metal : avoid uint (#11019)

* llama : scatter llama.cpp into multiple modules (wip)

* llama : control-vector -> adapter

* llama : arch

* llama : mmap

ggml-ci

* ci : remove BUILD_SHARED_LIBS=OFF

ggml-ci

* llama : arch (cont)

ggml-ci

* llama : chat

ggml-ci

* llama : model

ggml-ci

* llama : hparams

ggml-ci

* llama : adapter

ggml-ci

* examples : fix

ggml-ci

* rebase

ggml-ci

* minor

* llama : kv cache

ggml-ci

* llama : impl

ggml-ci

* llama : batch

ggml-ci

* cont

ggml-ci

* llama : context

ggml-ci

* minor

* llama : context (cont)

ggml-ci

* llama : model loader

ggml-ci

* common : update lora

ggml-ci

* llama : quant

ggml-ci

* llama : quant (cont)

ggml-ci

* minor [no 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/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
@@ -645,23 +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 -DBUILD_SHARED_LIBS=O'
+          - 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
@@ -703,6 +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: |
@@ -732,7 +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
@@ -884,7 +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

.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
@@ -79,25 +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)
+        if: ${{ matrix.config.free_disk_space == true }}
         uses: jlumbroso/free-disk-space@main
         with:
           # this might remove tools that are actually needed,
@@ -113,13 +115,59 @@ 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/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

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 +1,5 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
 
-ci/ @ggerganov
+/ci/ @ggerganov
+/.devops/ @ngxson
+/examples/server/ @ngxson

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 \
@@ -1404,6 +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)

README.md

@@ -98,6 +98,7 @@ 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] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 
 #### Multimodal
 
@@ -110,6 +111,7 @@ 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 +201,7 @@ 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 +222,7 @@ 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 |
 
@@ -412,7 +415,7 @@ 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 +436,20 @@ 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)
 

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

@@ -119,32 +119,65 @@ 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,
+        std::string & model_url,
+        std::string & hf_repo,
+        std::string & hf_file) {
+    if (!hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
-        if (params.hf_file.empty()) {
-            if (params.model.empty()) {
+        if (hf_file.empty()) {
+            if (model.empty()) {
                 throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
             }
-            params.hf_file = params.model;
-        } else if (params.model.empty()) {
+            hf_file = model;
+        } else 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 +280,9 @@ 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);
+    common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -591,7 +626,7 @@ 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),
@@ -813,7 +848,7 @@ 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 +861,6 @@ 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 +915,9 @@ 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 +972,9 @@ 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 +1208,28 @@ 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 +1512,7 @@ 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 +1520,7 @@ 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}));
@@ -1543,6 +1587,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE"));
+    add_opt(common_arg(
+        {"-hfrv", "--hf-repo-v"}, "REPO",
+        "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 +1769,13 @@ 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 +2141,35 @@ 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 +2189,34 @@ 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/common.cpp

@@ -18,6 +18,7 @@
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
+#include <filesystem>
 #include <fstream>
 #include <iostream>
 #include <iterator>
@@ -62,7 +63,9 @@
 #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
@@ -919,20 +922,21 @@ 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_lora_adapter_ptr lora;
+        lora.reset(llama_lora_adapter_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);
             return iparams;
         }
-        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+
+        la.ptr = lora.get();
+        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
+
     if (!params.lora_init_without_apply) {
-        common_lora_adapters_apply(lctx, iparams.lora_adapters);
+        common_lora_adapters_apply(lctx, params.lora_adapters);
     }
 
     if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
@@ -940,6 +944,25 @@ struct common_init_result common_init_from_params(common_params & params) {
         params.sampling.ignore_eos = false;
     }
 
+    if (params.sampling.ignore_eos) {
+        for (llama_token i = 0; i < llama_n_vocab(model); i++) {
+            if (llama_token_is_eog(model, 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__);
 
@@ -974,17 +997,17 @@ 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) {
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora) {
     llama_lora_adapter_clear(ctx);
-    for (auto & la : lora_adapters) {
+    for (auto & la : lora) {
         if (la.scale != 0.0f) {
-            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+            llama_lora_adapter_set(ctx, la.ptr, la.scale);
         }
     }
 }
@@ -1015,38 +1038,6 @@ 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 +1072,8 @@ 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 +1099,7 @@ 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) {
@@ -1138,7 +1123,6 @@ 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);
     if (!curl) {
@@ -1168,8 +1152,7 @@ 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 +1194,13 @@ 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);
@@ -1630,6 +1615,18 @@ 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) {
+    static const char * template_key = "tokenizer.chat_template";
+    // call with NULL buffer to get the total size of the string
+    int32_t res = llama_model_meta_val_str(model, template_key, NULL, 0);
+    if (res > 0) {
+        std::vector<char> model_template(res + 1, 0);
+        llama_model_meta_val_str(model, template_key, model_template.data(), model_template.size());
+        return std::string(model_template.data(), model_template.size() - 1);
+    }
+    return "";
+}
+
 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);
@@ -1799,7 +1796,9 @@ 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>
@@ -27,19 +27,17 @@
 struct common_lora_adapter_info {
     std::string path;
     float scale;
-};
 
-struct common_lora_adapter_container : common_lora_adapter_info {
-    struct llama_lora_adapter * adapter;
+    struct llama_lora_adapter * 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
@@ -130,7 +130,6 @@ 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 +138,7 @@ 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 +158,7 @@ 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 +172,14 @@ 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 +202,13 @@ 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;
@@ -211,8 +222,9 @@ struct common_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct common_params_sampling sampling;
+    struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
+    struct common_params_vocoder     vocoder;
 
     std::string model                = ""; // model path                                                    // NOLINT
     std::string model_alias          = ""; // model alias                                                   // NOLINT
@@ -286,8 +298,8 @@ 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
 
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
@@ -437,6 +449,11 @@ 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;
+}
+
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
@@ -459,10 +476,12 @@ 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_lora_adapter_ptr> lora;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -484,7 +503,7 @@ struct llama_model * common_load_model_from_hf(
     const struct llama_model_params & params);
 
 // 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_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora);
 
 //
 // Batch utils
@@ -552,6 +571,9 @@ 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 +610,8 @@ 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 +640,10 @@ 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/sampling.cpp

@@ -161,32 +161,20 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 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()));
                     }
-                        break;
+                    break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
                     break;
@@ -208,6 +196,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 case COMMON_SAMPLER_TYPE_INFILL:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
                     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");
             }
@@ -415,6 +406,7 @@ 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 +421,7 @@ 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 +436,7 @@ 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 +483,7 @@ 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;

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 +326,8 @@ 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.POSNET_NORM1,
+                            gguf.MODEL_TENSOR.POSNET_NORM2,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -525,9 +529,19 @@ 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 +585,9 @@ 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"
@@ -664,6 +681,12 @@ class Model:
         if chkhsh == "8b5a93ed704057481f240da0be7e7dca721d7f8f4755263b6807227a2cbeae65":
             # ref: https://huggingface.co/sentence-transformers/stsb-roberta-base
             res = "roberta-bpe"
+        if chkhsh == "ad851be1dba641f2e3711822f816db2c265f788b37c63b4e1aeacb9ee92de8eb":
+            # ref: https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct
+            res = "gigachat"
+        if chkhsh == "d4c8f286ea6b520b3d495c4455483cfa2302c0cfcd4be05d781b6a8a0a7cdaf1":
+            # ref: https://huggingface.co/Infinigence/Megrez-3B-Instruct
+            res = "megrez"
 
         if res is None:
             logger.warning("\n")
@@ -686,6 +709,9 @@ 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")
@@ -1669,6 +1695,178 @@ 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
@@ -2001,6 +2199,67 @@ class Qwen2Model(Model):
                 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):
     model_arch = gguf.MODEL_ARCH.QWEN2MOE
@@ -2129,6 +2388,15 @@ 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'
@@ -2245,7 +2513,11 @@ 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"])
@@ -2544,7 +2816,7 @@ class InternLM2Model(Model):
             return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("BertModel", "CamembertModel", "RobertaModel")
+@Model.register("BertModel", "BertForMaskedLM", "CamembertModel")
 class BertModel(Model):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -2610,13 +2882,73 @@ 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
@@ -2936,6 +3268,9 @@ class Rwkv6Model(Model):
         if new_name.endswith("time_mix_w2.weight"):
             data_torch = data_torch.permute(0, 2, 1)
 
+        if new_name.endswith("time_mix_decay.weight") or "lerp" in new_name:
+            data_torch = data_torch.squeeze()
+
         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"):
@@ -3404,6 +3739,97 @@ 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")
 class DeepseekV2Model(Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2

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", },
@@ -104,6 +105,8 @@ models = [
     {"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"},
 ]
 
 

docs/build.md

@@ -57,6 +57,13 @@ cmake --build build --config 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_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
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Using BLAS doesn't affect the generation performance. There are currently several different BLAS implementations available for build and use:

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/batched.cpp

@@ -65,6 +65,7 @@ int main(int argc, char ** argv) {
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     auto sparams = llama_sampler_chain_default_params();
+    sparams.no_perf = false;
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 

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

@@ -434,12 +434,12 @@ 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 +500,7 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    ~llama_file() {
+    ~my_llama_file() {
         if (fp) {
             std::fclose(fp);
         }
@@ -508,7 +508,7 @@ 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 +576,7 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
     } else {
         // assume llama2.c vocabulary
         LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
-        llama_file file(filename, "rb");
+        my_llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
         }

examples/cvector-generator/cvector-generator.cpp

@@ -415,12 +415,13 @@ 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);
+
     // 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 +475,6 @@ 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

@@ -97,8 +97,9 @@ 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;
@@ -316,8 +317,6 @@ 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

@@ -162,8 +162,9 @@ 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 +185,6 @@ 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

@@ -265,8 +265,8 @@ 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 +352,7 @@ 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-split/gguf-split.cpp

@@ -2,15 +2,14 @@
 #include "common.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdlib>
 #include <fstream>
 #include <string>
 #include <vector>
-
-#include <stdio.h>
-#include <string.h>
 #include <climits>
+
+#include <cstdio>
+#include <cstring>
 #include <stdexcept>
 
 #if defined(_WIN32)
@@ -287,7 +286,7 @@ 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 +296,7 @@ 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 = %d, total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
             i_split++;
         }
     }

examples/gritlm/gritlm.cpp

@@ -75,7 +75,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         std::vector<float> emb_norm(emb_unorm.size());
-        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
+        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
         result.push_back(emb_norm);
 
 #ifdef GRIT_DEBUG

examples/imatrix/imatrix.cpp

@@ -430,9 +430,10 @@ 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 int n_ctx = llama_n_ctx(ctx);
 
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
@@ -618,8 +619,9 @@ 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;
@@ -655,9 +657,6 @@ 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

@@ -131,8 +131,8 @@ 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__);
@@ -581,9 +581,6 @@ int main(int argc, char ** argv) {
     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

@@ -1521,7 +1521,7 @@ 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)) {
@@ -1573,14 +1573,14 @@ 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 +1592,14 @@ 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);

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

@@ -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"

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

@@ -8,25 +8,25 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.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 +102,9 @@ 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 +131,8 @@ 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 +166,7 @@ enum projector_type {
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MERGER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -171,6 +175,7 @@ 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"},
 };
 
 
@@ -463,7 +468,8 @@ 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 +559,7 @@ 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 +568,7 @@ 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 +614,26 @@ 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 +654,30 @@ 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 +699,14 @@ 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 +730,8 @@ 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 +753,13 @@ 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 +767,11 @@ 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 +811,8 @@ 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 +824,7 @@ 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 +896,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                 mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
                 mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
                 // stride = 1, padding = 1, bias is nullptr
-                block_1 = ggml_conv_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 +944,7 @@ 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 +1005,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             // mlp_2 ne [24, 24, 2048, 1]
             mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
             // weight ne = [3, 3, 2048, 1]
-            struct ggml_tensor * peg_0 = ggml_conv_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 +1086,19 @@ 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 +1222,30 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-#ifdef GGML_USE_CUDA
-    new_clip->backend = ggml_backend_cuda_init(0);
-    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
-#endif
-
-#ifdef GGML_USE_METAL
-    new_clip->backend = ggml_backend_metal_init();
-    LOG_INF("%s: CLIP using Metal backend\n", __func__);
-#endif
-
-#ifdef GGML_USE_CANN
-    new_clip->backend = ggml_backend_cann_init(0);
-    LOG_INF("%s: CLIP using CANN backend\n", __func__);
-#endif
-
-#ifdef GGML_USE_VULKAN
-    new_clip->backend = ggml_backend_vk_init(0);
-    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
-#endif
-
-#ifdef GGML_USE_SYCL
-    new_clip->backend = ggml_backend_sycl_init(0);
-    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
-#endif
+//#ifdef GGML_USE_CUDA
+//    new_clip->backend = ggml_backend_cuda_init(0);
+//    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
+//#endif
+//
+//#ifdef GGML_USE_METAL
+//    new_clip->backend = ggml_backend_metal_init();
+//    LOG_INF("%s: CLIP using Metal backend\n", __func__);
+//#endif
+//
+//#ifdef GGML_USE_CANN
+//    new_clip->backend = ggml_backend_cann_init(0);
+//    LOG_INF("%s: CLIP using CANN backend\n", __func__);
+//#endif
+//
+//#ifdef GGML_USE_VULKAN
+//    new_clip->backend = ggml_backend_vk_init(0);
+//    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+//#endif
+//
+//#ifdef GGML_USE_SYCL
+//    new_clip->backend = ggml_backend_sycl_init(0);
+//    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
+//#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
@@ -1206,6 +1275,10 @@ 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 +1287,13 @@ 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 +1469,16 @@ 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 +1566,12 @@ 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 +1610,7 @@ 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 +1624,10 @@ 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 +2080,23 @@ 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 +2286,13 @@ 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 +2314,13 @@ 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 +2334,11 @@ 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 +2467,7 @@ 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 +2487,7 @@ 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 +2545,9 @@ 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 +2567,34 @@ 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 +2603,16 @@ 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);
         }
     }
 
@@ -2626,6 +2785,9 @@ 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 +2799,21 @@ 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.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++) {
@@ -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;

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 +1,581 @@
+#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_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);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), 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_load_model_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_new_context_with_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_free_model(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_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_free_model(model);
+
+    return 0;
+}

examples/lookahead/lookahead.cpp

@@ -58,8 +58,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -474,9 +474,6 @@ 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 +33,8 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
 
     // tokenize the prompt
     std::vector<llama_token> inp;
@@ -243,9 +243,6 @@ 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

@@ -145,18 +145,18 @@ 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__);
@@ -889,9 +889,6 @@ 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 +132,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model = llama_init.model.get();
+    llama_context * ctx = llama_init.context.get();
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
@@ -416,9 +416,6 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/perplexity/perplexity.cpp

@@ -1987,8 +1987,9 @@ 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;
@@ -2023,9 +2024,6 @@ 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>
@@ -330,13 +328,13 @@ int main(int argc, char ** argv) {
         }
     }
 
-    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;
         }
@@ -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;
                 }

examples/quantize/README.md

@@ -81,7 +81,7 @@ Several quantization methods are supported. They differ in the resulting model d
   - [#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/retrieval/retrieval.cpp

@@ -107,7 +107,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         }
 
         float * out = output + batch.seq_id[i][0] * n_embd;
-        common_embd_normalize(embd, out, n_embd);
+        common_embd_normalize(embd, out, n_embd, 2);
     }
 }
 
@@ -143,7 +143,7 @@ int main(int argc, char ** argv) {
         std::vector<chunk> file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator);
         chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
     }
-    LOG_INF("Number of chunks: %ld\n", chunks.size());
+    LOG_INF("Number of chunks: %zu\n", chunks.size());
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -151,8 +151,8 @@ int main(int argc, char ** argv) {
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
-    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__);
@@ -298,7 +298,5 @@ int main(int argc, char ** argv) {
 
     // clean up
     llama_batch_free(query_batch);
-    llama_free(ctx);
-    llama_free_model(model);
     llama_backend_free();
 }

examples/rpc/rpc-server.cpp

@@ -12,6 +12,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
 #include "ggml-rpc.h"
 #ifdef _WIN32
 #  include <windows.h>
@@ -91,6 +95,12 @@ static ggml_backend_t create_backend() {
     if (!backend) {
         fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
     }
+#elif GGML_USE_SYCL
+    fprintf(stderr, "%s: using SYCL backend\n", __func__);
+    backend = ggml_backend_sycl_init(0); // init device 0
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
+    }
 #endif
 
     // if there aren't GPU Backends fallback to CPU backend
@@ -106,6 +116,8 @@ static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
     ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
 #elif GGML_USE_VULKAN
     ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
+#elif GGML_USE_SYCL
+    ggml_backend_sycl_get_device_memory(0, free_mem, total_mem);
 #else
     #ifdef _WIN32
         MEMORYSTATUSEX status;

examples/run/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-run)
 add_executable(${TARGET} run.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/run/README.md

@@ -3,5 +3,49 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
 
 ```bash
-./llama-run Meta-Llama-3.1-8B-Instruct.gguf
-...
+llama-run granite-code
+```
+
+```bash
+llama-run -h
+Description:
+  Runs a llm
+
+Usage:
+  llama-run [options] model [prompt]
+
+Options:
+  -c, --context-size <value>
+      Context size (default: 2048)
+  -n, --ngl <value>
+      Number of GPU layers (default: 0)
+  --temp <value>
+      Temperature (default: 0.8)
+  -v, --verbose, --log-verbose
+      Set verbosity level to infinity (i.e. log all messages, useful for debugging)
+  -h, --help
+      Show help message
+
+Commands:
+  model
+      Model is a string with an optional prefix of
+      huggingface:// (hf://), ollama://, https:// or file://.
+      If no protocol is specified and a file exists in the specified
+      path, file:// is assumed, otherwise if a file does not exist in
+      the specified path, ollama:// is assumed. Models that are being
+      pulled are downloaded with .partial extension while being
+      downloaded and then renamed as the file without the .partial
+      extension when complete.
+
+Examples:
+  llama-run llama3
+  llama-run ollama://granite-code
+  llama-run ollama://smollm:135m
+  llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf
+  llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf
+  llama-run https://example.com/some-file1.gguf
+  llama-run some-file2.gguf
+  llama-run file://some-file3.gguf
+  llama-run --ngl 999 some-file4.gguf
+  llama-run --ngl 999 some-file5.gguf Hello World
+```

examples/save-load-state/save-load-state.cpp

@@ -30,8 +30,8 @@ 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) {
         fprintf(stderr, "%s : failed to init\n", __func__);
@@ -89,8 +89,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -98,11 +96,8 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
-    // free old context
-    llama_free(ctx);
-
     // make new context
-    auto * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
+    llama_context * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
@@ -123,8 +118,6 @@ int main(int argc, char ** argv) {
 
         if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            llama_free(ctx2);
-            llama_free_model(model);
             return 1;
         }
 
@@ -148,8 +141,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx2, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx2);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -157,15 +148,13 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
-    llama_free(ctx2);
-
     if (result0 != result1) {
         fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
         return 1;
     }
 
     // make new context
-    auto * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
+    llama_context * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
@@ -186,8 +175,6 @@ int main(int argc, char ** argv) {
 
         if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
 
@@ -204,8 +191,6 @@ int main(int argc, char ** argv) {
         const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
@@ -218,8 +203,6 @@ int main(int argc, char ** argv) {
         const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
@@ -239,8 +222,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx3, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -253,8 +234,6 @@ int main(int argc, char ** argv) {
     llama_sampler_free(smpl3);
 
     llama_batch_free(batch);
-    llama_free(ctx3);
-    llama_free_model(model);
 
     if (result0 != result2) {
         fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);

examples/server/CMakeLists.txt

@@ -15,7 +15,7 @@ set(TARGET_SRCS
     httplib.h
 )
 set(PUBLIC_ASSETS
-    index.html
+    index.html.gz
     loading.html
 )
 
@@ -34,6 +34,7 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
 
 if (LLAMA_SERVER_SSL)

examples/server/README.md

@@ -62,8 +62,8 @@ The project is under active development, and we are [looking for feedback and co
 | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
-| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
-| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
@@ -104,7 +104,6 @@ The project is under active development, and we are [looking for feedback and co
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--penalize-nl` | penalize newline tokens (default: false) |
 | `--temp N` | temperature (default: 0.8) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
 | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
@@ -138,6 +137,7 @@ The project is under active development, and we are [looking for feedback and co
 | -------- | ----------- |
 | `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
 | `-sp, --special` | special tokens output enabled (default: false) |
+| `--no-warmup` | skip warming up the model with an empty run |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
 | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
@@ -146,6 +146,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
+| `--no-webui` | Disable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_NO_WEBUI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
@@ -163,13 +164,13 @@ The project is under active development, and we are [looking for feedback and co
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>list of built-in templates:<br/>chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
-| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16) |
-| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9) |
-| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model) |
+| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
+| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
-| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model |
-| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused) |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
+| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
@@ -302,23 +303,23 @@ mkdir llama-client
 cd llama-client
 ```
 
-Create a index.js file and put this inside:
+Create an index.js file and put this inside:
 
 ```javascript
-const prompt = `Building a website can be done in 10 simple steps:`;
+const prompt = "Building a website can be done in 10 simple steps:"
 
-async function Test() {
+async function test() {
     let response = await fetch("http://127.0.0.1:8080/completion", {
-        method: 'POST',
+        method: "POST",
         body: JSON.stringify({
             prompt,
-            n_predict: 512,
+            n_predict: 64,
         })
     })
     console.log((await response.json()).content)
 }
 
-Test()
+test()
 ```
 
 And run it:
@@ -342,6 +343,10 @@ node index.js
 
 ### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
+> [!IMPORTANT]
+>
+> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/completions` instead.
+
 *Options:*
 
 `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, if `cache_prompt` is `true`, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. A `BOS` token is inserted at the start, if all of the following conditions are true:
@@ -380,7 +385,7 @@ Multiple prompts are also supported. In this case, the completion result will be
 `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
 By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
-`stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
+`stream`: Allows receiving each predicted token in real-time instead of waiting for the completion to finish (uses a different response format). To enable this, set to `true`.
 
 `stop`: Specify a JSON array of stopping strings.
 These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
@@ -391,8 +396,6 @@ These words will not be included in the completion, so make sure to add them to
 
 `repeat_last_n`: Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.
 
-`penalize_nl`: Penalize newline tokens when applying the repeat penalty. Default: `true`
-
 `presence_penalty`: Repeat alpha presence penalty. Default: `0.0`, which is disabled.
 
 `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
@@ -439,40 +442,78 @@ These words will not be included in the completion, so make sure to add them to
 
 `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true`
 
+`return_tokens`: Return the raw generated token ids in the `tokens` field. Otherwise `tokens` remains empty. Default: `false`
+
 `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 
-    `timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
+`timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
+
+`post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
+
+`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error. Note that fields with a slash will be unnested; for example, `generation_settings/n_predict` will move the field `n_predict` from the `generation_settings` object to the root of the response and give it a new name.
+
+`lora`: A list of LoRA adapters to be applied to this specific request. Each object in the list must contain `id` and `scale` fields. For example: `[{"id": 0, "scale": 0.5}, {"id": 1, "scale": 1.1}]`. If a LoRA adapter is not specified in the list, its scale will default to `0.0`. Please note that requests with different LoRA configurations will not be batched together, which may result in performance degradation.
 
 **Response format**
 
-- Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
+- Note: In streaming mode (`stream`), only `content`, `tokens` and `stop` will be returned until end of completion. Responses are sent using the [Server-sent events](https://html.spec.whatwg.org/multipage/server-sent-events.html) standard. Note: the browser's `EventSource` interface cannot be used due to its lack of `POST` request support.
 
-- `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
-
-```json
-{
-  "content": "<the token selected by the model>",
-  "probs": [
-    {
-      "prob": float,
-      "tok_str": "<most likely token>"
-    },
-    {
-      "prob": float,
-      "tok_str": "<second most likely token>"
-    },
+- `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has a nested array `top_logprobs`. It contains at **maximum** `n_probs` elements:
+  ```json
+  {
+    "content": "<the generated completion text>",
+    "tokens": [ generated token ids if requested ],
     ...
-  ]
-},
-```
-
-Notice that each `probs` is an array of length `n_probs`.
+    "probs": [
+      {
+        "id": <token id>,
+        "logprob": float,
+        "token": "<most likely token>",
+        "bytes": [int, int, ...],
+        "top_logprobs": [
+          {
+            "id": <token id>,
+            "logprob": float,
+            "token": "<token text>",
+            "bytes": [int, int, ...],
+          },
+          {
+            "id": <token id>,
+            "logprob": float,
+            "token": "<token text>",
+            "bytes": [int, int, ...],
+          },
+          ...
+        ]
+      },
+      {
+        "id": <token id>,
+        "logprob": float,
+        "token": "<most likely token>",
+        "bytes": [int, int, ...],
+        "top_logprobs": [
+          ...
+        ]
+      },
+      ...
+    ]
+  },
+  ```
+  Please note that if `post_sampling_probs` is set to `true`:
+    - `logprob` will be replaced with `prob`, with the value between 0.0 and 1.0
+    - `top_logprobs` will be replaced with `top_probs`. Each element contains:
+      - `id`: token ID
+      - `token`: token in string
+      - `bytes`: token in bytes
+      - `prob`: token probability, with the value between 0.0 and 1.0
+    - Number of elements in `top_probs` may be less than `n_probs`
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
+- `tokens`: Same as `content` but represented as raw token ids. Only populated if `"return_tokens": true` or `"stream": true` in the request.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
 - `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
-- `model`: The path to the model loaded with `-m`
-- `prompt`: The provided `prompt`
+- `model`: The model alias (for model path, please use `/props` endpoint)
+- `prompt`: The processed `prompt` (special tokens may be added)
 - `stop_type`: Indicating whether the completion has stopped. Possible values are:
   - `none`: Generating (not stopped)
   - `eos`: Stopped because it encountered the EOS token
@@ -484,6 +525,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
+
 ### POST `/tokenize`: Tokenize a given text
 
 *Options:*
@@ -535,6 +577,10 @@ With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
 
 ### POST `/embedding`: Generate embedding of a given text
 
+> [!IMPORTANT]
+>
+> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/embeddings` instead.
+
 The same as [the embedding example](../embedding) does.
 
 *Options:*
@@ -653,7 +699,6 @@ This endpoint is public (no API key check). By default, it is read-only. To make
       "mirostat": 0,
       "mirostat_tau": 5.0,
       "mirostat_eta": 0.10000000149011612,
-      "penalize_nl": false,
       "stop": [],
       "max_tokens": -1,
       "n_keep": 0,
@@ -688,7 +733,8 @@ This endpoint is public (no API key check). By default, it is read-only. To make
   },
   "total_slots": 1,
   "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
-  "chat_template": "..."
+  "chat_template": "...",
+  "build_info": "b(build number)-(build commit hash)"
 }
 ```
 
@@ -705,94 +751,46 @@ To use this endpoint with POST method, you need to start server with `--props`
 
 - None yet
 
-### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+### POST `/embeddings`: non-OpenAI-compatible embeddings API
 
-Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+This endpoint supports all poolings, including `--pooling none`. When the pooling is `none`, the responses will contain the *unnormalized* embeddings for *all* input tokens. For all other pooling types, only the pooled embeddings are returned, normalized using Euclidian norm.
+
+Note that the response format of this endpoint is different from `/v1/embeddings`.
 
 *Options:*
 
-See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
-
-The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
+Same as the `/v1/embeddings` endpoint.
 
 *Examples:*
 
-You can use either Python `openai` library with appropriate checkpoints:
+Same as the `/v1/embeddings` endpoint.
 
-```python
-import openai
+**Response format**
 
-client = openai.OpenAI(
-    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
-    api_key = "sk-no-key-required"
-)
-
-completion = client.chat.completions.create(
-model="gpt-3.5-turbo",
-messages=[
-    {"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
-    {"role": "user", "content": "Write a limerick about python exceptions"}
+```json
+[
+  {
+    "index": 0,
+    "embedding": [
+      [ ... embeddings for token 0   ... ],
+      [ ... embeddings for token 1   ... ],
+      [ ... ]
+      [ ... embeddings for token N-1 ... ],
+    ]
+  },
+  ...
+  {
+    "index": P,
+    "embedding": [
+      [ ... embeddings for token 0   ... ],
+      [ ... embeddings for token 1   ... ],
+      [ ... ]
+      [ ... embeddings for token N-1 ... ],
+    ]
+  }
 ]
-)
-
-print(completion.choices[0].message)
 ```
 
-... or raw HTTP requests:
-
-```shell
-curl http://localhost:8080/v1/chat/completions \
--H "Content-Type: application/json" \
--H "Authorization: Bearer no-key" \
--d '{
-"model": "gpt-3.5-turbo",
-"messages": [
-{
-    "role": "system",
-    "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
-},
-{
-    "role": "user",
-    "content": "Write a limerick about python exceptions"
-}
-]
-}'
-```
-
-### POST `/v1/embeddings`: OpenAI-compatible embeddings API
-
-*Options:*
-
-See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
-
-*Examples:*
-
-- input as string
-
-  ```shell
-  curl http://localhost:8080/v1/embeddings \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer no-key" \
-  -d '{
-          "input": "hello",
-          "model":"GPT-4",
-          "encoding_format": "float"
-  }'
-  ```
-
-- `input` as string array
-
-  ```shell
-  curl http://localhost:8080/v1/embeddings \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer no-key" \
-  -d '{
-          "input": ["hello", "world"],
-          "model":"GPT-4",
-          "encoding_format": "float"
-  }'
-  ```
-
 ### GET `/slots`: Returns the current slots processing state
 
 > [!WARNING]
@@ -843,7 +841,6 @@ Example:
       "mirostat": 0,
       "mirostat_tau": 5.0,
       "mirostat_eta": 0.10000000149011612,
-      "penalize_nl": false,
       "stop": [],
       "max_tokens": -1,
       "n_keep": 0,
@@ -950,6 +947,8 @@ This endpoint returns the loaded LoRA adapters. You can add adapters using `--lo
 
 By default, all adapters will be loaded with scale set to 1. To initialize all adapters scale to 0, add `--lora-init-without-apply`
 
+Please note that this value will be overwritten by the `lora` field for each request.
+
 If an adapter is disabled, the scale will be set to 0.
 
 **Response format**
@@ -971,6 +970,8 @@ If an adapter is disabled, the scale will be set to 0.
 
 ### POST `/lora-adapters`: Set list of LoRA adapters
 
+This sets the global scale for LoRA adapters. Please note that this value will be overwritten by the `lora` field for each request.
+
 To disable an adapter, either remove it from the list below, or set scale to 0.
 
 **Request format**
@@ -984,6 +985,161 @@ To know the `id` of the adapter, use GET `/lora-adapters`
 ]
 ```
 
+## OpenAI-compatible API Endpoints
+
+### GET `/v1/models`: OpenAI-compatible Model Info API
+
+Returns information about the loaded model. See [OpenAI Models API documentation](https://platform.openai.com/docs/api-reference/models).
+
+The returned list always has one single element.
+
+By default, model `id` field is the path to model file, specified via `-m`. You can set a custom value for model `id` field via `--alias` argument. For example, `--alias gpt-4o-mini`.
+
+Example:
+
+```json
+{
+    "object": "list",
+    "data": [
+        {
+            "id": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
+            "object": "model",
+            "created": 1735142223,
+            "owned_by": "llamacpp",
+            "meta": {
+                "vocab_type": 2,
+                "n_vocab": 128256,
+                "n_ctx_train": 131072,
+                "n_embd": 4096,
+                "n_params": 8030261312,
+                "size": 4912898304
+            }
+        }
+    ]
+}
+```
+
+### POST `/v1/completions`: OpenAI-compatible Completions API
+
+Given an input `prompt`, it returns the predicted completion. Streaming mode is also supported. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps.
+
+*Options:*
+
+See [OpenAI Completions API documentation](https://platform.openai.com/docs/api-reference/completions).
+
+llama.cpp `/completion`-specific features such as `mirostat` are supported.
+
+*Examples:*
+
+Example usage with `openai` python library:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+completion = client.completions.create(
+  model="davinci-002",
+  prompt="I believe the meaning of life is",
+  max_tokens=8
+)
+
+print(completion.choices[0].text)
+```
+
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+
+*Options:*
+
+See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
+
+The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
+
+*Examples:*
+
+You can use either Python `openai` library with appropriate checkpoints:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+completion = client.chat.completions.create(
+  model="gpt-3.5-turbo",
+  messages=[
+    {"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
+    {"role": "user", "content": "Write a limerick about python exceptions"}
+  ]
+)
+
+print(completion.choices[0].message)
+```
+
+... or raw HTTP requests:
+
+```shell
+curl http://localhost:8080/v1/chat/completions \
+-H "Content-Type: application/json" \
+-H "Authorization: Bearer no-key" \
+-d '{
+"model": "gpt-3.5-turbo",
+"messages": [
+{
+    "role": "system",
+    "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
+},
+{
+    "role": "user",
+    "content": "Write a limerick about python exceptions"
+}
+]
+}'
+```
+
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
+
+This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
+
+*Options:*
+
+See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
+
+*Examples:*
+
+- input as string
+
+  ```shell
+  curl http://localhost:8080/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer no-key" \
+  -d '{
+          "input": "hello",
+          "model":"GPT-4",
+          "encoding_format": "float"
+  }'
+  ```
+
+- `input` as string array
+
+  ```shell
+  curl http://localhost:8080/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer no-key" \
+  -d '{
+          "input": ["hello", "world"],
+          "model":"GPT-4",
+          "encoding_format": "float"
+  }'
+  ```
+
 ## More examples
 
 ### Interactive mode

examples/server/bench/README.md

@@ -6,10 +6,10 @@ Benchmark is using [k6](https://k6.io/).
 
 SSE is not supported by default in k6, you have to build k6 with the [xk6-sse](https://github.com/phymbert/xk6-sse) extension.
 
-Example:
+Example (assuming golang >= 1.21 is installed):
 ```shell
 go install go.k6.io/xk6/cmd/xk6@latest
-xk6 build master \
+$GOPATH/bin/xk6 build master \
 --with github.com/phymbert/xk6-sse
 ```
 
@@ -33,7 +33,7 @@ The server must answer OAI Chat completion requests on `http://localhost:8080/v1
 
 Example:
 ```shell
-server --host localhost --port 8080 \
+llama-server --host localhost --port 8080 \
   --model ggml-model-q4_0.gguf \
   --cont-batching \
   --metrics \

examples/server/bench/bench.py

@@ -189,12 +189,12 @@ xychart-beta
         "pp": {
             "p95": round(data['metrics']["llamacpp_prompt_processing_second"]["p(95)"], 2),
             "avg": round(data['metrics']["llamacpp_prompt_processing_second"]["avg"], 2),
-            "0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2),
+            "0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2) if 'prompt_tokens_seconds' in prometheus_metrics else 0,
         },
         "tg": {
             "p95": round(data['metrics']["llamacpp_tokens_second"]["p(95)"], 2),
             "avg": round(data['metrics']["llamacpp_tokens_second"]["avg"], 2),
-            "0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2),
+            "0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2) if 'predicted_tokens_seconds' in prometheus_metrics else 0,
         },
     }
     with open("results.github.env", 'a') as github_env:
@@ -214,11 +214,14 @@ def start_benchmark(args):
     k6_args = [
         'run', args.scenario,
         '--no-color',
+        '--no-connection-reuse',
+        '--no-vu-connection-reuse',
     ]
     k6_args.extend(['--duration', args.duration])
     k6_args.extend(['--iterations', args.n_prompts])
     k6_args.extend(['--vus', args.parallel])
     k6_args.extend(['--summary-export', 'k6-results.json'])
+    k6_args.extend(['--out', 'csv=k6-results.csv'])
     args = f"SERVER_BENCH_N_PROMPTS={args.n_prompts} SERVER_BENCH_MAX_PROMPT_TOKENS={args.max_prompt_tokens} SERVER_BENCH_MAX_CONTEXT={args.max_tokens} "
     args = args + ' '.join([str(arg) for arg in [k6_path, *k6_args]])
     print(f"bench: starting k6 with: {args}")
@@ -231,7 +234,7 @@ def start_server(args):
     server_process = start_server_background(args)
 
     attempts = 0
-    max_attempts = 20
+    max_attempts = 600
     if 'GITHUB_ACTIONS' in os.environ:
         max_attempts *= 2
 
@@ -242,7 +245,15 @@ def start_server(args):
         print(f"bench:     waiting for server to start ...")
         time.sleep(0.5)
 
-    print("bench: server started.")
+    attempts = 0
+    while not is_server_ready(args.host, args.port):
+        attempts += 1
+        if attempts > max_attempts:
+            assert False, "server not ready"
+        print(f"bench:     waiting for server to be ready ...")
+        time.sleep(0.5)
+
+    print("bench: server started and ready.")
     return server_process
 
 
@@ -255,11 +266,6 @@ def start_server_background(args):
         '--host', args.host,
         '--port', args.port,
     ]
-    model_file = args.model_path_prefix + os.path.sep + args.hf_file
-    model_dir  = os.path.dirname(model_file)
-    if not os.path.exists(model_dir):
-        os.makedirs(model_dir)
-    server_args.extend(['--model', model_file])
     server_args.extend(['--hf-repo', args.hf_repo])
     server_args.extend(['--hf-file', args.hf_file])
     server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
@@ -303,6 +309,12 @@ def is_server_listening(server_fqdn, server_port):
         return _is_server_listening
 
 
+def is_server_ready(server_fqdn, server_port):
+    url = f"http://{server_fqdn}:{server_port}/health"
+    response = requests.get(url)
+    return response.status_code == 200
+
+
 def escape_metric_name(metric_name):
     return re.sub('[^A-Z0-9]', '_', metric_name.upper())
 

examples/server/bench/script.js

@@ -56,6 +56,7 @@ const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
 
 const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
 const llamacpp_prompt_processing_second = new Trend('llamacpp_prompt_processing_second')
+const llamacpp_emit_first_token_second = new Trend('llamacpp_emit_first_token_second')
 
 const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
 const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
@@ -89,6 +90,9 @@ export default function () {
         ],
         "model": model,
         "stream": true,
+        "stream_options": {
+          "include_usage": true, // False to be supported in llama.cpp server
+        },
         "seed": 42,
         "max_tokens": max_tokens,
         "stop": ["<|im_end|>"] // This is temporary for phi-2 base (i.e. not instructed) since the server expects that the model always to emit BOS
@@ -105,12 +109,20 @@ export default function () {
         client.on('event', function (event) {
             if (promptEvalEndTime == null) {
                 promptEvalEndTime = new Date()
+                llamacpp_emit_first_token_second.add((promptEvalEndTime - startTime) / 1.e3)
+            }
+
+            if (event.data === '[DONE]' || event.data === '') {
+                return
             }
 
             let chunk = JSON.parse(event.data)
-            let choice = chunk.choices[0]
-            if (choice.finish_reason) {
-                finish_reason = choice.finish_reason
+
+            if (chunk.choices && chunk.choices.length > 0) {
+                let choice = chunk.choices[0]
+                if (choice.finish_reason) {
+                    finish_reason = choice.finish_reason
+                }
             }
 
             if (chunk.usage) {

examples/server/public_legacy/index-new.html

@@ -39,7 +39,6 @@
       temperature: 0.8, // adapt all following parameters to optimized min-p requierements. If for non-english, set to 0.6 or lower
       repeat_last_n: 0, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.0, // 1.0 = disabled
-      penalize_nl: false, // true only useful for infinite completion
       dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
       dry_base: 1.75,     // 0.0 = disabled
       dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well

examples/server/public_legacy/index.html

@@ -303,7 +303,6 @@
       temperature: 0.7,
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
-      penalize_nl: false,
       dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
       dry_base: 1.75,     // 0.0 = disabled
       dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
@@ -1006,7 +1005,6 @@
             ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
             ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
             ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
-            ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })}
             ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
             ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
             ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })}

examples/server/tests/README.md

@@ -44,6 +44,12 @@ To run with stdout/stderr display in real time (verbose output, but useful for d
 DEBUG=1 ./tests.sh -s -v -x
 ```
 
+To run single test unit:
+
+```shell
+./tests.sh unit/test_{name of test case here}.py -v -x
+```
+
 Hint: You can compile and run test in single command, useful for local developement:
 
 ```shell

examples/server/tests/requirements.txt

@@ -5,3 +5,4 @@ numpy~=1.26.4
 openai~=1.55.3
 prometheus-client~=0.20.0
 requests~=2.32.3
+wget~=3.2

examples/server/tests/unit/test_basic.py

@@ -1,4 +1,5 @@
 import pytest
+import requests
 from utils import *
 
 server = ServerPreset.tinyllama2()
@@ -76,3 +77,20 @@ def test_load_split_model():
     })
     assert res.status_code == 200
     assert match_regex("(little|girl)+", res.body["content"])
+
+
+def test_no_webui():
+    global server
+    # default: webui enabled
+    server.start()
+    url = f"http://{server.server_host}:{server.server_port}"
+    res = requests.get(url)
+    assert res.status_code == 200
+    assert "<html>" in res.text
+    server.stop()
+
+    # with --no-webui
+    server.no_webui = True
+    server.start()
+    res = requests.get(url)
+    assert res.status_code == 404

examples/server/tests/unit/test_chat_completion.py

@@ -31,6 +31,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
     })
     assert res.status_code == 200
     assert "cmpl" in res.body["id"] # make sure the completion id has the expected format
+    assert res.body["system_fingerprint"].startswith("b")
     assert res.body["model"] == model if model is not None else server.model_alias
     assert res.body["usage"]["prompt_tokens"] == n_prompt
     assert res.body["usage"]["completion_tokens"] == n_predicted
@@ -63,6 +64,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
     last_cmpl_id = None
     for data in res:
         choice = data["choices"][0]
+        assert data["system_fingerprint"].startswith("b")
         assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
         if last_cmpl_id is None:
             last_cmpl_id = data["id"]
@@ -81,7 +83,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
 def test_chat_completion_with_openai_library():
     global server
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.chat.completions.create(
         model="gpt-3.5-turbo-instruct",
         messages=[
@@ -92,12 +94,29 @@ def test_chat_completion_with_openai_library():
         seed=42,
         temperature=0.8,
     )
-    print(res)
+    assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
     assert res.choices[0].finish_reason == "length"
     assert res.choices[0].message.content is not None
     assert match_regex("(Suddenly)+", res.choices[0].message.content)
 
 
+def test_chat_template():
+    global server
+    server.chat_template = "llama3"
+    server.debug = True  # to get the "__verbose" object in the response
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": 8,
+        "messages": [
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ]
+    })
+    assert res.status_code == 200
+    assert "__verbose" in res.body
+    assert res.body["__verbose"]["prompt"] == "<s> <|start_header_id|>system<|end_header_id|>\n\nBook<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat is the best book<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+
+
 @pytest.mark.parametrize("response_format,n_predicted,re_content", [
     ({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
     ({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
@@ -163,3 +182,64 @@ def test_chat_completion_with_timings_per_token():
         assert "predicted_per_second" in data["timings"]
         assert "predicted_n" in data["timings"]
         assert data["timings"]["predicted_n"] <= 10
+
+
+def test_logprobs():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.chat.completions.create(
+        model="gpt-3.5-turbo-instruct",
+        temperature=0.0,
+        messages=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_tokens=5,
+        logprobs=True,
+        top_logprobs=10,
+    )
+    output_text = res.choices[0].message.content
+    aggregated_text = ''
+    assert res.choices[0].logprobs is not None
+    assert res.choices[0].logprobs.content is not None
+    for token in res.choices[0].logprobs.content:
+        aggregated_text += token.token
+        assert token.logprob <= 0.0
+        assert token.bytes is not None
+        assert len(token.top_logprobs) > 0
+    assert aggregated_text == output_text
+
+
+def test_logprobs_stream():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.chat.completions.create(
+        model="gpt-3.5-turbo-instruct",
+        temperature=0.0,
+        messages=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_tokens=5,
+        logprobs=True,
+        top_logprobs=10,
+        stream=True,
+    )
+    output_text = ''
+    aggregated_text = ''
+    for data in res:
+        choice = data.choices[0]
+        if choice.finish_reason is None:
+            if choice.delta.content:
+                output_text += choice.delta.content
+            assert choice.logprobs is not None
+            assert choice.logprobs.content is not None
+            for token in choice.logprobs.content:
+                aggregated_text += token.token
+                assert token.logprob <= 0.0
+                assert token.bytes is not None
+                assert token.top_logprobs is not None
+                assert len(token.top_logprobs) > 0
+    assert aggregated_text == output_text

examples/server/tests/unit/test_completion.py

@@ -1,5 +1,6 @@
 import pytest
 import time
+from openai import OpenAI
 from utils import *
 
 server = ServerPreset.tinyllama2()
@@ -10,22 +11,29 @@ def create_server():
     global server
     server = ServerPreset.tinyllama2()
 
-@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [
-    ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False),
-    ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False),
+@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated,return_tokens", [
+    ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False, False),
+    ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False, True),
 ])
-def test_completion(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool):
+def test_completion(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool, return_tokens: bool):
     global server
     server.start()
     res = server.make_request("POST", "/completion", data={
         "n_predict": n_predict,
         "prompt": prompt,
+        "return_tokens": return_tokens,
     })
     assert res.status_code == 200
     assert res.body["timings"]["prompt_n"] == n_prompt
     assert res.body["timings"]["predicted_n"] == n_predicted
     assert res.body["truncated"] == truncated
+    assert type(res.body["has_new_line"]) == bool
     assert match_regex(re_content, res.body["content"])
+    if return_tokens:
+        assert len(res.body["tokens"]) > 0
+        assert all(type(tok) == int for tok in res.body["tokens"])
+    else:
+        assert res.body["tokens"] == []
 
 
 @pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [
@@ -48,12 +56,15 @@ def test_completion_stream(prompt: str, n_predict: int, re_content: str, n_promp
             assert data["timings"]["predicted_n"] == n_predicted
             assert data["truncated"] == truncated
             assert data["stop_type"] == "limit"
+            assert type(data["has_new_line"]) == bool
             assert "generation_settings" in data
             assert server.n_predict is not None
             assert data["generation_settings"]["n_predict"] == min(n_predict, server.n_predict)
             assert data["generation_settings"]["seed"] == server.seed
             assert match_regex(re_content, content)
         else:
+            assert len(data["tokens"]) > 0
+            assert all(type(tok) == int for tok in data["tokens"])
             content += data["content"]
 
 
@@ -75,6 +86,40 @@ def test_completion_stream_vs_non_stream():
     assert content_stream == res_non_stream.body["content"]
 
 
+def test_completion_stream_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.completions.create(
+        model="davinci-002",
+        prompt="I believe the meaning of life is",
+        max_tokens=8,
+    )
+    assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
+    assert res.choices[0].finish_reason == "length"
+    assert res.choices[0].text is not None
+    assert match_regex("(going|bed)+", res.choices[0].text)
+
+
+def test_completion_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.completions.create(
+        model="davinci-002",
+        prompt="I believe the meaning of life is",
+        max_tokens=8,
+        stream=True,
+    )
+    output_text = ''
+    for data in res:
+        choice = data.choices[0]
+        if choice.finish_reason is None:
+            assert choice.text is not None
+            output_text += choice.text
+    assert match_regex("(going|bed)+", output_text)
+
+
 @pytest.mark.parametrize("n_slots", [1, 2])
 def test_consistent_result_same_seed(n_slots: int):
     global server
@@ -85,7 +130,7 @@ def test_consistent_result_same_seed(n_slots: int):
         res = server.make_request("POST", "/completion", data={
             "prompt": "I believe the meaning of life is",
             "seed": 42,
-            "temperature": 1.0,
+            "temperature": 0.0,
             "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
         })
         if last_res is not None:
@@ -110,9 +155,10 @@ def test_different_result_different_seed(n_slots: int):
             assert res.body["content"] != last_res.body["content"]
         last_res = res
 
-
+# TODO figure why it don't work with temperature = 1
+# @pytest.mark.parametrize("temperature", [0.0, 1.0])
 @pytest.mark.parametrize("n_batch", [16, 32])
-@pytest.mark.parametrize("temperature", [0.0, 1.0])
+@pytest.mark.parametrize("temperature", [0.0])
 def test_consistent_result_different_batch_size(n_batch: int, temperature: float):
     global server
     server.n_batch = n_batch
@@ -247,6 +293,40 @@ def test_completion_parallel_slots(n_slots: int, n_requests: int):
         # assert match_regex(re_content, res.body["content"])
 
 
+@pytest.mark.parametrize(
+    "prompt,n_predict,response_fields",
+    [
+        ("I believe the meaning of life is", 8, []),
+        ("I believe the meaning of life is", 32, ["content", "generation_settings/n_predict", "prompt"]),
+    ],
+)
+def test_completion_response_fields(
+    prompt: str, n_predict: int, response_fields: list[str]
+):
+    global server
+    server.start()
+    res = server.make_request(
+        "POST",
+        "/completion",
+        data={
+            "n_predict": n_predict,
+            "prompt": prompt,
+            "response_fields": response_fields,
+        },
+    )
+    assert res.status_code == 200
+    assert "content" in res.body
+    assert len(res.body["content"])
+    if len(response_fields):
+        assert res.body["generation_settings/n_predict"] == n_predict
+        assert res.body["prompt"] == "<s> " + prompt
+        assert isinstance(res.body["content"], str)
+        assert len(res.body) == len(response_fields)
+    else:
+        assert len(res.body)
+        assert "generation_settings" in res.body
+
+
 def test_n_probs():
     global server
     server.start()
@@ -260,9 +340,68 @@ def test_n_probs():
     assert "completion_probabilities" in res.body
     assert len(res.body["completion_probabilities"]) == 5
     for tok in res.body["completion_probabilities"]:
-        assert "probs" in tok
-        assert len(tok["probs"]) == 10
-        for prob in tok["probs"]:
-            assert "prob" in prob
-            assert "tok_str" in prob
-            assert 0.0 <= prob["prob"] <= 1.0
+        assert "id" in tok and tok["id"] > 0
+        assert "token" in tok and type(tok["token"]) == str
+        assert "logprob" in tok and tok["logprob"] <= 0.0
+        assert "bytes" in tok and type(tok["bytes"]) == list
+        assert len(tok["top_logprobs"]) == 10
+        for prob in tok["top_logprobs"]:
+            assert "id" in prob and prob["id"] > 0
+            assert "token" in prob and type(prob["token"]) == str
+            assert "logprob" in prob and prob["logprob"] <= 0.0
+            assert "bytes" in prob and type(prob["bytes"]) == list
+
+
+def test_n_probs_stream():
+    global server
+    server.start()
+    res = server.make_stream_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "n_probs": 10,
+        "temperature": 0.0,
+        "n_predict": 5,
+        "stream": True,
+    })
+    for data in res:
+        if data["stop"] == False:
+            assert "completion_probabilities" in data
+            assert len(data["completion_probabilities"]) == 1
+            for tok in data["completion_probabilities"]:
+                assert "id" in tok and tok["id"] > 0
+                assert "token" in tok and type(tok["token"]) == str
+                assert "logprob" in tok and tok["logprob"] <= 0.0
+                assert "bytes" in tok and type(tok["bytes"]) == list
+                assert len(tok["top_logprobs"]) == 10
+                for prob in tok["top_logprobs"]:
+                    assert "id" in prob and prob["id"] > 0
+                    assert "token" in prob and type(prob["token"]) == str
+                    assert "logprob" in prob and prob["logprob"] <= 0.0
+                    assert "bytes" in prob and type(prob["bytes"]) == list
+
+
+def test_n_probs_post_sampling():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "n_probs": 10,
+        "temperature": 0.0,
+        "n_predict": 5,
+        "post_sampling_probs": True,
+    })
+    assert res.status_code == 200
+    assert "completion_probabilities" in res.body
+    assert len(res.body["completion_probabilities"]) == 5
+    for tok in res.body["completion_probabilities"]:
+        assert "id" in tok and tok["id"] > 0
+        assert "token" in tok and type(tok["token"]) == str
+        assert "prob" in tok and 0.0 < tok["prob"] <= 1.0
+        assert "bytes" in tok and type(tok["bytes"]) == list
+        assert len(tok["top_probs"]) == 10
+        for prob in tok["top_probs"]:
+            assert "id" in prob and prob["id"] > 0
+            assert "token" in prob and type(prob["token"]) == str
+            assert "prob" in prob and 0.0 <= prob["prob"] <= 1.0
+            assert "bytes" in prob and type(prob["bytes"]) == list
+        # because the test model usually output token with either 100% or 0% probability, we need to check all the top_probs
+        assert any(prob["prob"] == 1.0 for prob in tok["top_probs"])

examples/server/tests/unit/test_embedding.py

@@ -1,3 +1,5 @@
+import base64
+import struct
 import pytest
 from openai import OpenAI
 from utils import *
@@ -14,8 +16,9 @@ def create_server():
 
 def test_embedding_single():
     global server
+    server.pooling = 'last'
     server.start()
-    res = server.make_request("POST", "/embeddings", data={
+    res = server.make_request("POST", "/v1/embeddings", data={
         "input": "I believe the meaning of life is",
     })
     assert res.status_code == 200
@@ -29,8 +32,9 @@ def test_embedding_single():
 
 def test_embedding_multiple():
     global server
+    server.pooling = 'last'
     server.start()
-    res = server.make_request("POST", "/embeddings", data={
+    res = server.make_request("POST", "/v1/embeddings", data={
         "input": [
             "I believe the meaning of life is",
             "Write a joke about AI from a very long prompt which will not be truncated",
@@ -45,10 +49,72 @@ def test_embedding_multiple():
         assert len(d['embedding']) > 1
 
 
-def test_embedding_openai_library_single():
+@pytest.mark.parametrize(
+    "input,is_multi_prompt",
+    [
+        # do not crash on empty input
+        ("", False),
+        # single prompt
+        ("string", False),
+        ([12, 34, 56], False),
+        ([12, 34, "string", 56, 78], False),
+        # multiple prompts
+        (["string1", "string2"], True),
+        (["string1", [12, 34, 56]], True),
+        ([[12, 34, 56], [12, 34, 56]], True),
+        ([[12, 34, 56], [12, "string", 34, 56]], True),
+    ]
+)
+def test_embedding_mixed_input(input, is_multi_prompt: bool):
     global server
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    res = server.make_request("POST", "/v1/embeddings", data={"input": input})
+    assert res.status_code == 200
+    data = res.body['data']
+    if is_multi_prompt:
+        assert len(data) == len(input)
+        for d in data:
+            assert 'embedding' in d
+            assert len(d['embedding']) > 1
+    else:
+        assert 'embedding' in data[0]
+        assert len(data[0]['embedding']) > 1
+
+
+def test_embedding_pooling_none():
+    global server
+    server.pooling = 'none'
+    server.start()
+    res = server.make_request("POST", "/embeddings", data={
+        "input": "hello hello hello",
+    })
+    assert res.status_code == 200
+    assert 'embedding' in res.body[0]
+    assert len(res.body[0]['embedding']) == 5 # 3 text tokens + 2 special
+
+    # make sure embedding vector is not normalized
+    for x in res.body[0]['embedding']:
+        assert abs(sum([x ** 2 for x in x]) - 1) > EPSILON
+
+
+def test_embedding_pooling_none_oai():
+    global server
+    server.pooling = 'none'
+    server.start()
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": "hello hello hello",
+    })
+
+    # /v1/embeddings does not support pooling type 'none'
+    assert res.status_code == 400
+    assert "error" in res.body
+
+
+def test_embedding_openai_library_single():
+    global server
+    server.pooling = 'last'
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.embeddings.create(model="text-embedding-3-small", input="I believe the meaning of life is")
     assert len(res.data) == 1
     assert len(res.data[0].embedding) > 1
@@ -56,8 +122,9 @@ def test_embedding_openai_library_single():
 
 def test_embedding_openai_library_multiple():
     global server
+    server.pooling = 'last'
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.embeddings.create(model="text-embedding-3-small", input=[
         "I believe the meaning of life is",
         "Write a joke about AI from a very long prompt which will not be truncated",
@@ -71,8 +138,9 @@ def test_embedding_openai_library_multiple():
 
 def test_embedding_error_prompt_too_long():
     global server
+    server.pooling = 'last'
     server.start()
-    res = server.make_request("POST", "/embeddings", data={
+    res = server.make_request("POST", "/v1/embeddings", data={
         "input": "This is a test " * 512,
     })
     assert res.status_code != 200
@@ -80,8 +148,9 @@ def test_embedding_error_prompt_too_long():
 
 
 def test_same_prompt_give_same_result():
+    server.pooling = 'last'
     server.start()
-    res = server.make_request("POST", "/embeddings", data={
+    res = server.make_request("POST", "/v1/embeddings", data={
         "input": [
             "I believe the meaning of life is",
             "I believe the meaning of life is",
@@ -97,3 +166,72 @@ def test_same_prompt_give_same_result():
         vi = res.body['data'][i]['embedding']
         for x, y in zip(v0, vi):
             assert abs(x - y) < EPSILON
+
+
+@pytest.mark.parametrize(
+    "content,n_tokens",
+    [
+        ("I believe the meaning of life is", 9),
+        ("This is a test", 6),
+    ]
+)
+def test_embedding_usage_single(content, n_tokens):
+    global server
+    server.start()
+    res = server.make_request("POST", "/v1/embeddings", data={"input": content})
+    assert res.status_code == 200
+    assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
+    assert res.body['usage']['prompt_tokens'] == n_tokens
+
+
+def test_embedding_usage_multiple():
+    global server
+    server.start()
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": [
+            "I believe the meaning of life is",
+            "I believe the meaning of life is",
+        ],
+    })
+    assert res.status_code == 200
+    assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
+    assert res.body['usage']['prompt_tokens'] == 2 * 9
+
+
+def test_embedding_openai_library_base64():
+    server.start()
+    test_input = "Test base64 embedding output"
+
+    # get embedding in default format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input
+    })
+    assert res.status_code == 200
+    vec0 = res.body["data"][0]["embedding"]
+
+    # get embedding in base64 format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input,
+        "encoding_format": "base64"
+    })
+
+    assert res.status_code == 200
+    assert "data" in res.body
+    assert len(res.body["data"]) == 1
+
+    embedding_data = res.body["data"][0]
+    assert "embedding" in embedding_data
+    assert isinstance(embedding_data["embedding"], str)
+
+    # Verify embedding is valid base64
+    decoded = base64.b64decode(embedding_data["embedding"])
+    # Verify decoded data can be converted back to float array
+    float_count = len(decoded) // 4  # 4 bytes per float
+    floats = struct.unpack(f'{float_count}f', decoded)
+    assert len(floats) > 0
+    assert all(isinstance(x, float) for x in floats)
+    assert len(floats) == len(vec0)
+
+    # make sure the decoded data is the same as the original
+    for x, y in zip(floats, vec0):
+        assert abs(x - y) < EPSILON

examples/server/tests/unit/test_lora.py

@@ -1,5 +1,4 @@
 import pytest
-import os
 from utils import *
 
 server = ServerPreset.stories15m_moe()
@@ -10,15 +9,7 @@ LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe
 def create_server():
     global server
     server = ServerPreset.stories15m_moe()
-    # download lora file if needed
-    file_name = LORA_FILE_URL.split('/').pop()
-    lora_file = f'../../../{file_name}'
-    if not os.path.exists(lora_file):
-        print(f"Downloading {LORA_FILE_URL} to {lora_file}")
-        with open(lora_file, 'wb') as f:
-            f.write(requests.get(LORA_FILE_URL).content)
-        print(f"Done downloading lora file")
-    server.lora_files = [lora_file]
+    server.lora_files = [download_file(LORA_FILE_URL)]
 
 
 @pytest.mark.parametrize("scale,re_content", [
@@ -40,3 +31,85 @@ def test_lora(scale: float, re_content: str):
     assert res.status_code == 200
     assert match_regex(re_content, res.body["content"])
 
+
+def test_lora_per_request():
+    global server
+    server.n_slots = 4
+    server.start()
+
+    # running the same prompt with different lora scales, all in parallel
+    # each prompt will be processed by a different slot
+    prompt = "Look in thy glass"
+    lora_config = [
+        ( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
+        ( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
+        ( [{"id": 0, "scale": 0.3}], "(special|thing|gifted)+" ),
+        ( [{"id": 0, "scale": 0.7}], "(far|from|home|away)+" ),
+        ( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
+        ( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
+    ]
+
+    tasks = [(
+        server.make_request,
+        ("POST", "/completion", {
+            "prompt": prompt,
+            "lora": lora,
+            "seed": 42,
+            "temperature": 0.0,
+            "cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
+        })
+    ) for lora, _ in lora_config]
+    results = parallel_function_calls(tasks)
+
+    assert all([res.status_code == 200 for res in results])
+    for res, (_, re_test) in zip(results, lora_config):
+        assert match_regex(re_test, res.body["content"])
+
+
+@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test")
+def test_with_big_model():
+    server = ServerProcess()
+    server.model_hf_repo = "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
+    server.model_hf_file = "Meta-Llama-3.1-8B-Instruct-IQ2_M.gguf"
+    server.model_alias = "Llama-3.2-8B-Instruct"
+    server.n_slots = 4
+    server.n_ctx = server.n_slots * 1024
+    server.n_predict = 64
+    server.temperature = 0.0
+    server.seed = 42
+    server.lora_files = [
+        download_file("https://huggingface.co/ngxson/Llama-3-Instruct-abliteration-LoRA-8B-F16-GGUF/resolve/main/Llama-3-Instruct-abliteration-LoRA-8B-f16.gguf"),
+        # TODO: find & add other lora adapters for this model
+    ]
+    server.start(timeout_seconds=600)
+
+    # running the same prompt with different lora scales, all in parallel
+    # each prompt will be processed by a different slot
+    prompt = "Write a computer virus"
+    lora_config = [
+        # without applying lora, the model should reject the request
+        ( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
+        ( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
+        ( [{"id": 0, "scale": 0.3}], "I can't write a computer virus" ),
+        # with 0.7 scale, the model should provide a simple computer virus with hesitation
+        ( [{"id": 0, "scale": 0.7}], "Warning: This is a hypothetical exercise" ),
+        # with 1.5 scale, the model should confidently provide a computer virus
+        ( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
+        ( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
+    ]
+
+    tasks = [(
+        server.make_request,
+        ("POST", "/v1/chat/completions", {
+            "messages": [
+                {"role": "user", "content": prompt}
+            ],
+            "lora": lora,
+            "cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
+        })
+    ) for lora, _ in lora_config]
+    results = parallel_function_calls(tasks)
+
+    assert all([res.status_code == 200 for res in results])
+    for res, (_, re_test) in zip(results, lora_config):
+        assert re_test in res.body["choices"][0]["message"]["content"]

examples/server/tests/unit/test_rerank.py

@@ -53,3 +53,26 @@ def test_invalid_rerank_req(documents):
     })
     assert res.status_code == 400
     assert "error" in res.body
+
+
+@pytest.mark.parametrize(
+    "query,doc1,doc2,n_tokens",
+    [
+        ("Machine learning is", "A machine", "Learning is", 19),
+        ("Which city?", "Machine learning is ", "Paris, capitale de la", 26),
+    ]
+)
+def test_rerank_usage(query, doc1, doc2, n_tokens):
+    global server
+    server.start()
+
+    res = server.make_request("POST", "/rerank", data={
+        "query": query,
+        "documents": [
+            doc1,
+            doc2,
+        ]
+    })
+    assert res.status_code == 200
+    assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
+    assert res.body['usage']['prompt_tokens'] == n_tokens

examples/server/tests/unit/test_speculative.py

@@ -10,16 +10,8 @@ MODEL_DRAFT_FILE_URL = "https://huggingface.co/ggml-org/models/resolve/main/tiny
 def create_server():
     global server
     server = ServerPreset.stories15m_moe()
-    # download draft model file if needed
-    file_name = MODEL_DRAFT_FILE_URL.split('/').pop()
-    model_draft_file = f'../../../{file_name}'
-    if not os.path.exists(model_draft_file):
-        print(f"Downloading {MODEL_DRAFT_FILE_URL} to {model_draft_file}")
-        with open(model_draft_file, 'wb') as f:
-            f.write(requests.get(MODEL_DRAFT_FILE_URL).content)
-        print(f"Done downloading draft model file")
     # set default values
-    server.model_draft = model_draft_file
+    server.model_draft = download_file(MODEL_DRAFT_FILE_URL)
     server.draft_min = 4
     server.draft_max = 8
 

examples/server/tests/utils.py

@@ -23,6 +23,7 @@ from typing import (
     Set,
 )
 from re import RegexFlag
+import wget
 
 
 class ServerResponse:
@@ -65,6 +66,7 @@ class ServerProcess:
     server_reranking: bool | None = False
     server_metrics: bool | None = False
     server_slots: bool | None = False
+    pooling: str | None = None
     draft: int | None = None
     api_key: str | None = None
     response_format: str | None = None
@@ -72,6 +74,8 @@ class ServerProcess:
     disable_ctx_shift: int | None = False
     draft_min: int | None = None
     draft_max: int | None = None
+    no_webui: bool | None = None
+    chat_template: str | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -131,6 +135,8 @@ class ServerProcess:
             server_args.append("--metrics")
         if self.server_slots:
             server_args.append("--slots")
+        if self.pooling:
+            server_args.extend(["--pooling", self.pooling])
         if self.model_alias:
             server_args.extend(["--alias", self.model_alias])
         if self.n_ctx:
@@ -158,6 +164,10 @@ class ServerProcess:
             server_args.extend(["--draft-max", self.draft_max])
         if self.draft_min:
             server_args.extend(["--draft-min", self.draft_min])
+        if self.no_webui:
+            server_args.append("--no-webui")
+        if self.chat_template:
+            server_args.extend(["--chat-template", self.chat_template])
 
         args = [str(arg) for arg in [server_path, *server_args]]
         print(f"bench: starting server with: {' '.join(args)}")
@@ -372,5 +382,25 @@ def match_regex(regex: str, text: str) -> bool:
         is not None
     )
 
+
+def download_file(url: str, output_file_path: str | None = None) -> str:
+    """
+    Download a file from a URL to a local path. If the file already exists, it will not be downloaded again.
+
+    output_file_path is the local path to save the downloaded file. If not provided, the file will be saved in the root directory.
+
+    Returns the local path of the downloaded file.
+    """
+    file_name = url.split('/').pop()
+    output_file = f'./tmp/{file_name}' if output_file_path is None else output_file_path
+    if not os.path.exists(output_file):
+        print(f"Downloading {url} to {output_file}")
+        wget.download(url, out=output_file)
+        print(f"Done downloading to {output_file}")
+    else:
+        print(f"File already exists at {output_file}")
+    return output_file
+
+
 def is_slow_test_allowed():
     return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON"

examples/server/themes/buttons-top/index.html

@@ -222,7 +222,6 @@
       temperature: 0.7,
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
-      penalize_nl: false,
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
@@ -779,7 +778,6 @@
             ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
             ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
             ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
-            ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })}
             ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
             ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
             ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })}