;
                     int max_times = std::numeric_limits<int>::max();
                     try {
@@ -809,7 +854,7 @@ public:
                             return;
                         }
                         std::string pointer = ref.substr(ref.find('#') + 1);
-                        std::vector<std::string> tokens = string_split(pointer, "/");
+                        std::vector<std::string> tokens = split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
                             if (target.is_null() || !target.contains(sel)) {
@@ -860,7 +905,7 @@ public:
             for (const auto & v : schema["enum"]) {
                 enum_values.push_back(_generate_constant_rule(v));
             }
-            return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ") space");
+            return _add_rule(rule_name, "(" + join(enum_values.begin(), enum_values.end(), " | ") + ") space");
         } else if ((schema_type.is_null() || schema_type == "object")
                 && (schema.contains("properties") ||
                     (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
@@ -974,10 +1019,10 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
         }
         if (!_warnings.empty()) {
-            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());
+            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
         }
     }
 
@@ -991,27 +1036,10 @@ public:
 };
 
 std::string json_schema_to_grammar(const json & schema) {
-    return build_grammar([&](const llama_grammar_builder & callbacks) {
-        auto copy = schema;
-        callbacks.resolve_refs(copy);
-        callbacks.add_schema("", copy);
-    });
-}
-
-std::string build_grammar(const std::function<void(const llama_grammar_builder &)> & cb) {
-    SchemaConverter converter([&](const std::string &) { return json(); }, /* dotall= */ false);
-    llama_grammar_builder builder {
-        /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
-            return converter._add_rule(name, rule);
-        },
-        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
-            return converter.visit(schema, name == "root" ? "" : name);
-        },
-        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
-            converter.resolve_refs(schema, "");
-        }
-    };
-    cb(builder);
+    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
+    auto copy = schema;
+    converter.resolve_refs(copy, "input");
+    converter.visit(copy, "");
     converter.check_errors();
     return converter.format_grammar();
 }

common/json-schema-to-grammar.h

@@ -5,12 +5,4 @@
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json & schema);
-
-struct llama_grammar_builder {
-    std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
-    std::function<void(nlohmann::ordered_json &)> resolve_refs;
-};
-
-std::string build_grammar(const std::function<void(const llama_grammar_builder &)> & cb);
+std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);

common/ngram-cache.cpp

@@ -65,13 +65,13 @@ constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
 static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
     common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
     if (part_static_it == nc_static.end()) {
-        return LLAMA_TOKEN_NULL;
+        return -1;
     }
     const common_ngram_cache_part part_static = part_static_it->second;
 
     int max_count_static  = 0;
     int sum_count_static  = 0;
-    llama_token max_token = LLAMA_TOKEN_NULL;
+    llama_token max_token = -1;
 
     for (std::pair<llama_token, int> token_count_static : part_static) {
         const llama_token token = token_count_static.first;
@@ -85,10 +85,10 @@ static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram
     }
 
     if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
-        return LLAMA_TOKEN_NULL;
+        return -1;
     }
     if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
-        return LLAMA_TOKEN_NULL;
+        return -1;
     }
     return max_token;
 }
@@ -98,9 +98,9 @@ static llama_token try_draft(
     common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
     const int * min_sample_size, const int * min_percent) {
 
-    llama_token drafted_token = LLAMA_TOKEN_NULL;
+    llama_token drafted_token = -1;
 
-    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == LLAMA_TOKEN_NULL; --i) {
+    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
         const common_ngram ngram_primary = ngrams_primary[i];
 
         common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
@@ -112,7 +112,7 @@ static llama_token try_draft(
         int max_count_primary = 0;
         int max_count_static  = 0;
         int sum_count_primary = 0;
-        llama_token max_token = LLAMA_TOKEN_NULL;
+        llama_token max_token = -1;
 
         for (std::pair<llama_token, int> token_count_primary : part_primary) {
             const llama_token token = token_count_primary.first;
@@ -154,7 +154,7 @@ void common_ngram_cache_draft(
     }
 
     while ((int) draft.size()-1 < n_draft) {
-        llama_token drafted_token = LLAMA_TOKEN_NULL;
+        llama_token drafted_token = -1;
 
         const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
         common_ngram ngram_static;
@@ -177,17 +177,17 @@ void common_ngram_cache_draft(
             }
             ngrams_cd.push_back(ngram_cd);
         }
-        if (drafted_token == LLAMA_TOKEN_NULL) {
+        if (drafted_token == -1) {
             drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
         }
-        if (drafted_token == LLAMA_TOKEN_NULL) {
+        if (drafted_token == -1) {
             drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
         }
-        if (drafted_token == LLAMA_TOKEN_NULL) {
+        if (drafted_token == -1) {
             drafted_token = try_draft(nc_static, ngram_static);
         }
 
-        if (drafted_token == LLAMA_TOKEN_NULL) {
+        if (drafted_token == -1) {
             break;
         }
 

common/ngram-cache.h

@@ -17,13 +17,13 @@ struct common_ngram {
 
     common_ngram() {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            tokens[i] = LLAMA_TOKEN_NULL;
+            tokens[i] = -1;
         }
     }
 
     common_ngram(const llama_token * input, const int ngram_size) {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            tokens[i] = i < ngram_size ? input[i] : LLAMA_TOKEN_NULL;
+            tokens[i] = i < ngram_size ? input[i] : -1;
         }
     }
 

common/sampling.cpp

@@ -113,10 +113,7 @@ struct common_sampler {
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
-        const llama_model * model = llama_get_model(ctx);
-        const llama_vocab * vocab = llama_model_get_vocab(model);
-
-        const int n_vocab = llama_vocab_n_tokens(vocab);
+        const int n_vocab = llama_n_vocab(llama_get_model(ctx));
 
         cur.resize(n_vocab);
 
@@ -145,15 +142,13 @@ std::string common_params_sampling::print() const {
 }
 
 struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
 
     auto * result = new common_sampler {
         /* .params = */ params,
-        /* .grmr   = */ llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"),
+        /* .grmr   = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
         /* .chain  = */ llama_sampler_chain_init(lparams),
         /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur    = */ {},
@@ -162,24 +157,36 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     llama_sampler_chain_add(result->chain,
             llama_sampler_init_logit_bias(
-                llama_vocab_n_tokens(vocab),
+                llama_n_vocab(model),
                 params.logit_bias.size(),
                 params.logit_bias.data()));
 
+    llama_sampler_chain_add(result->chain,
+            llama_sampler_init_penalties(
+                llama_n_vocab  (model),
+                llama_token_eos(model),
+                llama_token_nl (model),
+                params.penalty_last_n,
+                params.penalty_repeat,
+                params.penalty_freq,
+                params.penalty_present,
+                params.penalize_nl,
+                params.ignore_eos));
+
     if (params.mirostat == 0) {
         for (const auto & cnstr : params.samplers) {
             switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
+                    case COMMON_SAMPLER_TYPE_DRY:
                     {
-                        std::vector<const char *> c_breakers;
+                        std::vector<const char*> c_breakers;
                         c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
+                        for (const auto& str : params.dry_sequence_breakers) {
                             c_breakers.push_back(str.c_str());
                         }
 
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                     }
-                    break;
+                        break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
                     break;
@@ -199,10 +206,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                     llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
                     break;
                 case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
                     break;
                 default:
                     GGML_ASSERT(false && "unknown sampler type");
@@ -211,7 +215,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
     } else if (params.mirostat == 1) {
         llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
     } else if (params.mirostat == 2) {
         llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
         llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
@@ -411,7 +415,6 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
         case COMMON_SAMPLER_TYPE_XTC:         return 'x';
         case COMMON_SAMPLER_TYPE_INFILL:      return 'i';
-        case COMMON_SAMPLER_TYPE_PENALTIES:   return 'e';
         default : return '?';
     }
 }
@@ -426,7 +429,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
         case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
         case COMMON_SAMPLER_TYPE_INFILL:      return "infill";
-        case COMMON_SAMPLER_TYPE_PENALTIES:   return "penalties";
         default : return "";
     }
 }
@@ -441,7 +443,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
         { "xtc",         COMMON_SAMPLER_TYPE_XTC },
         { "infill",      COMMON_SAMPLER_TYPE_INFILL },
-        { "penalties",   COMMON_SAMPLER_TYPE_PENALTIES },
     };
 
     // since samplers names are written multiple ways
@@ -488,7 +489,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL),      COMMON_SAMPLER_TYPE_INFILL },
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_PENALTIES),   COMMON_SAMPLER_TYPE_PENALTIES },
     };
 
     std::vector<common_sampler_type> samplers;

common/speculative.cpp

@@ -79,13 +79,10 @@ bool common_speculative_are_compatible(
     const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
     const struct llama_model * model_dft = llama_get_model(ctx_dft);
 
-    const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
-    const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
-
-    const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
+    const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
 
-    const bool vocab_type_dft = llama_vocab_type(vocab_dft);
+    const bool vocab_type_dft = llama_vocab_type(model_dft);
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
@@ -94,34 +91,34 @@ bool common_speculative_are_compatible(
         return false;
     }
 
-    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
-        llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
-        llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
-        LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
-        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
-        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
+    if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
+        llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
+        llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
+        LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
         return false;
     }
 
     {
-        const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
-        const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
 
         const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
 
         if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
             LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
                          "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
-                    __func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+                    __func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
             return false;
         }
 
         for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
-            const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
-            const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
-                LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
+                LOG_ERR("%s: draft model vocab must match target model to use speculation but "
                              "token %d content differs - target '%s', draft '%s'\n", __func__, i,
                         common_token_to_piece(ctx_tgt, i).c_str(),
                         common_token_to_piece(ctx_dft, i).c_str());

convert_hf_to_gguf.py

@@ -221,17 +221,17 @@ class Model:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
-        if (n_embd := self.find_hparam(["hidden_size", "n_embd"], optional=True)) is not None:
-            self.gguf_writer.add_embedding_length(n_embd)
-            logger.info(f"gguf: embedding length = {n_embd}")
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+        logger.info(f"gguf: embedding length = {n_embd}")
 
         if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
             logger.info(f"gguf: feed forward length = {n_ff}")
 
-        if (n_head := self.find_hparam(["num_attention_heads", "n_head"], optional=True)) is not None:
-            self.gguf_writer.add_head_count(n_head)
-            logger.info(f"gguf: head count = {n_head}")
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+        logger.info(f"gguf: head count = {n_head}")
 
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
@@ -296,9 +296,7 @@ class Model:
                     break
 
             for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
-                # TODO: why do we squeeze here?
-                # data = data_torch.squeeze().numpy()
-                data = data_torch.numpy()
+                data = data_torch.squeeze().numpy()
 
                 # if data ends up empty, it means data_torch was a scalar tensor -> restore
                 if len(data.shape) == 0:
@@ -326,9 +324,6 @@ class Model:
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
-                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
-                            gguf.MODEL_TENSOR.POSNET_NORM1,
-                            gguf.MODEL_TENSOR.POSNET_NORM2,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -478,11 +473,6 @@ class Model:
             return modelcls
         return func
 
-    @classmethod
-    def print_registered_models(cls):
-        for name in sorted(cls._model_classes.keys()):
-            logger.error(f"- {name}")
-
     @classmethod
     def from_model_architecture(cls, arch: str) -> type[Model]:
         try:
@@ -535,19 +525,9 @@ class Model:
             else:
                 token: str = reverse_vocab[i]
                 if token in added_vocab:
-                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
-                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
-                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
-                        if previous_token != token:
-                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
-
                     if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
-                        # NOTE: this was added for Gemma.
-                        # Encoding and decoding the tokens above isn't sufficient for this case.
                         token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
                         toktypes.append(gguf.TokenType.USER_DEFINED)
                 else:
@@ -591,9 +571,6 @@ class Model:
         if chkhsh == "8aeee3860c56296a157a1fe2fad249ec40aa59b1bb5709f4ade11c4e6fe652ed":
             # ref: https://huggingface.co/tiiuae/falcon-7b
             res = "falcon"
-        if chkhsh == "9d032fcbd5501f4a38150912590928bfb36091efb5df11b8e2124b0390e3fb1e":
-            # ref: https://huggingface.co/tiiuae/Falcon3-7B-Base
-            res = "falcon3"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
@@ -687,18 +664,6 @@ class Model:
         if chkhsh == "8b5a93ed704057481f240da0be7e7dca721d7f8f4755263b6807227a2cbeae65":
             # ref: https://huggingface.co/sentence-transformers/stsb-roberta-base
             res = "roberta-bpe"
-        if chkhsh == "ad851be1dba641f2e3711822f816db2c265f788b37c63b4e1aeacb9ee92de8eb":
-            # ref: https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct
-            res = "gigachat"
-        if chkhsh == "d4c8f286ea6b520b3d495c4455483cfa2302c0cfcd4be05d781b6a8a0a7cdaf1":
-            # ref: https://huggingface.co/Infinigence/Megrez-3B-Instruct
-            res = "megrez"
-        if chkhsh == "877081d19cf6996e2c4ff0e1236341e9b7bde288f5311a56a937f0afbbb3aeb5":
-            # ref: https://huggingface.co/deepseek-ai/DeepSeek-V3
-            res = "deepseek-v3"
-        if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5":
-            # ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
-            res = "deepseek-r1-qwen"
 
         if res is None:
             logger.warning("\n")
@@ -721,9 +686,6 @@ class Model:
         return res
         # Marker: End get_vocab_base_pre
 
-    def _set_vocab_none(self) -> None:
-        self.gguf_writer.add_tokenizer_model("none")
-
     def _set_vocab_gpt2(self) -> None:
         tokens, toktypes, tokpre = self.get_vocab_base()
         self.gguf_writer.add_tokenizer_model("gpt2")
@@ -1707,178 +1669,6 @@ class LlamaModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@Model.register("DeciLMForCausalLM")
-class DeciModel(Model):
-    model_arch = gguf.MODEL_ARCH.DECI
-
-    @staticmethod
-    def _ffn_mult_to_intermediate_size(ffn_mult: float, n_embd: int) -> int:
-        # DeciLM-specific code
-        intermediate_size = int(2 * ffn_mult * n_embd / 3)
-        return DeciModel._find_multiple(intermediate_size, 256)
-
-    @staticmethod
-    def _find_multiple(n: int, k: int) -> int:
-        # DeciLM-specific code
-        if n % k == 0:
-            return n
-        return n + k - (n % k)
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        if "block_configs" in self.hparams: # Llama-3_1-Nemotron-51B
-            _block_configs: list[dict[str,Any]] = self.hparams["block_configs"]
-            assert self.block_count == len(_block_configs)
-            self._num_kv_heads = list()
-            self._num_heads = list()
-            _ffn_multipliers = list()
-            # ***linear attention layer***
-            # if n_heads_in_group is None and replace_with_linear is True
-            # then _num_kv_heads[il] is 0 and _num_heads[il] is num_attention_heads
-            # ***attention-free layer***
-            # if n_heads_in_group is None and replace_with_linear is False
-            # then _num_kv_heads[il] is 0 and _num_heads[il] is 0
-            # ***normal attention-layer***
-            # if n_heads_in_group is not None, then
-            # _num_kv_heads[il] is num_attention_head // n_heads_in_group and
-            # _num_heads[il] is num_attention_head
-            for il in range(len(_block_configs)):
-                if _block_configs[il]["attention"]["n_heads_in_group"] is None:
-                    if _block_configs[il]["attention"]["replace_with_linear"] is True:
-                        self._num_kv_heads.append(0)
-                        self._num_heads.append(self.hparams["num_attention_heads"])
-                    else:
-                        self._num_kv_heads.append(0)
-                        self._num_heads.append(0)
-                else:
-                    self._num_kv_heads.append(self.hparams["num_attention_heads"] // _block_configs[il]["attention"]["n_heads_in_group"])
-                    self._num_heads.append(self.hparams["num_attention_heads"])
-                _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
-            assert self.block_count == len(self._num_kv_heads)
-            assert self.block_count == len(self._num_heads)
-            assert self.block_count == len(_ffn_multipliers)
-            assert isinstance(self._num_kv_heads, list) and isinstance(self._num_kv_heads[0], int)
-            assert isinstance(self._num_heads, list) and isinstance(self._num_heads[0], int)
-            assert isinstance(_ffn_multipliers, list) and isinstance(_ffn_multipliers[0], float)
-            self._ffn_dims: list[int] = [
-                DeciModel._ffn_mult_to_intermediate_size(multiplier, self.hparams["hidden_size"])
-                for multiplier in _ffn_multipliers
-            ]
-
-    def set_vocab(self):
-        # Please change tokenizer_config.json of Llama-3_1-Nemotron-51B's
-        # eos_token from '|eot_id|' to '|end_of_text|'
-        if self.hparams.get("vocab_size", 128256) == 128256:
-            tokens, toktypes, tokpre = self.get_vocab_base()
-            self.gguf_writer.add_tokenizer_model("gpt2")
-            self.gguf_writer.add_tokenizer_pre(tokpre)
-            self.gguf_writer.add_token_list(tokens)
-            self.gguf_writer.add_token_types(toktypes)
-
-            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-            special_vocab.add_to_gguf(self.gguf_writer)
-        else:
-            # DeciLM-7B
-            self._set_vocab_llama_hf()
-
-    def set_gguf_parameters(self):
-        if "block_configs" in self.hparams: # Llama-3_1-Nemotron-51B
-            assert self.block_count == len(self._num_kv_heads)
-            assert self.block_count == len(self._num_heads)
-            assert self.block_count == len(self._ffn_dims)
-            if (rope_theta := self.hparams.get("rope_theta")) is not None:
-                self.gguf_writer.add_rope_freq_base(rope_theta)
-            self.gguf_writer.add_head_count_kv(self._num_kv_heads)
-            self.gguf_writer.add_head_count(self._num_heads)
-            self.gguf_writer.add_feed_forward_length(self._ffn_dims)
-            self.gguf_writer.add_block_count(self.block_count)
-            self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-            self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-            self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
-            self.gguf_writer.add_key_length(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
-            self.gguf_writer.add_value_length(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
-            self.gguf_writer.add_file_type(self.ftype)
-        else: # DeciLM-7B
-            super().set_gguf_parameters()
-            if "num_key_value_heads_per_layer" in self.hparams: # DeciLM-7B
-                self._num_kv_heads: list[int] = self.hparams["num_key_value_heads_per_layer"]
-                assert self.block_count == len(self._num_kv_heads)
-                self.gguf_writer.add_head_count_kv(self._num_kv_heads)
-        hparams = self.hparams
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
-
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
-            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(rope_dim)
-
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-
-    @staticmethod
-    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
-        if n_head_kv is not None and n_head != n_head_kv:
-            n_head = n_head_kv
-        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                .swapaxes(1, 2)
-                .reshape(weights.shape))
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        n_head = self.hparams["num_attention_heads"]
-        if bid is not None:
-            if "num_key_value_heads_per_layer" in self.hparams:
-                n_kv_head = self.hparams["num_key_value_heads_per_layer"][bid]
-            elif "block_configs" in self.hparams:
-                n_kv_head = self._num_kv_heads[bid]
-                n_head = self._num_heads[bid]
-            else:
-                n_kv_head = self.hparams.get("num_key_value_heads")
-        else:
-            n_kv_head = self.hparams.get("num_key_value_heads")
-
-        if name.endswith(("q_proj.weight", "q_proj.bias")):
-            data_torch = DeciModel.permute(data_torch, n_head, n_head)
-        if name.endswith(("k_proj.weight", "k_proj.bias")):
-            data_torch = DeciModel.permute(data_torch, n_head, n_kv_head)
-        return [(self.map_tensor_name(name), data_torch)]
-
-    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
-            if rope_scaling.get("rope_type", '').lower() == "llama3":
-                base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
-                freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
-
-                factor = rope_scaling.get("factor", 8.0)
-                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
-                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
-                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
-
-                low_freq_wavelen = old_context_len / low_freq_factor
-                high_freq_wavelen = old_context_len / high_freq_factor
-                assert low_freq_wavelen != high_freq_wavelen
-
-                rope_factors = []
-                for freq in freqs:
-                    wavelen = 2 * math.pi / freq
-                    if wavelen < high_freq_wavelen:
-                        rope_factors.append(1)
-                    elif wavelen > low_freq_wavelen:
-                        rope_factors.append(factor)
-                    else:
-                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
-                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
-
-                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-
-
 @Model.register("BitnetForCausalLM")
 class BitnetModel(Model):
     model_arch = gguf.MODEL_ARCH.BITNET
@@ -2234,44 +2024,6 @@ class Qwen2VLModel(Model):
             yield name, data
 
 
-@Model.register("WavTokenizerDec")
-class WavTokenizerDecModel(Model):
-    model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
-        if \
-                name.endswith("codebook.cluster_size") or \
-                name.endswith("codebook.embed_avg") or \
-                name.endswith("codebook.inited"):
-            logger.debug(f"Skipping {name!r}")
-            return []
-
-        logger.info(f"{self.map_tensor_name(name)} -> {data_torch.shape}")
-
-        return [(self.map_tensor_name(name), data_torch)]
-
-    def set_vocab(self):
-        self._set_vocab_none()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_vocab_size         (self.hparams["vocab_size"])
-        self.gguf_writer.add_features_length    (self.hparams["n_embd_features"])
-        self.gguf_writer.add_feed_forward_length(self.hparams["n_ff"])
-        self.gguf_writer.add_group_norm_eps     (self.hparams["group_norm_epsilon"])
-        self.gguf_writer.add_group_norm_groups  (self.hparams["group_norm_groups"])
-
-        self.gguf_writer.add_posnet_embedding_length(self.hparams["posnet"]["n_embd"])
-        self.gguf_writer.add_posnet_block_count     (self.hparams["posnet"]["n_layer"])
-
-        self.gguf_writer.add_convnext_embedding_length(self.hparams["convnext"]["n_embd"])
-        self.gguf_writer.add_convnext_block_count     (self.hparams["convnext"]["n_layer"])
-
-        self.gguf_writer.add_causal_attention(False)
-
-
 @Model.register("Qwen2MoeForCausalLM")
 class Qwen2MoeModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2MOE
@@ -2400,15 +2152,6 @@ class Phi3MiniModel(Model):
     model_arch = gguf.MODEL_ARCH.PHI3
 
     def set_vocab(self):
-        # Phi-4 model uses GPT2Tokenizer
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                tokenizer_class = tokenizer_config_json['tokenizer_class']
-                if tokenizer_class == 'GPT2Tokenizer':
-                    return self._set_vocab_gpt2()
-
         from sentencepiece import SentencePieceProcessor
 
         tokenizer_path = self.dir_model / 'tokenizer.model'
@@ -2525,11 +2268,7 @@ class Phi3MiniModel(Model):
         self.gguf_writer.add_rope_dimension_count(rope_dims)
         self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
         self.gguf_writer.add_file_type(self.ftype)
-        sliding_window = self.hparams.get("sliding_window")
-        # use zero value of sliding_window to distinguish Phi-4 from other PHI3 models
-        if sliding_window is None:
-            sliding_window = 0
-        self.gguf_writer.add_sliding_window(sliding_window)
+        self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
@@ -2571,63 +2310,6 @@ class Phi3MiniModel(Model):
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
 
-@Model.register("PhiMoEForCausalLM")
-class PhiMoeModel(Phi3MiniModel):
-    model_arch = gguf.MODEL_ARCH.PHIMOE
-
-    _experts: list[dict[str, Tensor]] | None = None
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
-        self.gguf_writer.add_expert_count(self.hparams["num_local_experts"])
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # process the experts separately
-        if name.find("block_sparse_moe.experts") != -1:
-            n_experts = self.hparams["num_local_experts"]
-            assert bid is not None
-
-            if self._experts is None:
-                self._experts = [{} for _ in range(self.block_count)]
-
-            self._experts[bid][name] = data_torch
-
-            if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
-                # merge the experts into a single 3d tensor
-                for w_name in ["w1", "w2", "w3"]:
-                    datas: list[Tensor] = []
-
-                    for xid in range(n_experts):
-                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"
-                        datas.append(self._experts[bid][ename])
-                        del self._experts[bid][ename]
-
-                    data_torch = torch.stack(datas, dim=0)
-
-                    merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
-
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
-            else:
-                return []
-
-        return [(self.map_tensor_name(name), data_torch)]
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-
-        if self._experts is not None:
-            # flatten `list[dict[str, Tensor]]` into `list[str]`
-            experts = [k for d in self._experts for k in d.keys()]
-            if len(experts) > 0:
-                raise ValueError(f"Unprocessed experts: {experts}")
-
-
 @Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
     model_arch = gguf.MODEL_ARCH.PLAMO
@@ -2885,67 +2567,7 @@ class InternLM2Model(Model):
             return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("InternLM3ForCausalLM")
-class InternLM3Model(Model):
-    model_arch = gguf.MODEL_ARCH.LLAMA
-
-    def set_vocab(self):
-        tokens, scores, toktypes = self._create_vocab_sentencepiece()
-
-        self.gguf_writer.add_tokenizer_model("llama")
-        self.gguf_writer.add_tokenizer_pre("default")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-                if "added_tokens_decoder" in tokenizer_config_json:
-                    for token_id, token_data in tokenizer_config_json["added_tokens_decoder"].items():
-                        if token_data.get("special"):
-                            token_id = int(token_id)
-                            token = token_data["content"]
-                            special_vocab._set_special_token(token, token_id)
-                            # update eos token
-                            if token == '<|im_end|>' and "eos" in special_vocab.special_token_ids:
-                                special_vocab.special_token_ids["eos"] = token_id
-
-        special_vocab.add_to_gguf(self.gguf_writer)
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hparams = self.hparams
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
-
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
-            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(rope_dim)
-
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        n_head = self.hparams["num_attention_heads"]
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        if name.endswith(("q_proj.weight", "q_proj.bias")):
-            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
-        if name.endswith(("k_proj.weight", "k_proj.bias")):
-            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
-        return [(self.map_tensor_name(name), data_torch)]
-
-
-@Model.register("BertModel", "BertForMaskedLM", "CamembertModel")
+@Model.register("BertModel", "CamembertModel", "RobertaModel")
 class BertModel(Model):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -3011,73 +2633,13 @@ class BertModel(Model):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-        if name.startswith("bert."):
-            name = name[5:]
-
-        if name.endswith(".gamma"):
-            name = name[:-6] + ".weight"
-
-        if name.endswith(".beta"):
-            name = name[:-5] + ".bias"
-
         # we are only using BERT for embeddings so we don't need the pooling layer
         if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
             return [] # we don't need these
 
-        if name.startswith("cls.predictions"):
-            return []
-
-        if name.startswith("cls.seq_relationship"):
-            return []
-
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("RobertaModel")
-class RobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # we need the pad_token_id to know how to chop down position_embd matrix
-        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
-            self._position_offset = 1 + pad_token_id
-            if "max_position_embeddings" in self.hparams:
-                self.hparams["max_position_embeddings"] -= self._position_offset
-        else:
-            self._position_offset = None
-
-    def set_vocab(self):
-        """Support BPE tokenizers for roberta models"""
-        bpe_tok_path = self.dir_model / "tokenizer.json"
-        if bpe_tok_path.exists():
-            self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
-
-            # we need this to validate the size of the token_type embeddings
-            # though currently we are passing all zeros to the token_type embeddings
-            # "Sequence A" or "Sequence B"
-            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
-
-        else:
-            return super().set_vocab()
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "roberta.", remove the prefix
-        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
-        if name.startswith("roberta."):
-            name = name[8:]
-
-        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
-        if name == "embeddings.position_embeddings.weight":
-            if self._position_offset is not None:
-                data_torch = data_torch[self._position_offset:,:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
     model_arch = gguf.MODEL_ARCH.NOMIC_BERT
@@ -3385,8 +2947,6 @@ class Rwkv6Model(Model):
         # required by llama.cpp, unused
         self.gguf_writer.add_head_count(0)
 
-    lerp_weights: dict[int, dict[str, Tensor]] = {}
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         new_name = self.map_tensor_name(name)
 
@@ -3399,87 +2959,14 @@ class Rwkv6Model(Model):
         if new_name.endswith("time_mix_w2.weight"):
             data_torch = data_torch.permute(0, 2, 1)
 
-        if new_name.endswith("time_mix_decay.weight") or "lerp" in new_name:
-            data_torch = data_torch.squeeze()
-
-        try:
-            rescale_every_n_layers = self.hparams["rescale_every"]
-            if rescale_every_n_layers > 0:
-                if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
-                    data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
-        except KeyError:
-            pass
-
-        # concat time_mix_lerp weights to reduce some cpu overhead
-        # also reduces the number of tensors in the model
-        if bid is not None and "time_mix_lerp" in new_name and "time_mix_lerp_x" not in new_name:
-            try:
-                self.lerp_weights[bid][new_name] = data_torch
-            except KeyError:
-                self.lerp_weights[bid] = {new_name: data_torch}
-            if all(f"blk.{bid}.time_mix_lerp_{i}.weight" in self.lerp_weights[bid].keys() for i in ["w", "k", "v", "r", "g"]):
-                new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
-                data = torch.stack([self.lerp_weights[bid][f"blk.{bid}.time_mix_lerp_{i}.weight"].unsqueeze(0) for i in ["w", "k", "v", "r", "g"]], dim=0).unsqueeze(1)
-                yield (new_name, data)
-            return
+        rescale_every_n_layers = self.hparams["rescale_every"]
+        if rescale_every_n_layers > 0:
+            if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
+                data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
 
         yield (new_name, data_torch)
 
 
-@Model.register("RWKV6Qwen2ForCausalLM")
-class RWKV6Qwen2Model(Rwkv6Model):
-    model_arch = gguf.MODEL_ARCH.RWKV6QWEN2
-
-    def set_vocab(self):
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-    def set_gguf_parameters(self):
-        block_count = self.hparams["num_hidden_layers"]
-        num_attention_heads = self.hparams["num_attention_heads"]
-        num_key_value_heads = self.hparams["num_key_value_heads"]
-        hidden_size = self.hparams["hidden_size"]
-        head_size = hidden_size // num_attention_heads
-        rms_norm_eps = self.hparams["rms_norm_eps"]
-        intermediate_size = self.hparams["intermediate_size"]
-        time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
-        time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
-
-        # RWKV isn't context limited
-        self.gguf_writer.add_context_length(1048576)
-        self.gguf_writer.add_embedding_length(hidden_size)
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_wkv_head_size(head_size)
-        self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
-        self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
-        self.gguf_writer.add_feed_forward_length(intermediate_size)
-        self.gguf_writer.add_file_type(self.ftype)
-
-        # special parameters for time_mixing in RWKV6QWEN2
-        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_token_shift_count(1)
-        # RWKV6QWEN2 use grouped key/value like GQA
-        self.gguf_writer.add_head_count_kv(num_key_value_heads)
-
-        # required by llama.cpp, unused
-        self.gguf_writer.add_head_count(0)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        for new_name, data in super().modify_tensors(data_torch, name, bid):
-            if "time_mix_w1" in new_name or "time_mix_w2" in new_name:
-                data = data.view(5, -1, data.shape[-1])
-                # rwkv6qwen2 has a different order of rkvwg instead of the original wkvrg
-                # permute them here to avoid code changes
-                data = torch.stack([data[3], data[1], data[2], data[0], data[4]], dim=0).view(-1, data.shape[-1])
-                if "w2" in new_name:
-                    data = data.view(5, -1, data.shape[-1])
-                yield (new_name, data)
-                continue
-            yield (new_name, data)
-
-
 @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
@@ -3574,24 +3061,6 @@ class CommandR2Model(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
 
 
-@Model.register("Cohere2ForCausalLM")
-class Cohere2Model(Model):
-    model_arch = gguf.MODEL_ARCH.COHERE2
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-
-        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
-        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
-        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
-
-        rotary_pct = self.hparams["rotary_pct"]
-        hidden_size = self.hparams["hidden_size"]
-        num_attention_heads = self.hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(int(rotary_pct * (hidden_size // num_attention_heads)))
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
-
-
 @Model.register("OlmoForCausalLM")
 @Model.register("OLMoForCausalLM")
 class OlmoModel(Model):
@@ -3958,99 +3427,7 @@ class ArcticModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@Model.register("DeepseekForCausalLM")
-class DeepseekModel(Model):
-    model_arch = gguf.MODEL_ARCH.DEEPSEEK
-
-    def set_vocab(self):
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hparams = self.hparams
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
-            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
-
-        self.gguf_writer.add_rope_dimension_count(rope_dim)
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
-        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
-        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
-        self.gguf_writer.add_expert_weights_scale(1.0)
-        self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
-        self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
-
-    _experts: list[dict[str, Tensor]] | None = None
-
-    @staticmethod
-    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
-        if n_head_kv is not None and n_head != n_head_kv:
-            n_head = n_head_kv
-        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                .swapaxes(1, 2)
-                .reshape(weights.shape))
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        n_head = self.hparams["num_attention_heads"]
-        n_kv_head = self.hparams.get("num_key_value_heads")
-
-        if name.endswith(("q_proj.weight", "q_proj.bias")):
-            data_torch = DeepseekModel.permute(data_torch, n_head, n_head)
-        if name.endswith(("k_proj.weight", "k_proj.bias")):
-            data_torch = DeepseekModel.permute(data_torch, n_head, n_kv_head)
-
-        # process the experts separately
-        if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["n_routed_experts"]
-            assert bid is not None
-
-            if self._experts is None:
-                self._experts = [{} for _ in range(self.block_count)]
-
-            self._experts[bid][name] = data_torch
-
-            if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
-                # merge the experts into a single 3d tensor
-                for w_name in ["down_proj", "gate_proj", "up_proj"]:
-                    datas: list[Tensor] = []
-
-                    for xid in range(n_experts):
-                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                        datas.append(self._experts[bid][ename])
-                        del self._experts[bid][ename]
-
-                    data_torch = torch.stack(datas, dim=0)
-
-                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
-            else:
-                return []
-
-        return [(self.map_tensor_name(name), data_torch)]
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-
-        if self._experts is not None:
-            # flatten `list[dict[str, Tensor]]` into `list[str]`
-            experts = [k for d in self._experts for k in d.keys()]
-            if len(experts) > 0:
-                raise ValueError(f"Unprocessed experts: {experts}")
-
-
 @Model.register("DeepseekV2ForCausalLM")
-@Model.register("DeepseekV3ForCausalLM")
 class DeepseekV2Model(Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -4072,15 +3449,6 @@ class DeepseekV2Model(Model):
         self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
         self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
-
-        if hparams["scoring_func"] == "sigmoid":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        elif hparams["scoring_func"] == "softmax":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {hparams['scoring_func']}")
-
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
@@ -4093,16 +3461,6 @@ class DeepseekV2Model(Model):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # rename e_score_correction_bias tensors
-        if name.endswith("e_score_correction_bias"):
-            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
-
-        # skip Multi-Token Prediction (MTP) layers
-        block_count = self.hparams["num_hidden_layers"]
-        match = re.match(r"model.layers.(\d+)", name)
-        if match and int(match.group(1)) >= block_count:
-            return []
-
         # process the experts separately
         if name.find("mlp.experts") != -1:
             n_experts = self.hparams["n_routed_experts"]
@@ -4997,7 +4355,6 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "model", type=Path,
         help="directory containing model file",
-        nargs="?",
     )
     parser.add_argument(
         "--use-temp-file", action="store_true",
@@ -5035,15 +4392,8 @@ def parse_args() -> argparse.Namespace:
         "--metadata", type=Path,
         help="Specify the path for an authorship metadata override file"
     )
-    parser.add_argument(
-        "--print-supported-models", action="store_true",
-        help="Print the supported models"
-    )
 
-    args = parser.parse_args()
-    if not args.print_supported_models and args.model is None:
-        parser.error("the following arguments are required: model")
-    return args
+    return parser.parse_args()
 
 
 def split_str_to_n_bytes(split_str: str) -> int:
@@ -5067,11 +4417,6 @@ def split_str_to_n_bytes(split_str: str) -> int:
 def main() -> None:
     args = parse_args()
 
-    if args.print_supported_models:
-        logger.error("Supported models:")
-        Model.print_registered_models()
-        sys.exit(0)
-
     if args.verbose:
         logging.basicConfig(level=logging.DEBUG)
     else:

convert_hf_to_gguf_update.py

@@ -65,50 +65,45 @@ else:
 
 # TODO: add models here, base models preferred
 models = [
-    {"name": "llama-spm",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
-    {"name": "llama-bpe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
-    {"name": "phi-3",            "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
-    {"name": "deepseek-llm",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
-    {"name": "deepseek-coder",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
-    {"name": "falcon",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
-    {"name": "bert-bge",         "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
-    {"name": "falcon3",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon3-7B-Base", },
-    {"name": "bert-bge-large",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
-    {"name": "mpt",              "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
-    {"name": "starcoder",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
-    {"name": "gpt-2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
-    {"name": "stablelm2",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
-    {"name": "refact",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
-    {"name": "command-r",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
-    {"name": "qwen2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
-    {"name": "olmo",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
-    {"name": "dbrx",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
-    {"name": "jina-v1-en",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
-    {"name": "jina-v2-en",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
-    {"name": "jina-v2-es",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
-    {"name": "jina-v2-de",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
-    {"name": "smaug-bpe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
-    {"name": "poro-chat",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
-    {"name": "jina-v2-code",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
-    {"name": "viking",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
-    {"name": "gemma",            "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
-    {"name": "gemma-2",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
-    {"name": "jais",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
-    {"name": "t5",               "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
-    {"name": "codeshell",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
-    {"name": "tekken",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
-    {"name": "smollm",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
-    {'name': "bloom",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
-    {'name': "gpt3-finnish",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
-    {"name": "exaone",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
-    {"name": "phi-2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
-    {"name": "chameleon",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
-    {"name": "minerva-7b",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", },
-    {"name": "roberta-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"},
-    {"name": "gigachat",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct"},
-    {"name": "megrez",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
-    {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
-    {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
+    {"name": "llama-spm",      "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
+    {"name": "llama-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
+    {"name": "phi-3",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
+    {"name": "deepseek-llm",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
+    {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
+    {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
+    {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+    {"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
+    {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
+    {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
+    {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
+    {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
+    {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
+    {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
+    {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
+    {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v1-en",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
+    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
+    {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
+    {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
+    {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
+    {"name": "gemma",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
+    {"name": "gemma-2",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
+    {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "t5",             "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
+    {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
+    {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
+    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
+    {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
+    {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
+    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
+    {"name": "minerva-7b",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", },
+    {"name": "roberta-bpe",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"},
 ]
 
 

convert_lora_to_gguf.py

@@ -226,9 +226,6 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
     base_name = lora_tensor_name.replace("base_model.model.", "")
     base_name = base_name.replace(".lora_A.weight", ".weight")
     base_name = base_name.replace(".lora_B.weight", ".weight")
-    # models produced by mergekit-extract-lora have token embeddings in the adapter
-    base_name = base_name.replace(".lora_embedding_A", ".weight")
-    base_name = base_name.replace(".lora_embedding_B", ".weight")
     return base_name
 
 
@@ -263,10 +260,6 @@ def parse_args() -> argparse.Namespace:
         "--base", type=Path,
         help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
     )
-    parser.add_argument(
-        "--base-model-id", type=str,
-        help="the model ID of the base model, if it is not available locally or in the adapter config. If specified, it will ignore --base and load the base model config from the Hugging Face hub (Example: 'meta-llama/Llama-3.2-1B-Instruct')",
-    )
     parser.add_argument(
         "lora_path", type=Path,
         help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
@@ -297,7 +290,6 @@ if __name__ == '__main__':
 
     dir_base_model: Path | None = args.base
     dir_lora: Path = args.lora_path
-    base_model_id: str | None = args.base_model_id
     lora_config = dir_lora / "adapter_config.json"
     input_model = dir_lora / "adapter_model.safetensors"
 
@@ -321,10 +313,7 @@ if __name__ == '__main__':
         lparams: dict[str, Any] = json.load(f)
 
     # load base model
-    if base_model_id is not None:
-        logger.info(f"Loading base model from Hugging Face: {base_model_id}")
-        hparams = load_hparams_from_hf(base_model_id)
-    elif dir_base_model is None:
+    if dir_base_model is None:
         if "base_model_name_or_path" in lparams:
             model_id = lparams["base_model_name_or_path"]
             logger.info(f"Loading base model from Hugging Face: {model_id}")
@@ -382,16 +371,11 @@ if __name__ == '__main__':
                     if self.lazy:
                         tensor = LazyTorchTensor.from_eager(tensor)
                     base_name = get_base_tensor_name(name)
-                    # note: mergekit-extract-lora also adds token embeddings to the adapter
-                    is_lora_a = ".lora_A.weight" in name or ".lora_embedding_A" in name
-                    is_lora_b = ".lora_B.weight" in name or ".lora_embedding_B" in name
+                    is_lora_a = ".lora_A.weight" in name
+                    is_lora_b = ".lora_B.weight" in name
                     if not is_lora_a and not is_lora_b:
                         if ".base_layer.weight" in name:
                             continue
-                        # mergekit-extract-lora add these layernorm to the adapter, we need to keep them
-                        if "_layernorm" in name or ".norm" in name:
-                            yield (base_name, tensor)
-                            continue
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
                         if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
                             logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
@@ -423,21 +407,9 @@ if __name__ == '__main__':
                 if name == "lm_head.weight" and len(dest) == 0:
                     raise ValueError("lm_head is present in adapter, but is ignored in base model")
                 for dest_name, dest_data in dest:
-                    # mergekit-extract-lora add these layernorm to the adapter
-                    if "_norm" in dest_name:
-                        assert dest_data.dim() == 1
-                        yield (dest_name, dest_data)
-                        continue
-
-                    # otherwise, we must get the lora_A and lora_B tensors
                     assert isinstance(dest_data, LoraTorchTensor)
                     lora_a, lora_b = dest_data.get_lora_A_B()
 
-                    # note: mergekit-extract-lora flip and transpose A and B
-                    # here we only need to transpose token_embd.lora_a, see llm_build_inp_embd()
-                    if "token_embd.weight" in dest_name:
-                        lora_a = lora_a.T
-
                     yield (dest_name + ".lora_a", lora_a)
                     yield (dest_name + ".lora_b", lora_b)
 

docs/backend/SYCL.md

@@ -133,7 +133,7 @@ The docker build option is currently limited to *intel GPU* targets.
 ### Build image
 ```sh
 # Using FP16
-docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
+docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" -f .devops/llama-cli-intel.Dockerfile .
 ```
 
 *Notes*:

docs/build.md

@@ -127,8 +127,6 @@ For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
 
 This provides GPU acceleration using an NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from the [NVIDIA developer site](https://developer.nvidia.com/cuda-downloads).
 
-If you are using Fedora (using Fedora Workstation, or an 'Atomic' variant such as Silverblue), or would like to set up CUDA in a toolbox, please consider our [Fedora CUDA guide](./cuda-fedora.md). Unfortunately, the process is not as simple as one might expect.
-
 - Using `CMake`:
 
   ```bash
@@ -286,7 +284,7 @@ You don't need to install Vulkan SDK. It will be installed inside the container.
 
 ```sh
 # Build the image
-docker build -t llama-cpp-vulkan --target light -f .devops/vulkan.Dockerfile .
+docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
 
 # Then, use it:
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33

docs/cuda-fedora.md

@@ -1,317 +0,0 @@
-# Setting Up CUDA on Fedora
-
-In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox container. This guide is applicable for:
-- [Fedora Workstation](https://fedoraproject.org/workstation/)
-- [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/)
-- [Fedora Spins](https://fedoraproject.org/spins)
-- [Other Distributions](https://containertoolbx.org/distros/), including `Red Hat Enterprise Linux >= 8.`, `Arch Linux`, and `Ubuntu`.
-
-
-## Table of Contents
-
-- [Prerequisites](#prerequisites)
-- [Monitoring NVIDIA CUDA Repositories](#monitoring-nvidia-cuda-repositories)
-- [Using the Fedora 39 CUDA Repository](#using-the-fedora-39-cuda-repository)
-- [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
-- [Installing Essential Development Tools](#installing-essential-development-tools)
-- [Adding the CUDA Repository](#adding-the-cuda-repository)
-- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
-- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
-- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
-- [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
-- [Configuring the Environment](#configuring-the-environment)
-- [Verifying the Installation](#verifying-the-installation)
-- [Conclusion](#conclusion)
-- [Troubleshooting](#troubleshooting)
-- [Additional Notes](#additional-notes)
-- [References](#references)
-
-## Prerequisites
-
-- **Toolbox Installed on the Host System** `Fedora Silverblue` and `Fedora Workstation` both have toolbox by default, other distributions may need to install the [toolbox package](https://containertoolbx.org/install/).
-- **NVIDIA Drivers and Graphics Card installed on Host System (optional)** To run CUDA program, such as `llama.cpp`, the host should be setup to access your NVIDIA hardware. Fedora Hosts can use the [RPM Fusion Repository](https://rpmfusion.org/Howto/NVIDIA).
-- **Internet connectivity** to download packages.
-
-### Monitoring NVIDIA CUDA Repositories
-
-Before proceeding, it is advisable to check if NVIDIA has updated their CUDA repositories for your Fedora version. NVIDIA's repositories can be found at:
-
-- [Fedora 40 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora40/x86_64/)
-- [Fedora 41 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora41/x86_64/)
-
-As of the latest update, these repositories do not contain the `cuda` meta-package or are missing essential components.
-
-### Using the Fedora 39 CUDA Repository
-
-Since the newer repositories are incomplete, we'll use the Fedora 39 repository:
-
-- [Fedora 39 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora39/x86_64/)
-
-**Note:** Fedora 39 is no longer maintained, so we recommend using a toolbox environment to prevent system conflicts.
-
-## Creating a Fedora Toolbox Environment
-
-This guide focuses on Fedora hosts, but with small adjustments, it can work for other hosts. Using a Fedora 39 toolbox allows us to install the necessary packages without affecting the host system.
-
-**Note:** Toolbox is available for other systems, and even without Toolbox, it is possible to use Podman or Docker.
-
-We do not recommend installing on the host system, as Fedora 39 is out-of-maintenance, and instead you should upgrade to a maintained version of Fedora for your host.
-
-1. **Create a Fedora 39 Toolbox:**
-
-   ```bash
-   toolbox create --image registry.fedoraproject.org/fedora-toolbox:39 --container fedora-toolbox-39-cuda
-   ```
-
-2. **Enter the Toolbox:**
-
-   ```bash
-   toolbox enter --container fedora-toolbox-39-cuda
-   ```
-
-   Inside the toolbox, you have root privileges and can install packages without affecting the host system.
-
-## Installing Essential Development Tools
-
-1. **Synchronize the DNF Package Manager:**
-
-   ```bash
-   sudo dnf distro-sync
-   ```
-
-2. **Install the Default Text Editor (Optional):**
-
-   ```bash
-   sudo dnf install vim-default-editor --allowerasing
-   ```
-
-   The `--allowerasing` flag resolves any package conflicts.
-
-3. **Install Development Tools and Libraries:**
-
-   ```bash
-   sudo dnf install @c-development @development-tools cmake
-   ```
-
-   This installs essential packages for compiling software, including `gcc`, `make`, and other development headers.
-
-## Adding the CUDA Repository
-
-Add the NVIDIA CUDA repository to your DNF configuration:
-
-```bash
-sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/fedora39/x86_64/cuda-fedora39.repo
-```
-
-After adding the repository, synchronize the package manager again:
-
-```bash
-sudo dnf distro-sync
-```
-
-## Installing `nvidia-driver-libs`
-
-Attempt to install `nvidia-driver-libs`:
-
-```bash
-sudo dnf install nvidia-driver-libs
-```
-
-**Explanation:**
-
-- `nvidia-driver-libs` contains necessary NVIDIA driver libraries required by CUDA.
-- This step might fail due to conflicts with existing NVIDIA drivers on the host system.
-
-## Manually Resolving Package Conflicts
-
-If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
-
-### 1. Download the `nvidia-driver-libs` RPM
-
-```bash
-sudo dnf download --arch x86_64 nvidia-driver-libs
-```
-
-You should see a file similar to:
-
-```
-nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
-```
-
-### 2. Attempt to Install the RPM
-
-```bash
-sudo dnf install nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
-```
-
-**Expected Error:**
-
-Installation may fail with errors pointing to conflicts with `egl-gbm` and `egl-wayland`.
-
-**Note: It is important to carefully read the error messages to identify the exact paths that need to be excluded.**
-
-### 3. Download Dependencies
-
-```bash
-sudo dnf download --arch x86_64 egl-gbm egl-wayland
-```
-
-### 4. Install `egl-gbm` with Excluded Paths
-
-Exclude conflicting files during installation:
-
-```bash
-sudo rpm --install --verbose --hash \
-  --excludepath=/usr/lib64/libnvidia-egl-gbm.so.1.1.2 \
-  --excludepath=/usr/share/egl/egl_external_platform.d/15_nvidia_gbm.json \
-  egl-gbm-1.1.2^20240919gitb24587d-3.fc39.x86_64.rpm
-```
-
-**Explanation:**
-
-- The `--excludepath` option skips installing files that conflict with existing files.
-- Adjust the paths based on the error messages you receive.
-
-### 5. Install `egl-wayland` with Excluded Paths
-
-```bash
-sudo rpm --install --verbose --hash \
-  --excludepath=/usr/share/egl/egl_external_platform.d/10_nvidia_wayland.json \
-  egl-wayland-1.1.17^20241118giteeb29e1-5.fc39.x86_64.rpm
-```
-
-### 6. Install `nvidia-driver-libs` with Excluded Paths
-
-```bash
-sudo rpm --install --verbose --hash \
-  --excludepath=/usr/share/glvnd/egl_vendor.d/10_nvidia.json \
-  --excludepath=/usr/share/nvidia/nvoptix.bin \
-  nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
-```
-
-**Note:**
-
-- Replace the paths with the ones causing conflicts in your installation if they differ.
-- The `--verbose` and `--hash` options provide detailed output during installation.
-
-## Finalizing the Installation of `nvidia-driver-libs`
-
-After manually installing the dependencies, run:
-
-```bash
-sudo dnf install nvidia-driver-libs
-```
-
-You should receive a message indicating the package is already installed:
-
-```
-Package nvidia-driver-libs-3:560.35.05-1.fc39.x86_64 is already installed.
-Dependencies resolved.
-Nothing to do.
-Complete!
-```
-
-## Installing the CUDA Meta-Package
-
-Now that the driver libraries are installed, proceed to install CUDA:
-
-```bash
-sudo dnf install cuda
-```
-
-This installs the CUDA toolkit and associated packages.
-
-## Configuring the Environment
-
-To use CUDA, add its binary directory to your system's `PATH`.
-
-1. **Create a Profile Script:**
-
-   ```bash
-   sudo sh -c 'echo "export PATH=\$PATH:/usr/local/cuda/bin" >> /etc/profile.d/cuda.sh'
-   ```
-
-   **Explanation:**
-
-   - We add to  `/etc/profile.d/` as the `/etc/` folder is unique to this particular container, and is not shared with other containers or the host system.
-   - The backslash `\` before `$PATH` ensures the variable is correctly written into the script.
-
-2. **Make the Script Executable:**
-
-   ```bash
-   sudo chmod +x /etc/profile.d/cuda.sh
-   ```
-
-3. **Source the Script to Update Your Environment:**
-
-   ```bash
-   source /etc/profile.d/cuda.sh
-   ```
-
-   **Note:** This command updates your current shell session with the new `PATH`. The `/etc/profile.d/cuda.sh` script ensures that the CUDA binaries are available in your `PATH` for all future sessions.
-
-## Verifying the Installation
-
-To confirm that CUDA is correctly installed and configured, check the version of the NVIDIA CUDA Compiler (`nvcc`):
-
-```bash
-nvcc --version
-```
-
-You should see output similar to:
-
-```
-nvcc: NVIDIA (R) Cuda compiler driver
-Copyright (c) 2005-2024 NVIDIA Corporation
-Built on Tue_Oct_29_23:50:19_PDT_2024
-Cuda compilation tools, release 12.6, V12.6.85
-Build cuda_12.6.r12.6/compiler.35059454_0
-```
-
-This output confirms that the CUDA compiler is accessible and indicates the installed version.
-
-## Conclusion
-
-You have successfully set up CUDA on Fedora within a toolbox environment using the Fedora 39 CUDA repository. By manually resolving package conflicts and configuring the environment, you can develop CUDA applications without affecting your host system.
-
-## Troubleshooting
-
-- **Installation Failures:**
-  - If you encounter errors during installation, carefully read the error messages. They often indicate conflicting files or missing dependencies.
-  - Use the `--excludepath` option with `rpm` to exclude conflicting files during manual installations.
-
-- **Driver Conflicts:**
-  - Since the host system may already have NVIDIA drivers installed, conflicts can arise. Using the toolbox environment helps isolate these issues.
-
-- **Environment Variables Not Set:**
-  - If `nvcc` is not found after installation, ensure that `/usr/local/cuda/bin` is in your `PATH`.
-  - Run `echo $PATH` to check if the path is included.
-  - Re-source the profile script or open a new terminal session.
-
-## Additional Notes
-
-- **Updating CUDA in the Future:**
-  - Keep an eye on the official NVIDIA repositories for updates to your Fedora version.
-  - When an updated repository becomes available, adjust your `dnf` configuration accordingly.
-
-- **Building `llama.cpp`:**
-  - With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
-  - Ensure that any CUDA-specific build flags or paths are correctly set in your build configuration.
-
-- **Using the Toolbox Environment:**
-  - The toolbox environment is isolated from your host system, which helps prevent conflicts.
-  - Remember that system files and configurations inside the toolbox are separate from the host. By default the home directory of the user is shared between the host and the toolbox.
-
----
-
-**Disclaimer:** Manually installing and modifying system packages can lead to instability of the container. The above steps are provided as a guideline and may need adjustments based on your specific system configuration. Always back up important data before making significant system changes, especially as your home folder is writable and shared with he toolbox.
-
-**Acknowledgments:** Special thanks to the Fedora community and NVIDIA documentation for providing resources that assisted in creating this guide.
-
-## References
-
-- [Fedora Toolbox Documentation](https://docs.fedoraproject.org/en-US/fedora-silverblue/toolbox/)
-- [NVIDIA CUDA Installation Guide](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html)
-- [Podman Documentation](https://podman.io/get-started)
-
----

docs/development/HOWTO-add-model.md

@@ -28,7 +28,7 @@ The required steps to implement for an HF model are:
 ```python
 @Model.register("MyModelForCausalLM")
 class MyModel(Model):
-    model_arch = gguf.MODEL_ARCH.MYMODEL
+    model_arch = gguf.MODEL_ARCH.GROK
 ```
 
 2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)
@@ -79,14 +79,14 @@ Depending on the model configuration, tokenizer, code and tensors layout, you wi
 - `Model#set_vocab`
 - `Model#write_tensors`
 
-NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the convention and several tools like `quantize` expect this to proceed the weights.
+NOTE: Tensor names must end with `.weight` suffix, that is the convention and several tools like `quantize` expect this to proceed the weights.
 
 ### 2. Define the model architecture in `llama.cpp`
 
 The model params and tensors layout must be defined in `llama.cpp`:
 1. Define a new `llm_arch`
 2. Define the tensors layout in `LLM_TENSOR_NAMES`
-3. Add any non-standard metadata in `llm_load_hparams`
+3. Add any non standard metadata in `llm_load_hparams`
 4. Create the tensors for inference in `llm_load_tensors`
 5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
 
@@ -96,9 +96,9 @@ NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorc
 
 This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
 
-Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
+Have a look at existing implementation like `build_llama`, `build_dbrx` or `build_bert`.
 
-Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
+When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
 
 Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
 

docs/docker.md

@@ -60,9 +60,9 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ## Building Docker locally
 
 ```bash
-docker build -t local/llama.cpp:full-cuda --target full -f .devops/cuda.Dockerfile .
-docker build -t local/llama.cpp:light-cuda --target light -f .devops/cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda --target server -f .devops/cuda.Dockerfile .
+docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -95,9 +95,9 @@ Assuming one has the [mt-container-toolkit](https://developer.mthreads.com/musa/
 ## Building Docker locally
 
 ```bash
-docker build -t local/llama.cpp:full-musa --target full -f .devops/musa.Dockerfile .
-docker build -t local/llama.cpp:light-musa --target light -f .devops/musa.Dockerfile .
-docker build -t local/llama.cpp:server-musa --target server -f .devops/musa.Dockerfile .
+docker build -t local/llama.cpp:full-musa -f .devops/full-musa.Dockerfile .
+docker build -t local/llama.cpp:light-musa -f .devops/llama-cli-musa.Dockerfile .
+docker build -t local/llama.cpp:server-musa -f .devops/llama-server-musa.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the MUSA environment supported by your container host, as well as the GPU architecture.

examples/CMakeLists.txt

@@ -51,7 +51,6 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
-    add_subdirectory(tts)
     add_subdirectory(gen-docs)
     if (NOT GGML_BACKEND_DL)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/batched-bench/batched-bench.cpp

@@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@@ -50,7 +50,7 @@ int main(int argc, char ** argv) {
     // ensure enough sequences are available
     ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
-    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
@@ -194,7 +194,7 @@ int main(int argc, char ** argv) {
     llama_batch_free(batch);
 
     llama_free(ctx);
-    llama_model_free(model);
+    llama_free_model(model);
 
     llama_backend_free();
 

examples/batched.swift/Sources/main.swift

@@ -23,12 +23,12 @@ defer {
 }
 
 let model_params = llama_model_default_params()
-guard let model = llama_model_load_from_file(modelPath.cString(using: .utf8), model_params) else {
+guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
     print("Failed to load model")
     exit(1)
 }
 defer {
-    llama_model_free(model)
+    llama_free_model(model)
 }
 
 var tokens = tokenize(text: prompt, add_bos: true)
@@ -141,7 +141,7 @@ while n_cur <= n_len {
         let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
 
         // is it an end of stream? -> mark the stream as finished
-        if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
+        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {

examples/batched/batched.cpp

@@ -41,19 +41,17 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
-    tokens_list = common_tokenize(vocab, params.prompt, true);
+    tokens_list = common_tokenize(model, params.prompt, true);
 
     const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
@@ -64,10 +62,9 @@ int main(int argc, char ** argv) {
     ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_predict, n_parallel);
 
-    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     auto sparams = llama_sampler_chain_default_params();
-    sparams.no_perf = false;
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
@@ -122,8 +119,8 @@ int main(int argc, char ** argv) {
         }
 
         llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
-            decoder_start_token_id = llama_vocab_bos(vocab);
+        if (decoder_start_token_id == -1) {
+            decoder_start_token_id = llama_token_bos(model);
         }
 
         common_batch_clear(batch);
@@ -176,7 +173,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
 
             // is it an end of generation? -> mark the stream as finished
-            if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_predict) {
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
                 i_batch[i] = -1;
                 LOG("\n");
                 if (n_parallel > 1) {
@@ -238,7 +235,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_model_free(model);
+    llama_free_model(model);
 
     llama_backend_free();
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -1,6 +1,4 @@
 #include "ggml.h"
-#include "gguf.h"
-
 #include "llama.h"
 #include "common.h"
 #include "log.h"
@@ -436,12 +434,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-struct my_llama_file {
+struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
     size_t size;
 
-    my_llama_file(const char * fname, const char * mode) {
+    llama_file(const char * fname, const char * mode) {
         fp = std::fopen(fname, mode);
         if (fp == NULL) {
             size = 0;
@@ -502,7 +500,7 @@ struct my_llama_file {
         return std::string(chars.data(), len);
     }
 
-    ~my_llama_file() {
+    ~llama_file() {
         if (fp) {
             std::fclose(fp);
         }
@@ -510,7 +508,7 @@ struct my_llama_file {
 };
 
 static bool is_ggml_file(const char * filename) {
-    my_llama_file file(filename, "rb");
+    llama_file file(filename, "rb");
     if (file.size < 4) {
         return false;
     }
@@ -578,7 +576,7 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
     } else {
         // assume llama2.c vocabulary
         LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
-        my_llama_file file(filename, "rb");
+        llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
         }
@@ -691,8 +689,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
     gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
     gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
-    gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, LLAMA_TOKEN_NULL);
-    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, LLAMA_TOKEN_NULL);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
 
     gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
@@ -911,7 +909,7 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab   = config.vocab_size; //llama_vocab_n_vocab(lctx);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
     model.hparams.n_ctx     = params.n_ctx;
     model.hparams.n_embd    = config.dim; //params.n_embd;
     model.hparams.n_ff      = config.hidden_dim;

examples/cvector-generator/cvector-generator.cpp

@@ -1,9 +1,7 @@
-#include "ggml.h"
-#include "gguf.h"
-
 #include "arg.h"
 #include "common.h"
 #include "llama.h"
+#include "ggml.h"
 #include "pca.hpp"
 #include "mean.hpp"
 
@@ -273,9 +271,7 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const llama_model * model = llama_get_model(ctx);
-        const llama_vocab * vocab = llama_model_get_vocab(model);
-        const bool add_bos = llama_vocab_get_add_bos(vocab);
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
         tokens_pos = common_tokenize(ctx, pos, add_bos, true);
         tokens_neg = common_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
@@ -419,13 +415,12 @@ int main(int argc, char ** argv) {
     // load the model to get hparams
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
-    int n_layers = llama_model_n_layer(model);
-    int n_embd = llama_model_n_embd(model);
-
+    int n_layers = llama_n_layer(model);
+    int n_embd = llama_n_embd(model);
     // get model hint param (a.k.a model arch name)
     char model_hint[128];
     llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
@@ -479,6 +474,8 @@ int main(int argc, char ** argv) {
 
     // done with the model, we can now free it to make gain some memory
     printf("Done evaluate prompts, unload model...\n");
+    llama_free(ctx);
+    llama_free_model(model);
 
     bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
 

examples/cvector-generator/mean.hpp

@@ -15,7 +15,7 @@ static void run(
     for (size_t il = 0; il < v_input.size(); ++il) {
         // prepare output vector
         struct ggml_tensor * ctrl_out = v_output[il];
-        ggml_format_name(ctrl_out, "direction.%zu", il+1);
+        ggml_format_name(ctrl_out, "direction.%ld", il+1);
 
         // calculate mean vector
         struct ggml_tensor * t_layer = v_input[il];

examples/cvector-generator/pca.hpp

@@ -302,7 +302,7 @@ static void run_pca(
 
         // prepare output vector
         struct ggml_tensor * ctrl_out = v_output[il];
-        ggml_format_name(ctrl_out, "direction.%zu", il+1);
+        ggml_format_name(ctrl_out, "direction.%ld", il+1);
 
         // run power_iteration
         params.i_layer = il;

examples/embedding/embedding.cpp

@@ -97,17 +97,14 @@ int main(int argc, char ** argv) {
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
-
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    const int n_ctx_train = llama_model_n_ctx_train(model);
+    const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
@@ -150,7 +147,7 @@ int main(int argc, char ** argv) {
     // check if the last token is SEP
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
+        if (inp.empty() || inp.back() != llama_token_sep(model)) {
             LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
@@ -183,7 +180,7 @@ int main(int argc, char ** argv) {
     }
 
     // allocate output
-    const int n_embd = llama_model_n_embd(model);
+    const int n_embd = llama_n_embd(model);
     std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
@@ -319,6 +316,8 @@ int main(int argc, char ** argv) {
 
     // clean up
     llama_batch_free(batch);
+    llama_free(ctx);
+    llama_free_model(model);
     llama_backend_free();
 
     return 0;

examples/eval-callback/eval-callback.cpp

@@ -127,10 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool run(llama_context * ctx, const common_params & params) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    const bool add_bos = llama_vocab_get_add_bos(vocab);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
@@ -165,9 +162,8 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
-
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);
         return 1;
@@ -188,6 +184,9 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

@@ -1,13 +1,12 @@
-#include "ggml.h"
-#include "ggml-alloc.h"
-#include "gguf.h"
-
 #include "arg.h"
 #include "common.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
 
 #include <map>
 #include <vector>
 #include <string>
+#include <thread>
 #include <fstream>
 
 static bool g_verbose = false;
@@ -129,7 +128,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<common_adapter_lora_info> & lora_files,
+            std::vector<common_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -266,8 +265,8 @@ struct lora_merge_ctx {
             fout.write((const char *)data.data(), data.size());
         }
 
-        printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %zu tensors to output file\n", __func__, trans.size());
+        printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -345,25 +344,15 @@ struct lora_merge_ctx {
             gf = ggml_new_graph(ctx0);
             struct ggml_tensor * cur = inp_base;
             for (size_t i = 0; i < adapters.size(); ++i) {
-                struct ggml_tensor * delta;
-                bool is_tok_embd = string_starts_with(name_base, "token_embd");
-                if (is_tok_embd) {
-                    printf("%s :     detected token embeddings tensor\n", __func__);
-                    delta = ggml_mul_mat(ctx0,
-                        ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32),
-                        ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32));
-                } else {
-                    delta = ggml_mul_mat(ctx0,
-                        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32))),
-                        ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
-                }
+                struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32)));
+                struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
                 // scale
                 const float alpha = adapters[i]->alpha;
                 const float rank  = (float) inp_b[i]->ne[0];
                 const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
                 delta = ggml_scale(ctx0, delta, scale);
                 cur = ggml_add(ctx0, delta, cur);
-                printf("%s :   + merging from adapter[%zu] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
+                printf("%s :   + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
                 printf("%s :     input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
             }
             cur = ggml_cast(ctx0, cur, out->type);

examples/gbnf-validator/gbnf-validator.cpp

@@ -11,15 +11,19 @@
 static bool llama_grammar_validate(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
     const auto cpts = unicode_cpts_from_utf8(input_str);
 
-    auto & stacks_cur = llama_grammar_get_stacks(grammar);
+    const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
+          llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
 
     size_t pos = 0;
     for (const auto & cpt : cpts) {
-        llama_grammar_accept(grammar, cpt);
+        const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
+
+        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
 
         if (stacks_cur.empty()) {
             error_pos = pos;
             error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
+            stacks_cur = stacks_prev;
             return false;
         }
         ++pos;
@@ -78,8 +82,7 @@ int main(int argc, char** argv) {
 
     llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
     if (grammar == nullptr) {
-        fprintf(stdout, "Failed to initialize llama_grammar\n");
-        return 1;
+        throw std::runtime_error("Failed to initialize llama_grammar");
     }
     // Read the input file
     std::string input_str;

examples/gguf-hash/gguf-hash.cpp

@@ -1,5 +1,4 @@
 #include "ggml.h"
-#include "gguf.h"
 
 #include <cstdlib>   /* abort() */
 #include <cstddef>

examples/gguf-split/gguf-split.cpp

@@ -1,19 +1,18 @@
-#include "ggml.h"
-#include "gguf.h"
 #include "llama.h"
 #include "common.h"
 
 #include <algorithm>
-#include <cinttypes>
-#include <climits>
-#include <cstdio>
+#include <cmath>
 #include <cstdlib>
-#include <stdexcept>
-#include <cstring>
 #include <fstream>
 #include <string>
 #include <vector>
 
+#include <stdio.h>
+#include <string.h>
+#include <climits>
+#include <stdexcept>
+
 #if defined(_WIN32)
     #include <windows.h>
     #ifndef PATH_MAX
@@ -298,7 +297,7 @@ struct split_strategy {
                 total_size += ggml_nbytes(t);
             }
             total_size = total_size / 1000 / 1000; // convert to megabytes
-            printf("split %05d: n_tensors = %" PRIi64 ", total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            printf("split %05d: n_tensors = %d, total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
             i_split++;
         }
     }

examples/gguf-split/tests.sh

@@ -41,7 +41,7 @@ echo PASS
 echo
 
 # 2b. Test the sharded model is loading properly
-$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
 echo PASS
 echo
 
@@ -51,7 +51,7 @@ echo PASS
 echo
 
 # 3b. Test the merged model is loading properly
-$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
 echo PASS
 echo
 
@@ -61,7 +61,7 @@ echo PASS
 echo
 
 # 4b. Test the sharded model is loading properly
-$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
 echo PASS
 echo
 
@@ -71,7 +71,7 @@ echo
 #echo
 
 # 5b. Test the merged model is loading properly
-#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
+#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
 #echo PASS
 #echo
 
@@ -81,7 +81,7 @@ echo PASS
 echo
 
 # 6b. Test the sharded model is loading properly
-$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
 echo PASS
 echo
 

examples/gguf/gguf.cpp

@@ -1,9 +1,10 @@
 #include "ggml.h"
-#include "gguf.h"
 
 #include <cstdio>
+#include <cinttypes>
 #include <string>
 #include <sstream>
+#include <fstream>
 #include <vector>
 
 #undef MIN
@@ -134,10 +135,9 @@ static bool gguf_ex_read_0(const std::string & fname) {
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
-            const size_t size   = gguf_get_tensor_size  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -182,10 +182,9 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
-            const size_t size   = gguf_get_tensor_size  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -200,8 +199,7 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            printf("%s: tensor[%d]: n_dims = %d, ne = (%d, %d, %d, %d), name = %s, data = %p\n",
-                __func__, i, ggml_n_dims(cur), int(cur->ne[0]), int(cur->ne[1]), int(cur->ne[2]), int(cur->ne[3]), cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;
@@ -217,7 +215,7 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
                 const float * data = (const float *) cur->data;
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
-                        fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));
+                        fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
                         gguf_free(ctx);
                         return false;
                     }
@@ -247,8 +245,6 @@ int main(int argc, char ** argv) {
         check_data = false;
     }
 
-    srand(123456);
-
     const std::string fname(argv[1]);
     const std::string mode (argv[2]);
 

examples/gritlm/gritlm.cpp

@@ -11,7 +11,6 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     std::vector<std::vector<float>> result;
 
     const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
 
     llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
@@ -20,16 +19,16 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         const std::string input_string = instruction + sentences[i];
 
-        std::vector<llama_token> inputs = common_tokenize(vocab, input_string, true, false);
+        std::vector<llama_token> inputs = common_tokenize(model, input_string, true, false);
 
         const int32_t n_toks = inputs.size();
 
         // GritLM seems to have EOS = ""
         // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_vocab_eos(vocab));
+        // inputs.push_back(llama_token_eos(model));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
+        const int32_t n_inst = common_tokenize(model, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -53,7 +52,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         llama_decode(ctx, batch);
 
         // get embedding dimensions
-        uint64_t n_embd = llama_model_n_embd(model);
+        uint64_t n_embd = llama_n_embd(model);
 
         // allocate embedding output
         std::vector<float> emb_unorm(n_embd, 0.0f);
@@ -76,7 +75,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         std::vector<float> emb_norm(emb_unorm.size());
-        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
+        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
         result.push_back(emb_norm);
 
 #ifdef GRIT_DEBUG
@@ -98,9 +97,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     std::string result;
 
     const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    llama_token eos_token = llama_vocab_eos(vocab);
+    llama_token eos_token = llama_token_eos(model);
 
     llama_kv_cache_clear(ctx);
     llama_set_embeddings(ctx, false);
@@ -108,7 +105,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
-    std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
+    std::vector<llama_token> inputs = common_tokenize(model, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
@@ -168,10 +165,10 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
 
     // create generation context
-    llama_context * ctx = llama_init_from_model(model, cparams);
+    llama_context * ctx = llama_new_context_with_model(model, cparams);
 
     auto sparams = llama_sampler_chain_default_params();
 
@@ -200,7 +197,7 @@ int main(int argc, char * argv[]) {
         const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
         const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
 
-        const int n_embd = llama_model_n_embd(model);
+        const int n_embd = llama_n_embd(model);
 
         const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
         const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
@@ -222,7 +219,7 @@ int main(int argc, char * argv[]) {
 
     llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_model_free(model);
+    llama_free_model(model);
     llama_backend_free();
 
     return 0;

examples/imatrix/imatrix.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <sstream>
 #include <thread>
 #include <mutex>
 #include <vector>
@@ -39,7 +40,7 @@ public:
     void set_params(common_params params) { m_params = std::move(params); }
     bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
     void save_imatrix(int ncall = -1) const;
-    bool load_imatrix(const char * fname);
+    bool load_imatrix(const char * file_name);
 private:
     std::unordered_map<std::string, Stats> m_stats;
     common_params                          m_params;
@@ -428,14 +429,10 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const common_params & params) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    const bool add_bos = llama_vocab_get_add_bos(vocab);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
-    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
-
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
@@ -470,7 +467,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
     const int n_chunk_max = tokens.size() / n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_vocab_n_tokens(vocab);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -510,7 +507,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_vocab_bos(vocab);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             common_batch_clear(batch);
@@ -621,15 +618,14 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
-
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);
         return 1;
     }
 
-    const int n_ctx_train = llama_model_n_ctx_train(model);
+    const int n_ctx_train = llama_n_ctx_train(model);
     if (params.n_ctx > n_ctx_train) {
         LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, params.n_ctx);
@@ -659,6 +655,9 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     llama_backend_free();
 
     return 0;

examples/infill/infill.cpp

@@ -131,17 +131,15 @@ int main(int argc, char ** argv) {
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
     common_init_result llama_init = common_init_from_params(params);
 
-    model = llama_init.model.get();
-    ctx = llama_init.context.get();
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    const int n_ctx_train = llama_model_n_ctx_train(model);
+    const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
     LOG_DBG("n_ctx: %d\n", n_ctx);
 
@@ -154,28 +152,28 @@ int main(int argc, char ** argv) {
         LOG_INF("\n");
         LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_vocab_get_add_bos(vocab);
-    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
 
     std::vector<llama_token> embd_inp;
     std::vector<llama_token> embd_end;
     std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
     std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-    GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
-    GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
+    GGML_ASSERT(llama_token_fim_pre(model) >= 0);
+    GGML_ASSERT(llama_token_fim_suf(model) >= 0);
 
-    inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
-    inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
     embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
     embd_end = params.spm_infill ? inp_pfx : inp_sfx;
     if (add_bos) {
-        embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
+        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
     }
     embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
 
-    const llama_token middle_token = llama_vocab_fim_mid(vocab);
+    const llama_token middle_token = llama_token_fim_mid(model);
     if (middle_token >= 0) {
         embd_inp.push_back(middle_token);
     }
@@ -187,7 +185,7 @@ int main(int argc, char ** argv) {
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_vocab_bos(vocab));
+        embd_inp.push_back(llama_token_bos(model));
         LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
     }
 
@@ -422,10 +420,10 @@ int main(int argc, char ** argv) {
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
             // deal with eot token in infill mode
-            if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
+            if ((common_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
                     // print an eot token
-                    LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
+                    LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 LOG("\n");
                 console::set_display(console::user_input);
@@ -465,13 +463,13 @@ int main(int argc, char ** argv) {
                 std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
                 std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-                inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
-                inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
                 embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
                 embd_end = params.spm_infill ? inp_pfx : inp_sfx;
                 if (add_bos) {
-                    embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
+                    embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
                 }
                 embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
 
@@ -486,7 +484,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
+            else if (llama_token_is_eog(model, common_sampler_last(smpl))) {
                 LOG_DBG("found EOS token\n");
 
                 if (params.interactive) {
@@ -502,7 +500,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG_DBG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_vocab_bos(vocab));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -565,7 +563,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of generation
-        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) {
+        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
             break;
         }
 
@@ -577,12 +575,15 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
+        LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
     }
 
     LOG("\n");
     common_perf_print(ctx, smpl);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     common_sampler_free(smpl);
     llama_backend_free();
 

examples/llama-bench/llama-bench.cpp

@@ -683,7 +683,7 @@ struct cmd_params_instance {
     bool               cpu_strict;
     int                poll;
     int                n_gpu_layers;
-    std::string        rpc_servers_str;
+    std::string        rpc_servers;
     llama_split_mode   split_mode;
     int                main_gpu;
     bool               no_kv_offload;
@@ -696,37 +696,8 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
-        if (!rpc_servers_str.empty()) {
-            auto rpc_servers = string_split<std::string>(rpc_servers_str, ',');
-
-            // add RPC devices
-            if (!rpc_servers.empty()) {
-                ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
-                if (!rpc_reg) {
-                    fprintf(stderr, "%s: failed to find RPC backend\n", __func__);
-                    exit(1);
-                }
-
-                typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
-                ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
-                if (!ggml_backend_rpc_add_device_fn) {
-                    fprintf(stderr, "%s: failed to find RPC device add function\n", __func__);
-                    exit(1);
-                }
-                static std::vector<ggml_backend_dev_t> devices;
-                devices.clear();
-                for (const std::string & server : rpc_servers) {
-                    ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
-                    if (dev) {
-                        devices.push_back(dev);
-                    } else {
-                        fprintf(stderr, "%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
-                        exit(1);
-                    }
-                }
-                devices.push_back(nullptr);
-                mparams.devices = devices.data();
-            }
+        if (!rpc_servers.empty()) {
+            mparams.rpc_servers = rpc_servers.c_str();
         }
         mparams.split_mode   = split_mode;
         mparams.main_gpu     = main_gpu;
@@ -737,7 +708,7 @@ struct cmd_params_instance {
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
-        return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str &&
+        return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers == other.rpc_servers &&
                split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
                tensor_split == other.tensor_split;
     }
@@ -1430,8 +1401,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model   = llama_get_model(ctx);
-    const llama_vocab * vocab   = llama_model_get_vocab(model);
-    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
+    const int32_t       n_vocab = llama_n_vocab(model);
 
     std::vector<llama_token> tokens(n_batch);
 
@@ -1439,7 +1409,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
 
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        tokens[0]    = n_processed == 0 && llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
+        tokens[0]    = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
@@ -1454,10 +1424,9 @@ static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model   = llama_get_model(ctx);
-    const llama_vocab * vocab   = llama_model_get_vocab(model);
-    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
+    const int32_t       n_vocab = llama_n_vocab(model);
 
-    llama_token token = llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
+    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1));
@@ -1557,10 +1526,10 @@ int main(int argc, char ** argv) {
         // keep the same model between tests when possible
         if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
             if (lmodel) {
-                llama_model_free(lmodel);
+                llama_free_model(lmodel);
             }
 
-            lmodel = llama_model_load_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
             if (lmodel == NULL) {
                 fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
                 return 1;
@@ -1568,10 +1537,10 @@ int main(int argc, char ** argv) {
             prev_inst = &inst;
         }
 
-        llama_context * ctx = llama_init_from_model(lmodel, inst.to_llama_cparams());
+        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
-            llama_model_free(lmodel);
+            llama_free_model(lmodel);
             return 1;
         }
 
@@ -1657,7 +1626,7 @@ int main(int argc, char ** argv) {
         ggml_threadpool_free_fn(threadpool);
     }
 
-    llama_model_free(lmodel);
+    llama_free_model(lmodel);
 
     if (p) {
         p->print_footer();

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

@@ -19,7 +19,6 @@ android {
         externalNativeBuild {
             cmake {
                 arguments += "-DLLAMA_BUILD_COMMON=ON"
-                arguments += "-DGGML_LLAMAFILE=OFF"
                 arguments += "-DCMAKE_BUILD_TYPE=Release"
                 cppFlags += listOf()
                 arguments += listOf()

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

@@ -87,7 +87,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
     auto path_to_model = env->GetStringUTFChars(filename, 0);
     LOGi("Loading model from %s", path_to_model);
 
-    auto model = llama_model_load_from_file(path_to_model, model_params);
+    auto model = llama_load_model_from_file(path_to_model, model_params);
     env->ReleaseStringUTFChars(filename, path_to_model);
 
     if (!model) {
@@ -102,7 +102,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
-    llama_model_free(reinterpret_cast<llama_model *>(model));
+    llama_free_model(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
@@ -305,9 +305,7 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
-    //llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
-    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
-    delete batch;
+    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }
 
 extern "C"
@@ -347,7 +345,6 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
         jlong context_pointer,
         jlong batch_pointer,
         jstring jtext,
-        jboolean format_chat,
         jint n_len
     ) {
 
@@ -357,8 +354,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
 
-    bool parse_special = (format_chat == JNI_TRUE);
-    const auto tokens_list = common_tokenize(context, text, true, parse_special);
+    const auto tokens_list = common_tokenize(context, text, 1);
 
     auto n_ctx = llama_n_ctx(context);
     auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@@ -370,7 +366,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     }
 
     for (auto id : tokens_list) {
-        LOGi("token: `%s`-> %d ", common_token_to_piece(context, id).c_str(), id);
+        LOGi("%s", common_token_to_piece(context, id).c_str());
     }
 
     common_batch_clear(*batch);
@@ -407,7 +403,6 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
     const auto batch   = reinterpret_cast<llama_batch   *>(batch_pointer);
     const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
     const auto model = llama_get_model(context);
-    const auto vocab = llama_model_get_vocab(model);
 
     if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
     if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
@@ -417,7 +412,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
     const auto new_token_id = llama_sampler_sample(sampler, context, -1);
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
-    if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
+    if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
         return nullptr;
     }
 

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -65,7 +65,6 @@ class LLamaAndroid {
         context: Long,
         batch: Long,
         text: String,
-        formatChat: Boolean,
         nLen: Int
     ): Int
 
@@ -116,10 +115,10 @@ class LLamaAndroid {
         }
     }
 
-    fun send(message: String, formatChat: Boolean = false): Flow<String> = flow {
+    fun send(message: String): Flow<String> = flow {
         when (val state = threadLocalState.get()) {
             is State.Loaded -> {
-                val ncur = IntVar(completion_init(state.context, state.batch, message, formatChat, nlen))
+                val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
                     val str = completion_loop(state.context, state.batch, state.sampler, nlen, ncur)
                     if (str == null) {

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

@@ -52,8 +52,8 @@ actor LlamaContext {
     deinit {
         llama_sampler_free(sampling)
         llama_batch_free(batch)
-        llama_model_free(model)
         llama_free(context)
+        llama_free_model(model)
         llama_backend_free()
     }
 
@@ -65,7 +65,7 @@ actor LlamaContext {
         model_params.n_gpu_layers = 0
         print("Running on simulator, force use n_gpu_layers = 0")
 #endif
-        let model = llama_model_load_from_file(path, model_params)
+        let model = llama_load_model_from_file(path, model_params)
         guard let model else {
             print("Could not load model at \(path)")
             throw LlamaError.couldNotInitializeContext
@@ -151,7 +151,7 @@ actor LlamaContext {
 
         new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
 
-        if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
+        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             print("\n")
             is_done = true
             let new_token_str = String(cString: temporary_invalid_cchars + [0])

examples/llava/README-minicpmo2.6.md

@@ -1,46 +0,0 @@
-## MiniCPM-o 2.6
-Currently, this readme only supports minicpm-omni's image capabilities, and we will update the full-mode support as soon as possible.
-
-### Prepare models and code
-
-Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch model from huggingface to "MiniCPM-o-2_6" folder.
-
-Clone llama.cpp:
-```bash
-git clone git@github.com:OpenBMB/llama.cpp.git
-cd llama.cpp
-git checkout minicpm-omni
-```
-
-### Usage of MiniCPM-o 2.6
-
-Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
-
-```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
-python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
-
-# quantize int4 version
-./llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
-```
-
-Build llama.cpp using `CMake`:
-https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md
-
-```bash
-cmake -B build
-cmake --build build --config Release
-```
-
-Inference on Linux or Mac
-```
-# run f16 version
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
-
-# run quantized int4 version
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
-
-# or run in interactive mode
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
-```

examples/llava/clip.cpp

@@ -7,27 +7,26 @@
 #include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "gguf.h"
 
-//#ifdef GGML_USE_CUDA
-//#include "ggml-cuda.h"
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//#include "ggml-sycl.h"
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//#include "ggml-metal.h"
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//#include "ggml-cann.h"
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//#include "ggml-vulkan.h"
-//#endif
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
 
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
@@ -263,7 +262,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             {
                 const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
                 int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {
@@ -718,9 +717,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         else if (ctx->minicpmv_version == 3) {
             pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
         }
-        else if (ctx->minicpmv_version == 4) {
-            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
-        }
         ggml_set_name(pos_embed, "pos_embed");
         ggml_set_input(pos_embed);
     }
@@ -900,7 +896,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                 mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
                 mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
                 // stride = 1, padding = 1, bias is nullptr
-                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
 
                 // layer norm
                 // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
@@ -948,7 +944,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             // block_2
             {
                 // stride = 2
-                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
 
                 // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
                 // layer norm
@@ -1009,7 +1005,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             // mlp_2 ne [24, 24, 2048, 1]
             mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
             // weight ne = [3, 3, 2048, 1]
-            struct ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
             peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
             peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
             mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
@@ -1056,11 +1052,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                     n_head = hidden_size/d_head;
                     num_query = 64;
                 }
-                else if (ctx->minicpmv_version == 4) {
-                    hidden_size = 3584;
-                    n_head = hidden_size/d_head;
-                    num_query = 64;
-                }
 
                 struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
                 Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
@@ -1231,30 +1222,30 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-//#ifdef GGML_USE_CUDA
-//    new_clip->backend = ggml_backend_cuda_init(0);
-//    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//    new_clip->backend = ggml_backend_metal_init();
-//    LOG_INF("%s: CLIP using Metal backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//    new_clip->backend = ggml_backend_cann_init(0);
-//    LOG_INF("%s: CLIP using CANN backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//    new_clip->backend = ggml_backend_vk_init(0);
-//    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//    new_clip->backend = ggml_backend_sycl_init(0);
-//    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
-//#endif
+#ifdef GGML_USE_CUDA
+    new_clip->backend = ggml_backend_cuda_init(0);
+    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+    new_clip->backend = ggml_backend_metal_init();
+    LOG_INF("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_CANN
+    new_clip->backend = ggml_backend_cann_init(0);
+    LOG_INF("%s: CLIP using CANN backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_SYCL
+    new_clip->backend = ggml_backend_sycl_init(0);
+    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
+#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
@@ -2049,7 +2040,6 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
                 images[images.size()-1].push_back(patch);
             }
         }
-        clip_image_u8_free(refine_image);
     }
     return images;
 }
@@ -2088,13 +2078,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
                 clip_image_f32_free(res);
             }
         }
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            for (size_t j = 0; j < imgs[i].size(); ++j) {
-                if (imgs[i][j] != nullptr) {
-                    clip_image_u8_free(imgs[i][j]);
-                }
-            }
-        }
         return true;
     }
     else if (ctx->has_qwen2vl_merger) {
@@ -2351,9 +2334,6 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
         else if (ctx->minicpmv_version == 3) {
             n_patches = 64;
         }
-        else if (ctx->minicpmv_version == 4) {
-            n_patches = 64;
-        }
     } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         int patch_size = params.patch_size * 2;
         int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
@@ -2533,8 +2513,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
-            int bucket_coords_h[1024];
-            int bucket_coords_w[1024];
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
             for (int i = 0; i < pos_h; i++){
                 bucket_coords_h[i] = std::floor(70.0*i/pos_h);
             }
@@ -2562,9 +2542,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             else if (ctx->minicpmv_version == 3) {
                 embed_dim = 3584;
             }
-            else if (ctx->minicpmv_version == 4) {
-                embed_dim = 3584;
-            }
             auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
             float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
@@ -2757,8 +2734,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         total_size_org += orig_size;
         total_size_new += new_size;
         gguf_set_tensor_type(ctx_out, name.c_str(), new_type);
-        GGML_ASSERT(gguf_get_tensor_size(ctx_out, gguf_find_tensor(ctx_out, name.c_str())) == new_size);
-        gguf_set_tensor_data(ctx_out, name.c_str(), new_data);
+        gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
         fout.write((const char *)new_data, new_size);
         size_t pad = GGML_PAD(new_size, gguf_get_alignment(ctx_out)) - new_size;
         for (size_t j = 0; j < pad; ++j) {
@@ -2808,9 +2784,6 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         else if (ctx->minicpmv_version == 3) {
             return 3584;
         }
-        else if (ctx->minicpmv_version == 4) {
-            return 3584;
-        }
     }
     if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         return ctx->vision_model.mm_1_b->ne[0];

examples/llava/llava-cli.cpp

@@ -47,12 +47,8 @@ static const char * sample(struct common_sampler * smpl,
                            int * n_past) {
     const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
     common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
         ret = common_token_to_piece(ctx_llama, id);
@@ -225,7 +221,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -243,10 +239,11 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
+
     llama_context_params ctx_params = common_context_params_to_llama(*params);
     ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
 
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
         LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@@ -268,7 +265,7 @@ static void llava_free(struct llava_context * ctx_llava) {
     }
 
     llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
+    llama_free_model(ctx_llava->model);
     llama_backend_free();
 }
 
@@ -326,7 +323,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    llama_model_free(model);
+    llama_free_model(model);
 
     return 0;
 }

examples/llava/llava.cpp

@@ -216,7 +216,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     return true;
 }
 
-static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) {
+static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
     int width = image->nx;
     int height = image->ny;
     int num_patches = (height / patch_size) * (width / patch_size);
@@ -277,7 +277,13 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
                 encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
             }
             else {
-                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+                int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
+                if (has_minicpmv_projector == 2) {
+                    encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+                }
+                else if (has_minicpmv_projector == 3) {
+                    encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+                }
             }
 
             if (!encoded) {
@@ -307,9 +313,6 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         load_image_size->height = img->ny;
         clip_add_load_image_size(ctx_clip, load_image_size);
         LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
@@ -381,7 +384,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
 bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
         // make sure that the correct mmproj was used, i.e., compare apples to apples
-    int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
+    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
         LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
@@ -453,7 +456,7 @@ struct llava_embd_batch {
 };
 
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
 
     for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
         int n_eval = image_embed->n_image_pos - i;

examples/llava/minicpmv-cli.cpp

@@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -54,7 +54,7 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
         ctx_params.n_ctx = params->n_ctx;
     }
 
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
         LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@@ -75,7 +75,7 @@ static void llava_free(struct llava_context * ctx_llava) {
     }
 
     llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
+    llama_free_model(ctx_llava->model);
     llama_backend_free();
 }
 
@@ -140,9 +140,6 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     else if (has_minicpmv_projector == 3) {
         system_prompt = "<|im_start|>user\n";
     }
-    else if (has_minicpmv_projector == 4) {
-        system_prompt = "<|im_start|>user\n";
-    }
     LOG_INF("%s: image token past: %d\n", __func__, n_past);
     eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
     process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
@@ -170,12 +167,8 @@ static const char * sample(struct common_sampler * smpl,
                            int * n_past) {
     const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
     common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
         ret = common_token_to_piece(ctx_llama, id);
@@ -230,9 +223,6 @@ static struct common_sampler * llama_init(struct llava_context * ctx_llava, comm
         else if (has_minicpmv_projector == 3) {
             user_prompt = "<|im_start|>user\n" + prompt;
         }
-        else if (has_minicpmv_projector == 4) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
     }
 
     eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
@@ -242,9 +232,6 @@ static struct common_sampler * llama_init(struct llava_context * ctx_llava, comm
     else if (has_minicpmv_projector == 3) {
         eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
     }
-    else if (has_minicpmv_projector == 4) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
 
     // generate the response
 
@@ -317,6 +304,7 @@ int main(int argc, char ** argv) {
                     const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
                     response += tmp;
                     if (strcmp(tmp, "</s>") == 0) break;
+                    if (strstr(tmp, "###")) break; // Yi-VL behavior
                     printf("%s", tmp);// mistral llava-1.6
                     if (strstr(response.c_str(), "<user>")) break; // minicpm-v
                     fflush(stdout);

examples/llava/minicpmv-convert-image-encoder-to-gguf.py

@@ -501,7 +501,7 @@ default_image_mean = [0.48145466, 0.4578275, 0.40821073]
 default_image_std = [0.26862954, 0.26130258, 0.27577711]
 ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
 ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
-ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4', default=2)
+ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3', default=2)
 
 # with proper
 args = ap.parse_args()
@@ -545,19 +545,12 @@ if args.use_f32:
 
 minicpmv_version = args.minicpmv_version
 emb_dim = 4096
-block_count = 26
 if minicpmv_version == 1:
     emb_dim = 2304
-    block_count = 26
 elif minicpmv_version == 2:
     emb_dim = 4096
-    block_count = 27
 elif minicpmv_version == 3:
     emb_dim = 3584
-    block_count = 27
-elif minicpmv_version == 4:
-    emb_dim = 3584
-    block_count = 27
 
 default_vision_config = {
         "hidden_size": 1152,
@@ -574,9 +567,6 @@ model = Idefics2VisionTransformer(vision_config)
 if minicpmv_version == 3:
     vision_config = SiglipVisionConfig(**default_vision_config)
     model = SiglipVisionTransformer(vision_config)
-elif minicpmv_version == 4:
-    vision_config = SiglipVisionConfig(**default_vision_config)
-    model = SiglipVisionTransformer(vision_config)
 
 processor = None
 # if model.attn_pool is not None:
@@ -597,7 +587,7 @@ elif args.minicpmv_projector is not None:
     fname_middle = "mmproj-"
     has_text_encoder = False
     has_minicpmv_projector = True
-    minicpmv_version = 4
+    minicpmv_version = 3
 elif args.vision_only:
     fname_middle = "vision-"
     has_text_encoder = False
@@ -635,6 +625,7 @@ if has_vision_encoder:
     fout.add_uint32("clip.vision.projection_dim", 0)
     fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
     fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    block_count = 26
     fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
 
     if processor is not None:

examples/llava/minicpmv-surgery.py

@@ -8,7 +8,7 @@ ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
 args = ap.parse_args()
 
 # find the model part that includes the the multimodal projector weights
-model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True, torch_dtype=torch.bfloat16)
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
 checkpoint = model.state_dict()
 
 # get a list of mm tensor names

examples/llava/qwen2_vl_surgery.py

@@ -88,8 +88,6 @@ def main(args):
     else:
         raise ValueError()
 
-    local_model = False
-    model_path = ""
     model_name = args.model_name
     print("model_name: ", model_name)
     qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
@@ -99,10 +97,8 @@ def main(args):
     vcfg = cfg.vision_config
 
     if os.path.isdir(model_name):
-        local_model = True
         if model_name.endswith(os.sep):
             model_name = model_name[:-1]
-        model_path = model_name
         model_name = os.path.basename(model_name)
     fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
 
@@ -143,10 +139,7 @@ def main(args):
             it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
     """
 
-    if local_model:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path)
-    else:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
+    processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
     fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
     fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
 

examples/llava/qwen2vl-cli.cpp

@@ -27,7 +27,7 @@
 
 static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
                                      int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
     const int patch_size = 14 * 2;
     const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0);
     const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0);
@@ -132,12 +132,8 @@ static const char * sample(struct common_sampler * smpl,
                            int * n_past, int * st_pos_id) {
     const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
     common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
         ret = common_token_to_piece(ctx_llama, id);
@@ -314,7 +310,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -332,10 +328,11 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
+
     llama_context_params ctx_params = common_context_params_to_llama(*params);
     ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
 
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
         LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@@ -357,7 +354,7 @@ static void llava_free(struct llava_context * ctx_llava) {
     }
 
     llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
+    llama_free_model(ctx_llava->model);
     llama_backend_free();
 }
 
@@ -484,7 +481,7 @@ static void debug_test_mrope_2d() {
 }
 
 static void debug_dump_img_embed(struct llava_context * ctx_llava) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llava->ctx_llama));
     int ne = n_embd * 4;
     float vals[56 * 56 * 3];
     // float embd[ne];
@@ -578,7 +575,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    llama_model_free(model);
+    llama_free_model(model);
 
     return 0;
 }

examples/lookahead/lookahead.cpp

@@ -58,10 +58,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -149,7 +147,7 @@ int main(int argc, char ** argv) {
     }
 
     // here we keep adding new n-grams as we go
-    ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);
+    ngram_container ngrams_observed(llama_n_vocab(model), N, G);
 
     // debug
     struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
@@ -299,7 +297,7 @@ int main(int argc, char ** argv) {
                 }
                 fflush(stdout);
 
-                if (llama_vocab_is_eog(vocab, id)) {
+                if (llama_token_is_eog(model, id)) {
                     has_eos = true;
                 }
 
@@ -476,6 +474,9 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     llama_backend_free();
 
     LOG("\n\n");

examples/lookup/lookup-create.cpp

@@ -1,9 +1,14 @@
 #include "arg.h"
 #include "common.h"
 #include "ngram-cache.h"
+#include "ggml.h"
 #include "llama.h"
 
+#include <cstdint>
+#include <fstream>
+#include <iostream>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 int main(int argc, char ** argv){
@@ -20,16 +25,16 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model_ptr & model = llama_init.model;
-    llama_context_ptr & ctx = llama_init.context;
-
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx.get(), params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
     fprintf(stderr, "%s: tokenization done\n", __func__);
 
+
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
     fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());

examples/lookup/lookup-stats.cpp

@@ -30,11 +30,12 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_context_ptr & ctx = llama_init.context;
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx.get(), params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
 
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
@@ -65,7 +66,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = llama_n_ctx(ctx.get());
+    const int n_ctx = llama_n_ctx(ctx);
 
     int n_drafted = 0;
     int n_accept  = 0;
@@ -149,6 +150,9 @@ int main(int argc, char ** argv){
     LOG_INF("n_accept     = %d\n", n_accept);
     LOG_INF("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     llama_backend_free();
 
     LOG("\n\n");

examples/lookup/lookup.cpp

@@ -33,10 +33,8 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
@@ -138,7 +136,7 @@ int main(int argc, char ** argv){
                 LOG("%s", token_str.c_str());
             }
 
-            if (llama_vocab_is_eog(vocab, id)) {
+            if (llama_token_is_eog(model, id)) {
                 has_eos = true;
             }
 
@@ -245,6 +243,9 @@ int main(int argc, char ** argv){
 
     llama_batch_free(batch_tgt);
 
+    llama_free(ctx);
+    llama_free_model(model);
+
     llama_backend_free();
 
     LOG("\n\n");

examples/main-cmake-pkg/CMakeLists.txt

@@ -0,0 +1,32 @@
+cmake_minimum_required(VERSION 3.12)
+project("llama-cli-cmake-pkg" C CXX)
+set(TARGET llama-cli-cmake-pkg)
+
+find_package(Llama 0.0.1 REQUIRED)
+
+# Bake common functionality in with target. Because applications
+# using the relocatable Llama package should be outside of the
+# source tree, llama-cli-cmake-pkg pretends the dependencies are built-in.
+set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
+add_library(common OBJECT)
+file(GLOB _common_files
+    "${_common_path}/*.h"
+    "${_common_path}/*.cpp"
+)
+target_sources(common PRIVATE ${_common_files})
+
+# If the common project was part of "llama-cli-cmake-pkg" the transient
+# defines would automatically be attached. Because the common func-
+# tionality is separate, but dependent upon the defines, it must be
+# explicitly extracted from the "llama" target.
+#
+get_target_property(_llama_transient_defines llama
+    INTERFACE_COMPILE_DEFINITIONS)
+
+target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
+target_include_directories(${TARGET} PRIVATE ${_common_path})
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/main-cmake-pkg/README.md

@@ -0,0 +1,31 @@
+# llama.cpp/example/main-cmake-pkg
+
+This program builds [llama-cli](../main) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+
+## Building
+
+Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
+
+### Considerations
+
+When hardware acceleration libraries are used (e.g. CUDA, Metal, etc.), CMake must be able to locate the associated CMake package.
+
+### Build llama.cpp and install to C:\LlamaCPP directory
+
+```cmd
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+cmake -B build -DBUILD_SHARED_LIBS=OFF -G "Visual Studio 17 2022" -A x64
+cmake --build build --config Release
+cmake --install build --prefix C:/LlamaCPP
+```
+
+### Build llama-cli-cmake-pkg
+
+
+```cmd
+cd ..\examples\main-cmake-pkg
+cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake --build build --config Release
+cmake --install build --prefix C:/MyLlamaApp
+```

examples/main/README.md

@@ -177,11 +177,16 @@ Example usage: `--temp 0`
 
 -   `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
 -   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+-   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
 
 The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.
 
 The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
 
+Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
+
+Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
+
 ### DRY Repetition Penalty
 
 DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
@@ -310,9 +315,9 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Batch Size
 
-- `-ub N`, `--ubatch-size N`: Physical batch size. This is the maximum number of tokens that may be processed at a time. Increasing this value may improve performance during prompt processing, at the expense of higher memory usage. Default: `512`.
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
 
-- `-b N`, `--batch-size N`: Logical batch size. Increasing this value above the value of the physical batch size may improve prompt processing performance when using multiple GPUs with pipeline parallelism. Default: `2048`.
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
 
 ### Prompt Caching