ci : change python3 -> python

ggml-ci

.devops/cpu.Dockerfile

@@ -0,0 +1,81 @@
+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 $(nproc)
+
+RUN mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+RUN mkdir -p /app/full \
+    && cp build/bin/* /app/full \
+    && cp *.py /app/full \
+    && cp -r gguf-py /app/full \
+    && cp -r requirements /app/full \
+    && cp requirements.txt /app/full \
+    && cp .devops/tools.sh /app/full/tools.sh
+
+## Base image
+FROM ubuntu:$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

@@ -0,0 +1,94 @@
+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

@@ -1,33 +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}
-
-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

@@ -1,33 +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}
-
-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

@@ -1,50 +0,0 @@
-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

@@ -1,38 +0,0 @@
-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

@@ -0,0 +1,91 @@
+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

@@ -1,38 +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}
-# 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

@@ -1,28 +0,0 @@
-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

@@ -1,38 +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}
-# 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

@@ -1,45 +0,0 @@
-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

@@ -1,27 +0,0 @@
-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

@@ -1,29 +0,0 @@
-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

@@ -1,43 +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}
-# 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

@@ -1,34 +0,0 @@
-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

@@ -1,43 +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}
-# 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

@@ -1,54 +0,0 @@
-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

@@ -1,31 +0,0 @@
-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

@@ -1,33 +0,0 @@
-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

@@ -0,0 +1,108 @@
+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/nix/package.nix

@@ -31,6 +31,7 @@
   # Increases the runtime closure size by ~700M
   useMpi ? false,
   useRocm ? config.rocmSupport,
+  rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets,
   enableCurl ? true,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
@@ -188,7 +189,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ]
     ++ optionals useRocm [
       (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
-      (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
+      (cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmGpuTargets)
     ]
     ++ optionals useMetalKit [
       (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")

.devops/rocm.Dockerfile

@@ -0,0 +1,113 @@
+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

@@ -8,11 +8,11 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert_hf_to_gguf.py "$@"
+    exec python3 ./convert_hf_to_gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./llama-quantize "$@"
+    exec ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./llama-cli "$@"
+    exec ./llama-cli "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -20,11 +20,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
             echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
         else
             echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..."
-            ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
+            exec ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./llama-server "$@"
+    exec ./llama-server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "

.devops/vulkan.Dockerfile

@@ -0,0 +1,88 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK and cURL
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+
+# Build it
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+    cmake --build build --config Release -j$(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" ]

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

@@ -65,12 +65,22 @@ 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 the command that you entered and any generated text.
+          Please copy and paste any relevant log output, including 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

@@ -52,6 +52,16 @@ 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:
@@ -74,7 +84,7 @@ body:
     attributes:
       label: Relevant log output
       description: >
-          If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
+          If applicable, please copy and paste any relevant log output, including any generated text.
           This will be automatically formatted into code, so no need for backticks.
       render: shell
     validations:

.github/workflows/build.yml

@@ -60,8 +60,7 @@ jobs:
             -DLLAMA_CURL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DGGML_RPC=ON \
-            -DBUILD_SHARED_LIBS=OFF
+            -DGGML_RPC=ON
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -123,8 +122,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_CURL=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
-            -DBUILD_SHARED_LIBS=OFF
+            -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -181,7 +179,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -317,7 +315,7 @@ jobs:
           wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
           sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential vulkan-sdk
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk
 
       - name: Build
         id: cmake_build
@@ -327,6 +325,12 @@ jobs:
           cmake -DGGML_VULKAN=ON ..
           cmake --build . --config Release -j $(nproc)
 
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.0.2
@@ -552,35 +556,44 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-# TODO: tmp disabled. see for possible re-enable:
-#       https://github.com/ggerganov/llama.cpp/pull/10525
-#  macOS-latest-swift:
-#    runs-on: macos-latest
-#
-#    strategy:
-#      matrix:
-#        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        id: depends
-#        continue-on-error: true
-#        run: |
-#          brew update
-#
-#      - name: xcodebuild for swift package
-#        id: xcodebuild
-#        run: |
-#          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
-#
-#      - name: Build Swift Example
-#        id: make_build_swift_example
-#        run: |
-#            make swift
+  macOS-latest-swift:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build llama.cpp with CMake
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+          sudo cmake --install . --config Release
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}"
 
   windows-msys2:
     runs-on: windows-latest
@@ -636,23 +649,25 @@ 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 -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
           - build: 'openblas-x64'
-            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"'
+            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"'
           - build: 'kompute-x64'
-            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'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=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 -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=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 -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=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'
 
     steps:
       - name: Clone
@@ -694,6 +709,28 @@ jobs:
         run: |
           choco install ninja
 
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.build == 'llvm-arm64-opencl-adreno' }}
+        run: |
+          git clone https://github.com/KhronosGroup/OpenCL-Headers
+          cd OpenCL-Headers
+          mkdir build && cd build
+          cmake .. `
+            -DBUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build . --target install
+          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+          cd OpenCL-ICD-Loader
+          mkdir build-arm64-release && cd build-arm64-release
+          cmake .. `
+            -A arm64 `
+            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build . --target install --config release
+
       - name: Build
         id: cmake_build
         run: |
@@ -723,7 +760,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -875,7 +912,7 @@ 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" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=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
@@ -1104,6 +1141,29 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+          sudo cmake --install . --config Release
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme llama-Package -destination 'generic/platform=iOS'
+
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
@@ -1131,23 +1191,6 @@ jobs:
 
           ./gradlew build --no-daemon
 
-#  freeBSD-latest:
-#    runs-on: macos-12
-#    steps:
-#    - name: Clone
-#      uses: actions/checkout@v4
-#
-#    - name: Build
-#      uses: cross-platform-actions/action@v0.19.0
-#      with:
-#        operating_system: freebsd
-#        version: '13.2'
-#        hypervisor: 'qemu'
-#        run: |
-#            sudo pkg update
-#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas
-#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
-
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -1194,7 +1237,7 @@ jobs:
 
       - name: Create release
         id: create_release
-        uses: anzz1/action-create-release@v1
+        uses: ggml-org/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:

.github/workflows/docker.yml

@@ -34,21 +34,14 @@ jobs:
     strategy:
       matrix:
         config:
-          - { 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" }
+          # 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}
           # 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: "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" }
+          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
@@ -56,10 +49,10 @@ jobs:
           fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
 
       - name: Log in to Docker Hub
         uses: docker/login-action@v2
@@ -79,26 +72,34 @@ 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}-${SHORT_HASH}"
           fi
-
           # list all tags possible
-          TAGS=""
-          TAGS="${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
+          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
         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)
-        uses: jlumbroso/free-disk-space@main
+        if: ${{ matrix.config.free_disk_space == true }}
+        uses: ggml-org/free-disk-space@v1.3.1
         with:
           # this might remove tools that are actually needed,
           # if set to "true" but frees about 6 GB
@@ -113,13 +114,59 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Build and push Docker image (tagged + versioned)
-        if: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
+      - 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 }}
         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.output_tags }}
+          tags: ${{ steps.tag.outputs.full_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

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

.github/workflows/server.yml

@@ -79,7 +79,7 @@ jobs:
       # Setup nodejs (to be used for verifying bundled index.html)
       - uses: actions/setup-node@v4
         with:
-          node-version: 22
+          node-version: '22.11.0'
 
       - name: Verify bundled index.html
         id: verify_server_index_html

CMakeLists.txt

@@ -46,11 +46,9 @@ if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 endif()
 
-if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
-    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/source-charset:utf-8>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/source-charset:utf-8>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/execution-charset:utf-8>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/execution-charset:utf-8>")
+if (MSVC)
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/utf-8>")
 endif()
 
 #

CMakePresets.json

@@ -31,6 +31,13 @@
     { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
     { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "ON" } },
 
+    {
+        "name": "x64-windows-llvm", "hidden": true,
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/x64-windows-llvm.cmake"
+        }
+    },
+
     {
         "name": "arm64-windows-msvc", "hidden": true,
         "architecture": { "value": "arm64",    "strategy": "external" },
@@ -70,6 +77,11 @@
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
+    { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] },
+    { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] },
+    { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] },
+    { "name": "x64-windows-llvm+static-release", "inherits": [ "base", "x64-windows-llvm", "reldbg", "static" ] },
+
     { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
     { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },

CODEOWNERS

@@ -1,3 +1,11 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
 
-ci/ @ggerganov
+/ci/ @ggerganov
+/.devops/*.Dockerfile @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
 
@@ -20,14 +20,104 @@
 - 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
-- 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
+- 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`
-- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
+- 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)
 - 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

@@ -22,6 +22,7 @@ BUILD_TARGETS = \
 	llama-infill \
 	llama-llava-cli \
 	llama-minicpmv-cli\
+	llama-qwen2vl-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -445,6 +446,10 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	MK_CFLAGS     += -march=native -mtune=native
 	HOST_CXXFLAGS += -march=native -mtune=native
 
+	# Usage AMX build test
+	#MK_CFLAGS     += -march=graniterapids -mtune=graniterapids
+	#HOST_CXXFLAGS += -march=graniterapids -mtune=graniterapids
+
 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
 	#MK_CXXFLAGS += -mfma -mf16c -mavx
@@ -948,7 +953,6 @@ DIR_COMMON = common
 
 OBJ_GGML = \
 	$(DIR_GGML)/src/ggml.o \
-	$(DIR_GGML)/src/ggml-aarch64.o \
 	$(DIR_GGML)/src/ggml-alloc.o \
 	$(DIR_GGML)/src/ggml-backend.o \
 	$(DIR_GGML)/src/ggml-backend-reg.o \
@@ -956,9 +960,11 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-quants.o \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \
 	$(OBJ_GGML_EXT)
 
 OBJ_LLAMA = \
@@ -1098,17 +1104,10 @@ DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
 # Default target
 all: $(BUILD_TARGETS)
 
+# force c++ build for source file that have same name as c file
 # Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
-#       g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
-$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
-	ggml/src/ggml-cpu/ggml-cpu.cpp \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/include/ggml-cpu.h \
-	ggml/src/ggml-impl.h
-	$(CXX) $(CXXFLAGS)   -c $< -o $@
+$(DIR_GGML)/%_cpp.o: $(DIR_GGML)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
 
 # Rules for building object files
 $(DIR_GGML)/%.o: $(DIR_GGML)/%.c
@@ -1406,6 +1405,14 @@ llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
+llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)

Package.swift

@@ -2,59 +2,6 @@
 
 import PackageDescription
 
-var sources = [
-    "src/llama.cpp",
-    "src/llama-vocab.cpp",
-    "src/llama-grammar.cpp",
-    "src/llama-sampling.cpp",
-    "src/unicode.cpp",
-    "src/unicode-data.cpp",
-    "ggml/src/ggml.c",
-    "ggml/src/ggml-aarch64.c",
-    "ggml/src/ggml-alloc.c",
-    "ggml/src/ggml-backend.cpp",
-    "ggml/src/ggml-backend-reg.cpp",
-    "ggml/src/ggml-cpu/ggml-cpu.c",
-    "ggml/src/ggml-cpu/ggml-cpu.cpp",
-    "ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
-    "ggml/src/ggml-cpu/ggml-cpu-quants.c",
-    "ggml/src/ggml-threading.cpp",
-    "ggml/src/ggml-quants.c",
-]
-
-var resources: [Resource] = []
-var linkerSettings: [LinkerSetting] = []
-var cSettings: [CSetting] =  [
-    .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-    .unsafeFlags(["-fno-objc-arc"]),
-    .headerSearchPath("ggml/src"),
-    .headerSearchPath("ggml/src/ggml-cpu"),
-    // NOTE: NEW_LAPACK will required iOS version 16.4+
-    // We should consider add this in the future when we drop support for iOS 14
-    // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-    // .define("ACCELERATE_NEW_LAPACK"),
-    // .define("ACCELERATE_LAPACK_ILP64")
-    .define("GGML_USE_CPU"),
-]
-
-
-#if canImport(Darwin)
-sources.append("ggml/src/ggml-common.h")
-sources.append("ggml/src/ggml-metal/ggml-metal.m")
-resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
-linkerSettings.append(.linkedFramework("Accelerate"))
-cSettings.append(
-    contentsOf: [
-        .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL"),
-    ]
-)
-#endif
-
-#if os(Linux)
-    cSettings.append(.define("_GNU_SOURCE"))
-#endif
-
 let package = Package(
     name: "llama",
     platforms: [
@@ -67,26 +14,6 @@ let package = Package(
         .library(name: "llama", targets: ["llama"]),
     ],
     targets: [
-        .target(
-            name: "llama",
-            path: ".",
-            exclude: [
-               "build",
-               "cmake",
-               "examples",
-               "scripts",
-               "models",
-               "tests",
-               "CMakeLists.txt",
-               "Makefile",
-               "ggml/src/ggml-metal-embed.metal"
-            ],
-            sources: sources,
-            resources: resources,
-            publicHeadersPath: "spm-headers",
-            cSettings: cSettings,
-            linkerSettings: linkerSettings
-        )
-    ],
-    cxxLanguageStandard: .cxx17
+        .systemLibrary(name: "llama", pkgConfig: "llama"),
+    ]
 )

README.md

@@ -69,6 +69,7 @@ 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)
@@ -98,6 +99,8 @@ 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
 
@@ -110,6 +113,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
+- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
 
 </details>
 
@@ -199,6 +203,7 @@ 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
 
 </details>
 
@@ -219,7 +224,7 @@ 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 |
-| [hipBLAS](docs/build.md#hipblas) | AMD GPU |
+| [HIP](docs/build.md#hip) | AMD GPU |
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 
@@ -240,6 +245,8 @@ 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.
@@ -258,21 +265,12 @@ 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 simple text completion</summary>
-
-    ```bash
-    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>
 
+    Models with a built-in chat template will automatically activate conversation mode. If this doesn't occur, you can manually enable it by adding `-cnv` and specifying a suitable chat template with `--chat-template NAME`
+
     ```bash
-    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv
+    llama-cli -m model.gguf
 
     # > hi, who are you?
     # Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
@@ -284,17 +282,28 @@ To learn more about model quantization, [read this documentation](examples/quant
     </details>
 
 - <details>
-    <summary>Run with custom chat template</summary>
+    <summary>Run in conversation mode with custom chat template</summary>
 
     ```bash
-    # use the "chatml" template
-    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
+    # use the "chatml" template (use -h to see the list of supported templates)
+    llama-cli -m model.gguf -cnv --chat-template chatml
 
     # use a custom template
-    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
+    llama-cli -m model.gguf -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
     ```
 
-    [Supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+    </details>
+
+- <details>
+    <summary>Run simple text completion</summary>
+
+    To disable conversation mode explicitly, use `-no-cnv`
+
+    ```bash
+    llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128 -no-cnv
+
+    # I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
+    ```
 
     </details>
 
@@ -412,7 +421,7 @@ To learn more about model quantization, [read this documentation](examples/quant
 [^1]: [examples/perplexity/README.md](examples/perplexity/README.md)
 [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
 
-## [`llama-bench`](example/bench)
+## [`llama-bench`](examples/llama-bench)
 
 #### Benchmark the performance of the inference for various parameters.
 
@@ -433,6 +442,20 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
+## [`llama-run`](examples/run)
+
+#### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].
+
+- <details>
+    <summary>Run a model with a specific prompt (by default it's pulled from Ollama registry)</summary>
+
+    ```bash
+    llama-run granite-code
+    ```
+
+    </details>
+
+[^3]: [RamaLama](https://github.com/containers/ramalama)
 
 ## [`llama-simple`](examples/simple)
 

Sources/llama/llama.h

@@ -0,0 +1,4 @@
+#pragma once
+
+#include <llama.h>
+

Sources/llama/module.modulemap

@@ -0,0 +1,5 @@
+module llama [system] {
+    header "llama.h"
+    link "llama"
+    export *
+}

ci/run.sh

@@ -299,7 +299,7 @@ function gg_run_open_llama_7b_v2 {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -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 --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-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -433,7 +433,7 @@ function gg_run_pythia_1_4b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -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 --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-cli -no-cnv --model ${model_f16}  -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -564,7 +564,7 @@ function gg_run_pythia_2_8b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -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 --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-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -699,7 +699,7 @@ function gg_run_embd_bge_small {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -747,7 +747,7 @@ function gg_run_rerank_tiny {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
@@ -814,8 +814,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    # Create a fresh python3 venv and enter it
-    if ! python3 -m venv "$MNT/venv"; then
+    # Create a fresh python venv and enter it
+    if ! python -m venv "$MNT/venv"; then
         echo "Error: Failed to create Python virtual environment at $MNT/venv."
         exit 1
     fi

cmake/llama.pc.in

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

cmake/x64-windows-llvm.cmake

@@ -0,0 +1,11 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR x86_64 )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( arch_c_flags "-march=native" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" )
+

common/CMakeLists.txt

@@ -81,7 +81,7 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info)
 # Use curl to download model url
 if (LLAMA_CURL)
     find_package(CURL REQUIRED)
-    add_definitions(-DLLAMA_USE_CURL)
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
     find_library(CURL_LIBRARY curl REQUIRED)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})

common/arg.cpp

@@ -22,6 +22,11 @@ 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;
@@ -37,6 +42,10 @@ 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);
@@ -119,32 +128,74 @@ std::string common_arg::to_string() {
 // utils
 //
 
-static void common_params_handle_model_default(common_params & params) {
-    if (!params.hf_repo.empty()) {
+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) {
+    if (!hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
-        if (params.hf_file.empty()) {
-            if (params.model.empty()) {
-                throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
+        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;
             }
-            params.hf_file = params.model;
-        } else if (params.model.empty()) {
+        }
+        // make sure model path is present (for caching purposes)
+        if (model.empty()) {
             // this is to avoid different repo having same file name, or same file name in different subdirs
-            std::string filename = params.hf_repo + "_" + params.hf_file;
+            std::string filename = hf_repo + "_" + hf_file;
             // to make sure we don't have any slashes in the filename
             string_replace_all(filename, "/", "_");
-            params.model = fs_get_cache_file(filename);
+            model = fs_get_cache_file(filename);
         }
-    } else if (!params.model_url.empty()) {
-        if (params.model.empty()) {
-            auto f = string_split<std::string>(params.model_url, '#').front();
+    } else if (!model_url.empty()) {
+        if (model.empty()) {
+            auto f = string_split<std::string>(model_url, '#').front();
             f = string_split<std::string>(f, '?').front();
-            params.model = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            model = fs_get_cache_file(string_split<std::string>(f, '/').back());
         }
-    } else if (params.model.empty()) {
-        params.model = DEFAULT_MODEL_PATH;
+    } else if (model.empty()) {
+        model = DEFAULT_MODEL_PATH;
     }
 }
 
+const std::vector<ggml_type> kv_cache_types = {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_BF16,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_IQ4_NL,
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+};
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    for (const auto & type : kv_cache_types) {
+        if (ggml_type_name(type) == s) {
+            return type;
+        }
+    }
+    throw std::runtime_error("Unsupported cache type: " + s);
+}
+
+static std::string get_all_kv_cache_types() {
+    std::ostringstream msg;
+    for (const auto & type : kv_cache_types) {
+        msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 //
 // CLI argument parsing functions
 //
@@ -247,7 +298,9 @@ 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");
     }
 
-    common_params_handle_model_default(params);
+    // 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);
+    common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file, params.hf_token);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -385,7 +438,7 @@ 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)) {
+        if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
             ctx_arg.options.push_back(std::move(arg));
         }
     };
@@ -591,7 +644,7 @@ 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}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).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),
@@ -614,7 +667,7 @@ 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"),
@@ -638,7 +691,7 @@ 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)",
@@ -665,7 +718,7 @@ 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"),
@@ -724,15 +777,19 @@ 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"},
-        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"
-        ),
+        "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)",
         [](common_params & params) {
-            params.conversation = true;
+            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;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
@@ -786,7 +843,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -813,7 +870,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--sampling-seq"}, "SEQUENCE",
+        {"--sampling-seq", "--sampler-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);
@@ -826,13 +883,6 @@ 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),
@@ -887,6 +937,9 @@ 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);
         }
@@ -941,6 +994,9 @@ 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());
@@ -1174,18 +1230,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
-        string_format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()),
+        string_format(
+            "KV cache data type for K\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.cache_type_k)
+        ),
         [](common_params & params, const std::string & value) {
-            // TODO: get the type right here
-            params.cache_type_k = value;
+            params.cache_type_k = kv_cache_type_from_str(value);
         }
     ).set_env("LLAMA_ARG_CACHE_TYPE_K"));
     add_opt(common_arg(
         {"-ctv", "--cache-type-v"}, "TYPE",
-        string_format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()),
+        string_format(
+            "KV cache data type for V\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.cache_type_v)
+        ),
         [](common_params & params, const std::string & value) {
-            // TODO: get the type right here
-            params.cache_type_v = value;
+            params.cache_type_v = kv_cache_type_from_str(value);
         }
     ).set_env("LLAMA_ARG_CACHE_TYPE_V"));
     add_opt(common_arg(
@@ -1468,7 +1534,7 @@ 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 });
+            params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
         }
         // 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}));
@@ -1476,7 +1542,7 @@ 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) });
+            params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
         }
         // 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}));
@@ -1530,19 +1596,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
-        {"-hfr", "--hf-repo"}, "REPO",
-        "Hugging Face model repository (default: unused)",
+        {"-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)",
         [](common_params & params, const std::string & value) {
             params.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
-        "Hugging Face model file (default: unused)",
+        "Hugging Face model file. If specified, it will override the quant in --hf-repo (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)",
@@ -1711,6 +1793,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.public_path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
+    add_opt(common_arg(
+        {"--no-webui"},
+        string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.webui = false;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_WEBUI"));
     add_opt(common_arg(
         {"--embedding", "--embeddings"},
         string_format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"),
@@ -2076,35 +2165,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.speculative.n_max = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MAX"));
     add_opt(common_arg(
         {"--draft-min", "--draft-n-min"}, "N",
         string_format("minimum number of draft tokens to use for speculative decoding (default: %d)", params.speculative.n_min),
         [](common_params & params, int value) {
             params.speculative.n_min = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MIN"));
     add_opt(common_arg(
         {"--draft-p-split"}, "P",
         string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
         [](common_params & params, const std::string & value) {
             params.speculative.p_split = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT"));
     add_opt(common_arg(
         {"--draft-p-min"}, "P",
         string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
         [](common_params & params, const std::string & value) {
             params.speculative.p_min = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_P_MIN"));
     add_opt(common_arg(
         {"-cd", "--ctx-size-draft"}, "N",
         string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx),
         [](common_params & params, int value) {
             params.speculative.n_ctx = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CTX_SIZE_DRAFT"));
     add_opt(common_arg(
         {"-devd", "--device-draft"}, "<dev1,dev2,..>",
         "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
@@ -2124,14 +2213,34 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 fprintf(stderr, "warning: consult docs/build.md for compilation instructions\n");
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_GPU_LAYERS_DRAFT"));
     add_opt(common_arg(
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
             params.speculative.model = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).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}));
+
+    // 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

@@ -12,6 +12,7 @@
 
 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
@@ -53,9 +54,11 @@ 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/common.cpp

@@ -2,6 +2,9 @@
 #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:
@@ -18,6 +21,7 @@
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
+#include <filesystem>
 #include <fstream>
 #include <iostream>
 #include <iterator>
@@ -62,11 +66,29 @@
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
-#define PATH_MAX MAX_PATH
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
 #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;
@@ -843,7 +865,7 @@ 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_load_model_from_file(params.model.c_str(), mparams);
+        model = llama_model_load_from_file(params.model.c_str(), mparams);
     }
 
     if (model == NULL) {
@@ -851,26 +873,28 @@ 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_token_bos(model) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
+        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__);
             ok = false;
         }
 
-        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__);
+        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__);
             ok = false;
         }
 
-        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__);
+        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__);
             ok = false;
         }
 
         if (!ok) {
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
@@ -878,40 +902,40 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     auto cparams = common_context_params_to_llama(params);
 
-    llama_context * lctx = llama_new_context_with_model(model, cparams);
+    llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
-        llama_free_model(model);
+        llama_model_free(model);
         return iparams;
     }
 
     if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        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;
+        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+        params.ctx_shift = false;
     }
 
     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_n_layer(model);
+        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_model_n_layer(model);
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
 
-        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);
+        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);
         if (err) {
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
@@ -919,33 +943,54 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
-        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) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
             return iparams;
         }
-        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
-    }
-    if (!params.lora_init_without_apply) {
-        common_lora_adapters_apply(lctx, iparams.lora_adapters);
+
+        la.ptr = lora.get();
+        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
-    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__);
+    if (!params.lora_init_without_apply) {
+        common_set_adapter_lora(lctx, params.lora_adapters);
+    }
+
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
     }
 
+    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_token_bos(model);
-        llama_token eos = llama_token_eos(model);
+        llama_token bos = llama_vocab_bos(vocab);
+        llama_token eos = llama_vocab_eos(vocab);
+
         // some models (e.g. T5) don't have a BOS token
         if (bos != LLAMA_TOKEN_NULL) {
             tmp.push_back(bos);
@@ -960,7 +1005,7 @@ 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 == -1) {
+            if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
                 decoder_start_token_id = bos;
             }
             tmp.clear();
@@ -974,17 +1019,17 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_perf_context_reset(lctx);
     }
 
-    iparams.model   = model;
-    iparams.context = lctx;
+    iparams.model.reset(model);
+    iparams.context.reset(lctx);
 
     return iparams;
 }
 
-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) {
+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) {
         if (la.scale != 0.0f) {
-            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+            llama_set_adapter_lora(ctx, la.ptr, la.scale);
         }
     }
 }
@@ -1015,38 +1060,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     return mparams;
 }
 
-static ggml_type kv_cache_type_from_str(const std::string & s) {
-    if (s == "f32") {
-        return GGML_TYPE_F32;
-    }
-    if (s == "f16") {
-        return GGML_TYPE_F16;
-    }
-    if (s == "bf16") {
-        return GGML_TYPE_BF16;
-    }
-    if (s == "q8_0") {
-        return GGML_TYPE_Q8_0;
-    }
-    if (s == "q4_0") {
-        return GGML_TYPE_Q4_0;
-    }
-    if (s == "q4_1") {
-        return GGML_TYPE_Q4_1;
-    }
-    if (s == "iq4_nl") {
-        return GGML_TYPE_IQ4_NL;
-    }
-    if (s == "q5_0") {
-        return GGML_TYPE_Q5_0;
-    }
-    if (s == "q5_1") {
-        return GGML_TYPE_Q5_1;
-    }
-
-    throw std::runtime_error("Unsupported cache type: " + s);
-}
-
 struct llama_context_params common_context_params_to_llama(const common_params & params) {
     auto cparams = llama_context_default_params();
 
@@ -1081,8 +1094,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
         cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
     }
 
-    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
-    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
+    cparams.type_k = params.cache_type_k;
+    cparams.type_v = params.cache_type_v;
 
     return cparams;
 }
@@ -1108,13 +1121,7 @@ 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 starts_with(const std::string & str, const std::string & prefix) {
-    // While we wait for C++20's std::string::starts_with...
-    return str.rfind(prefix, 0) == 0;
-}
-
-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) {
@@ -1138,9 +1145,9 @@ static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_
 }
 
 static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
-
     // Initialize libcurl
-    std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
     if (!curl) {
         LOG_ERR("%s: error initializing libcurl\n", __func__);
         return false;
@@ -1154,11 +1161,9 @@ 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 ";
-      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);
+        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);
     }
 
 #if defined(_WIN32)
@@ -1168,8 +1173,7 @@ static bool common_download_file(const std::string & url, const std::string & pa
 #endif
 
     // Check if the file already exists locally
-    struct stat model_file_info;
-    auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
+    auto file_exists = std::filesystem::exists(path);
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
@@ -1211,11 +1215,13 @@ 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);
@@ -1427,7 +1433,7 @@ struct llama_model * common_load_model_from_url(
         }
     }
 
-    return llama_load_model_from_file(local_path.c_str(), params);
+    return llama_model_load_from_file(local_path.c_str(), params);
 }
 
 struct llama_model * common_load_model_from_hf(
@@ -1453,6 +1459,80 @@ 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(
@@ -1474,6 +1554,11 @@ 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
 
 //
@@ -1572,21 +1657,23 @@ std::vector<llama_token> common_tokenize(
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
-    return common_tokenize(llama_get_model(ctx), text, add_special, 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);
 }
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_model * model,
+    const struct llama_vocab * vocab,
            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(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -1595,12 +1682,18 @@ 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(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+    const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
     if (n_chars < 0) {
         piece.resize(-n_chars);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+        int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
         GGML_ASSERT(check == -n_chars);
     }
     else {
@@ -1610,13 +1703,19 @@ std::string common_token_to_piece(const struct llama_context * ctx, llama_token
     return piece;
 }
 
-std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
+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 text;
     text.resize(std::max(text.capacity(), tokens.size()));
-    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    int32_t n_chars = llama_detokenize(vocab, 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(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        n_chars = llama_detokenize(vocab, 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
     }
 
@@ -1630,9 +1729,14 @@ std::string common_detokenize(llama_context * ctx, const std::vector<llama_token
 // Chat template utils
 //
 
+std::string common_get_builtin_chat_template(const struct llama_model * model) {
+    const char * ptr_tmpl = llama_model_chat_template(model);
+    return ptr_tmpl == nullptr ? "" : ptr_tmpl;
+}
+
 bool common_chat_verify_template(const std::string & tmpl) {
     llama_chat_message chat[] = {{"user", "test"}};
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
+    const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);
     return res >= 0;
 }
 
@@ -1643,16 +1747,16 @@ std::string common_chat_apply_template(const struct llama_model * model,
     int alloc_size = 0;
     bool fallback = false; // indicate if we must fallback to default chatml
     std::vector<llama_chat_message> chat;
-    for (auto & msg : msgs) {
+    for (const 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();
+    const char * ptr_tmpl = tmpl.empty() ? llama_model_chat_template(model) : 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(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+    int32_t res = llama_chat_apply_template(ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
 
     // error: chat template is not supported
     if (res < 0) {
@@ -1660,18 +1764,17 @@ std::string common_chat_apply_template(const struct llama_model * model,
             // 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 the built-in template is not supported, we default to chatml
+        res = llama_chat_apply_template("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(
-            fallback ? nullptr : model,
             fallback ? "chatml" : ptr_tmpl,
             chat.data(), chat.size(), add_ass, buf.data(), buf.size());
     }
@@ -1799,7 +1902,9 @@ 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.h"
+#include "llama-cpp.h"
 
 #include <string>
 #include <vector>
@@ -24,22 +24,20 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct common_lora_adapter_info {
+struct common_adapter_lora_info {
     std::string path;
     float scale;
-};
 
-struct common_lora_adapter_container : common_lora_adapter_info {
-    struct llama_lora_adapter * adapter;
+    struct llama_adapter_lora * ptr;
 };
 
 using llama_tokens = std::vector<llama_token>;
 
 // build info
 extern int LLAMA_BUILD_NUMBER;
-extern char const * LLAMA_COMMIT;
-extern char const * LLAMA_COMPILER;
-extern char const * LLAMA_BUILD_TARGET;
+extern const char * LLAMA_COMMIT;
+extern const char * LLAMA_COMPILER;
+extern const char * LLAMA_BUILD_TARGET;
 
 struct common_control_vector_load_info;
 
@@ -80,6 +78,7 @@ enum llama_example {
     LLAMA_EXAMPLE_LLAVA,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
+    LLAMA_EXAMPLE_TTS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -95,6 +94,7 @@ 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
@@ -103,6 +103,12 @@ 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
@@ -130,7 +136,6 @@ 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;
@@ -139,6 +144,7 @@ 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,
@@ -158,6 +164,7 @@ 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
@@ -171,6 +178,14 @@ struct common_params_speculative {
     std::string model = ""; // draft model for speculative decoding                          // 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
+};
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -193,11 +208,13 @@ 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;
@@ -211,11 +228,12 @@ 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          = "unknown"; // model alias                                            // NOLINT
+    std::string model_alias          = ""; // model alias                                                   // NOLINT
     std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
     std::string hf_repo              = ""; // HF repo                                                       // NOLINT
@@ -234,8 +252,8 @@ struct common_params {
     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_lora_adapter_apply)
-    std::vector<common_lora_adapter_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_adapter_lora_apply)
+    std::vector<common_adapter_lora_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
 
@@ -263,7 +281,6 @@ 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
 
@@ -286,8 +303,10 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
 
-    std::string cache_type_k = "f16"; // KV cache data type for the K
-    std::string cache_type_v = "f16"; // KV cache data type for the V
+    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
@@ -437,6 +456,16 @@ std::vector<std::string> string_split<std::string>(const std::string & input, ch
     return parts;
 }
 
+static bool string_starts_with(const std::string & str,
+                               const std::string & prefix) {  // While we wait for C++20's std::string::starts_with...
+    return str.rfind(prefix, 0) == 0;
+}
+
+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);
 
@@ -459,10 +488,12 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
+// note: defines object's lifetime
 struct common_init_result {
-    struct llama_model   * model   = nullptr;
-    struct llama_context * context = nullptr;
-    std::vector<common_lora_adapter_container> lora_adapters;
+    llama_model_ptr   model;
+    llama_context_ptr context;
+
+    std::vector<llama_adapter_lora_ptr> lora;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -482,9 +513,12 @@ struct llama_model * common_load_model_from_hf(
     const std::string & local_path,
     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_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
+void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
 //
 // Batch utils
@@ -522,7 +556,7 @@ std::vector<llama_token> common_tokenize(
                         bool   parse_special = false);
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_model * model,
+    const struct llama_vocab * vocab,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
@@ -534,11 +568,21 @@ 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(
-                         llama_context * ctx,
+            const struct llama_context * ctx,
+        const std::vector<llama_token> & tokens,
+                                  bool   special = true);
+
+std::string common_detokenize(
+              const struct llama_vocab * vocab,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
@@ -552,6 +596,9 @@ struct common_chat_msg {
     std::string content;
 };
 
+// Get the built-in chat template for the model. Return empty string if not present.
+std::string common_get_builtin_chat_template(const struct llama_model * model);
+
 // 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);
 
@@ -588,7 +635,8 @@ void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_si
 // Embedding utils
 //
 
-void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
+// TODO: repace embd_norm with an enum
+void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
 
 float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 
@@ -617,6 +665,10 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 // Split utils
 //
 
-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";
+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";
+
+}

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 -1;
+        return LLAMA_TOKEN_NULL;
     }
     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 = -1;
+    llama_token max_token = LLAMA_TOKEN_NULL;
 
     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 -1;
+        return LLAMA_TOKEN_NULL;
     }
     if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
-        return -1;
+        return LLAMA_TOKEN_NULL;
     }
     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 = -1;
+    llama_token drafted_token = LLAMA_TOKEN_NULL;
 
-    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
+    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == LLAMA_TOKEN_NULL; --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 = -1;
+        llama_token max_token = LLAMA_TOKEN_NULL;
 
         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 = -1;
+        llama_token drafted_token = LLAMA_TOKEN_NULL;
 
         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 == -1) {
+        if (drafted_token == LLAMA_TOKEN_NULL) {
             drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
         }
-        if (drafted_token == -1) {
+        if (drafted_token == LLAMA_TOKEN_NULL) {
             drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
         }
-        if (drafted_token == -1) {
+        if (drafted_token == LLAMA_TOKEN_NULL) {
             drafted_token = try_draft(nc_static, ngram_static);
         }
 
-        if (drafted_token == -1) {
+        if (drafted_token == LLAMA_TOKEN_NULL) {
             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] = -1;
+            tokens[i] = LLAMA_TOKEN_NULL;
         }
     }
 
     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] : -1;
+            tokens[i] = i < ngram_size ? input[i] : LLAMA_TOKEN_NULL;
         }
     }
 

common/sampling.cpp

@@ -113,7 +113,10 @@ struct common_sampler {
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
-        const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+        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);
 
         cur.resize(n_vocab);
 
@@ -142,13 +145,15 @@ 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(model, params.grammar.c_str(), "root"),
+        /* .grmr   = */ llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"),
         /* .chain  = */ llama_sampler_chain_init(lparams),
         /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur    = */ {},
@@ -157,36 +162,24 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     llama_sampler_chain_add(result->chain,
             llama_sampler_init_logit_bias(
-                llama_n_vocab(model),
+                llama_vocab_n_tokens(vocab),
                 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      (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      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                     }
-                        break;
+                    break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
                     break;
@@ -206,7 +199,10 @@ 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   (model));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                    break;
+                case COMMON_SAMPLER_TYPE_PENALTIES:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                     break;
                 default:
                     GGML_ASSERT(false && "unknown sampler type");
@@ -215,7 +211,7 @@ 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_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), 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));
@@ -415,6 +411,7 @@ 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 '?';
     }
 }
@@ -429,6 +426,7 @@ 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 "";
     }
 }
@@ -443,6 +441,7 @@ 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
@@ -489,6 +488,7 @@ 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

@@ -62,6 +62,10 @@ struct common_speculative * common_speculative_init(
 }
 
 void common_speculative_free(struct common_speculative * spec) {
+    if (spec == nullptr) {
+        return;
+    }
+
     common_sampler_free(spec->smpl);
 
     llama_batch_free(spec->batch);
@@ -75,10 +79,13 @@ 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 bool vocab_type_tgt = llama_vocab_type(model_tgt);
+    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);
     LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
 
-    const bool vocab_type_dft = llama_vocab_type(model_dft);
+    const bool vocab_type_dft = llama_vocab_type(vocab_dft);
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
@@ -87,34 +94,34 @@ bool common_speculative_are_compatible(
         return false;
     }
 
-    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));
+    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));
         return false;
     }
 
     {
-        const int n_vocab_tgt = llama_n_vocab(model_tgt);
-        const int n_vocab_dft = llama_n_vocab(model_dft);
+        const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
+        const int n_vocab_dft = llama_vocab_n_tokens(vocab_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_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+                    __func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_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_token_get_text(model_tgt, i);
-            const char * token_text_dft = llama_token_get_text(model_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);
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
-                LOG_ERR("%s: draft model vocab must match target model to use speculation but "
+                LOG_ERR("%s: draft vocab vocab must match target vocab 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}")
 
-        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_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}")
 
         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}")
 
-        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 := 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}")
 
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
@@ -296,7 +296,9 @@ class Model:
                     break
 
             for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
-                data = data_torch.squeeze().numpy()
+                # TODO: why do we squeeze here?
+                # data = data_torch.squeeze().numpy()
+                data = data_torch.numpy()
 
                 # if data ends up empty, it means data_torch was a scalar tensor -> restore
                 if len(data.shape) == 0:
@@ -324,6 +326,9 @@ 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")
@@ -473,6 +478,11 @@ 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:
@@ -525,9 +535,19 @@ 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:
@@ -571,6 +591,9 @@ 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"
@@ -658,6 +681,21 @@ class Model:
         if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
             # ref: https://huggingface.co/facebook/chameleon-7b
             res = "chameleon"
+        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
+            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
+            res = "minerva-7b"
+        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 res is None:
             logger.warning("\n")
@@ -680,6 +718,9 @@ 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")
@@ -1663,6 +1704,178 @@ 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
@@ -1986,6 +2199,75 @@ class Qwen2Model(Model):
         except FileNotFoundError:
             self._set_vocab_gpt2()
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+
+
+@Model.register("Qwen2VLForConditionalGeneration")
+class Qwen2VLModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        mrope_section = self.hparams["rope_scaling"]["mrope_section"]
+        mrope_section += [0] * max(0, 4 - len(mrope_section))
+        self.gguf_writer.add_rope_dimension_sections(mrope_section)
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for name, data in super().get_tensors():
+            if name.startswith("visual."):
+                continue
+            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):
@@ -2115,6 +2397,15 @@ 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'
@@ -2231,7 +2522,11 @@ 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)
-        self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
+        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)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
@@ -2273,6 +2568,63 @@ 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
@@ -2530,7 +2882,7 @@ class InternLM2Model(Model):
             return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("BertModel", "CamembertModel")
+@Model.register("BertModel", "BertForMaskedLM", "CamembertModel")
 class BertModel(Model):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -2571,7 +2923,8 @@ class BertModel(Model):
 
         # we need this to validate the size of the token_type embeddings
         # though currently we are passing all zeros to the token_type embeddings
-        self.gguf_writer.add_token_type_count(2)  # "Sequence A" or "Sequence B"
+        # "Sequence A" or "Sequence B"
+        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
 
         # convert to phantom space vocab
         def phantom(tok):
@@ -2595,13 +2948,73 @@ 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
@@ -2909,6 +3322,8 @@ 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)
 
@@ -2921,14 +3336,87 @@ class Rwkv6Model(Model):
         if new_name.endswith("time_mix_w2.weight"):
             data_torch = data_torch.permute(0, 2, 1)
 
-        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))
+        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
 
         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
@@ -3023,6 +3511,24 @@ 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):
@@ -3389,7 +3895,99 @@ 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
 
@@ -3411,6 +4009,15 @@ 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"]:
@@ -3423,6 +4030,16 @@ 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"]
@@ -4317,6 +4934,7 @@ 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",
@@ -4354,8 +4972,15 @@ 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"
+    )
 
-    return parser.parse_args()
+    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
 
 
 def split_str_to_n_bytes(split_str: str) -> int:
@@ -4379,6 +5004,11 @@ 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

@@ -72,6 +72,7 @@ models = [
     {"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", },
@@ -102,6 +103,11 @@ models = [
     {"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"},
 ]
 
 

convert_lora_to_gguf.py

@@ -226,6 +226,9 @@ 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
 
 
@@ -260,6 +263,10 @@ 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)",
@@ -290,6 +297,7 @@ 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"
 
@@ -313,7 +321,10 @@ if __name__ == '__main__':
         lparams: dict[str, Any] = json.load(f)
 
     # load base model
-    if dir_base_model is None:
+    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 "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}")
@@ -371,11 +382,16 @@ if __name__ == '__main__':
                     if self.lazy:
                         tensor = LazyTorchTensor.from_eager(tensor)
                     base_name = get_base_tensor_name(name)
-                    is_lora_a = ".lora_A.weight" in name
-                    is_lora_b = ".lora_B.weight" in 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
                     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")
@@ -407,9 +423,21 @@ 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/build.md

@@ -55,7 +55,14 @@ cmake --build build --config Release
     cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
     cmake --build build-arm64-windows-llvm-release
     ```
-    Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
+    Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels.
+
+    For building with ninja generator and clang compiler as default:
+      -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
+      ```bash
+      cmake --preset x64-windows-llvm-release
+      cmake --build build-x64-windows-llvm-release
+      ```
 
 ## BLAS Build
 
@@ -120,6 +127,8 @@ 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

docs/cuda-fedora.md

@@ -0,0 +1,317 @@
+# 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.GROK
+    model_arch = gguf.MODEL_ARCH.MYMODEL
 ```
 
 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` suffix, that is the convention and several tools like `quantize` expect this to proceed the weights.
+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.
 
 ### 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 implementation like `build_llama`, `build_dbrx` or `build_bert`.
+Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
 
-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.
+Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
 
 Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
 

examples/CMakeLists.txt

@@ -20,7 +20,12 @@ else()
     add_subdirectory(batched)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
-    add_subdirectory(gbnf-validator)
+
+    if (NOT WIN32)
+        # disabled on Windows because it uses internal functions not exported with LLAMA_API
+        add_subdirectory(gbnf-validator)
+    endif()
+
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -46,12 +51,17 @@ 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
         add_subdirectory(convert-llama2c-to-ggml)
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
-        add_subdirectory(quantize-stats)
+        if (NOT WIN32)
+            # disabled on Windows because it uses internal functions not exported with LLAMA_API
+            add_subdirectory(quantize-stats)
+        endif()
         add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)

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_load_model_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_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_new_context_with_model(model, ctx_params);
+    llama_context * ctx = llama_init_from_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_free_model(model);
+    llama_model_free(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_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
+guard let model = llama_model_load_from_file(modelPath.cString(using: .utf8), model_params) else {
     print("Failed to load model")
     exit(1)
 }
 defer {
-    llama_free_model(model)
+    llama_model_free(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_token_is_eog(model, new_token_id) || n_cur == n_len {
+        if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {

examples/batched/batched.cpp

@@ -41,17 +41,19 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_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(model, params.prompt, true);
+    tokens_list = common_tokenize(vocab, params.prompt, true);
 
     const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
@@ -62,9 +64,10 @@ 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_new_context_with_model(model, ctx_params);
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
 
     auto sparams = llama_sampler_chain_default_params();
+    sparams.no_perf = false;
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
@@ -119,8 +122,8 @@ int main(int argc, char ** argv) {
         }
 
         llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-        if (decoder_start_token_id == -1) {
-            decoder_start_token_id = llama_token_bos(model);
+        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            decoder_start_token_id = llama_vocab_bos(vocab);
         }
 
         common_batch_clear(batch);
@@ -173,7 +176,7 @@ 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_token_is_eog(model, new_token_id) || n_cur == n_predict) {
+            if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_predict) {
                 i_batch[i] = -1;
                 LOG("\n");
                 if (n_parallel > 1) {
@@ -235,7 +238,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_free_model(model);
+    llama_model_free(model);
 
     llama_backend_free();
 

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

@@ -1,4 +1,6 @@
 #include "ggml.h"
+#include "gguf.h"
+
 #include "llama.h"
 #include "common.h"
 #include "log.h"
@@ -434,12 +436,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-struct llama_file {
+struct my_llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
     size_t size;
 
-    llama_file(const char * fname, const char * mode) {
+    my_llama_file(const char * fname, const char * mode) {
         fp = std::fopen(fname, mode);
         if (fp == NULL) {
             size = 0;
@@ -500,7 +502,7 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    ~llama_file() {
+    ~my_llama_file() {
         if (fp) {
             std::fclose(fp);
         }
@@ -508,7 +510,7 @@ struct llama_file {
 };
 
 static bool is_ggml_file(const char * filename) {
-    llama_file file(filename, "rb");
+    my_llama_file file(filename, "rb");
     if (file.size < 4) {
         return false;
     }
@@ -576,7 +578,7 @@ 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);
-        llama_file file(filename, "rb");
+        my_llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
         }
@@ -689,8 +691,8 @@ 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, -1);
-    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
+    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_CONTEXT_LENGTH, model->hparams.n_ctx);
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
@@ -909,7 +911,7 @@ 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_n_vocab(lctx);
+    model.hparams.n_vocab   = config.vocab_size; //llama_vocab_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

@@ -1,7 +1,9 @@
+#include "ggml.h"
+#include "gguf.h"
+
 #include "arg.h"
 #include "common.h"
 #include "llama.h"
-#include "ggml.h"
 #include "pca.hpp"
 #include "mean.hpp"
 
@@ -271,7 +273,9 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+        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);
         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());
@@ -415,12 +419,13 @@ 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;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
 
     // int n_ctx = llama_n_ctx(ctx);
-    int n_layers = llama_n_layer(model);
-    int n_embd = llama_n_embd(model);
+    int n_layers = llama_model_n_layer(model);
+    int n_embd = llama_model_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);
@@ -474,8 +479,6 @@ 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.%ld", il+1);
+        ggml_format_name(ctrl_out, "direction.%zu", 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.%ld", il+1);
+        ggml_format_name(ctrl_out, "direction.%zu", il+1);
 
         // run power_iteration
         params.i_layer = il;

examples/embedding/embedding.cpp

@@ -97,14 +97,17 @@ 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;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
+
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(model);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    const int n_ctx_train = llama_model_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
@@ -147,7 +150,7 @@ 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_token_sep(model)) {
+        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
             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__);
         }
@@ -180,7 +183,7 @@ int main(int argc, char ** argv) {
     }
 
     // allocate output
-    const int n_embd = llama_n_embd(model);
+    const int n_embd = llama_model_n_embd(model);
     std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
@@ -316,8 +319,6 @@ 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

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

@@ -1,12 +1,13 @@
-#include "arg.h"
-#include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "gguf.h"
+
+#include "arg.h"
+#include "common.h"
 
 #include <map>
 #include <vector>
 #include <string>
-#include <thread>
 #include <fstream>
 
 static bool g_verbose = false;
@@ -128,7 +129,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<common_lora_adapter_info> & lora_files,
+            std::vector<common_adapter_lora_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
@@ -265,8 +266,8 @@ struct lora_merge_ctx {
             fout.write((const char *)data.data(), data.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());
+        printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged);
+        printf("%s : wrote %zu tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -352,7 +353,7 @@ struct lora_merge_ctx {
                 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[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
+                printf("%s :   + merging from adapter[%zu] 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

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

examples/gguf-hash/gguf-hash.cpp

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

examples/gguf-split/gguf-split.cpp

@@ -1,18 +1,19 @@
+#include "ggml.h"
+#include "gguf.h"
 #include "llama.h"
 #include "common.h"
 
 #include <algorithm>
-#include <cmath>
+#include <cinttypes>
+#include <climits>
+#include <cstdio>
 #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
@@ -287,7 +288,7 @@ struct split_strategy {
     }
 
     void print_info() {
-        printf("n_split: %ld\n", ctx_outs.size());
+        printf("n_split: %zu\n", ctx_outs.size());
         int i_split = 0;
         for (auto & ctx_out : ctx_outs) {
             // re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
@@ -297,7 +298,7 @@ struct split_strategy {
                 total_size += ggml_nbytes(t);
             }
             total_size = total_size / 1000 / 1000; // convert to megabytes
-            printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            printf("split %05d: n_tensors = %" PRIi64 ", total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
             i_split++;
         }
     }

examples/gguf/gguf.cpp

@@ -1,10 +1,9 @@
 #include "ggml.h"
+#include "gguf.h"
 
 #include <cstdio>
-#include <cinttypes>
 #include <string>
 #include <sstream>
-#include <fstream>
 #include <vector>
 
 #undef MIN
@@ -135,9 +134,10 @@ 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, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
         }
     }
 
@@ -182,9 +182,10 @@ 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, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
         }
     }
 
@@ -199,7 +200,8 @@ 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, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
+            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);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;
@@ -215,7 +217,7 @@ 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] = %f\n", __func__, i, j, data[j]);
+                        fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));
                         gguf_free(ctx);
                         return false;
                     }
@@ -245,6 +247,8 @@ 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

@@ -11,6 +11,7 @@ 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);
 
@@ -19,16 +20,16 @@ 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(model, input_string, true, false);
+        std::vector<llama_token> inputs = common_tokenize(vocab, 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_token_eos(model));
+        // inputs.push_back(llama_vocab_eos(vocab));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = common_tokenize(model, instruction, true, false).size();
+        const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -52,7 +53,7 @@ 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_n_embd(model);
+        uint64_t n_embd = llama_model_n_embd(model);
 
         // allocate embedding output
         std::vector<float> emb_unorm(n_embd, 0.0f);
@@ -75,7 +76,7 @@ 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);
+        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
         result.push_back(emb_norm);
 
 #ifdef GRIT_DEBUG
@@ -97,7 +98,9 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     std::string result;
 
     const llama_model * model = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(model);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    llama_token eos_token = llama_vocab_eos(vocab);
 
     llama_kv_cache_clear(ctx);
     llama_set_embeddings(ctx, false);
@@ -105,7 +108,7 @@ 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(model, prompt, false, true);
+    std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
@@ -165,10 +168,10 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
 
     // create generation context
-    llama_context * ctx = llama_new_context_with_model(model, cparams);
+    llama_context * ctx = llama_init_from_model(model, cparams);
 
     auto sparams = llama_sampler_chain_default_params();
 
@@ -197,7 +200,7 @@ 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_n_embd(model);
+        const int n_embd = llama_model_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);
@@ -219,7 +222,7 @@ int main(int argc, char * argv[]) {
 
     llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_free_model(model);
+    llama_model_free(model);
     llama_backend_free();
 
     return 0;

examples/imatrix/imatrix.cpp

@@ -7,7 +7,6 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
-#include <sstream>
 #include <thread>
 #include <mutex>
 #include <vector>
@@ -40,7 +39,7 @@ 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 * file_name);
+    bool load_imatrix(const char * fname);
 private:
     std::unordered_map<std::string, Stats> m_stats;
     common_params                          m_params;
@@ -429,10 +428,14 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const common_params & params) {
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    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 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__);
 
@@ -467,7 +470,7 @@ 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_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -507,7 +510,7 @@ 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_token_bos(llama_get_model(ctx));
+                tokens[batch_start] = llama_vocab_bos(vocab);
             }
 
             common_batch_clear(batch);
@@ -618,14 +621,15 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
+
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx_train = llama_model_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);
@@ -655,9 +659,6 @@ 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

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

@@ -1401,7 +1401,8 @@ 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 int32_t       n_vocab = llama_n_vocab(model);
+    const llama_vocab * vocab   = llama_model_get_vocab(model);
+    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
 
     std::vector<llama_token> tokens(n_batch);
 
@@ -1409,7 +1410,7 @@ 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_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        tokens[0]    = n_processed == 0 && llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
@@ -1424,9 +1425,10 @@ 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 int32_t       n_vocab = llama_n_vocab(model);
+    const llama_vocab * vocab   = llama_model_get_vocab(model);
+    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
 
-    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+    llama_token token = llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1));
@@ -1521,15 +1523,15 @@ int main(int argc, char ** argv) {
     for (const auto & inst : params_instances) {
         params_idx++;
         if (params.progress) {
-            fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count);
+            fprintf(stderr, "llama-bench: benchmark %d/%zu: starting\n", params_idx, params_count);
         }
         // keep the same model between tests when possible
         if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
             if (lmodel) {
-                llama_free_model(lmodel);
+                llama_model_free(lmodel);
             }
 
-            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            lmodel = llama_model_load_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;
@@ -1537,10 +1539,10 @@ int main(int argc, char ** argv) {
             prev_inst = &inst;
         }
 
-        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
+        llama_context * ctx = llama_init_from_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_free_model(lmodel);
+            llama_model_free(lmodel);
             return 1;
         }
 
@@ -1573,14 +1575,14 @@ int main(int argc, char ** argv) {
         // warmup run
         if (t.n_prompt > 0) {
             if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
             }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
             test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
             }
             test_gen(ctx, 1, t.n_threads);
         }
@@ -1592,14 +1594,14 @@ int main(int argc, char ** argv) {
 
             if (t.n_prompt > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: prompt run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
                 test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
                 test_gen(ctx, t.n_gen, t.n_threads);
@@ -1626,7 +1628,7 @@ int main(int argc, char ** argv) {
         ggml_threadpool_free_fn(threadpool);
     }
 
-    llama_free_model(lmodel);
+    llama_model_free(lmodel);
 
     if (p) {
         p->print_footer();

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

@@ -19,6 +19,7 @@ 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_load_model_from_file(path_to_model, model_params);
+    auto model = llama_model_load_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_free_model(reinterpret_cast<llama_model *>(model));
+    llama_model_free(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
@@ -305,7 +305,9 @@ 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));
+    //llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
+    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
+    delete batch;
 }
 
 extern "C"
@@ -403,6 +405,7 @@ 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");
@@ -412,7 +415,7 @@ 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_token_is_eog(model, new_token_id) || n_cur == n_len) {
+    if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
         return nullptr;
     }
 

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_load_model_from_file(path, model_params)
+        let model = llama_model_load_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_token_is_eog(model, new_token_id) || n_cur == n_len {
+        if llama_vocab_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])
@@ -210,20 +210,20 @@ actor LlamaContext {
 
             llama_kv_cache_clear(context)
 
-            let t_pp_start = ggml_time_us()
+            let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
             llama_synchronize(context)
 
-            let t_pp_end = ggml_time_us()
+            let t_pp_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
             // bench text generation
 
             llama_kv_cache_clear(context)
 
-            let t_tg_start = ggml_time_us()
+            let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
             for i in 0..<tg {
                 llama_batch_clear(&batch)
@@ -238,7 +238,7 @@ actor LlamaContext {
                 llama_synchronize(context)
             }
 
-            let t_tg_end = ggml_time_us()
+            let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
             llama_kv_cache_clear(context)
 

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -7,6 +7,7 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
+		1809696D2D05A39F00400EE8 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = 1809696C2D05A39F00400EE8 /* llama */; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
 		79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
@@ -17,7 +18,6 @@
 		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
 		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
 		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
-		DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
 		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
 /* End PBXBuildFile section */
 
@@ -42,7 +42,7 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				DF810E132B4A5BA200301144 /* llama in Frameworks */,
+				1809696D2D05A39F00400EE8 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
 			);
@@ -151,7 +151,7 @@
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
-				DF810E122B4A5BA200301144 /* llama */,
+				1809696C2D05A39F00400EE8 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -429,7 +429,7 @@
 /* End XCConfigurationList section */
 
 /* Begin XCSwiftPackageProductDependency section */
-		DF810E122B4A5BA200301144 /* llama */ = {
+		1809696C2D05A39F00400EE8 /* llama */ = {
 			isa = XCSwiftPackageProductDependency;
 			productName = llama;
 		};

examples/llava/CMakeLists.txt

@@ -43,3 +43,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+set(TARGET llama-qwen2vl-cli)
+add_executable(${TARGET} qwen2vl-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llava/clip.cpp

@@ -7,26 +7,27 @@
 #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"
@@ -102,7 +103,9 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
 #define KEY_N_EMBD              "clip.%s.embedding_length"
 #define KEY_N_FF                "clip.%s.feed_forward_length"
 #define KEY_N_BLOCK             "clip.%s.block_count"
@@ -129,7 +132,8 @@ static std::string format(const char * fmt, ...) {
 #define TN_TOKEN_EMBD      "%s.token_embd.weight"
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
-#define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
+#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
 #define TN_PATCH_BIAS      "v.patch_embd.bias"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
@@ -163,6 +167,7 @@ enum projector_type {
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MERGER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -171,6 +176,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
     { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
 };
 
 
@@ -257,7 +263,7 @@ 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 = gguf_get_arr_data(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {
@@ -463,7 +469,8 @@ struct clip_vision_model {
 
     // embeddings
     struct ggml_tensor * class_embedding;
-    struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_embeddings_0;
+    struct ggml_tensor * patch_embeddings_1;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
     struct ggml_tensor * patch_bias;
     struct ggml_tensor * position_embeddings;
 
@@ -553,6 +560,7 @@ struct clip_ctx {
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
     bool has_minicpmv_projector = false;
+    bool has_qwen2vl_merger = false;
     int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
@@ -561,6 +569,7 @@ struct clip_ctx {
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
+    bool use_silu = false;
     int32_t ftype = 1;
 
     bool has_class_embedding = true;
@@ -606,14 +615,26 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             image_size_height = imgs->data->ny;
         }
     }
+    else if (ctx->has_qwen2vl_merger) {
+        // use the image's native resolution when image is avaible
+        if (is_inf) {
+        // if (imgs->data->nx && imgs->data->ny) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
     const int patch_size           = hparams.patch_size;
     const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int patches_w            = image_size_width / patch_size;
+    const int patches_h            = image_size_height / patch_size;
     const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    const int num_position_ids     = ctx->has_qwen2vl_merger ? num_positions * 4 : num_positions;
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
     int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
 
     const int batch_size = imgs->size;
 
@@ -634,10 +655,30 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
-    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
 
-    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
-    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+    if (ctx->has_qwen2vl_merger) {
+        GGML_ASSERT(image_size_width % (patch_size * 2) == 0);
+        GGML_ASSERT(image_size_height % (patch_size * 2) == 0);
+
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            hidden_size * 2, patches_w / 2, patches_h, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            hidden_size * 2, patches_w / 2, 2, batch_size * (patches_h / 2));
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
+        inp = ggml_reshape_3d(
+            ctx0, inp,
+            hidden_size, patches_w * patches_h, batch_size);
+    }
+    else {
+        inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+    }
 
     if (ctx->has_patch_bias) {
         // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
@@ -659,12 +700,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         }
     }
 
-    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
     ggml_set_name(positions, "positions");
     ggml_set_input(positions);
 
-    embeddings =
-        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
+    if (!ctx->has_qwen2vl_merger) { // qwen2vl use rope position embedding
+        embeddings =
+            ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
+    }
 
     if (ctx->has_minicpmv_projector) {
         int pos_w = image_size_width/patch_size;
@@ -688,7 +731,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // loop over layers
-    if (ctx->has_minicpmv_projector) {
+    if (ctx->has_minicpmv_projector || ctx->has_qwen2vl_merger) {
+        // TODO: figure out why we doing thing in this way ???
         n_layer += 1;
     }
     for (int il = 0; il < n_layer - 1; il++) {
@@ -710,8 +754,13 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             struct ggml_tensor * Q =
                 ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
 
-            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
+            if (ctx->has_qwen2vl_merger) {
+                Q = ggml_rope_multi(
+                    ctx0, Q, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            }
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
 
@@ -719,6 +768,11 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                 ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
 
             K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+            if (ctx->has_qwen2vl_merger) {
+                K = ggml_rope_multi(
+                    ctx0, K, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            }
             K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
             K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
 
@@ -758,6 +812,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
         if (ctx->use_gelu) {
             cur = ggml_gelu_inplace(ctx0, cur);
+        } else if (ctx->use_silu) {
+            cur = ggml_silu_inplace(ctx0, cur);
         } else {
             cur = ggml_gelu_quick_inplace(ctx0, cur);
         }
@@ -769,6 +825,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         cur = ggml_add(ctx0, embeddings, cur);
 
         embeddings = cur;
+
     }
 
     // post-layernorm
@@ -840,7 +897,7 @@ 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_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+                block_1 = ggml_conv_2d_dw(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]
@@ -888,7 +945,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             // block_2
             {
                 // stride = 2
-                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 = ggml_conv_2d_dw(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
@@ -949,7 +1006,7 @@ 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_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 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);
             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));
@@ -1030,6 +1087,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ASSERT(false);
         }
     }
+    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+        embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
+
+        embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+        // GELU activation
+        embeddings = ggml_gelu(ctx0, embeddings);
+
+        // Second linear layer
+        embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+    }
 
     // build the graph
     ggml_build_forward_expand(gf, embeddings);
@@ -1153,30 +1223,30 @@ 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();
@@ -1206,6 +1276,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
         }
 
+        idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER);
+        if (idx != -1) {
+            new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx);
+        }
         // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
 
         GGML_ASSERT(new_clip->has_vision_encoder);
@@ -1214,6 +1288,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         idx = get_key_idx(ctx, KEY_USE_GELU);
         new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
 
+        try {
+            idx = get_key_idx(ctx, KEY_USE_SILU);
+            new_clip->use_silu = gguf_get_val_bool(ctx, idx);
+        } catch (std::runtime_error & /*e*/) {
+            new_clip->use_silu = false;
+        }
+
         if (verbosity >= 1) {
             LOG_INF("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
@@ -1389,11 +1470,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         try {
-            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.patch_embeddings_0    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& /*e*/) {
             LOG_ERR("%s: failed to load vision model tensors\n", __func__);
         }
+        try {
+            vision_model.patch_embeddings_1    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1);
+        } catch(const std::exception& /*e*/) {
+            new_clip->has_qwen2vl_merger = false;
+        }
 
         // LLaVA projection
         if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
@@ -1481,6 +1567,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
             vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) {
+            vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+            vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+            vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+            vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1519,6 +1611,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
+        batch.data = nullptr;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
@@ -1532,6 +1625,10 @@ void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size
     ctx_clip->load_image_size = load_image_size;
 }
 
+struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) {
+    return ctx_clip->load_image_size;
+}
+
 struct clip_image_size * clip_image_size_init() {
     struct clip_image_size * load_image_size = new struct clip_image_size();
     load_image_size->width = 448;
@@ -1984,6 +2081,23 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         }
         return true;
     }
+    else if (ctx->has_qwen2vl_merger) {
+        clip_image_u8 * resized = clip_image_u8_init();
+        auto patch_size = clip_patch_size(ctx) * 2;
+        int nx = ceil((float)img->nx / patch_size) * patch_size;
+        int ny = ceil((float)img->ny / patch_size) * patch_size;
+        bicubic_resize(*img, *resized, nx, ny);
+
+        res_imgs->data = new clip_image_f32[1];
+        // clip_image_f32 * res = clip_image_f32_init();
+        normalize_image_u8_to_f32(resized, res_imgs->data, ctx->image_mean, ctx->image_std);
+        // res_imgs->data[0] = *res;
+        res_imgs->size = 1;
+
+        // clip_image_f32_free(res);
+        clip_image_u8_free(resized);
+        return true;
+    }
 
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
@@ -2173,6 +2287,13 @@ size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
     return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }
 
+size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
+    clip_image_f32 img;
+    img.nx = img_w;
+    img.ny = img_h;
+    return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
+}
+
 int32_t clip_image_size(const struct clip_ctx * ctx) {
     return ctx->vision_model.hparams.image_size;
 }
@@ -2194,6 +2315,13 @@ const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
 }
 
 int clip_n_patches(const struct clip_ctx * ctx) {
+    clip_image_f32 img;
+    img.nx = ctx->vision_model.hparams.image_size;
+    img.ny = ctx->vision_model.hparams.image_size;
+    return clip_n_patches_by_img(ctx, &img);
+}
+
+int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->vision_model.hparams;
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
@@ -2207,6 +2335,11 @@ int clip_n_patches(const struct clip_ctx * ctx) {
         else if (ctx->minicpmv_version == 3) {
             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);
+        int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
+        n_patches = x_patch * y_patch;
     }
 
     return n_patches;
@@ -2335,7 +2468,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int image_size = hparams.image_size;
     int image_size_width  = image_size;
     int image_size_height = image_size;
-    if (ctx->has_minicpmv_projector) {
+    if (ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger) {
         image_size_width  = imgs->data[0].nx;
         image_size_height = imgs->data[0].ny;
     }
@@ -2355,7 +2488,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
             const int ny = imgs->data[i].ny;
-            if (!ctx->has_minicpmv_projector) {
+            if (!(ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger)) {
                 GGML_ASSERT(nx == image_size && ny == image_size);
             }
 
@@ -2413,9 +2546,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             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));
-            for(int i=0;i<pos_w * pos_h;++i){
-                for(int j=0;j<embed_dim;++j){
-                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+            for(int i=0;i < pos_w * pos_h; ++i){
+                for(int j=0; j < embed_dim; ++j){
+                    pos_embed_data[i * embed_dim + j] = pos_embed_t[i][j];
                 }
             }
 
@@ -2435,7 +2568,34 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             }
         }
 
-        {
+        if (ctx->has_qwen2vl_merger) {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+            const int pw = image_size_width / patch_size;
+            const int ph = image_size_height / patch_size;
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+
+            int ptr = 0;
+            for (int y = 0; y < ph; y+=2)
+            {
+                for (int x = 0; x < pw; x+=2)
+                {
+                    for (int dy = 0; dy < 2; dy++) {
+                        for (int dx = 0; dx < 2; dx++) {
+                            positions_data[ptr]                 = y + dy;
+                            positions_data[num_patches + ptr]     = x + dx;
+                            positions_data[num_patches * 2 + ptr] = y + dy;
+                            positions_data[num_patches * 3 + ptr] = x + dx;
+                            ptr++;
+                        }
+                    }
+                }
+            }
+
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+        else {
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
 
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
@@ -2444,16 +2604,16 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             }
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
-        }
 
-        {
-            struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-            int* patches_data = (int*)malloc(ggml_nbytes(patches));
-            for (int i = 0; i < num_patches; i++) {
-                patches_data[i] = i + 1;
+            {
+                struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+                int* patches_data = (int*)malloc(ggml_nbytes(patches));
+                for (int i = 0; i < num_patches; i++) {
+                    patches_data[i] = i + 1;
+                }
+                ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+                free(patches_data);
             }
-            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-            free(patches_data);
         }
     }
 
@@ -2575,7 +2735,8 @@ 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);
-        gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
+        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);
         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) {
@@ -2626,6 +2787,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return 3584;
         }
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+        return ctx->vision_model.mm_1_b->ne[0];
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -2637,3 +2801,21 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
     }
     return 0;
 }
+
+bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
+    return ctx->has_qwen2vl_merger;
+}
+
+
+bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
+    clip_image_f32 clip_img;
+    clip_img.buf.resize(h * w * 3);
+    for (int i = 0; i < h*w*3; i++)
+    {
+        clip_img.buf[i] = img[i];
+    }
+    clip_img.nx = w;
+    clip_img.ny = h;
+    clip_image_encode(ctx, n_threads, &clip_img, vec);
+    return true;
+}

examples/llava/clip.h

@@ -45,6 +45,7 @@ CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
 
 CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
 CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
@@ -55,11 +56,13 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
+CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
 
 CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
 
 CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
@@ -86,6 +89,9 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
 #ifdef __cplusplus
 }

examples/llava/llava-cli.cpp

@@ -47,8 +47,12 @@ 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_token_is_eog(llama_get_model(ctx_llama), id)) {
+    if (llama_vocab_is_eog(vocab, id)) {
         ret = "</s>";
     } else {
         ret = common_token_to_piece(ctx_llama, id);
@@ -221,7 +225,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -239,11 +243,10 @@ 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_new_context_with_model(model, ctx_params);
+    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
         LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@@ -265,7 +268,7 @@ static void llava_free(struct llava_context * ctx_llava) {
     }
 
     llama_free(ctx_llava->ctx_llama);
-    llama_free_model(ctx_llava->model);
+    llama_model_free(ctx_llava->model);
     llama_backend_free();
 }
 
@@ -323,7 +326,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    llama_free_model(model);
+    llama_model_free(model);
 
     return 0;
 }

examples/llava/llava.cpp

@@ -259,25 +259,33 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    if (clip_is_minicpmv(ctx_clip)) {
+    if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
         std::vector<float *> image_embd_v;
         image_embd_v.resize(img_res_v.size);
         struct clip_image_size * load_image_size = clip_image_size_init();
+
         for (size_t i = 0; i < img_res_v.size; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
             int patch_size=14;
             load_image_size->width = img_res_v.data[i].nx;
             load_image_size->height = img_res_v.data[i].ny;
             clip_add_load_image_size(ctx_clip, load_image_size);
+
             bool encoded = false;
-            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) {
+            if (clip_is_qwen2vl(ctx_clip)) {
                 encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
             }
+            else {
+                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) {
                 LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
@@ -290,8 +298,11 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
-            std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
-            n_img_pos_out += clip_n_patches(ctx_clip);
+            std::memcpy(
+                image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
+                image_embd_v[i],
+                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -373,7 +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_n_embd(llama_get_model(ctx_llama));
+    int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
         LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
@@ -387,7 +398,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     if (clip_is_minicpmv(ctx_clip)) {
         num_max_patches = 10;
     }
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
+    float * image_embd;
+    if (clip_is_qwen2vl(ctx_clip)) {
+        // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
+        image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
+    } else {
+        image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
+    }
     if (!image_embd) {
         LOG_ERR("Unable to allocate memory for image embeddings\n");
         return false;
@@ -439,7 +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_n_embd(llama_get_model(ctx_llama));
+    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
 
     for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
         int n_eval = image_embed->n_image_pos - i;

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_load_model_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_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_new_context_with_model(model, ctx_params);
+    llama_context * ctx_llama = llama_init_from_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_free_model(ctx_llava->model);
+    llama_model_free(ctx_llava->model);
     llama_backend_free();
 }
 
@@ -167,8 +167,12 @@ 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_token_is_eog(llama_get_model(ctx_llama), id)) {
+    if (llama_vocab_is_eog(vocab, id)) {
         ret = "</s>";
     } else {
         ret = common_token_to_piece(ctx_llama, id);

examples/llava/qwen2_vl_surgery.py

@@ -0,0 +1,165 @@
+import argparse
+from typing import Dict
+
+import torch
+import numpy as np
+from gguf import *
+from transformers import (
+    Qwen2VLForConditionalGeneration,
+    Qwen2VLProcessor,
+    AutoProcessor,
+    Qwen2VLConfig
+)
+
+
+VISION = "clip.vision"
+
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def to_gguf_name(name: str) -> str:
+    og = name
+    name = name.replace("text_model", "t").replace("vision_model", "v")
+    name = name.replace("blocks", "blk").replace("embeddings.", "")
+    name = name.replace("attn.", "attn_")
+    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
+    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
+    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+    name = name.replace("merger.mlp", 'mm')
+    print(f"[to_gguf_name] {og} --> {name}")
+    return name
+
+
+def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
+    vision_model = qwen2vl.visual
+    tensor_map = {}
+    for name, ten in vision_model.state_dict().items():
+        ten = ten.numpy()
+        if 'qkv' in name:
+            if ten.ndim == 2: # weight
+                c3, _ = ten.shape
+            else:             # bias
+                c3 = ten.shape[0]
+            assert c3 % 3 == 0
+            c = c3 // 3
+            wq = ten[:c]
+            wk = ten[c: c * 2]
+            wv = ten[c * 2:]
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
+        elif 'merger' in name:
+            if name.endswith("ln_q.weight"):
+                tensor_map['v.post_ln.weight'] = ten
+            elif name.endswith("ln_q.bias"):
+                tensor_map['v.post_ln.bias'] = ten
+            else:
+                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
+                tensor_map[to_gguf_name(name)] = ten
+        elif 'patch_embed.proj.weight' in name:
+            # NOTE: split Conv3D into Conv2Ds
+            c1, c2, kt, kh, kw = ten.shape
+            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
+            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
+        else:
+            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
+
+    for new_name, ten in tensor_map.items():
+        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
+            tensor_map[new_name] = ten.astype(np.float32)
+        else:
+            tensor_map[new_name] = ten.astype(dtype)
+    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
+    return tensor_map
+
+
+def main(args):
+    if args.data_type == 'fp32':
+        dtype = torch.float32
+        np_dtype = np.float32
+        ftype = 0
+    elif args.data_type == 'fp16':
+        dtype = torch.float32
+        np_dtype = np.float16
+        ftype = 1
+    else:
+        raise ValueError()
+
+    local_model = False
+    model_path = ""
+    model_name = args.model_name
+    print("model_name: ", model_name)
+    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
+        model_name, torch_dtype=dtype, device_map="cpu"
+    )
+    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+    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"
+
+    fout = GGUFWriter(path=fname_out, arch="clip")
+    fout.add_description("image encoder for Qwen2VL")
+
+    fout.add_file_type(ftype)
+    fout.add_bool("clip.has_text_encoder", False)
+    fout.add_bool("clip.has_vision_encoder", True)
+    fout.add_bool("clip.has_qwen2vl_merger", True)
+    fout.add_string("clip.projector_type", "qwen2vl_merger")
+
+    print(cfg.vision_config)
+    if 'silu' in cfg.vision_config.hidden_act.lower():
+        fout.add_bool("clip.use_silu", True)
+        fout.add_bool("clip.use_gelu", False)
+    elif 'gelu' in cfg.vision_config.hidden_act.lower():
+        fout.add_bool("clip.use_silu", False)
+        fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
+    else:
+        raise ValueError()
+
+    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
+    for name, data in tensor_map.items():
+        fout.add_tensor(name, data)
+
+    fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
+    fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
+    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0)  # not sure what this does, put 0 here as a placeholder
+    fout.add_name(model_name)
+    """
+    HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
+            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)
+    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]
+
+    fout.write_header_to_file()
+    fout.write_kv_data_to_file()
+    fout.write_tensors_to_file()
+    fout.close()
+    print("save model as: ", fname_out)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
+    parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
+    args = parser.parse_args()
+    main(args)

examples/llava/qwen2vl-cli.cpp

@@ -0,0 +1,584 @@
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+#ifdef NDEBUG
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#endif
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
+
+
+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));
+    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);
+    auto img_tokens = image_embed->n_image_pos;
+    // llama_pos mrope_pos[img_tokens * 4];
+    std::vector<llama_pos> mrope_pos;
+    mrope_pos.resize(img_tokens * 4);
+
+    for (int y = 0; y < ph; y++)
+    {
+        for (int x = 0; x < pw; x++)
+        {
+            int i = y * pw + x;
+            mrope_pos[i] = *st_pos_id;
+            mrope_pos[i + img_tokens] = *st_pos_id + y;
+            mrope_pos[i + img_tokens * 2] = *st_pos_id + x;
+            mrope_pos[i + img_tokens * 3] = 0;
+        }
+    }
+    *st_pos_id += std::max(pw, ph);
+
+    int processed = 0;
+    std::vector<llama_pos> batch_mrope_pos;
+    batch_mrope_pos.resize(img_tokens * 4);
+
+    for (int i = 0; i < img_tokens; i += n_batch) {
+        int n_eval = img_tokens - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+
+        // llama_pos batch_mrope_pos[n_eval * 4];
+        std::fill(batch_mrope_pos.begin(), batch_mrope_pos.end(), 0);
+        memcpy(batch_mrope_pos.data(), &mrope_pos[processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 1], &mrope_pos[img_tokens * 1 + processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos));
+
+        llama_batch batch = {
+            int32_t(n_eval),                // n_tokens
+            nullptr,                        // token
+            (image_embed->embed+i*n_embd),  // embed
+            batch_mrope_pos.data(),         // pos
+            nullptr,  // n_seq_id
+            nullptr,  // seq_id
+            nullptr,  // logits
+        };
+
+        if (llama_decode(ctx_llama, batch)) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+        processed += n_eval;
+    }
+    return true;
+}
+
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
+    int N = (int) tokens.size();
+    std::vector<llama_pos> pos;
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        auto batch = llama_batch_get_one(&tokens[i], n_eval);
+        // TODO: add mrope pos ids somewhere else
+        pos.resize(batch.n_tokens * 4);
+        std::fill(pos.begin(), pos.end(), 0);
+        for (int j = 0; j < batch.n_tokens * 3; j ++) {
+            pos[j] = *st_pos_id + (j % batch.n_tokens);
+        }
+        batch.pos = pos.data();
+
+        if (llama_decode(ctx_llama, batch)) {
+            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+        *st_pos_id += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past, int * st_pos_id) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past, st_pos_id);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, int * st_pos_id, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past, st_pos_id);
+    return true;
+}
+
+static const char * sample(struct common_sampler * smpl,
+                           struct llama_context * ctx_llama,
+                           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)) {
+        ret = "</s>";
+    } else {
+        ret = common_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past, st_pos_id);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void print_usage(int, char ** argv) {
+    LOG("\n example usage:\n");
+    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            LOG_INF("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
+        if (!embed) {
+            LOG_ERR("%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
+        if (!embed) {
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
+    int n_past = 0;
+    int cur_pos_id = 0;
+
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+
+    std::string system_prompt, user_prompt;
+    size_t image_pos = prompt.find("<|vision_start|>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<|vision_pad|>").length());
+        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|>";
+        user_prompt = "<|vision_end|>" + prompt + "<|im_end|>\n<|im_start|>assistant\n";
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, true);
+    if (image_embed != nullptr) {
+        auto image_size = clip_get_load_image_size(ctx_llava->ctx_clip);
+        qwen2vl_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past, &cur_pos_id, image_size);
+    }
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, false);
+
+    // generate the response
+
+    LOG("\n");
+
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    if (!smpl) {
+        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
+    std::string response = "";
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past, &cur_pos_id);
+        response += tmp;
+        if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
+        LOG("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
+        fflush(stdout);
+    }
+
+    common_sampler_free(smpl);
+    LOG("\n");
+}
+
+static struct llama_model * llava_init(common_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    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);
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    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);
+
+    if (ctx_llama == NULL) {
+        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_model_free(ctx_llava->model);
+    llama_backend_free();
+}
+
+#ifndef NDEBUG
+
+static void debug_test_mrope_2d() {
+    // 1. Initialize backend
+    ggml_backend_t backend = NULL;
+    std::string backend_name = "";
+#ifdef GGML_USE_CUDA
+    fprintf(stderr, "%s: using CUDA backend\n", __func__);
+    backend = ggml_backend_cuda_init(0); // init device 0
+    backend_name = "cuda";
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+    }
+#endif
+    // if there aren't GPU Backends fallback to CPU backend
+    if (!backend) {
+        backend = ggml_backend_cpu_init();
+        backend_name = "cpu";
+    }
+
+    // Calculate the size needed to allocate
+    size_t ctx_size = 0;
+    ctx_size += 2 * ggml_tensor_overhead(); // tensors
+    // no need to allocate anything else!
+
+    // 2. Allocate `ggml_context` to store tensor data
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_backend_alloc_ctx_tensors()
+    };
+    struct ggml_context * ctx = ggml_init(params);
+
+    struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 128, 12, 30);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 30 * 4);
+    ggml_set_name(pos, "pos");
+    ggml_set_input(pos);
+
+    std::vector<float> dummy_q;
+    dummy_q.resize(128 * 12 * 30);
+    std::fill(dummy_q.begin(), dummy_q.end(), 0.1);
+    // memcpy(inp_raw->data, dummy_q.data(), 128 * 12 * 30 * ggml_element_size(inp_raw));
+
+    std::vector<int> pos_id;
+    pos_id.resize(30 * 4);
+    for (int i = 0; i < 30; i ++) {
+        pos_id[i] = i;
+        pos_id[i + 30] = i + 10;
+        pos_id[i + 60] = i + 20;
+        pos_id[i + 90] = i + 30;
+    }
+    int sections[4] = {32, 32, 0, 0};
+
+    // 4. Allocate a `ggml_backend_buffer` to store all tensors
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+
+    // 5. Copy tensor data from main memory (RAM) to backend buffer
+    ggml_backend_tensor_set(inp_raw, dummy_q.data(), 0, ggml_nbytes(inp_raw));
+    ggml_backend_tensor_set(pos, pos_id.data(), 0, ggml_nbytes(pos));
+
+    // 6. Create a `ggml_cgraph` for mul_mat operation
+    struct ggml_cgraph * gf = NULL;
+    struct ggml_context * ctx_cgraph = NULL;
+
+    // create a temporally context to build the graph
+    struct ggml_init_params params0 = {
+        /*.mem_size   =*/ ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
+    };
+    ctx_cgraph = ggml_init(params0);
+    gf = ggml_new_graph(ctx_cgraph);
+
+    struct ggml_tensor * result0 = ggml_rope_multi(
+        ctx_cgraph, inp_raw, pos, nullptr,
+        128/2, sections, LLAMA_ROPE_TYPE_VISION, 32768, 1000000, 1,
+        0, 1, 32, 1);
+
+    // Add "result" tensor and all of its dependencies to the cgraph
+    ggml_build_forward_expand(gf, result0);
+
+    // 7. Create a `ggml_gallocr` for cgraph computation
+    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+    ggml_gallocr_alloc_graph(allocr, gf);
+
+    // 9. Run the computation
+    int n_threads = 1; // Optional: number of threads to perform some operations with multi-threading
+    if (ggml_backend_is_cpu(backend)) {
+        ggml_backend_cpu_set_n_threads(backend, n_threads);
+    }
+    ggml_backend_graph_compute(backend, gf);
+
+    // 10. Retrieve results (output tensors)
+    // in this example, output tensor is always the last tensor in the graph
+    struct ggml_tensor * result = result0;
+    // struct ggml_tensor * result = gf->nodes[gf->n_nodes - 1];
+    float * result_data = (float *)malloc(ggml_nbytes(result));
+    // because the tensor data is stored in device buffer, we need to copy it back to RAM
+    ggml_backend_tensor_get(result, result_data, 0, ggml_nbytes(result));
+    const std::string bin_file = "mrope_2d_" + backend_name +".bin";
+    std::ofstream outFile(bin_file, std::ios::binary);
+
+    if (outFile.is_open()) {
+        outFile.write(reinterpret_cast<const char*>(result_data), ggml_nbytes(result));
+        outFile.close();
+        std::cout << "Data successfully written to " + bin_file << std::endl;
+    } else {
+        std::cerr << "Error opening file!" << std::endl;
+    }
+
+    free(result_data);
+    // 11. Free memory and exit
+    ggml_free(ctx_cgraph);
+    ggml_gallocr_free(allocr);
+    ggml_free(ctx);
+    ggml_backend_buffer_free(buffer);
+    ggml_backend_free(backend);
+}
+
+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 ne = n_embd * 4;
+    float vals[56 * 56 * 3];
+    // float embd[ne];
+    std::vector<float> embd;
+    embd.resize(ne);
+
+    for (int i = 0; i < 56*56; i++)
+    {
+        for (int c = 0; c < 3; c++)
+            vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
+    }
+
+    clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
+
+    std::ofstream outFile("img_embed.bin", std::ios::binary);
+    if (outFile.is_open()) {
+        outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
+
+        outFile.close();
+        std::cout << "Data successfully written to mrope.bin" << std::endl;
+    } else {
+        std::cerr << "Error opening file!" << std::endl;
+    }
+}
+
+#endif
+
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        print_usage(argc, argv);
+        return 1;
+    }
+
+    auto * model = llava_init(&params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
+        return 1;
+    }
+
+    if (prompt_contains_image(params.prompt)) {
+        auto * ctx_llava = llava_init_context(&params, model);
+
+        auto * image_embed = load_image(ctx_llava, &params, "");
+
+        // process the prompt
+        process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+        llama_perf_context_print(ctx_llava->ctx_llama);
+        llava_image_embed_free(image_embed);
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+#ifndef NDEBUG
+    } else if (params.image[0].empty()) {
+        auto ctx_llava = llava_init_context(&params, model);
+
+        debug_test_mrope_2d();
+        debug_dump_img_embed(ctx_llava);
+
+        llama_perf_context_print(ctx_llava->ctx_llama);
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+#endif
+    } else {
+        for (auto & image : params.image) {
+            auto * ctx_llava = llava_init_context(&params, model);
+
+            auto * image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_perf_context_print(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
+    }
+
+    llama_model_free(model);
+
+    return 0;
+}

examples/lookahead/lookahead.cpp

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

@@ -1,14 +1,9 @@
 #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){
@@ -25,16 +20,16 @@ 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;
-    llama_context * ctx = llama_init.context;
+    llama_model_ptr & model = llama_init.model;
+    llama_context_ptr & ctx = llama_init.context;
+
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx.get(), 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

@@ -30,12 +30,11 @@ 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;
-    llama_context * ctx = llama_init.context;
+    llama_context_ptr & ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx.get(), params.prompt, true, true);
 
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
@@ -66,7 +65,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx = llama_n_ctx(ctx.get());
 
     int n_drafted = 0;
     int n_accept  = 0;
@@ -150,9 +149,6 @@ 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

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

@@ -177,16 +177,11 @@ 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)).

examples/main/main.cpp

@@ -5,7 +5,6 @@
 #include "sampling.h"
 #include "llama.h"
 
-#include <cassert>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
@@ -31,6 +30,8 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static const char * DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant";
+
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
 static common_sampler          ** g_smpl;
@@ -145,24 +146,26 @@ int main(int argc, char ** argv) {
     llama_context * ctx = nullptr;
     common_sampler * smpl = nullptr;
 
-    std::vector<common_chat_msg> chat_msgs;
-
     g_model = &model;
     g_ctx = &ctx;
     g_smpl = &smpl;
 
+    std::vector<common_chat_msg> chat_msgs;
+
     // load the model and apply lora adapter, if any
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
     common_init_result llama_init = common_init_from_params(params);
 
-    model = llama_init.model;
-    ctx = llama_init.context;
+    model = llama_init.model.get();
+    ctx = llama_init.context.get();
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n", __func__);
         return 1;
     }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
 
     auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
@@ -196,15 +199,31 @@ int main(int argc, char ** argv) {
 
     llama_attach_threadpool(ctx, threadpool, threadpool_batch);
 
-    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx_train = llama_model_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
     if (n_ctx > n_ctx_train) {
         LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
     }
 
+    // auto enable conversation mode if chat template is available
+    const bool has_chat_template = !common_get_builtin_chat_template(model).empty() || !params.chat_template.empty();
+    if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
+        if (has_chat_template) {
+            LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
+            params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
+        } else {
+            params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
+        }
+    }
+
+    // in case user force-activate conversation mode (via -cnv) without proper chat template, we show a warning
+    if (params.conversation_mode && !has_chat_template) {
+        LOG_WRN("%s: chat template is not available or is not supported. This may cause the model to output suboptimal responses\n", __func__);
+    }
+
     // print chat template example in conversation mode
-    if (params.conversation) {
+    if (params.conversation_mode) {
         if (params.enable_chat_template) {
             LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(model, params.chat_template).c_str());
         } else {
@@ -241,9 +260,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool add_bos = llama_add_bos_token(model);
+    const bool add_bos = llama_vocab_get_add_bos(vocab);
     if (!llama_model_has_encoder(model)) {
-        GGML_ASSERT(!llama_add_eos_token(model));
+        GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
     }
 
     LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
@@ -251,8 +270,10 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd_inp;
 
     {
-        auto prompt = (params.conversation && params.enable_chat_template && !params.prompt.empty())
-            ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
+        auto prompt = (params.conversation_mode && params.enable_chat_template)
+            // format the system prompt in conversation mode (fallback to default if empty)
+            ? chat_add_and_format(model, chat_msgs, "system", params.prompt.empty() ? DEFAULT_SYSTEM_MESSAGE : params.prompt)
+            // otherwise use the prompt as is
             : params.prompt;
         if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
             LOG_DBG("tokenize the prompt\n");
@@ -269,7 +290,7 @@ int main(int argc, char ** argv) {
     // Should not run without any tokens
     if (embd_inp.empty()) {
         if (add_bos) {
-            embd_inp.push_back(llama_token_bos(model));
+            embd_inp.push_back(llama_vocab_bos(vocab));
             LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
         } else {
             LOG_ERR("input is empty\n");
@@ -326,7 +347,7 @@ int main(int argc, char ** argv) {
         params.n_keep += add_bos; // always keep the BOS token
     }
 
-    if (params.conversation) {
+    if (params.conversation_mode) {
         params.interactive_first = true;
     }
 
@@ -450,7 +471,11 @@ int main(int argc, char ** argv) {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
         LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-        LOG_INF(       "%s\n", control_message);
+        LOG_INF(       "%s", control_message);
+        if (params.conversation_mode && params.enable_chat_template && params.prompt.empty()) {
+            LOG_INF(   " - Using default system message. To change it, set a different value via -p PROMPT or -f FILE argument.\n");
+        }
+        LOG_INF("\n");
 
         is_interacting = params.interactive_first;
     }
@@ -494,8 +519,8 @@ int main(int argc, char ** argv) {
         }
 
         llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-        if (decoder_start_token_id == -1) {
-            decoder_start_token_id = llama_token_bos(model);
+        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            decoder_start_token_id = llama_vocab_bos(vocab);
         }
 
         embd_inp.clear();
@@ -742,7 +767,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of generation tokens in interactive mode
-            if (llama_token_is_eog(model, common_sampler_last(smpl))) {
+            if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
                 LOG_DBG("found an EOG token\n");
 
                 if (params.interactive) {
@@ -762,7 +787,7 @@ int main(int argc, char ** argv) {
             }
 
             // if current token is not EOG, we add it to current assistant message
-            if (params.conversation) {
+            if (params.conversation_mode) {
                 const auto id = common_sampler_last(smpl);
                 assistant_ss << common_token_to_piece(ctx, id, false);
             }
@@ -770,17 +795,17 @@ int main(int argc, char ** argv) {
             if (n_past > 0 && is_interacting) {
                 LOG_DBG("waiting for user input\n");
 
-                if (params.conversation) {
+                if (params.conversation_mode) {
                     LOG("\n> ");
                 }
 
                 if (params.input_prefix_bos) {
                     LOG_DBG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(model));
+                    embd_inp.push_back(llama_vocab_bos(vocab));
                 }
 
                 std::string buffer;
-                if (!params.input_prefix.empty() && !params.conversation) {
+                if (!params.input_prefix.empty() && !params.conversation_mode) {
                     LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     LOG("%s", params.input_prefix.c_str());
                 }
@@ -804,7 +829,7 @@ int main(int argc, char ** argv) {
                 // Entering a empty line lets the user pass control back
                 if (buffer.length() > 1) {
                     // append input suffix if any
-                    if (!params.input_suffix.empty() && !params.conversation) {
+                    if (!params.input_suffix.empty() && !params.conversation_mode) {
                         LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         LOG("%s", params.input_suffix.c_str());
                     }
@@ -817,7 +842,7 @@ int main(int argc, char ** argv) {
                         string_process_escapes(buffer);
                     }
 
-                    bool format_chat = params.conversation && params.enable_chat_template;
+                    bool format_chat = params.conversation_mode && params.enable_chat_template;
                     std::string user_inp = format_chat
                         ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
                         : std::move(buffer);
@@ -830,8 +855,8 @@ int main(int argc, char ** argv) {
 
                     // if user stop generation mid-way, we must add EOT to finish model's last response
                     if (need_insert_eot && format_chat) {
-                        llama_token eot = llama_token_eot(model);
-                        embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot);
+                        llama_token eot = llama_vocab_eot(vocab);
+                        embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);
                         need_insert_eot = false;
                     }
 
@@ -866,7 +891,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of generation
-        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
+        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {
             LOG(" [end of text]\n");
             break;
         }
@@ -889,9 +914,6 @@ int main(int argc, char ** argv) {
 
     common_sampler_free(smpl);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     ggml_threadpool_free_fn(threadpool);

examples/parallel/parallel.cpp

@@ -132,8 +132,10 @@ 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;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
@@ -358,7 +360,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (llama_token_is_eog(model, id) ||
+                        (llama_vocab_is_eog(vocab, id) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {
@@ -416,9 +418,6 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/passkey/passkey.cpp

@@ -63,22 +63,24 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return 1;
     }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     // initialize the context
 
     llama_context_params ctx_params = common_context_params_to_llama(params);
 
-    ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
+    ctx_params.n_ctx = llama_model_n_ctx_train(model)*n_grp + n_keep;
 
     GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
 
-    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
     if (ctx == NULL) {
         LOG_ERR("%s: failed to create the llama_context\n" , __func__);
         return 1;
@@ -223,7 +225,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
 
             // is it an end of generation?
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
                 LOG("\n");
 
                 break;
@@ -266,7 +268,7 @@ int main(int argc, char ** argv) {
     llama_batch_free(batch);
 
     llama_free(ctx);
-    llama_free_model(model);
+    llama_model_free(model);
 
     llama_backend_free();
 

examples/perplexity/perplexity.cpp

@@ -296,8 +296,11 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    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);
+    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
 
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
@@ -338,7 +341,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_batch = params.n_batch;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
 
     int count = 0;
     double nll = 0.0;
@@ -382,7 +385,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+                tokens[batch_start] = llama_vocab_bos(vocab);
             }
 
             const auto * batch_logits = llama_get_logits(ctx);
@@ -444,8 +447,11 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    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);
+    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -485,7 +491,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_batch = params.n_batch;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
 
     int count = 0;
     double nll = 0.0;
@@ -557,7 +563,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
 
                 // add BOS token for the first batch of each chunk
                 if (add_bos && j == 0) {
-                    tokens[seq_start] = llama_token_bos(llama_get_model(ctx));
+                    tokens[seq_start] = llama_vocab_bos(vocab);
                 }
 
                 for (int k = 0; k < batch_size; ++k) {
@@ -732,6 +738,9 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
 }
 
 static void hellaswag_score(llama_context * ctx, const common_params & params) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     // Calculates hellaswag score (acc_norm) from prompt
     //
     // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@@ -765,7 +774,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     LOG_INF("%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
-    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(vocab) == LLAMA_VOCAB_TYPE_SPM;
     LOG_INF("================================= is_spm = %d\n", is_spm);
 
     // The tasks should be randomized so the score stabilizes quickly.
@@ -848,7 +857,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
 
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@@ -1072,6 +1081,8 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
  *
  */
 static void winogrande_score(llama_context * ctx, const common_params & params) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
 
     constexpr int k_min_trailing_ctx = 3;
 
@@ -1130,7 +1141,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
 
     const int max_tasks_per_batch = 128;
     const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@@ -1374,6 +1385,8 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
 //     https://huggingface.co/datasets/truthful_qa
 //
 static void multiple_choice_score(llama_context * ctx, const common_params & params) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
 
     std::istringstream strstream(params.prompt);
     uint32_t n_task;
@@ -1482,7 +1495,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_vocab = llama_vocab_n_tokens(vocab);
 
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@@ -1655,6 +1668,9 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
 }
 
 static void kl_divergence(llama_context * ctx, const common_params & params) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     if (params.logits_file.empty()) {
         LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
         return;
@@ -1688,8 +1704,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         LOG_ERR("%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
         return;
     }
-    if (n_vocab != llama_n_vocab(llama_get_model(ctx))) {
-        LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
+    if (n_vocab != llama_vocab_n_tokens(vocab)) {
+        LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_vocab_n_tokens(vocab));
     }
 
     std::vector<llama_token> tokens(size_t(n_ctx) * n_chunk);
@@ -1701,8 +1717,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
     const int n_batch = params.n_batch;
     const int num_batches = (n_ctx + n_batch - 1)/n_batch;
     const int nv = 2*((n_vocab + 1)/2) + 4;
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    const bool add_bos = llama_vocab_get_add_bos(vocab);
+    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
 
     std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
     std::vector<float>    kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
@@ -1761,7 +1777,7 @@ static void kl_divergence(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_token_bos(llama_get_model(ctx));
+                tokens[batch_start] = llama_vocab_bos(vocab);
             }
 
             common_batch_clear(batch);
@@ -1987,14 +2003,15 @@ int main(int argc, char ** argv) {
     // load the model and apply lora adapter, if any
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
+
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx_train = llama_model_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",
@@ -2023,9 +2040,6 @@ 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/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "llama-impl.h"
+#include "llama-context.h"
+#include "common.h"
 
 #include <algorithm>
 #include <cassert>
@@ -9,11 +9,9 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <map>
 #include <numeric>
 #include <regex>
 #include <string>
-#include <unordered_map>
 #include <vector>
 #include <thread>
 #include <mutex>
@@ -311,7 +309,7 @@ int main(int argc, char ** argv) {
         auto mparams = llama_model_default_params();
         mparams.use_mlock  = false;
 
-        model = llama_load_model_from_file(params.model.c_str(), mparams);
+        model = llama_model_load_from_file(params.model.c_str(), mparams);
 
         if (model == NULL) {
             fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
@@ -321,22 +319,22 @@ int main(int argc, char ** argv) {
         auto cparams = llama_context_default_params();
         cparams.n_ctx = 256;
 
-        ctx = llama_new_context_with_model(model, cparams);
+        ctx = llama_init_from_model(model, cparams);
 
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
-            llama_free_model(model);
+            llama_model_free(model);
             return 1;
         }
     }
 
-    const auto &tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(ctx);
 
     // check layer tensors
     int included_layers = 0;
     int64_t max_nelements = 0;
     bool is_f16 = false;
-    for (const auto& kv_tensor : tensors) {
+    for (const auto & kv_tensor : tensors) {
         if (!layer_included(params, kv_tensor.first)) {
             continue;
         }
@@ -349,7 +347,7 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: error: Quantization should be tested with a float model, "
                 "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
             llama_free(ctx);
-            llama_free_model(model);
+            llama_model_free(model);
             return 1;
         }
         included_layers++;
@@ -371,8 +369,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        const auto *  qfns     = ggml_get_type_traits(type);
-        const auto *  qfns_cpu = ggml_get_type_traits_cpu(type);
+        const auto * qfns     = ggml_get_type_traits(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
         if (qfns_cpu->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
@@ -382,7 +380,7 @@ int main(int argc, char ** argv) {
 
             error_stats global_stats {};
 
-            for (const auto& kv_tensor : tensors) {
+            for (const auto & kv_tensor : tensors) {
                 if (!layer_included(params, kv_tensor.first)) {
                     continue;
                 }
@@ -411,7 +409,7 @@ int main(int argc, char ** argv) {
 
 
     llama_free(ctx);
-    llama_free_model(model);
+    llama_model_free(model);
     // report timing
     {
         const int64_t t_main_end_us = ggml_time_us();

examples/quantize/README.md

@@ -54,8 +54,6 @@ As the models are currently fully loaded into memory, you will need adequate dis
 
 Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
 
-The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format.
-
 *(outdated)*
 
 | Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
@@ -83,7 +81,7 @@ The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interle
   - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
   - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
   - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
-  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
   - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
   - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
   - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)

examples/quantize/quantize.cpp

@@ -48,9 +48,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_M",   LLAMA_FTYPE_MOSTLY_Q5_K_M,   " 5.33G, +0.0569 ppl @ Llama-3-8B",  },
     { "Q6_K",     LLAMA_FTYPE_MOSTLY_Q6_K,     " 6.14G, +0.0217 ppl @ Llama-3-8B",  },
     { "Q8_0",     LLAMA_FTYPE_MOSTLY_Q8_0,     " 7.96G, +0.0026 ppl @ Llama-3-8B",  },
-    { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
-    { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
-    { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
     { "F16",      LLAMA_FTYPE_MOSTLY_F16,      "14.00G, +0.0020 ppl @ Mistral-7B",  },
     { "BF16",     LLAMA_FTYPE_MOSTLY_BF16,     "14.00G, -0.0050 ppl @ Mistral-7B",  },
     { "F32",      LLAMA_FTYPE_ALL_F32,         "26.00G              @ 7B",          },