Merge remote-tracking branch 'origin/master' into tool-call

This commit is contained in:
ochafik 2024-12-26 21:26:21 +00:00
commit e70ce3f613
127 changed files with 9301 additions and 3310 deletions

81
.devops/cpu.Dockerfile Normal file
View file

@ -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" ]

94
.devops/cuda.Dockerfile Normal file
View file

@ -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" ]

View file

@ -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"]

View file

@ -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"]

View file

@ -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"]

View file

@ -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"]

91
.devops/intel.Dockerfile Normal file
View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

View file

@ -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" ]

108
.devops/musa.Dockerfile Normal file
View file

@ -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" ]

View file

@ -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")

113
.devops/rocm.Dockerfile Normal file
View file

@ -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" ]

88
.devops/vulkan.Dockerfile Normal file
View file

@ -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" ]

View file

@ -317,7 +317,7 @@ 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 +327,12 @@ 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

View file

@ -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

View file

@ -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
@ -220,7 +221,7 @@ 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 |
@ -413,7 +414,7 @@ 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.
@ -447,7 +448,7 @@ To learn more about model quantization, [read this documentation](examples/quant
</details>
[^3]: [https://github.com/containers/ramalama](RamaLama)
[^3]: [RamaLama](https://github.com/containers/ramalama)
## [`llama-simple`](examples/simple)

View file

@ -119,29 +119,33 @@ 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;
}
}
@ -276,7 +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);
@ -620,7 +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, LLAMA_EXAMPLE_IMATRIX}).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),
@ -842,7 +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);
@ -855,13 +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),
@ -916,6 +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);
}
@ -970,6 +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());
@ -1582,6 +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)",
@ -2214,5 +2233,25 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
add_opt(common_arg(
{"-mv", "--model-vocoder"}, "FNAME",
"vocoder model for audio generation (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.model = value;
}
).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
// 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;
}

View file

@ -979,6 +979,25 @@ 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__);
@ -1115,7 +1134,7 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
#define CURL_MAX_RETRY 3
#define CURL_RETRY_DELAY_SECONDS 2
static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
int remaining_attempts = max_attempts;
while (remaining_attempts > 0) {
@ -1139,7 +1158,6 @@ 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) {
@ -1212,11 +1230,13 @@ 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);
@ -1858,7 +1878,9 @@ 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;

View file

@ -86,6 +86,7 @@ enum llama_example {
LLAMA_EXAMPLE_LLAVA,
LLAMA_EXAMPLE_LOOKUP,
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_TTS,
LLAMA_EXAMPLE_COUNT,
};
@ -101,6 +102,7 @@ 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
@ -136,7 +138,6 @@ 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;
@ -145,6 +146,7 @@ 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,
@ -165,6 +167,7 @@ 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
@ -178,6 +181,14 @@ 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
@ -200,11 +211,13 @@ 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;
@ -218,8 +231,9 @@ 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
@ -608,7 +622,8 @@ 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);

View file

@ -170,32 +170,20 @@ 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;
@ -217,6 +205,9 @@ 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");
}
@ -428,6 +419,7 @@ 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 '?';
}
}
@ -442,6 +434,7 @@ 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 "";
}
}
@ -456,6 +449,7 @@ 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
@ -502,6 +496,7 @@ 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;

View file

@ -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 +1695,184 @@ 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_token_types = ['bos', 'eos', 'eom', 'eot']
)
special_vocab._set_special_token("bos", 128000)
special_vocab._set_special_token("eos", 128001)
special_vocab._set_special_token("eom", 128008)
special_vocab._set_special_token("eot", 128009)
special_vocab.add_to_gguf(self.gguf_writer)
else:
# DeciLM-7B
self._set_vocab_llama_hf()
# self._set_vocab_gpt2()
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)
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
@ -2024,6 +2228,44 @@ class Qwen2VLModel(Model):
yield name, data
@Model.register("WavTokenizerDec")
class WavTokenizerDecModel(Model):
model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
if \
name.endswith("codebook.cluster_size") or \
name.endswith("codebook.embed_avg") or \
name.endswith("codebook.inited"):
logger.debug(f"Skipping {name!r}")
return []
logger.info(f"{self.map_tensor_name(name)} -> {data_torch.shape}")
return [(self.map_tensor_name(name), data_torch)]
def set_vocab(self):
self._set_vocab_none()
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_vocab_size (self.hparams["vocab_size"])
self.gguf_writer.add_features_length (self.hparams["n_embd_features"])
self.gguf_writer.add_feed_forward_length(self.hparams["n_ff"])
self.gguf_writer.add_group_norm_eps (self.hparams["group_norm_epsilon"])
self.gguf_writer.add_group_norm_groups (self.hparams["group_norm_groups"])
self.gguf_writer.add_posnet_embedding_length(self.hparams["posnet"]["n_embd"])
self.gguf_writer.add_posnet_block_count (self.hparams["posnet"]["n_layer"])
self.gguf_writer.add_convnext_embedding_length(self.hparams["convnext"]["n_embd"])
self.gguf_writer.add_convnext_block_count (self.hparams["convnext"]["n_layer"])
self.gguf_writer.add_causal_attention(False)
@Model.register("Qwen2MoeForCausalLM")
class Qwen2MoeModel(Model):
model_arch = gguf.MODEL_ARCH.QWEN2MOE
@ -2152,6 +2394,15 @@ 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'
@ -2268,7 +2519,11 @@ 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"])
@ -2567,7 +2822,7 @@ 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
@ -2633,13 +2888,73 @@ 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
@ -2959,6 +3274,9 @@ 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"):
@ -3427,6 +3745,97 @@ 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

View file

@ -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"},
]

View file

@ -51,6 +51,7 @@ 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

View file

@ -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);

View file

@ -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];

View file

@ -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;

View file

@ -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);

View file

@ -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;

View file

@ -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

View file

@ -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()

View file

@ -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"
@ -896,7 +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]
@ -944,7 +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
@ -1005,7 +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));
@ -1222,30 +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();

View file

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

View file

@ -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)).

View file

@ -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);
}
}

View file

@ -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;

View file

@ -4,7 +4,7 @@ The purpose of this example is to demonstrate a minimal usage of llama.cpp for r
```bash
llama-run granite-code
...
```
```bash
llama-run -h
@ -19,6 +19,10 @@ Options:
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
@ -42,6 +46,6 @@ Examples:
llama-run https://example.com/some-file1.gguf
llama-run some-file2.gguf
llama-run file://some-file3.gguf
llama-run --ngl 99 some-file4.gguf
llama-run --ngl 99 some-file5.gguf Hello World
...
llama-run --ngl 999 some-file4.gguf
llama-run --ngl 999 some-file5.gguf Hello World
```

View file

@ -1,6 +1,8 @@
#if defined(_WIN32)
# include <windows.h>
#else
# include <sys/file.h>
# include <sys/ioctl.h>
# include <unistd.h>
#endif
@ -8,6 +10,7 @@
# include <curl/curl.h>
#endif
#include <climits>
#include <cstdarg>
#include <cstdio>
#include <cstring>
@ -21,41 +24,150 @@
#include "json.hpp"
#include "llama-cpp.h"
#define printe(...) \
do { \
fprintf(stderr, __VA_ARGS__); \
} while (0)
GGML_ATTRIBUTE_FORMAT(1, 2)
static std::string fmt(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
const int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::string buf;
buf.resize(size);
const int size2 = vsnprintf(const_cast<char *>(buf.data()), buf.size() + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return buf;
}
GGML_ATTRIBUTE_FORMAT(1, 2)
static int printe(const char * fmt, ...) {
va_list args;
va_start(args, fmt);
const int ret = vfprintf(stderr, fmt, args);
va_end(args);
return ret;
}
class Opt {
public:
int init(int argc, const char ** argv) {
construct_help_str_();
ctx_params = llama_context_default_params();
model_params = llama_model_default_params();
context_size_default = ctx_params.n_batch;
ngl_default = model_params.n_gpu_layers;
common_params_sampling sampling;
temperature_default = sampling.temp;
if (argc < 2) {
printe("Error: No arguments provided.\n");
print_help();
return 1;
}
// Parse arguments
if (parse(argc, argv)) {
printe("Error: Failed to parse arguments.\n");
help();
print_help();
return 1;
}
// If help is requested, show help and exit
if (help_) {
help();
if (help) {
print_help();
return 2;
}
ctx_params.n_batch = context_size >= 0 ? context_size : context_size_default;
model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default;
temperature = temperature >= 0 ? temperature : temperature_default;
return 0; // Success
}
llama_context_params ctx_params;
llama_model_params model_params;
std::string model_;
std::string user_;
int context_size_ = 2048, ngl_ = -1;
std::string user;
int context_size = -1, ngl = -1;
float temperature = -1;
bool verbose = false;
private:
std::string help_str_;
bool help_ = false;
int context_size_default = -1, ngl_default = -1;
float temperature_default = -1;
bool help = false;
void construct_help_str_() {
help_str_ =
bool parse_flag(const char ** argv, int i, const char * short_opt, const char * long_opt) {
return strcmp(argv[i], short_opt) == 0 || strcmp(argv[i], long_opt) == 0;
}
int handle_option_with_value(int argc, const char ** argv, int & i, int & option_value) {
if (i + 1 >= argc) {
return 1;
}
option_value = std::atoi(argv[++i]);
return 0;
}
int handle_option_with_value(int argc, const char ** argv, int & i, float & option_value) {
if (i + 1 >= argc) {
return 1;
}
option_value = std::atof(argv[++i]);
return 0;
}
int parse(int argc, const char ** argv) {
bool options_parsing = true;
for (int i = 1, positional_args_i = 0; i < argc; ++i) {
if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
if (handle_option_with_value(argc, argv, i, context_size) == 1) {
return 1;
}
} else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) {
if (handle_option_with_value(argc, argv, i, ngl) == 1) {
return 1;
}
} else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
if (handle_option_with_value(argc, argv, i, temperature) == 1) {
return 1;
}
} else if (options_parsing &&
(parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
verbose = true;
} else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
help = true;
return 0;
} else if (options_parsing && strcmp(argv[i], "--") == 0) {
options_parsing = false;
} else if (positional_args_i == 0) {
if (!argv[i][0] || argv[i][0] == '-') {
return 1;
}
++positional_args_i;
model_ = argv[i];
} else if (positional_args_i == 1) {
++positional_args_i;
user = argv[i];
} else {
user += " " + std::string(argv[i]);
}
}
return 0;
}
void print_help() const {
printf(
"Description:\n"
" Runs a llm\n"
"\n"
@ -64,15 +176,13 @@ class Opt {
"\n"
"Options:\n"
" -c, --context-size <value>\n"
" Context size (default: " +
std::to_string(context_size_);
help_str_ +=
")\n"
" Context size (default: %d)\n"
" -n, --ngl <value>\n"
" Number of GPU layers (default: " +
std::to_string(ngl_);
help_str_ +=
")\n"
" Number of GPU layers (default: %d)\n"
" --temp <value>\n"
" Temperature (default: %.1f)\n"
" -v, --verbose, --log-verbose\n"
" Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n"
" -h, --help\n"
" Show help message\n"
"\n"
@ -92,67 +202,102 @@ class Opt {
" llama-run ollama://granite-code\n"
" llama-run ollama://smollm:135m\n"
" llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n"
" llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n"
" llama-run "
"huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n"
" llama-run https://example.com/some-file1.gguf\n"
" llama-run some-file2.gguf\n"
" llama-run file://some-file3.gguf\n"
" llama-run --ngl 99 some-file4.gguf\n"
" llama-run --ngl 99 some-file5.gguf Hello World\n";
" llama-run --ngl 999 some-file4.gguf\n"
" llama-run --ngl 999 some-file5.gguf Hello World\n",
context_size_default, ngl_default, temperature_default);
}
int parse(int argc, const char ** argv) {
int positional_args_i = 0;
for (int i = 1; i < argc; ++i) {
if (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0) {
if (i + 1 >= argc) {
return 1;
}
context_size_ = std::atoi(argv[++i]);
} else if (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0) {
if (i + 1 >= argc) {
return 1;
}
ngl_ = std::atoi(argv[++i]);
} else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0) {
help_ = true;
return 0;
} else if (!positional_args_i) {
++positional_args_i;
model_ = argv[i];
} else if (positional_args_i == 1) {
++positional_args_i;
user_ = argv[i];
} else {
user_ += " " + std::string(argv[i]);
}
}
return model_.empty(); // model_ is the only required value
}
void help() const { printf("%s", help_str_.c_str()); }
};
struct progress_data {
size_t file_size = 0;
size_t file_size = 0;
std::chrono::steady_clock::time_point start_time = std::chrono::steady_clock::now();
bool printed = false;
bool printed = false;
};
struct FileDeleter {
void operator()(FILE * file) const {
static int get_terminal_width() {
#if defined(_WIN32)
CONSOLE_SCREEN_BUFFER_INFO csbi;
GetConsoleScreenBufferInfo(GetStdHandle(STD_OUTPUT_HANDLE), &csbi);
return csbi.srWindow.Right - csbi.srWindow.Left + 1;
#else
struct winsize w;
ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
return w.ws_col;
#endif
}
#ifdef LLAMA_USE_CURL
class File {
public:
FILE * file = nullptr;
FILE * open(const std::string & filename, const char * mode) {
file = fopen(filename.c_str(), mode);
return file;
}
int lock() {
if (file) {
# ifdef _WIN32
fd = _fileno(file);
hFile = (HANDLE) _get_osfhandle(fd);
if (hFile == INVALID_HANDLE_VALUE) {
fd = -1;
return 1;
}
OVERLAPPED overlapped = { 0 };
if (!LockFileEx(hFile, LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, MAXDWORD, MAXDWORD,
&overlapped)) {
fd = -1;
return 1;
}
# else
fd = fileno(file);
if (flock(fd, LOCK_EX | LOCK_NB) != 0) {
fd = -1;
return 1;
}
# endif
}
return 0;
}
~File() {
if (fd >= 0) {
# ifdef _WIN32
if (hFile != INVALID_HANDLE_VALUE) {
OVERLAPPED overlapped = { 0 };
UnlockFileEx(hFile, 0, MAXDWORD, MAXDWORD, &overlapped);
}
# else
flock(fd, LOCK_UN);
# endif
}
if (file) {
fclose(file);
}
}
private:
int fd = -1;
# ifdef _WIN32
HANDLE hFile;
# endif
};
typedef std::unique_ptr<FILE, FileDeleter> FILE_ptr;
#ifdef LLAMA_USE_CURL
class CurlWrapper {
class HttpClient {
public:
int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
const bool progress, std::string * response_str = nullptr) {
@ -163,10 +308,20 @@ class CurlWrapper {
}
progress_data data;
FILE_ptr out;
File out;
if (!output_file.empty()) {
output_file_partial = output_file + ".partial";
out.reset(fopen(output_file_partial.c_str(), "ab"));
if (!out.open(output_file_partial, "ab")) {
printe("Failed to open file\n");
return 1;
}
if (out.lock()) {
printe("Failed to exclusively lock file\n");
return 1;
}
}
set_write_options(response_str, out);
@ -181,7 +336,7 @@ class CurlWrapper {
return 0;
}
~CurlWrapper() {
~HttpClient() {
if (chunk) {
curl_slist_free_all(chunk);
}
@ -195,13 +350,13 @@ class CurlWrapper {
CURL * curl = nullptr;
struct curl_slist * chunk = nullptr;
void set_write_options(std::string * response_str, const FILE_ptr & out) {
void set_write_options(std::string * response_str, const File & out) {
if (response_str) {
curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, capture_data);
curl_easy_setopt(curl, CURLOPT_WRITEDATA, response_str);
} else {
curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, write_data);
curl_easy_setopt(curl, CURLOPT_WRITEDATA, out.get());
curl_easy_setopt(curl, CURLOPT_WRITEDATA, out.file);
}
}
@ -219,7 +374,7 @@ class CurlWrapper {
if (progress) {
curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L);
curl_easy_setopt(curl, CURLOPT_XFERINFODATA, &data);
curl_easy_setopt(curl, CURLOPT_XFERINFOFUNCTION, progress_callback);
curl_easy_setopt(curl, CURLOPT_XFERINFOFUNCTION, update_progress);
}
}
@ -255,37 +410,31 @@ class CurlWrapper {
int mins = (static_cast<int>(seconds) % 3600) / 60;
int secs = static_cast<int>(seconds) % 60;
std::ostringstream out;
if (hrs > 0) {
out << hrs << "h " << std::setw(2) << std::setfill('0') << mins << "m " << std::setw(2) << std::setfill('0')
<< secs << "s";
return fmt("%dh %02dm %02ds", hrs, mins, secs);
} else if (mins > 0) {
out << mins << "m " << std::setw(2) << std::setfill('0') << secs << "s";
return fmt("%dm %02ds", mins, secs);
} else {
out << secs << "s";
return fmt("%ds", secs);
}
return out.str();
}
static std::string human_readable_size(curl_off_t size) {
static const char * suffix[] = { "B", "KB", "MB", "GB", "TB" };
char length = sizeof(suffix) / sizeof(suffix[0]);
int i = 0;
double dbl_size = size;
char length = sizeof(suffix) / sizeof(suffix[0]);
int i = 0;
double dbl_size = size;
if (size > 1024) {
for (i = 0; (size / 1024) > 0 && i < length - 1; i++, size /= 1024) {
dbl_size = size / 1024.0;
}
}
std::ostringstream out;
out << std::fixed << std::setprecision(2) << dbl_size << " " << suffix[i];
return out.str();
return fmt("%.2f %s", dbl_size, suffix[i]);
}
static int progress_callback(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
curl_off_t) {
static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
curl_off_t) {
progress_data * data = static_cast<progress_data *>(ptr);
if (total_to_download <= 0) {
return 0;
@ -293,27 +442,68 @@ class CurlWrapper {
total_to_download += data->file_size;
const curl_off_t now_downloaded_plus_file_size = now_downloaded + data->file_size;
const curl_off_t percentage = (now_downloaded_plus_file_size * 100) / total_to_download;
const curl_off_t pos = (percentage / 5);
std::string progress_bar;
for (int i = 0; i < 20; ++i) {
progress_bar.append((i < pos) ? "" : " ");
}
const curl_off_t percentage = calculate_percentage(now_downloaded_plus_file_size, total_to_download);
std::string progress_prefix = generate_progress_prefix(percentage);
// Calculate download speed and estimated time to completion
const auto now = std::chrono::steady_clock::now();
const std::chrono::duration<double> elapsed_seconds = now - data->start_time;
const double speed = now_downloaded / elapsed_seconds.count();
const double estimated_time = (total_to_download - now_downloaded) / speed;
printe("\r%ld%% |%s| %s/%s %.2f MB/s %s ", percentage, progress_bar.c_str(),
human_readable_size(now_downloaded).c_str(), human_readable_size(total_to_download).c_str(),
speed / (1024 * 1024), human_readable_time(estimated_time).c_str());
fflush(stderr);
const double speed = calculate_speed(now_downloaded, data->start_time);
const double tim = (total_to_download - now_downloaded) / speed;
std::string progress_suffix =
generate_progress_suffix(now_downloaded_plus_file_size, total_to_download, speed, tim);
int progress_bar_width = calculate_progress_bar_width(progress_prefix, progress_suffix);
std::string progress_bar;
generate_progress_bar(progress_bar_width, percentage, progress_bar);
print_progress(progress_prefix, progress_bar, progress_suffix);
data->printed = true;
return 0;
}
static curl_off_t calculate_percentage(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download) {
return (now_downloaded_plus_file_size * 100) / total_to_download;
}
static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", percentage); }
static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
const auto now = std::chrono::steady_clock::now();
const std::chrono::duration<double> elapsed_seconds = now - start_time;
return now_downloaded / elapsed_seconds.count();
}
static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download,
double speed, double estimated_time) {
const int width = 10;
return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width,
human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width,
human_readable_time(estimated_time).c_str());
}
static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) {
int progress_bar_width = get_terminal_width() - progress_prefix.size() - progress_suffix.size() - 3;
if (progress_bar_width < 1) {
progress_bar_width = 1;
}
return progress_bar_width;
}
static std::string generate_progress_bar(int progress_bar_width, curl_off_t percentage,
std::string & progress_bar) {
const curl_off_t pos = (percentage * progress_bar_width) / 100;
for (int i = 0; i < progress_bar_width; ++i) {
progress_bar.append((i < pos) ? "" : " ");
}
return progress_bar;
}
static void print_progress(const std::string & progress_prefix, const std::string & progress_bar,
const std::string & progress_suffix) {
printe("\r%*s\r%s%s| %s", get_terminal_width(), " ", progress_prefix.c_str(), progress_bar.c_str(),
progress_suffix.c_str());
}
// Function to write data to a file
static size_t write_data(void * ptr, size_t size, size_t nmemb, void * stream) {
FILE * out = static_cast<FILE *>(stream);
@ -344,12 +534,12 @@ class LlamaData {
return 1;
}
context = initialize_context(model, opt.context_size_);
context = initialize_context(model, opt);
if (!context) {
return 1;
}
sampler = initialize_sampler();
sampler = initialize_sampler(opt);
return 0;
}
@ -357,8 +547,8 @@ class LlamaData {
#ifdef LLAMA_USE_CURL
int download(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
const bool progress, std::string * response_str = nullptr) {
CurlWrapper curl;
if (curl.init(url, headers, output_file, progress, response_str)) {
HttpClient http;
if (http.init(url, headers, output_file, progress, response_str)) {
return 1;
}
@ -438,13 +628,17 @@ class LlamaData {
}
int resolve_model(std::string & model_) {
int ret = 0;
if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) {
remove_proto(model_);
return ret;
}
const std::string bn = basename(model_);
const std::vector<std::string> headers = { "--header",
"Accept: application/vnd.docker.distribution.manifest.v2+json" };
int ret = 0;
if (string_starts_with(model_, "file://") || std::filesystem::exists(bn)) {
remove_proto(model_);
} else if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
remove_proto(model_);
ret = huggingface_dl(model_, headers, bn);
} else if (string_starts_with(model_, "ollama://")) {
@ -464,23 +658,23 @@ class LlamaData {
// Initializes the model and returns a unique pointer to it
llama_model_ptr initialize_model(Opt & opt) {
ggml_backend_load_all();
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = opt.ngl_ >= 0 ? opt.ngl_ : model_params.n_gpu_layers;
resolve_model(opt.model_);
llama_model_ptr model(llama_load_model_from_file(opt.model_.c_str(), model_params));
printe(
"\r%*s"
"\rLoading model",
get_terminal_width(), " ");
llama_model_ptr model(llama_load_model_from_file(opt.model_.c_str(), opt.model_params));
if (!model) {
printe("%s: error: unable to load model from file: %s\n", __func__, opt.model_.c_str());
}
printe("\r%*s\r", static_cast<int>(sizeof("Loading model")), " ");
return model;
}
// Initializes the context with the specified parameters
llama_context_ptr initialize_context(const llama_model_ptr & model, const int n_ctx) {
llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = n_ctx;
ctx_params.n_batch = n_ctx;
llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params));
llama_context_ptr initialize_context(const llama_model_ptr & model, const Opt & opt) {
llama_context_ptr context(llama_new_context_with_model(model.get(), opt.ctx_params));
if (!context) {
printe("%s: error: failed to create the llama_context\n", __func__);
}
@ -489,10 +683,10 @@ class LlamaData {
}
// Initializes and configures the sampler
llama_sampler_ptr initialize_sampler() {
llama_sampler_ptr initialize_sampler(const Opt & opt) {
llama_sampler_ptr sampler(llama_sampler_chain_init(llama_sampler_chain_default_params()));
llama_sampler_chain_add(sampler.get(), llama_sampler_init_min_p(0.05f, 1));
llama_sampler_chain_add(sampler.get(), llama_sampler_init_temp(0.8f));
llama_sampler_chain_add(sampler.get(), llama_sampler_init_temp(opt.temperature));
llama_sampler_chain_add(sampler.get(), llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
return sampler;
@ -609,16 +803,20 @@ static int read_user_input(std::string & user) {
}
// Function to generate a response based on the prompt
static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response) {
static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response,
const bool stdout_a_terminal) {
// Set response color
printf("\033[33m");
if (stdout_a_terminal) {
printf("\033[33m");
}
if (generate(llama_data, prompt, response)) {
printe("failed to generate response\n");
return 1;
}
// End response with color reset and newline
printf("\n\033[0m");
printf("\n%s", stdout_a_terminal ? "\033[0m" : "");
return 0;
}
@ -635,29 +833,51 @@ static int apply_chat_template_with_error_handling(LlamaData & llama_data, const
}
// Helper function to handle user input
static int handle_user_input(std::string & user_input, const std::string & user_) {
if (!user_.empty()) {
user_input = user_;
static int handle_user_input(std::string & user_input, const std::string & user) {
if (!user.empty()) {
user_input = user;
return 0; // No need for interactive input
}
printf(
"\r "
"\r\033[32m> \033[0m");
"\r%*s"
"\r\033[32m> \033[0m",
get_terminal_width(), " ");
return read_user_input(user_input); // Returns true if input ends the loop
}
static bool is_stdin_a_terminal() {
#if defined(_WIN32)
HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE);
DWORD mode;
return GetConsoleMode(hStdin, &mode);
#else
return isatty(STDIN_FILENO);
#endif
}
static bool is_stdout_a_terminal() {
#if defined(_WIN32)
HANDLE hStdout = GetStdHandle(STD_OUTPUT_HANDLE);
DWORD mode;
return GetConsoleMode(hStdout, &mode);
#else
return isatty(STDOUT_FILENO);
#endif
}
// Function to tokenize the prompt
static int chat_loop(LlamaData & llama_data, const std::string & user_) {
static int chat_loop(LlamaData & llama_data, const std::string & user) {
int prev_len = 0;
llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
static const bool stdout_a_terminal = is_stdout_a_terminal();
while (true) {
// Get user input
std::string user_input;
while (handle_user_input(user_input, user_)) {
while (handle_user_input(user_input, user)) {
}
add_message("user", user_.empty() ? user_input : user_, llama_data);
add_message("user", user.empty() ? user_input : user, llama_data);
int new_len;
if (apply_chat_template_with_error_handling(llama_data, true, new_len) < 0) {
return 1;
@ -665,11 +885,11 @@ static int chat_loop(LlamaData & llama_data, const std::string & user_) {
std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
std::string response;
if (generate_response(llama_data, prompt, response)) {
if (generate_response(llama_data, prompt, response, stdout_a_terminal)) {
return 1;
}
if (!user_.empty()) {
if (!user.empty()) {
break;
}
@ -682,22 +902,13 @@ static int chat_loop(LlamaData & llama_data, const std::string & user_) {
return 0;
}
static void log_callback(const enum ggml_log_level level, const char * text, void *) {
if (level == GGML_LOG_LEVEL_ERROR) {
static void log_callback(const enum ggml_log_level level, const char * text, void * p) {
const Opt * opt = static_cast<Opt *>(p);
if (opt->verbose || level == GGML_LOG_LEVEL_ERROR) {
printe("%s", text);
}
}
static bool is_stdin_a_terminal() {
#if defined(_WIN32)
HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE);
DWORD mode;
return GetConsoleMode(hStdin, &mode);
#else
return isatty(STDIN_FILENO);
#endif
}
static std::string read_pipe_data() {
std::ostringstream result;
result << std::cin.rdbuf(); // Read all data from std::cin
@ -714,20 +925,20 @@ int main(int argc, const char ** argv) {
}
if (!is_stdin_a_terminal()) {
if (!opt.user_.empty()) {
opt.user_ += "\n\n";
if (!opt.user.empty()) {
opt.user += "\n\n";
}
opt.user_ += read_pipe_data();
opt.user += read_pipe_data();
}
llama_log_set(log_callback, nullptr);
llama_log_set(log_callback, &opt);
LlamaData llama_data;
if (llama_data.init(opt)) {
return 1;
}
if (chat_loop(llama_data, opt.user_)) {
if (chat_loop(llama_data, opt.user)) {
return 1;
}

View file

@ -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)

View file

@ -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) |
@ -344,6 +343,10 @@ node index.js
### POST `/completion`: Given a `prompt`, it returns the predicted completion.
> [!IMPORTANT]
>
> This endpoint is **not** OAI-compatible
*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:
@ -393,8 +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.
@ -441,40 +442,76 @@ 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`
`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.
**Response format**
- Note: In streaming mode (`stream`), only `content` 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.
- 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
@ -655,7 +692,6 @@ 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,
@ -690,7 +726,8 @@ 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)"
}
```
@ -842,6 +879,8 @@ curl http://localhost:8080/v1/chat/completions \
### 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).
@ -874,6 +913,46 @@ See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-r
}'
```
### POST `/embeddings`: non-OpenAI-compatible embeddings API
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:*
Same as the `/v1/embeddings` endpoint.
*Examples:*
Same as the `/v1/embeddings` endpoint.
**Response format**
```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 ... ],
]
}
]
```
### GET `/slots`: Returns the current slots processing state
> [!WARNING]
@ -924,7 +1003,6 @@ Example:
"mirostat": 0,
"mirostat_tau": 5.0,
"mirostat_eta": 0.10000000149011612,
"penalize_nl": false,
"stop": [],
"max_tokens": -1,
"n_keep": 0,

File diff suppressed because one or more lines are too long

Binary file not shown.

View file

@ -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

View file

@ -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 })}

View file

@ -15,7 +15,7 @@
#define MIMETYPE_JSON "application/json; charset=utf-8"
// auto generated files (update with ./deps.sh)
#include "index.html.hpp"
#include "index.html.gz.hpp"
#include "loading.html.hpp"
#include <atomic>
@ -79,8 +79,9 @@ enum error_type {
};
struct slot_params {
bool stream = true;
bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt
bool stream = true;
bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt
bool return_tokens = false;
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_discard = 0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
@ -93,7 +94,9 @@ struct slot_params {
json input_prefix;
json input_suffix;
std::vector<std::string> antiprompt;
std::vector<std::string> response_fields;
bool timings_per_token = false;
bool post_sampling_probs = false;
bool ignore_eos = false;
struct common_params_sampling sampling;
@ -139,7 +142,6 @@ struct slot_params {
{"mirostat", sampling.mirostat},
{"mirostat_tau", sampling.mirostat_tau},
{"mirostat_eta", sampling.mirostat_eta},
{"penalize_nl", sampling.penalize_nl},
{"stop", antiprompt},
{"max_tokens", n_predict}, // User configured n_predict
{"n_keep", n_keep},
@ -150,11 +152,13 @@ struct slot_params {
{"n_probs", sampling.n_probs},
{"min_keep", sampling.min_keep},
{"grammar", sampling.grammar},
{"grammar_trigger_words", sampling.grammar_trigger_words},
{"samplers", samplers},
{"speculative.n_max", speculative.n_max},
{"speculative.n_min", speculative.n_min},
{"speculative.p_min", speculative.p_min},
{"timings_per_token", timings_per_token},
{"post_sampling_probs", post_sampling_probs},
};
}
};
@ -188,6 +192,7 @@ struct server_task {
static slot_params params_from_json_cmpl(
const llama_model * model,
const llama_context * ctx,
const common_params & params_base,
const json & data) {
slot_params params;
@ -203,12 +208,14 @@ struct server_task {
params.stream = json_value(data, "stream", false);
params.cache_prompt = json_value(data, "cache_prompt", true);
params.return_tokens = json_value(data, "return_tokens", false);
params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", defaults.n_predict));
params.n_indent = json_value(data, "n_indent", defaults.n_indent);
params.n_keep = json_value(data, "n_keep", defaults.n_keep);
params.n_discard = json_value(data, "n_discard", defaults.n_discard);
//params.t_max_prompt_ms = json_value(data, "t_max_prompt_ms", defaults.t_max_prompt_ms); // TODO: implement
params.t_max_predict_ms = json_value(data, "t_max_predict_ms", defaults.t_max_predict_ms);
params.response_fields = json_value(data, "response_fields", std::vector<std::string>());
params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k);
params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p);
@ -230,10 +237,10 @@ struct server_task {
params.sampling.mirostat = json_value(data, "mirostat", defaults.sampling.mirostat);
params.sampling.mirostat_tau = json_value(data, "mirostat_tau", defaults.sampling.mirostat_tau);
params.sampling.mirostat_eta = json_value(data, "mirostat_eta", defaults.sampling.mirostat_eta);
params.sampling.penalize_nl = json_value(data, "penalize_nl", defaults.sampling.penalize_nl);
params.sampling.seed = json_value(data, "seed", defaults.sampling.seed);
params.sampling.n_probs = json_value(data, "n_probs", defaults.sampling.n_probs);
params.sampling.min_keep = json_value(data, "min_keep", defaults.sampling.min_keep);
params.post_sampling_probs = json_value(data, "post_sampling_probs", defaults.post_sampling_probs);
params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min);
params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max);
@ -243,8 +250,27 @@ struct server_task {
params.speculative.n_min = std::max(params.speculative.n_min, 2);
params.speculative.n_max = std::max(params.speculative.n_max, 0);
// TODO: add more sanity checks for the input parameters
if (params.sampling.penalty_last_n < -1) {
throw std::runtime_error("Error: repeat_last_n must be >= -1");
}
if (params.sampling.dry_penalty_last_n < -1) {
throw std::runtime_error("Error: dry_penalty_last_n must be >= -1");
}
if (params.sampling.penalty_last_n == -1) {
// note: should be the slot's context and not the full context, but it's ok
params.sampling.penalty_last_n = llama_n_ctx(ctx);
}
if (params.sampling.dry_penalty_last_n == -1) {
params.sampling.dry_penalty_last_n = llama_n_ctx(ctx);
}
if (params.sampling.dry_base < 1.0f) {
params.sampling.dry_base = defaults.sampling.dry_base;
params.sampling.dry_base = defaults.sampling.dry_base;
}
// sequence breakers for DRY
@ -430,41 +456,75 @@ inline std::string stop_type_to_str(stop_type type) {
struct completion_token_output {
llama_token tok;
float prob;
std::string text_to_send;
struct token_prob {
struct prob_info {
llama_token tok;
std::string tok_str;
std::string txt;
float prob;
};
std::vector<token_prob> probs;
std::vector<prob_info> probs;
json to_json() const {
json to_json(bool post_sampling_probs) const {
json probs_for_token = json::array();
for (const auto & p : probs) {
std::string txt(p.txt);
txt.resize(validate_utf8(txt));
probs_for_token.push_back(json {
{"tok_str", p.tok_str},
{"prob", p.prob},
{"id", p.tok},
{"token", txt},
{"bytes", str_to_bytes(p.txt)},
{
post_sampling_probs ? "prob" : "logprob",
post_sampling_probs ? p.prob : logarithm(p.prob)
},
});
}
return probs_for_token;
}
static json probs_vector_to_json(const std::vector<completion_token_output> & probs) {
static json probs_vector_to_json(const std::vector<completion_token_output> & probs, bool post_sampling_probs) {
json out = json::array();
for (const auto & prob : probs) {
const std::string tok_str = prob.text_to_send;
for (const auto & p : probs) {
std::string txt(p.text_to_send);
txt.resize(validate_utf8(txt));
out.push_back(json {
{"content", tok_str},
{"probs", prob.to_json()},
{"id", p.tok},
{"token", txt},
{"bytes", str_to_bytes(p.text_to_send)},
{
post_sampling_probs ? "prob" : "logprob",
post_sampling_probs ? p.prob : logarithm(p.prob)
},
{
post_sampling_probs ? "top_probs" : "top_logprobs",
p.to_json(post_sampling_probs)
},
});
}
return out;
}
static float logarithm(float x) {
// nlohmann::json converts -inf to null, so we need to prevent that
return x == 0.0f ? std::numeric_limits<float>::lowest() : std::log(x);
}
static std::vector<unsigned char> str_to_bytes(const std::string & str) {
std::vector<unsigned char> bytes;
for (unsigned char c : str) {
bytes.push_back(c);
}
return bytes;
}
};
struct server_task_result_cmpl_final : server_task_result {
int index = 0;
std::string content;
std::string content;
llama_tokens tokens;
bool stream;
result_timings timings;
std::string prompt;
@ -473,11 +533,13 @@ struct server_task_result_cmpl_final : server_task_result {
int32_t n_decoded;
int32_t n_prompt_tokens;
int32_t n_tokens_cached;
int32_t has_new_line;
bool has_new_line;
std::string stopping_word;
stop_type stop = STOP_TYPE_NONE;
bool post_sampling_probs;
std::vector<completion_token_output> probs_output;
std::vector<std::string> response_fields;
slot_params generation_params;
@ -508,6 +570,7 @@ struct server_task_result_cmpl_final : server_task_result {
json res = json {
{"index", index},
{"content", stream ? "" : content}, // in stream mode, content is already in last partial chunk
{"tokens", stream ? llama_tokens {} : tokens},
{"id_slot", id_slot},
{"stop", true},
{"model", oaicompat_model},
@ -522,10 +585,10 @@ struct server_task_result_cmpl_final : server_task_result {
{"tokens_cached", n_tokens_cached},
{"timings", timings.to_json()},
};
if (!probs_output.empty()) {
res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output);
if (!stream && !probs_output.empty()) {
res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs);
}
return res;
return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
}
json to_json_oaicompat_chat() {
@ -560,22 +623,30 @@ struct server_task_result_cmpl_final : server_task_result {
message_content = content;
}
json choices = json::array({json{
json choice {
{"finish_reason", finish_reason},
{"index", 0},
{"message", json{{"content", message_content},
{"tool_calls", tool_calls},
{"role", "assistant"}
}
}}});
{"message", {
{"content", message_content},
{"tool_calls", tool_calls},
{"role", "assistant"},
}},
};
if (!stream && probs_output.size() > 0) {
choice["logprobs"] = json{
{"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
};
}
std::time_t t = std::time(0);
json res = json {
{"choices", choices},
{"created", t},
{"model", oaicompat_model},
{"object", "chat.completion"},
json res {
{"choices", json::array({choice})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
@ -602,16 +673,19 @@ struct server_task_result_cmpl_final : server_task_result {
finish_reason = "stop";
}
json choices = json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}}});
json choice {
{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}
};
json ret = json {
{"choices", choices},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"object", "chat.completion.chunk"},
json ret {
{"choices", json::array({choice})},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion.chunk"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
@ -629,12 +703,15 @@ struct server_task_result_cmpl_final : server_task_result {
struct server_task_result_cmpl_partial : server_task_result {
int index = 0;
std::string content;
std::string content;
llama_tokens tokens;
int32_t n_decoded;
int32_t n_prompt_tokens;
std::vector<completion_token_output> probs_output;
bool post_sampling_probs;
completion_token_output prob_output;
result_timings timings;
// OAI-compat fields
@ -663,6 +740,7 @@ struct server_task_result_cmpl_partial : server_task_result {
json res = json {
{"index", index},
{"content", content},
{"tokens", tokens},
{"stop", false},
{"id_slot", id_slot},
{"tokens_predicted", n_decoded},
@ -672,8 +750,8 @@ struct server_task_result_cmpl_partial : server_task_result {
if (timings.prompt_n > 0) {
res.push_back({"timings", timings.to_json()});
}
if (!probs_output.empty()) {
res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output);
if (!prob_output.probs.empty()) {
res["completion_probabilities"] = completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs);
}
return res;
}
@ -744,7 +822,7 @@ struct server_task_result_cmpl_partial : server_task_result {
json second_ret = json{
{"choices", json::array({json{{"finish_reason", nullptr},
{"index", 0},
{"delta", json{
{"delta", json {
{"content", content}}}
}})},
{"created", t},
@ -759,18 +837,27 @@ struct server_task_result_cmpl_partial : server_task_result {
{"finish_reason", nullptr},
{"index", 0},
{"delta",
json{
json {
{"content", content},
}},
}});
}
GGML_ASSERT(choices.size() >= 1);
if (prob_output.probs.size() > 0) {
choices[0]["logprobs"] = json{
{"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
};
}
json ret = json {
{"choices", choices},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"object", "chat.completion.chunk"}
{"choices", choices},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion.chunk"}
};
if (timings.prompt_n >= 0) {
@ -783,32 +870,52 @@ struct server_task_result_cmpl_partial : server_task_result {
struct server_task_result_embd : server_task_result {
int index = 0;
std::vector<float> embedding;
std::vector<std::vector<float>> embedding;
int32_t n_tokens;
// OAI-compat fields
bool oaicompat = false;
virtual int get_index() override {
return index;
}
virtual json to_json() override {
return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
}
json to_json_non_oaicompat() {
return json {
{"index", index},
{"embedding", embedding},
};
}
json to_json_oaicompat() {
return json {
{"index", index},
{"embedding", embedding[0]},
{"tokens_evaluated", n_tokens},
};
}
};
struct server_task_result_rerank : server_task_result {
int index = 0;
float score = -1e6;
int32_t n_tokens;
virtual int get_index() override {
return index;
}
virtual json to_json() override {
return json {
{"index", index},
{"score", score},
{"index", index},
{"score", score},
{"tokens_evaluated", n_tokens},
};
}
};
@ -1015,8 +1122,11 @@ struct server_slot {
size_t last_nl_pos = 0;
std::string generated_text;
std::string generated_text;
llama_tokens generated_tokens;
llama_tokens cache_tokens;
std::vector<completion_token_output> generated_token_probs;
bool has_next_token = true;
@ -1037,7 +1147,6 @@ struct server_slot {
// stats
size_t n_sent_text = 0; // number of sent text character
size_t n_sent_token_probs = 0;
int64_t t_start_process_prompt;
int64_t t_start_generation;
@ -1059,9 +1168,9 @@ struct server_slot {
stopping_word = "";
n_past = 0;
n_sent_text = 0;
n_sent_token_probs = 0;
task_type = SERVER_TASK_TYPE_COMPLETION;
generated_tokens.clear();
generated_token_probs.clear();
}
@ -1526,7 +1635,7 @@ struct server_context {
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_add_bos_token(model);
has_eos_token = !llama_add_eos_token(model);
has_eos_token = llama_token_eos(model) != LLAMA_TOKEN_NULL;
if (!params_base.speculative.model.empty()) {
SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
@ -1791,7 +1900,8 @@ struct server_context {
auto match = slot.antiprompts.findSingleTokenMatch(result.tok);
// remember which tokens were sampled - used for repetition penalties during sampling
const std::string token_str = common_token_to_piece(ctx, result.tok, params_base.special || (match.pos != std::string::npos && match.is_grammar_trigger));
const std::string token_str = result.text_to_send;
// const std::string token_str = common_token_to_piece(ctx, result.tok, params_base.special || (match.pos != std::string::npos && match.is_grammar_trigger));
slot.sampled = result.tok;
if (match.pos != std::string::npos && !match.is_partial) {
@ -1807,30 +1917,15 @@ struct server_context {
// search stop word and delete it
slot.generated_text += token_str;
if (slot.params.return_tokens) {
slot.generated_tokens.push_back(result.tok);
}
slot.has_next_token = true;
// check if there is incomplete UTF-8 character at the end
bool incomplete = false;
for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i) {
unsigned char c = slot.generated_text[slot.generated_text.size() - i];
if ((c & 0xC0) == 0x80) {
// continuation byte: 10xxxxxx
continue;
}
if ((c & 0xE0) == 0xC0) {
// 2-byte character: 110xxxxx ...
incomplete = i < 2;
} else if ((c & 0xF0) == 0xE0) {
// 3-byte character: 1110xxxx ...
incomplete = i < 3;
} else if ((c & 0xF8) == 0xF0) {
// 4-byte character: 11110xxx ...
incomplete = i < 4;
}
// else 1-byte character or invalid byte
break;
}
bool incomplete = validate_utf8(slot.generated_text) < slot.generated_text.size();
// search stop word and delete it
if (!incomplete) {
size_t pos = std::min(slot.n_sent_text, slot.generated_text.size());
@ -1969,58 +2064,53 @@ struct server_context {
return slot.has_next_token; // continue
}
json get_formated_generation(const server_slot & slot) const {
std::vector<std::string> samplers;
samplers.reserve(slot.params.sampling.samplers.size());
for (const auto & sampler : slot.params.sampling.samplers) {
samplers.emplace_back(common_sampler_type_to_str(sampler));
}
void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) {
size_t n_probs = slot.params.sampling.n_probs;
size_t n_vocab = llama_n_vocab(llama_get_model(ctx));
if (post_sampling) {
const auto * cur_p = common_sampler_get_candidates(slot.smpl);
const size_t max_probs = cur_p->size;
return json {
{"n_ctx", slot.n_ctx},
{"n_predict", slot.n_predict}, // Server configured n_predict
{"model", params_base.model_alias},
{"seed", slot.params.sampling.seed},
{"seed_cur", slot.smpl ? common_sampler_get_seed(slot.smpl) : 0},
{"temperature", slot.params.sampling.temp},
{"dynatemp_range", slot.params.sampling.dynatemp_range},
{"dynatemp_exponent", slot.params.sampling.dynatemp_exponent},
{"top_k", slot.params.sampling.top_k},
{"top_p", slot.params.sampling.top_p},
{"min_p", slot.params.sampling.min_p},
{"xtc_probability", slot.params.sampling.xtc_probability},
{"xtc_threshold", slot.params.sampling.xtc_threshold},
{"typical_p", slot.params.sampling.typ_p},
{"repeat_last_n", slot.params.sampling.penalty_last_n},
{"repeat_penalty", slot.params.sampling.penalty_repeat},
{"presence_penalty", slot.params.sampling.penalty_present},
{"frequency_penalty", slot.params.sampling.penalty_freq},
{"dry_multiplier", slot.params.sampling.dry_multiplier},
{"dry_base", slot.params.sampling.dry_base},
{"dry_allowed_length", slot.params.sampling.dry_allowed_length},
{"dry_penalty_last_n", slot.params.sampling.dry_penalty_last_n},
{"dry_sequence_breakers", slot.params.sampling.dry_sequence_breakers},
{"mirostat", slot.params.sampling.mirostat},
{"mirostat_tau", slot.params.sampling.mirostat_tau},
{"mirostat_eta", slot.params.sampling.mirostat_eta},
{"penalize_nl", slot.params.sampling.penalize_nl},
{"stop", slot.params.antiprompt},
{"grammar_trigger_words", slot.params.sampling.grammar_trigger_words},
{"max_tokens", slot.params.n_predict}, // User configured n_predict
{"n_keep", slot.params.n_keep},
{"n_discard", slot.params.n_discard},
{"ignore_eos", slot.params.sampling.ignore_eos},
{"stream", slot.params.stream},
//{"logit_bias", slot.params.sampling.logit_bias},
{"n_probs", slot.params.sampling.n_probs},
{"min_keep", slot.params.sampling.min_keep},
{"grammar", slot.params.sampling.grammar},
{"samplers", samplers},
{"speculative", slot.can_speculate()},
{"speculative.n_max", slot.params.speculative.n_max},
{"speculative.n_min", slot.params.speculative.n_min},
{"speculative.p_min", slot.params.speculative.p_min},
};
// set probability for sampled token
for (size_t i = 0; i < max_probs; i++) {
if (cur_p->data[i].id == result.tok) {
result.prob = cur_p->data[i].p;
break;
}
}
// set probability for top n_probs tokens
result.probs.reserve(max_probs);
for (size_t i = 0; i < std::min(max_probs, n_probs); i++) {
result.probs.push_back({
cur_p->data[i].id,
common_detokenize(ctx, {cur_p->data[i].id}, special),
cur_p->data[i].p
});
}
} else {
// TODO: optimize this with min-p optimization
std::vector<llama_token_data> cur = get_token_probabilities(ctx, idx);
// set probability for sampled token
for (size_t i = 0; i < n_vocab; i++) {
// set probability for sampled token
if (cur[i].id == result.tok) {
result.prob = cur[i].p;
break;
}
}
// set probability for top n_probs tokens
result.probs.reserve(n_probs);
for (size_t i = 0; i < std::min(n_vocab, n_probs); i++) {
result.probs.push_back({
cur[i].id,
common_detokenize(ctx, {cur[i].id}, special),
cur[i].p
});
}
}
}
void send_error(const server_task & task, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) {
@ -2048,9 +2138,11 @@ struct server_context {
res->id = slot.id_task;
res->index = slot.index;
res->content = tkn.text_to_send;
res->tokens = { tkn.tok };
res->n_decoded = slot.n_decoded;
res->n_prompt_tokens = slot.n_prompt_tokens;
res->n_decoded = slot.n_decoded;
res->n_prompt_tokens = slot.n_prompt_tokens;
res->post_sampling_probs = slot.params.post_sampling_probs;
res->verbose = slot.params.verbose;
res->oaicompat = slot.params.oaicompat;
@ -2062,17 +2154,7 @@ struct server_context {
// populate res.probs_output
if (slot.params.sampling.n_probs > 0) {
const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size());
const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size());
std::vector<completion_token_output> probs_output;
if (probs_pos < probs_stop_pos) {
res->probs_output = std::vector<completion_token_output>(
slot.generated_token_probs.begin() + probs_pos,
slot.generated_token_probs.begin() + probs_stop_pos);
}
res->prob_output = tkn; // copy the token probs
}
// populate timings if this is final response or timings_per_token is enabled
@ -2090,16 +2172,19 @@ struct server_context {
res->index = slot.index;
res->content = slot.generated_text;
res->tokens = slot.generated_tokens;
res->timings = slot.get_timings();
res->prompt = common_detokenize(ctx, slot.prompt_tokens, true);
res->response_fields = slot.params.response_fields;
res->truncated = slot.truncated;
res->n_decoded = slot.n_decoded;
res->n_prompt_tokens = slot.n_prompt_tokens;
res->n_tokens_cached = slot.n_past;
res->has_new_line = slot.has_new_line;
res->stopping_word = slot.stopping_word;
res->stop = slot.stop;
res->truncated = slot.truncated;
res->n_decoded = slot.n_decoded;
res->n_prompt_tokens = slot.n_prompt_tokens;
res->n_tokens_cached = slot.n_past;
res->has_new_line = slot.has_new_line;
res->stopping_word = slot.stopping_word;
res->stop = slot.stop;
res->post_sampling_probs = slot.params.post_sampling_probs;
res->verbose = slot.params.verbose;
res->stream = slot.params.stream;
@ -2133,8 +2218,10 @@ struct server_context {
void send_embedding(const server_slot & slot, const llama_batch & batch) {
auto res = std::make_unique<server_task_result_embd>();
res->id = slot.id_task;
res->index = slot.index;
res->id = slot.id_task;
res->index = slot.index;
res->n_tokens = slot.n_prompt_tokens;
res->oaicompat = slot.params.oaicompat;
const int n_embd = llama_n_embd(model);
@ -2153,12 +2240,18 @@ struct server_context {
if (embd == NULL) {
SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]);
res->embedding = std::vector<float>(n_embd, 0.0f);
res->embedding.push_back(std::vector<float>(n_embd, 0.0f));
continue;
}
common_embd_normalize(embd, embd_res.data(), n_embd);
res->embedding = embd_res;
// normalize only when there is pooling
// TODO: configurable
if (llama_pooling_type(slot.ctx) != LLAMA_POOLING_TYPE_NONE) {
common_embd_normalize(embd, embd_res.data(), n_embd, 2);
res->embedding.push_back(embd_res);
} else {
res->embedding.push_back({ embd, embd + n_embd });
}
}
SLT_DBG(slot, "%s", "sending embeddings\n");
@ -2170,6 +2263,7 @@ struct server_context {
auto res = std::make_unique<server_task_result_rerank>();
res->id = slot.id_task;
res->index = slot.index;
res->n_tokens = slot.n_prompt_tokens;
for (int i = 0; i < batch.n_tokens; ++i) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) {
@ -2771,7 +2865,10 @@ struct server_context {
// add prompt tokens for processing in the current batch
while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) {
common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, false);
// without pooling, we want to output the embeddings for all the tokens in the batch
const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE;
common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, need_embd);
if (slot.params.cache_prompt) {
slot.cache_tokens.push_back(prompt_tokens[slot.n_past]);
@ -2886,7 +2983,9 @@ struct server_context {
continue; // continue loop of slots
}
llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
const int tok_idx = slot.i_batch - i;
llama_token id = common_sampler_sample(slot.smpl, ctx, tok_idx);
slot.i_batch = -1;
@ -2905,17 +3004,12 @@ struct server_context {
slot.t_token_generation = (t_current - slot.t_start_generation) / 1e3;
completion_token_output result;
result.tok = id;
result.tok = id;
result.text_to_send = common_token_to_piece(ctx, result.tok, params_base.special);
result.prob = 1.0f; // TODO: set it here instead of doing inside populate_token_probs
const auto * cur_p = common_sampler_get_candidates(slot.smpl);
for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
auto tok_id = cur_p->data[i].id;
result.probs.push_back({
tok_id,
tokens_to_output_formatted_string(ctx, tok_id),
i >= cur_p->size ? 0.0f : cur_p->data[i].p,
});
if (slot.params.sampling.n_probs > 0) {
populate_token_probs(slot, result, slot.params.post_sampling_probs, params_base.special, tok_idx);
}
if (!process_token(result, slot)) {
@ -2999,7 +3093,11 @@ struct server_context {
for (size_t i = 0; i < ids.size(); ++i) {
completion_token_output result;
result.tok = ids[i];
result.tok = ids[i];
result.text_to_send = common_token_to_piece(ctx, result.tok, params_base.special);
result.prob = 1.0f; // set later
// TODO: set result.probs
if (!process_token(result, slot)) {
// release slot because of stop condition
@ -3495,6 +3593,7 @@ int main(int argc, char ** argv) {
{ "bos_token", common_token_to_piece(ctx_server.ctx, llama_token_bos(ctx_server.model), true) },
{ "eos_token", common_token_to_piece(ctx_server.ctx, llama_token_eos(ctx_server.model), true) },
{ "chat_template", chat_template.source()},
{ "build_info", build_info },
};
if (ctx_server.params_base.use_jinja) {
auto tool_use_chat_template = llama_chat_template_from_model(ctx_server.model, ctx_server.params_base.chat_template, /* prefer_tool_use= */ true);
@ -3548,7 +3647,7 @@ int main(int argc, char ** argv) {
task.index = i;
task.prompt_tokens = std::move(tokenized_prompts[i]);
task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.params_base, data);
task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.ctx, ctx_server.params_base, data);
task.id_selected_slot = json_value(data, "id_slot", -1);
// OAI-compat
@ -3732,7 +3831,7 @@ int main(int argc, char ** argv) {
{"object", "list"},
{"data", {
{
{"id", params.model_alias},
{"id", params.model_alias.empty() ? params.model : params.model_alias},
{"object", "model"},
{"created", std::time(0)},
{"owned_by", "llamacpp"},
@ -3797,34 +3896,61 @@ int main(int argc, char ** argv) {
res_ok(res, data);
};
const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, bool oaicompat) {
const json body = json::parse(req.body);
bool oaicompat = false;
// an input prompt can be a string or a list of tokens (integer)
if (oaicompat && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
return;
}
// for the shape of input/content, see tokenize_input_prompts()
json prompt;
if (body.count("input") != 0) {
oaicompat = true;
prompt = body.at("input");
} else if (body.count("content") != 0) {
// with "content", we only support single prompt
prompt = std::vector<std::string>{body.at("content")};
} else if (body.contains("content")) {
oaicompat = false;
prompt = body.at("content");
} else {
res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
return;
}
bool use_base64 = false;
if (body.count("encoding_format") != 0) {
const std::string& format = body.at("encoding_format");
if (format == "base64") {
use_base64 = true;
} else if (format != "float") {
res_error(res, format_error_response("The format to return the embeddings in. Can be either float or base64", ERROR_TYPE_INVALID_REQUEST));
return;
}
}
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
for (const auto & tokens : tokenized_prompts) {
// this check is necessary for models that do not add BOS token to the input
if (tokens.empty()) {
res_error(res, format_error_response("Input content cannot be empty", ERROR_TYPE_INVALID_REQUEST));
return;
}
}
// create and queue the task
json responses = json::array();
bool error = false;
{
std::vector<server_task> tasks;
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, /* add_special */ false, true);
for (size_t i = 0; i < tokenized_prompts.size(); i++) {
server_task task = server_task(SERVER_TASK_TYPE_EMBEDDING);
server_task task = server_task(SERVER_TASK_TYPE_EMBEDDING);
task.id = ctx_server.queue_tasks.get_new_id();
task.index = i;
task.prompt_tokens = std::move(tokenized_prompts[i]);
// OAI-compat
task.params.oaicompat = oaicompat;
tasks.push_back(task);
}
@ -3852,12 +3978,18 @@ int main(int argc, char ** argv) {
}
// write JSON response
json root = oaicompat
? format_embeddings_response_oaicompat(body, responses)
: responses.size() == 1 ? responses[0] : json(responses);
json root = oaicompat ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses);
res_ok(res, root);
};
const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, false);
};
const auto handle_embeddings_oai = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, true);
};
const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) {
res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
@ -4004,8 +4136,13 @@ int main(int argc, char ** argv) {
}
} else {
// using embedded static index.html
svr->Get("/", [](const httplib::Request &, httplib::Response & res) {
res.set_content(reinterpret_cast<const char*>(index_html), index_html_len, "text/html; charset=utf-8");
svr->Get("/", [](const httplib::Request & req, httplib::Response & res) {
if (req.get_header_value("Accept-Encoding").find("gzip") == std::string::npos) {
res.set_content("Error: gzip is not supported by this browser", "text/plain");
} else {
res.set_header("Content-Encoding", "gzip");
res.set_content(reinterpret_cast<const char*>(index_html_gz), index_html_gz_len, "text/html; charset=utf-8");
}
return false;
});
}
@ -4026,7 +4163,7 @@ int main(int argc, char ** argv) {
svr->Post("/infill", handle_infill);
svr->Post("/embedding", handle_embeddings); // legacy
svr->Post("/embeddings", handle_embeddings);
svr->Post("/v1/embeddings", handle_embeddings);
svr->Post("/v1/embeddings", handle_embeddings_oai);
svr->Post("/rerank", handle_rerank);
svr->Post("/reranking", handle_rerank);
svr->Post("/v1/rerank", handle_rerank);

View file

@ -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"]
@ -92,7 +94,7 @@ 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)
@ -323,3 +325,64 @@ def test_hello_world_tool_call(tool: dict, expected_arguments: dict, hf_repo: st
assert tool["function"]["name"] == tool_call["function"]["name"]
actual_arguments = json.loads(tool_call["function"]["arguments"])
assert json.dumps(expected_arguments) == json.dumps(actual_arguments), f"tool arguments: {json.dumps(actual_arguments)}, expected: {json.dumps(expected_arguments)}"
def test_logprobs():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
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}")
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

View file

@ -10,22 +10,29 @@ 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 +55,15 @@ 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"]
@ -85,7 +95,7 @@ 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 +120,10 @@ 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 +258,40 @@ 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 +305,68 @@ 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"])

View file

@ -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

View file

@ -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

View file

@ -65,6 +65,7 @@ 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
@ -134,6 +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:

View file

@ -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 })}

View file

@ -225,7 +225,6 @@
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
@ -782,7 +781,6 @@
${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 })}

View file

@ -3,6 +3,7 @@
#include "common.h"
#include "log.h"
#include "llama.h"
#include "common/base64.hpp"
#ifndef NDEBUG
// crash the server in debug mode, otherwise send an http 500 error
@ -24,7 +25,7 @@
#include <vector>
#include <memory>
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo"
using json = nlohmann::ordered_json;
@ -58,6 +59,8 @@ static T json_value(const json & body, const std::string & key, const T & defaul
}
}
const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
//
// tokenizer and input processing utils
//
@ -90,6 +93,28 @@ static bool json_is_array_of_mixed_numbers_strings(const json & data) {
return false;
}
// get value by path(key1 / key2)
static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
json result = json::object();
for (const std::string & path : paths) {
json current = js;
const auto keys = string_split<std::string>(path, /*separator*/ '/');
bool valid_path = true;
for (const std::string & k : keys) {
if (valid_path && current.is_object() && current.contains(k)) {
current = current[k];
} else {
valid_path = false;
}
}
if (valid_path) {
result[path] = current;
}
}
return result;
}
/**
* this handles 2 cases:
* - only string, example: "string"
@ -140,6 +165,7 @@ static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_
* and multiple prompts (multi-tasks):
* - "prompt": ["string1", "string2"]
* - "prompt": ["string1", [12, 34, 56]]
* - "prompt": [[12, 34, 56], [78, 90, 12]]
* - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
*/
static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
@ -172,6 +198,36 @@ static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, con
return result;
}
// return the last index of character that can form a valid string
// if the last character is potentially cut in half, return the index before the cut
// if validate_utf8(text) == text.size(), then the whole text is valid utf8
static size_t validate_utf8(const std::string& text) {
size_t len = text.size();
if (len == 0) return 0;
// Check the last few bytes to see if a multi-byte character is cut off
for (size_t i = 1; i <= 4 && i <= len; ++i) {
unsigned char c = text[len - i];
// Check for start of a multi-byte sequence from the end
if ((c & 0xE0) == 0xC0) {
// 2-byte character start: 110xxxxx
// Needs at least 2 bytes
if (i < 2) return len - i;
} else if ((c & 0xF0) == 0xE0) {
// 3-byte character start: 1110xxxx
// Needs at least 3 bytes
if (i < 3) return len - i;
} else if ((c & 0xF8) == 0xF0) {
// 4-byte character start: 11110xxx
// Needs at least 4 bytes
if (i < 4) return len - i;
}
}
// If no cut-off multi-byte character is found, return full length
return len;
}
//
// template utils
//
@ -618,23 +674,41 @@ static json oaicompat_completion_params_parse(
return llama_params;
}
static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
static json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false) {
json data = json::array();
int32_t n_tokens = 0;
int i = 0;
for (const auto & elem : embeddings) {
data.push_back(json{
{"embedding", json_value(elem, "embedding", json::array())},
{"index", i++},
{"object", "embedding"}
});
json embedding_obj;
if (use_base64) {
const auto& vec = json_value(elem, "embedding", json::array()).get<std::vector<float>>();
const char* data_ptr = reinterpret_cast<const char*>(vec.data());
size_t data_size = vec.size() * sizeof(float);
embedding_obj = {
{"embedding", base64::encode(data_ptr, data_size)},
{"index", i++},
{"object", "embedding"},
{"encoding_format", "base64"}
};
} else {
embedding_obj = {
{"embedding", json_value(elem, "embedding", json::array())},
{"index", i++},
{"object", "embedding"}
};
}
data.push_back(embedding_obj);
n_tokens += json_value(elem, "tokens_evaluated", 0);
}
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json { // TODO: fill
{"prompt_tokens", 0},
{"total_tokens", 0}
{"usage", json {
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"data", data}
};
@ -644,20 +718,23 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
static json format_response_rerank(const json & request, const json & ranks) {
json data = json::array();
int32_t n_tokens = 0;
int i = 0;
for (const auto & rank : ranks) {
data.push_back(json{
{"index", i++},
{"relevance_score", json_value(rank, "score", 0.0)},
});
n_tokens += json_value(rank, "tokens_evaluated", 0);
}
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json { // TODO: fill
{"prompt_tokens", 0},
{"total_tokens", 0}
{"usage", json {
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"results", data}
};
@ -724,3 +801,33 @@ static json format_logit_bias(const std::vector<llama_logit_bias> & logit_bias)
static std::string safe_json_to_str(json data) {
return data.dump(-1, ' ', false, json::error_handler_t::replace);
}
static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
std::vector<llama_token_data> cur;
const auto * logits = llama_get_logits_ith(ctx, idx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
// sort tokens by logits
std::sort(cur.begin(), cur.end(), [](const llama_token_data & a, const llama_token_data & b) {
return a.logit > b.logit;
});
// apply softmax
float max_l = cur[0].logit;
float cum_sum = 0.0f;
for (size_t i = 0; i < cur.size(); ++i) {
float p = expf(cur[i].logit - max_l);
cur[i].p = p;
cum_sum += p;
}
for (size_t i = 0; i < cur.size(); ++i) {
cur[i].p /= cum_sum;
}
return cur;
}

View file

@ -201,6 +201,10 @@
<details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
<summary class="collapse-title font-bold">Advanced config</summary>
<div class="collapse-content">
<div class="flex flex-row items-center mb-2" v-if="isDev">
<!-- this button only shows in dev mode, used to import a demo conversation to test message rendering -->
<button class="btn" @click="debugImportDemoConv()">(debug) Import demo conversation</button>
</div>
<div class="flex flex-row items-center mb-2">
<input type="checkbox" class="checkbox" v-model="config.showTokensPerSecond" />
<span class="ml-4">Show tokens per second</span>

View file

@ -8,8 +8,12 @@
"name": "webui",
"version": "0.0.0",
"dependencies": {
"@sec-ant/readable-stream": "^0.6.0",
"@vscode/markdown-it-katex": "^1.1.1",
"autoprefixer": "^10.4.20",
"daisyui": "^4.12.14",
"highlight.js": "^11.10.0",
"katex": "^0.16.15",
"markdown-it": "^14.1.0",
"postcss": "^8.4.49",
"tailwindcss": "^3.4.15",
@ -18,6 +22,7 @@
"vue": "^3.5.13"
},
"devDependencies": {
"sass-embedded": "^1.83.0",
"vite": "^5.4.10"
}
},
@ -33,6 +38,13 @@
"url": "https://github.com/sponsors/sindresorhus"
}
},
"node_modules/@bufbuild/protobuf": {
"version": "2.2.3",
"resolved": "https://registry.npmjs.org/@bufbuild/protobuf/-/protobuf-2.2.3.tgz",
"integrity": "sha512-tFQoXHJdkEOSwj5tRIZSPNUuXK3RaR7T1nUrPgbYX1pUbvqqaaZAsfo+NXBPsz5rZMSKVFrgK1WL8Q/MSLvprg==",
"devOptional": true,
"license": "(Apache-2.0 AND BSD-3-Clause)"
},
"node_modules/@esbuild/aix-ppc64": {
"version": "0.21.5",
"resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.21.5.tgz",
@ -606,6 +618,21 @@
"win32"
]
},
"node_modules/@sec-ant/readable-stream": {
"version": "0.6.0",
"resolved": "https://registry.npmjs.org/@sec-ant/readable-stream/-/readable-stream-0.6.0.tgz",
"integrity": "sha512-uiBh8DrB5FN35gP6/o8JEhEQ7/ci1jUsOZO/VMUjyvTpjtV54VstOXVj1TvTj/wsT23pfX6butxxh3qufsW3+g==",
"license": "MIT"
},
"node_modules/@vscode/markdown-it-katex": {
"version": "1.1.1",
"resolved": "https://registry.npmjs.org/@vscode/markdown-it-katex/-/markdown-it-katex-1.1.1.tgz",
"integrity": "sha512-3KTlbsRBPJQLE2YmLL7K6nunTlU+W9T5+FjfNdWuIUKgxSS6HWLQHaO3L4MkJi7z7MpIPpY+g4N+cWNBPE/MSA==",
"license": "MIT",
"dependencies": {
"katex": "^0.16.4"
}
},
"node_modules/@vue/compiler-dom": {
"version": "3.5.13",
"resolved": "https://registry.npmjs.org/@vue/compiler-dom/-/compiler-dom-3.5.13.tgz",
@ -1004,6 +1031,13 @@
"browserslist": ">= 4.21.0"
}
},
"node_modules/buffer-builder": {
"version": "0.2.0",
"resolved": "https://registry.npmjs.org/buffer-builder/-/buffer-builder-0.2.0.tgz",
"integrity": "sha512-7VPMEPuYznPSoR21NE1zvd2Xna6c/CloiZCfcMXR1Jny6PjX0N4Nsa38zcBFo/FMK+BlA+FLKbJCQ0i2yxp+Xg==",
"devOptional": true,
"license": "MIT/X11"
},
"node_modules/caniuse-lite": {
"version": "1.0.30001684",
"resolved": "https://registry.npmjs.org/caniuse-lite/-/caniuse-lite-1.0.30001684.tgz",
@ -1166,6 +1200,22 @@
"node": ">=8.0"
}
},
"node_modules/colorjs.io": {
"version": "0.5.2",
"resolved": "https://registry.npmjs.org/colorjs.io/-/colorjs.io-0.5.2.tgz",
"integrity": "sha512-twmVoizEW7ylZSN32OgKdXRmo1qg+wT5/6C3xu5b9QsWzSFAhHLn2xd8ro0diCsKfCj1RdaTP/nrcW+vAoQPIw==",
"devOptional": true,
"license": "MIT"
},
"node_modules/commander": {
"version": "8.3.0",
"resolved": "https://registry.npmjs.org/commander/-/commander-8.3.0.tgz",
"integrity": "sha512-OkTL9umf+He2DZkUq8f8J9of7yL6RJKI24dVITBmNfZBmri9zYZQrKkuXiKhyfPSu8tUhnVBB1iKXevvnlR4Ww==",
"license": "MIT",
"engines": {
"node": ">= 12"
}
},
"node_modules/css-selector-tokenizer": {
"version": "0.8.0",
"resolved": "https://registry.npmjs.org/css-selector-tokenizer/-/css-selector-tokenizer-0.8.0.tgz",
@ -1473,6 +1523,31 @@
"node": ">=10.13.0"
}
},
"node_modules/has-flag": {
"version": "4.0.0",
"resolved": "https://registry.npmjs.org/has-flag/-/has-flag-4.0.0.tgz",
"integrity": "sha512-EykJT/Q1KjTWctppgIAgfSO0tKVuZUjhgMr17kqTumMl6Afv3EISleU7qZUzoXDFTAHTDC4NOoG/ZxU3EvlMPQ==",
"devOptional": true,
"license": "MIT",
"engines": {
"node": ">=8"
}
},
"node_modules/highlight.js": {
"version": "11.10.0",
"resolved": "https://registry.npmjs.org/highlight.js/-/highlight.js-11.10.0.tgz",
"integrity": "sha512-SYVnVFswQER+zu1laSya563s+F8VDGt7o35d4utbamowvUNLLMovFqwCLSocpZTz3MgaSRA1IbqRWZv97dtErQ==",
"engines": {
"node": ">=12.0.0"
}
},
"node_modules/immutable": {
"version": "5.0.3",
"resolved": "https://registry.npmjs.org/immutable/-/immutable-5.0.3.tgz",
"integrity": "sha512-P8IdPQHq3lA1xVeBRi5VPqUm5HDgKnx0Ru51wZz5mjxHr5n3RWhjIpOFU7ybkUxfB+5IToy+OLaHYDBIWsv+uw==",
"devOptional": true,
"license": "MIT"
},
"node_modules/is-glob": {
"version": "4.0.3",
"resolved": "https://registry.npmjs.org/is-glob/-/is-glob-4.0.3.tgz",
@ -1503,6 +1578,22 @@
"jiti": "bin/jiti.js"
}
},
"node_modules/katex": {
"version": "0.16.15",
"resolved": "https://registry.npmjs.org/katex/-/katex-0.16.15.tgz",
"integrity": "sha512-yE9YJIEAk2aZ+FL/G8r+UGw0CTUzEA8ZFy6E+8tc3spHUKq3qBnzCkI1CQwGoI9atJhVyFPEypQsTY7mJ1Pi9w==",
"funding": [
"https://opencollective.com/katex",
"https://github.com/sponsors/katex"
],
"license": "MIT",
"dependencies": {
"commander": "^8.3.0"
},
"bin": {
"katex": "cli.js"
}
},
"node_modules/lilconfig": {
"version": "2.1.0",
"resolved": "https://registry.npmjs.org/lilconfig/-/lilconfig-2.1.0.tgz",
@ -2022,6 +2113,381 @@
"integrity": "sha512-AYnb1nQyY49te+VRAVgmzfcgjYS91mY5P0TKUDCLEM+gNnA+3T6rWITXRLYCpahpqSQbN5cE+gHpnPyXjHWxcw==",
"license": "MIT"
},
"node_modules/rxjs": {
"version": "7.8.1",
"resolved": "https://registry.npmjs.org/rxjs/-/rxjs-7.8.1.tgz",
"integrity": "sha512-AA3TVj+0A2iuIoQkWEK/tqFjBq2j+6PO6Y0zJcvzLAFhEFIO3HL0vls9hWLncZbAAbK0mar7oZ4V079I/qPMxg==",
"devOptional": true,
"license": "Apache-2.0",
"dependencies": {
"tslib": "^2.1.0"
}
},
"node_modules/sass-embedded": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded/-/sass-embedded-1.83.0.tgz",
"integrity": "sha512-/8cYZeL39evUqe0o//193na51Q1VWZ61qhxioQvLJwOtWIrX+PgNhCyD8RSuTtmzc4+6+waFZf899bfp/MCUwA==",
"devOptional": true,
"license": "MIT",
"dependencies": {
"@bufbuild/protobuf": "^2.0.0",
"buffer-builder": "^0.2.0",
"colorjs.io": "^0.5.0",
"immutable": "^5.0.2",
"rxjs": "^7.4.0",
"supports-color": "^8.1.1",
"sync-child-process": "^1.0.2",
"varint": "^6.0.0"
},
"bin": {
"sass": "dist/bin/sass.js"
},
"engines": {
"node": ">=16.0.0"
},
"optionalDependencies": {
"sass-embedded-android-arm": "1.83.0",
"sass-embedded-android-arm64": "1.83.0",
"sass-embedded-android-ia32": "1.83.0",
"sass-embedded-android-riscv64": "1.83.0",
"sass-embedded-android-x64": "1.83.0",
"sass-embedded-darwin-arm64": "1.83.0",
"sass-embedded-darwin-x64": "1.83.0",
"sass-embedded-linux-arm": "1.83.0",
"sass-embedded-linux-arm64": "1.83.0",
"sass-embedded-linux-ia32": "1.83.0",
"sass-embedded-linux-musl-arm": "1.83.0",
"sass-embedded-linux-musl-arm64": "1.83.0",
"sass-embedded-linux-musl-ia32": "1.83.0",
"sass-embedded-linux-musl-riscv64": "1.83.0",
"sass-embedded-linux-musl-x64": "1.83.0",
"sass-embedded-linux-riscv64": "1.83.0",
"sass-embedded-linux-x64": "1.83.0",
"sass-embedded-win32-arm64": "1.83.0",
"sass-embedded-win32-ia32": "1.83.0",
"sass-embedded-win32-x64": "1.83.0"
}
},
"node_modules/sass-embedded-android-arm": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-android-arm/-/sass-embedded-android-arm-1.83.0.tgz",
"integrity": "sha512-uwFSXzJlfbd4Px189xE5l+cxN8+TQpXdQgJec7TIrb4HEY7imabtpYufpVdqUVwT1/uiis5V4+qIEC4Vl5XObQ==",
"cpu": [
"arm"
],
"license": "MIT",
"optional": true,
"os": [
"android"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-android-arm64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-android-arm64/-/sass-embedded-android-arm64-1.83.0.tgz",
"integrity": "sha512-GBiCvM4a2rkWBLdYDxI6XYnprfk5U5c81g69RC2X6kqPuzxzx8qTArQ9M6keFK4+iDQ5N9QTwFCr0KbZTn+ZNQ==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"android"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-android-ia32": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-android-ia32/-/sass-embedded-android-ia32-1.83.0.tgz",
"integrity": "sha512-5ATPdGo2SICqAhiJl/Z8KQ23zH4sGgobGgux0TnrNtt83uHZ+r+To/ubVJ7xTkZxed+KJZnIpolGD8dQyQqoTg==",
"cpu": [
"ia32"
],
"license": "MIT",
"optional": true,
"os": [
"android"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-android-riscv64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-android-riscv64/-/sass-embedded-android-riscv64-1.83.0.tgz",
"integrity": "sha512-aveknUOB8GZewOzVn2Uwk+DKcncTR50Q6vtzslNMGbYnxtgQNHzy8A1qVEviNUruex+pHofppeMK4iMPFAbiEQ==",
"cpu": [
"riscv64"
],
"license": "MIT",
"optional": true,
"os": [
"android"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-android-x64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-android-x64/-/sass-embedded-android-x64-1.83.0.tgz",
"integrity": "sha512-WqIay/72ncyf9Ph4vS742J3a73wZihWmzFUwpn1OD6lme1Aj4eWzWIve5IVnlTEJgcZcDHu6ECID9IZgehJKoA==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"android"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-darwin-arm64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-darwin-arm64/-/sass-embedded-darwin-arm64-1.83.0.tgz",
"integrity": "sha512-XQl9QqgxFFIPm/CzHhmppse5o9ocxrbaAdC2/DAnlAqvYWBBtgFqPjGoYlej13h9SzfvNoogx+y9r+Ap+e+hYg==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"darwin"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-darwin-x64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-darwin-x64/-/sass-embedded-darwin-x64-1.83.0.tgz",
"integrity": "sha512-ERQ7Tvp1kFOW3ux4VDFIxb7tkYXHYc+zJpcrbs0hzcIO5ilIRU2tIOK1OrNwrFO6Qxyf7AUuBwYKLAtIU/Nz7g==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"darwin"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-arm": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-arm/-/sass-embedded-linux-arm-1.83.0.tgz",
"integrity": "sha512-baG9RYBJxUFmqwDNC9h9ZFElgJoyO3jgHGjzEZ1wHhIS9anpG+zZQvO8bHx3dBpKEImX+DBeLX+CxsFR9n81gQ==",
"cpu": [
"arm"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-arm64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-arm64/-/sass-embedded-linux-arm64-1.83.0.tgz",
"integrity": "sha512-syEAVTJt4qhaMLxrSwOWa46zdqHJdnqJkLUK+t9aCr8xqBZLPxSUeIGji76uOehQZ1C+KGFj6n9xstHN6wzOJw==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-ia32": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-ia32/-/sass-embedded-linux-ia32-1.83.0.tgz",
"integrity": "sha512-RRBxQxMpoxu5+XcSSc6QR/o9asEwUzR8AbCS83RaXcdTIHTa/CccQsiAoDDoPlRsMTLqnzs0LKL4CfOsf7zBbA==",
"cpu": [
"ia32"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-musl-arm": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-musl-arm/-/sass-embedded-linux-musl-arm-1.83.0.tgz",
"integrity": "sha512-Yc7u2TelCfBab+PRob9/MNJFh3EooMiz4urvhejXkihTiKSHGCv5YqDdtWzvyb9tY2Jb7YtYREVuHwfdVn3dTQ==",
"cpu": [
"arm"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-musl-arm64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-musl-arm64/-/sass-embedded-linux-musl-arm64-1.83.0.tgz",
"integrity": "sha512-Y7juhPHClUO2H5O+u+StRy6SEAcwZ+hTEk5WJdEmo1Bb1gDtfHvJaWB/iFZJ2tW0W1e865AZeUrC4OcOFjyAQA==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-musl-ia32": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-musl-ia32/-/sass-embedded-linux-musl-ia32-1.83.0.tgz",
"integrity": "sha512-arQeYwGmwXV8byx5G1PtSzZWW1jbkfR5qrIHMEbTFSAvAxpqjgSvCvrHMOFd73FcMxVaYh4BX9LQNbKinkbEdg==",
"cpu": [
"ia32"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-musl-riscv64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-musl-riscv64/-/sass-embedded-linux-musl-riscv64-1.83.0.tgz",
"integrity": "sha512-E6uzlIWz59rut+Z3XR6mLG915zNzv07ISvj3GUNZENdHM7dF8GQ//ANoIpl5PljMQKp89GnYdvo6kj2gnaBf/g==",
"cpu": [
"riscv64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-musl-x64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-musl-x64/-/sass-embedded-linux-musl-x64-1.83.0.tgz",
"integrity": "sha512-eAMK6tyGqvqr21r9g8BnR3fQc1rYFj85RGduSQ3xkITZ6jOAnOhuU94N5fwRS852Hpws0lXhET+7JHXgg3U18w==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-riscv64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-riscv64/-/sass-embedded-linux-riscv64-1.83.0.tgz",
"integrity": "sha512-Ojpi78pTv02sy2fUYirRGXHLY3fPnV/bvwuC2i5LwPQw2LpCcFyFTtN0c5h4LJDk9P6wr+/ZB/JXU8tHIOlK+Q==",
"cpu": [
"riscv64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-linux-x64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-linux-x64/-/sass-embedded-linux-x64-1.83.0.tgz",
"integrity": "sha512-3iLjlXdoPfgZRtX4odhRvka1BQs5mAXqfCtDIQBgh/o0JnGPzJIWWl9bYLpHxK8qb+uyVBxXYgXpI0sCzArBOw==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-win32-arm64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-win32-arm64/-/sass-embedded-win32-arm64-1.83.0.tgz",
"integrity": "sha512-iOHw/8/t2dlTW3lOFwG5eUbiwhEyGWawivlKWJ8lkXH7fjMpVx2VO9zCFAm8RvY9xOHJ9sf1L7g5bx3EnNP9BQ==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"win32"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-win32-ia32": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-win32-ia32/-/sass-embedded-win32-ia32-1.83.0.tgz",
"integrity": "sha512-2PxNXJ8Pad4geVcTXY4rkyTr5AwbF8nfrCTDv0ulbTvPhzX2mMKEGcBZUXWn5BeHZTBc6whNMfS7d5fQXR9dDQ==",
"cpu": [
"ia32"
],
"license": "MIT",
"optional": true,
"os": [
"win32"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sass-embedded-win32-x64": {
"version": "1.83.0",
"resolved": "https://registry.npmjs.org/sass-embedded-win32-x64/-/sass-embedded-win32-x64-1.83.0.tgz",
"integrity": "sha512-muBXkFngM6eLTNqOV0FQi7Dv9s+YRQ42Yem26mosdan/GmJQc81deto6uDTgrYn+bzFNmiXcOdfm+0MkTWK3OQ==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"win32"
],
"engines": {
"node": ">=14.0.0"
}
},
"node_modules/sucrase": {
"version": "3.35.0",
"resolved": "https://registry.npmjs.org/sucrase/-/sucrase-3.35.0.tgz",
@ -2641,6 +3107,45 @@
"node": ">=8"
}
},
"node_modules/supports-color": {
"version": "8.1.1",
"resolved": "https://registry.npmjs.org/supports-color/-/supports-color-8.1.1.tgz",
"integrity": "sha512-MpUEN2OodtUzxvKQl72cUF7RQ5EiHsGvSsVG0ia9c5RbWGL2CI4C7EpPS8UTBIplnlzZiNuV56w+FuNxy3ty2Q==",
"devOptional": true,
"license": "MIT",
"dependencies": {
"has-flag": "^4.0.0"
},
"engines": {
"node": ">=10"
},
"funding": {
"url": "https://github.com/chalk/supports-color?sponsor=1"
}
},
"node_modules/sync-child-process": {
"version": "1.0.2",
"resolved": "https://registry.npmjs.org/sync-child-process/-/sync-child-process-1.0.2.tgz",
"integrity": "sha512-8lD+t2KrrScJ/7KXCSyfhT3/hRq78rC0wBFqNJXv3mZyn6hW2ypM05JmlSvtqRbeq6jqA94oHbxAr2vYsJ8vDA==",
"devOptional": true,
"license": "MIT",
"dependencies": {
"sync-message-port": "^1.0.0"
},
"engines": {
"node": ">=16.0.0"
}
},
"node_modules/sync-message-port": {
"version": "1.1.3",
"resolved": "https://registry.npmjs.org/sync-message-port/-/sync-message-port-1.1.3.tgz",
"integrity": "sha512-GTt8rSKje5FilG+wEdfCkOcLL7LWqpMlr2c3LRuKt/YXxcJ52aGSbGBAdI4L3aaqfrBt6y711El53ItyH1NWzg==",
"devOptional": true,
"license": "MIT",
"engines": {
"node": ">=16.0.0"
}
},
"node_modules/tailwindcss": {
"version": "3.4.15",
"resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-3.4.15.tgz",
@ -2684,12 +3189,26 @@
"integrity": "sha512-iBHbi7BQxrFmwZUQJsT0SjNzlLLsXhvW/kg7EyOMVMBIrlnj/qYofwo1LVLZi+3GbUEo96Iu2eqToI2+lZoAEQ==",
"license": "MIT"
},
"node_modules/tslib": {
"version": "2.8.1",
"resolved": "https://registry.npmjs.org/tslib/-/tslib-2.8.1.tgz",
"integrity": "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w==",
"devOptional": true,
"license": "0BSD"
},
"node_modules/uc.micro": {
"version": "2.1.0",
"resolved": "https://registry.npmjs.org/uc.micro/-/uc.micro-2.1.0.tgz",
"integrity": "sha512-ARDJmphmdvUk6Glw7y9DQ2bFkKBHwQHLi2lsaH6PPmz/Ka9sFOBsBluozhDltWmnv9u/cF6Rt87znRTPV+yp/A==",
"license": "MIT"
},
"node_modules/varint": {
"version": "6.0.0",
"resolved": "https://registry.npmjs.org/varint/-/varint-6.0.0.tgz",
"integrity": "sha512-cXEIW6cfr15lFv563k4GuVuW/fiwjknytD37jIOLSdSWuOI6WnO/oKwmP2FQTU2l01LP8/M5TSAJpzUaGe3uWg==",
"devOptional": true,
"license": "MIT"
},
"node_modules/vite": {
"version": "5.4.11",
"resolved": "https://registry.npmjs.org/vite/-/vite-5.4.11.tgz",

View file

@ -6,14 +6,20 @@
"scripts": {
"dev": "vite",
"build": "vite build",
"preview": "vite preview"
"preview": "vite preview",
"analyze": "ANALYZE=1 npx vite-bundle-visualizer"
},
"devDependencies": {
"sass-embedded": "^1.83.0",
"vite": "^5.4.10"
},
"dependencies": {
"@sec-ant/readable-stream": "^0.6.0",
"@vscode/markdown-it-katex": "^1.1.1",
"autoprefixer": "^10.4.20",
"daisyui": "^4.12.14",
"highlight.js": "^11.10.0",
"katex": "^0.16.15",
"markdown-it": "^14.1.0",
"postcss": "^8.4.49",
"tailwindcss": "^3.4.15",

View file

@ -0,0 +1,33 @@
{
"demo": true,
"id": "conv-1734086746930",
"lastModified": 1734087548943,
"messages": [
{
"id": 1734086764521,
"role": "user",
"content": "this is a demo conversation, used in dev mode"
},
{
"id": 1734087548327,
"role": "assistant",
"content": "This is the formula:\n\n$\\frac{e^{x_i}}{\\sum_{j=1}^{n}e^{x_j}}$\n\nGiven an input vector \\(\\mathbf{x} = [x_1, x_2, \\ldots, x_n]\\)\n\n\\[\ny_i = \\frac{e^{x_i}}{\\sum_{j=1}^n e^{x_j}}\n\\]\n\nCode block latex:\n```latex\n\\frac{e^{x_i}}{\\sum_{j=1}^{n}e^{x_j}}\n```\n\nTest dollar sign: $1234 $4567\n\nInvalid latex syntax: $E = mc^$ and $$E = mc^$$",
"timings": {
"prompt_n": 1,
"prompt_ms": 28.923,
"predicted_n": 25,
"predicted_ms": 573.016
}
},
{
"id": 1734087548328,
"role": "user",
"content": "this is a demo conversation, used in dev mode"
},
{
"id": 1734087548329,
"role": "assistant",
"content": "Code block:\n```js\nconsole.log('hello world')\n```\n```sh\nls -la /dev\n```"
}
]
}

View file

@ -0,0 +1,60 @@
import hljs from 'highlight.js/lib/core';
// only import commonly used languages to reduce bundle size
import python from 'highlight.js/lib/languages/python';
import javascript from 'highlight.js/lib/languages/javascript';
import json from 'highlight.js/lib/languages/json';
import bash from 'highlight.js/lib/languages/bash';
import yaml from 'highlight.js/lib/languages/yaml';
import markdown from 'highlight.js/lib/languages/markdown';
import scss from 'highlight.js/lib/languages/scss';
import xml from 'highlight.js/lib/languages/xml';
import ruby from 'highlight.js/lib/languages/ruby';
import go from 'highlight.js/lib/languages/go';
import java from 'highlight.js/lib/languages/java';
import rust from 'highlight.js/lib/languages/rust';
import scala from 'highlight.js/lib/languages/scala';
import cpp from 'highlight.js/lib/languages/cpp';
import csharp from 'highlight.js/lib/languages/csharp';
import swift from 'highlight.js/lib/languages/swift';
import dart from 'highlight.js/lib/languages/dart';
import elixir from 'highlight.js/lib/languages/elixir';
import kotlin from 'highlight.js/lib/languages/kotlin';
import lua from 'highlight.js/lib/languages/lua';
import php from 'highlight.js/lib/languages/php';
import latex from 'highlight.js/lib/languages/latex';
hljs.registerLanguage('python', python);
hljs.registerLanguage('javascript', javascript);
hljs.registerLanguage('json', json);
hljs.registerLanguage('yaml', yaml);
hljs.registerLanguage('markdown', markdown);
hljs.registerLanguage('xml', xml);
hljs.registerLanguage('ruby', ruby);
hljs.registerLanguage('go', go);
hljs.registerLanguage('java', java);
hljs.registerLanguage('rust', rust);
hljs.registerLanguage('scala', scala);
hljs.registerLanguage('csharp', csharp);
hljs.registerLanguage('swift', swift);
hljs.registerLanguage('dart', dart);
hljs.registerLanguage('elixir', elixir);
hljs.registerLanguage('kotlin', kotlin);
hljs.registerLanguage('lua', lua);
hljs.registerLanguage('php', php);
hljs.registerLanguage('latex', latex);
// reuse some languages to further reduce bundle size
hljs.registerLanguage('shell', bash);
hljs.registerLanguage('bash', bash);
hljs.registerLanguage('sh', bash);
hljs.registerLanguage('css', scss);
hljs.registerLanguage('scss', scss);
hljs.registerLanguage('c', cpp);
hljs.registerLanguage('cpp', cpp);
export default hljs;

View file

@ -0,0 +1,66 @@
import katex from 'katex';
// Adapted from https://github.com/SchneeHertz/markdown-it-katex-gpt
// MIT license
const defaultOptions = {
delimiters: [
{ left: '\\[', right: '\\]', display: true },
{ left: '\\(', right: '\\)', display: false },
],
};
export function renderLatexHTML(content, display = false) {
return katex.renderToString(content, {
throwOnError: false,
output: 'mathml',
displayMode: display,
});
}
function escapedBracketRule(options) {
return (state, silent) => {
const max = state.posMax;
const start = state.pos;
for (const { left, right, display } of options.delimiters) {
// Check if it starts with the left delimiter
if (!state.src.slice(start).startsWith(left)) continue;
// Skip the length of the left delimiter
let pos = start + left.length;
// Find the matching right delimiter
while (pos < max) {
if (state.src.slice(pos).startsWith(right)) {
break;
}
pos++;
}
// No matching right delimiter found, skip to the next match
if (pos >= max) continue;
// If not in silent mode, convert LaTeX formula to MathML
if (!silent) {
const content = state.src.slice(start + left.length, pos);
try {
const renderedContent = renderLatexHTML(content, display);
const token = state.push('html_inline', '', 0);
token.content = renderedContent;
} catch (e) {
console.error(e);
}
}
// Update position, skip the length of the right delimiter
state.pos = pos + right.length;
return true;
}
}
}
export default function (md, options = defaultOptions) {
md.inline.ruler.after('text', 'escaped_bracket', escapedBracketRule(options));
}

View file

@ -1,8 +1,20 @@
import './styles.css';
import './styles.scss';
import { createApp, defineComponent, shallowRef, computed, h } from 'vue/dist/vue.esm-bundler.js';
import MarkdownIt from 'markdown-it';
import TextLineStream from 'textlinestream';
// math formula rendering
import 'katex/dist/katex.min.css';
import markdownItKatexGpt from './katex-gpt';
import markdownItKatexNormal from '@vscode/markdown-it-katex';
// code highlighting
import hljs from './highlight-config';
import daisyuiThemes from 'daisyui/src/theming/themes';
// ponyfill for missing ReadableStream asyncIterator on Safari
import { asyncIterator } from '@sec-ant/readable-stream/ponyfill/asyncIterator';
const isDev = import.meta.env.MODE === 'development';
// utility functions
@ -10,18 +22,36 @@ const isString = (x) => !!x.toLowerCase;
const isBoolean = (x) => x === true || x === false;
const isNumeric = (n) => !isString(n) && !isNaN(n) && !isBoolean(n);
const escapeAttr = (str) => str.replace(/>/g, '&gt;').replace(/"/g, '&quot;');
const copyStr = (str) => navigator.clipboard.writeText(str);
const copyStr = (textToCopy) => {
// Navigator clipboard api needs a secure context (https)
if (navigator.clipboard && window.isSecureContext) {
navigator.clipboard.writeText(textToCopy);
} else {
// Use the 'out of viewport hidden text area' trick
const textArea = document.createElement('textarea');
textArea.value = textToCopy;
// Move textarea out of the viewport so it's not visible
textArea.style.position = 'absolute';
textArea.style.left = '-999999px';
document.body.prepend(textArea);
textArea.select();
document.execCommand('copy');
}
};
// constants
const BASE_URL = localStorage.getItem('base') // for debugging
|| (new URL('.', document.baseURI).href).toString().replace(/\/$/, ''); // for production
const BASE_URL = isDev
? (localStorage.getItem('base') || 'https://localhost:8080') // for debugging
: (new URL('.', document.baseURI).href).toString().replace(/\/$/, ''); // for production
console.log({ BASE_URL });
const CONFIG_DEFAULT = {
// Note: in order not to introduce breaking changes, please keep the same data type (number, string, etc) if you want to change the default value. Do not use null or undefined for default value.
apiKey: '',
systemMessage: 'You are a helpful assistant.',
showTokensPerSecond: false,
// make sure these default values are in sync with `common.h`
samplers: 'dkypmxt',
samplers: 'edkypmxt',
temperature: 0.8,
dynatemp_range: 0.0,
dynatemp_exponent: 1.0,
@ -69,12 +99,39 @@ const CONFIG_INFO = {
// config keys having numeric value (i.e. temperature, top_k, top_p, etc)
const CONFIG_NUMERIC_KEYS = Object.entries(CONFIG_DEFAULT).filter(e => isNumeric(e[1])).map(e => e[0]);
// list of themes supported by daisyui
const THEMES = ['light', 'dark', 'cupcake', 'bumblebee', 'emerald', 'corporate', 'synthwave', 'retro', 'cyberpunk', 'valentine', 'halloween', 'garden', 'forest', 'aqua', 'lofi', 'pastel', 'fantasy', 'wireframe', 'black', 'luxury', 'dracula', 'cmyk', 'autumn', 'business', 'acid', 'lemonade', 'night', 'coffee', 'winter', 'dim', 'nord', 'sunset'];
const THEMES = ['light', 'dark']
// make sure light & dark are always at the beginning
.concat(Object.keys(daisyuiThemes).filter(t => t !== 'light' && t !== 'dark'));
// markdown support
const VueMarkdown = defineComponent(
(props) => {
const md = shallowRef(new MarkdownIt({ breaks: true }));
const md = shallowRef(new MarkdownIt({
breaks: true,
highlight: function (str, lang) { // Add highlight.js
if (lang && hljs.getLanguage(lang)) {
try {
return '<pre><code class="hljs">' +
hljs.highlight(str, { language: lang, ignoreIllegals: true }).value +
'</code></pre>';
} catch (__) {}
}
return '<pre><code class="hljs">' + md.value.utils.escapeHtml(str) + '</code></pre>';
}
}));
// support latex with double dollar sign and square brackets
md.value.use(markdownItKatexGpt, {
delimiters: [
{ left: '\\[', right: '\\]', display: true },
{ left: '\\(', right: '\\)', display: false },
{ left: '$$', right: '$$', display: false },
// do not add single dollar sign here, other wise it will confused with dollar used for money symbol
],
throwOnError: false,
});
// support latex with single dollar sign
md.value.use(markdownItKatexNormal, { throwOnError: false });
// add copy button to code blocks
const origFenchRenderer = md.value.renderer.rules.fence;
md.value.renderer.rules.fence = (tokens, idx, ...args) => {
const content = tokens[idx].content;
@ -88,9 +145,9 @@ const VueMarkdown = defineComponent(
};
window.copyStr = copyStr;
const content = computed(() => md.value.render(props.source));
return () => h("div", { innerHTML: content.value });
return () => h('div', { innerHTML: content.value });
},
{ props: ["source"] }
{ props: ['source'] }
);
// input field to be used by settings modal
@ -244,7 +301,7 @@ async function* sendSSEPostRequest(url, fetchOptions) {
const lines = res.body
.pipeThrough(new TextDecoderStream())
.pipeThrough(new TextLineStream());
for await (const line of lines) {
for await (const line of asyncIterator(lines)) {
if (isDev) console.log({line});
if (line.startsWith('data:') && !line.endsWith('[DONE]')) {
const data = JSON.parse(line.slice(5));
@ -278,6 +335,7 @@ const mainApp = createApp({
themes: THEMES,
configDefault: {...CONFIG_DEFAULT},
configInfo: {...CONFIG_INFO},
isDev,
}
},
computed: {},
@ -289,6 +347,7 @@ const mainApp = createApp({
if (this.isGenerating) chatScrollToBottom(true);
});
resizeObserver.observe(pendingMsgElem);
this.setSelectedTheme(this.selectedTheme);
},
watch: {
viewingConvId: function(val, oldVal) {
@ -305,6 +364,8 @@ const mainApp = createApp({
},
setSelectedTheme(theme) {
this.selectedTheme = theme;
document.body.setAttribute('data-theme', theme);
document.body.setAttribute('data-color-scheme', daisyuiThemes[theme]?.['color-scheme'] ?? 'auto');
StorageUtils.setTheme(theme);
},
newConversation() {
@ -399,7 +460,7 @@ const mainApp = createApp({
method: 'POST',
headers: {
'Content-Type': 'application/json',
'Authorization': this.config.apiKey ? `Bearer ${this.config.apiKey}` : undefined,
...(this.config.apiKey ? {'Authorization': `Bearer ${this.config.apiKey}`} : {})
},
body: JSON.stringify(params),
signal: abortController.signal,
@ -513,6 +574,17 @@ const mainApp = createApp({
fetchMessages() {
this.messages = StorageUtils.getOneConversation(this.viewingConvId)?.messages ?? [];
},
// debug functions
async debugImportDemoConv() {
const res = await fetch('/demo-conversation.json');
const demoConv = await res.json();
StorageUtils.remove(demoConv.id);
for (const msg of demoConv.messages) {
StorageUtils.appendMsg(demoConv.id, msg);
}
this.fetchConversation();
}
},
});
mainApp.config.errorHandler = alert;

View file

@ -1,26 +0,0 @@
@tailwind base;
@tailwind components;
@tailwind utilities;
.markdown {
h1, h2, h3, h4, h5, h6, ul, ol, li { all: revert; }
pre {
@apply whitespace-pre-wrap rounded-lg p-2;
border: 1px solid currentColor;
}
/* TODO: fix markdown table */
}
.show-on-hover {
@apply md:opacity-0 md:group-hover:opacity-100;
}
.btn-mini {
@apply cursor-pointer hover:shadow-md;
}
.chat-screen { max-width: 900px; }
.chat-bubble-base-300 {
--tw-bg-opacity: 1;
--tw-text-opacity: 1;
@apply bg-base-300 text-base-content;
}

View file

@ -0,0 +1,48 @@
@use "sass:meta";
@tailwind base;
@tailwind components;
@tailwind utilities;
.markdown {
h1, h2, h3, h4, h5, h6, ul, ol, li { all: revert; }
pre {
@apply whitespace-pre-wrap rounded-lg p-2;
border: 1px solid currentColor;
}
/* TODO: fix markdown table */
}
.show-on-hover {
@apply md:opacity-0 md:group-hover:opacity-100;
}
.btn-mini {
@apply cursor-pointer hover:shadow-md;
}
.chat-screen { max-width: 900px; }
.chat-bubble-base-300 {
--tw-bg-opacity: 1;
--tw-text-opacity: 1;
@apply bg-base-300 text-base-content;
}
/* Highlight.js */
[data-color-scheme='light'] {
@include meta.load-css('highlight.js/styles/stackoverflow-light');
}
[data-color-scheme='dark'] {
@include meta.load-css('highlight.js/styles/stackoverflow-dark');
}
[data-color-scheme='auto'] {
@media (prefers-color-scheme: light) {
@include meta.load-css('highlight.js/styles/stackoverflow-light');
}
@media (prefers-color-scheme: dark) {
@include meta.load-css('highlight.js/styles/stackoverflow-dark');
}
}
.hljs {
background: transparent !important;
padding: 0.5em !important;
}

View file

@ -2,6 +2,9 @@
import { viteSingleFile } from 'vite-plugin-singlefile';
import path from 'path';
import fs from 'fs';
import zlib from 'zlib';
const MAX_BUNDLE_SIZE = 1.5 * 1024 * 1024; // only increase when absolutely necessary
const GUIDE_FOR_FRONTEND = `
<!--
@ -12,25 +15,45 @@ const GUIDE_FOR_FRONTEND = `
-->
`.trim();
export default {
plugins: [
viteSingleFile(),
(function llamaCppPlugin() {
let config;
return {
name: 'llamacpp:build',
apply: 'build',
async configResolved(_config) {
config = _config;
},
writeBundle() {
const outputIndexHtml = path.join(config.build.outDir, 'index.html');
const content = fs.readFileSync(outputIndexHtml, 'utf-8');
const BUILD_PLUGINS = [
viteSingleFile(),
(function llamaCppPlugin() {
let config;
return {
name: 'llamacpp:build',
apply: 'build',
async configResolved(_config) {
config = _config;
},
writeBundle() {
const outputIndexHtml = path.join(config.build.outDir, 'index.html');
const content = GUIDE_FOR_FRONTEND + '\n' + fs.readFileSync(outputIndexHtml, 'utf-8');
const compressed = zlib.gzipSync(Buffer.from(content, 'utf-8'), { level: 9 });
const targetOutputFile = path.join(config.build.outDir, '../../public/index.html');
fs.writeFileSync(targetOutputFile, GUIDE_FOR_FRONTEND + '\n' + content);
// because gzip header contains machine-specific info, we must remove these data from the header
// timestamp
compressed[0x4] = 0;
compressed[0x5] = 0;
compressed[0x6] = 0;
compressed[0x7] = 0;
// OS
compressed[0x9] = 0;
if (compressed.byteLength > MAX_BUNDLE_SIZE) {
throw new Error(
`Bundle size is too large (${Math.ceil(compressed.byteLength / 1024)} KB).\n` +
`Please reduce the size of the frontend or increase MAX_BUNDLE_SIZE in vite.config.js.\n`,
);
}
const targetOutputFile = path.join(config.build.outDir, '../../public/index.html.gz');
fs.writeFileSync(targetOutputFile, compressed);
}
})(),
],
}
})(),
];
/** @type {import('vite').UserConfig} */
export default {
plugins: process.env.ANALYZE ? [] : BUILD_PLUGINS,
};

View file

@ -0,0 +1,5 @@
set(TARGET llama-tts)
add_executable(${TARGET} tts.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)

View file

@ -0,0 +1,180 @@
# convert the https://huggingface.co/novateur/WavTokenizer-large-speech-75token to HF format
# the goal is to be able to reuse the convert_hf_to_gguf.py after that to create a GGUF file with the WavTokenizer decoder
#
# TODO: this script is LLM-generated and probably very inefficient and should be rewritten
import torch
import json
import os
import sys
import re
from safetensors.torch import save_file
# default
model_path = './model.pt';
# read from CLI
if len(sys.argv) > 1:
model_path = sys.argv[1]
# get the directory of the input model
path_dst = os.path.dirname(model_path)
print(f"Loading model from {model_path}")
model = torch.load(model_path, map_location='cpu')
#print(model)
# print all keys
for key in model.keys():
print(key)
if key == 'hyper_parameters':
#print(model[key])
# dump as json pretty
print(json.dumps(model[key], indent=4))
#if key != 'state_dict' and key != 'optimizer_states':
# print(model[key])
# Check if the loaded model is a state_dict or a model instance
if isinstance(model, torch.nn.Module):
state_dict = model.state_dict()
else:
state_dict = model
# Print the structure of the state_dict to understand its format
print("State dictionary keys:")
for key in state_dict.keys():
print(key)
# Ensure the state_dict is flat and contains only torch.Tensor objects
def flatten_state_dict(state_dict, parent_key='', sep='.'):
items = []
items_new = []
for k, v in state_dict.items():
new_key = f"{parent_key}{sep}{k}" if parent_key else k
if isinstance(v, torch.Tensor):
items.append((new_key, v))
elif isinstance(v, dict):
items.extend(flatten_state_dict(v, new_key, sep=sep).items())
return dict(items)
size_total_mb = 0
for key, value in list(items):
# keep only what we need for inference
if not key.startswith('state_dict.feature_extractor.encodec.quantizer.') and \
not key.startswith('state_dict.backbone.') and \
not key.startswith('state_dict.head.out'):
print('Skipping key: ', key)
continue
new_key = key
new_key = new_key.replace('state_dict.', '')
new_key = new_key.replace('pos_net', 'posnet')
# check if matches "backbone.posnet.%d.bias" or "backbone.posnet.%d.weight"
if new_key.startswith("backbone.posnet."):
match = re.match(r"backbone\.posnet\.(\d+)\.(bias|weight)", new_key)
if match:
new_key = f"backbone.posnet.{match.group(1)}.norm.{match.group(2)}"
# "feature_extractor.encodec.quantizer.vq.layers.0._codebook.embed" -> "backbone.embedding.weight"
if new_key == "feature_extractor.encodec.quantizer.vq.layers.0._codebook.embed":
new_key = "backbone.embedding.weight"
# these are the only rows used
# ref: https://github.com/edwko/OuteTTS/blob/a613e79c489d8256dd657ea9168d78de75895d82/outetts/wav_tokenizer/audio_codec.py#L100
if new_key.endswith("norm.scale.weight"):
new_key = new_key.replace("norm.scale.weight", "norm.weight")
value = value[0]
if new_key.endswith("norm.shift.weight"):
new_key = new_key.replace("norm.shift.weight", "norm.bias")
value = value[0]
if new_key.endswith("gamma"):
new_key = new_key.replace("gamma", "gamma.weight")
# convert from 1D [768] to 2D [768, 1] so that ggml_add can broadcast the bias
if (new_key.endswith("norm.weight") or new_key.endswith("norm1.weight") or new_key.endswith("norm2.weight") or new_key.endswith(".bias")) and (new_key.startswith("backbone.posnet") or new_key.startswith("backbone.embed.bias")):
value = value.unsqueeze(1)
if new_key.endswith("dwconv.bias"):
value = value.unsqueeze(1)
size_mb = value.element_size() * value.nelement() / (1024 * 1024)
print(f"{size_mb:8.2f} MB - {new_key}: {value.shape}")
size_total_mb += size_mb
#print(key, '->', new_key, ': ', value)
#print(key, '->', new_key)
items_new.append((new_key, value))
print(f"Total size: {size_total_mb:8.2f} MB")
return dict(items_new)
flattened_state_dict = flatten_state_dict(state_dict)
# Convert the model to the safetensors format
output_path = path_dst + '/model.safetensors'
save_file(flattened_state_dict, output_path)
print(f"Model has been successfully converted and saved to {output_path}")
# Calculate the total size of the .safetensors file
total_size = os.path.getsize(output_path)
# Create the weight map
weight_map = {
"model.safetensors": ["*"] # Assuming all weights are in one file
}
# Create metadata for the index.json file
metadata = {
"total_size": total_size,
"weight_map": weight_map
}
# Save the metadata to index.json
index_path = path_dst + '/index.json'
with open(index_path, 'w') as f:
json.dump(metadata, f, indent=4)
print(f"Metadata has been saved to {index_path}")
config = {
"architectures": [
"WavTokenizerDec"
],
"hidden_size": 1282,
"n_embd_features": 512,
"n_ff": 2304,
"vocab_size": 4096,
"n_head": 1,
"layer_norm_epsilon": 1e-6,
"group_norm_epsilon": 1e-6,
"group_norm_groups": 32,
"max_position_embeddings": 8192, # ?
"n_layer": 12,
"posnet": {
"n_embd": 768,
"n_layer": 6
},
"convnext": {
"n_embd": 768,
"n_layer": 12
},
}
with open(path_dst + '/config.json', 'w') as f:
json.dump(config, f, indent=4)
print(f"Config has been saved to {path_dst + 'config.json'}")

175
examples/tts/tts-outetts.py Normal file
View file

@ -0,0 +1,175 @@
import sys
#import json
#import struct
import requests
import re
def process_text(text: str):
text = re.sub(r'\d+(\.\d+)?', lambda x: x.group(), text.lower()) # TODO this needs to be fixed
text = re.sub(r'[-_/,\.\\]', ' ', text)
text = re.sub(r'[^a-z\s]', '', text)
text = re.sub(r'\s+', ' ', text).strip()
return text.split()
# usage:
# python tts-outetts.py http://server-llm:port http://server-dec:port "text"
if len(sys.argv) <= 3:
print("usage: python tts-outetts.py http://server-llm:port http://server-dec:port \"text\"")
exit(1)
host_llm = sys.argv[1]
host_dec = sys.argv[2]
text = sys.argv[3]
prefix = """<|im_start|>
<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>"""
words = process_text(text)
words = "<|text_sep|>".join([i.strip() for i in words])
words += "<|text_end|>\n"
# voice data
# TODO: load from json
#suffix = """<|audio_start|>
#the<|t_0.08|><|code_start|><|257|><|740|><|636|><|913|><|788|><|1703|><|code_end|>
#overall<|t_0.36|><|code_start|><|127|><|201|><|191|><|774|><|700|><|532|><|1056|><|557|><|798|><|298|><|1741|><|747|><|1662|><|1617|><|1702|><|1527|><|368|><|1588|><|1049|><|1008|><|1625|><|747|><|1576|><|728|><|1019|><|1696|><|1765|><|code_end|>
#package<|t_0.56|><|code_start|><|935|><|584|><|1319|><|627|><|1016|><|1491|><|1344|><|1117|><|1526|><|1040|><|239|><|1435|><|951|><|498|><|723|><|1180|><|535|><|789|><|1649|><|1637|><|78|><|465|><|1668|><|901|><|595|><|1675|><|117|><|1009|><|1667|><|320|><|840|><|79|><|507|><|1762|><|1508|><|1228|><|1768|><|802|><|1450|><|1457|><|232|><|639|><|code_end|>
#from<|t_0.19|><|code_start|><|604|><|782|><|1682|><|872|><|1532|><|1600|><|1036|><|1761|><|647|><|1554|><|1371|><|653|><|1595|><|950|><|code_end|>
#just<|t_0.25|><|code_start|><|1782|><|1670|><|317|><|786|><|1748|><|631|><|599|><|1155|><|1364|><|1524|><|36|><|1591|><|889|><|1535|><|541|><|440|><|1532|><|50|><|870|><|code_end|>
#two<|t_0.24|><|code_start|><|1681|><|1510|><|673|><|799|><|805|><|1342|><|330|><|519|><|62|><|640|><|1138|><|565|><|1552|><|1497|><|1552|><|572|><|1715|><|1732|><|code_end|>
#people<|t_0.39|><|code_start|><|593|><|274|><|136|><|740|><|691|><|633|><|1484|><|1061|><|1138|><|1485|><|344|><|428|><|397|><|1562|><|645|><|917|><|1035|><|1449|><|1669|><|487|><|442|><|1484|><|1329|><|1832|><|1704|><|600|><|761|><|653|><|269|><|code_end|>
#is<|t_0.16|><|code_start|><|566|><|583|><|1755|><|646|><|1337|><|709|><|802|><|1008|><|485|><|1583|><|652|><|10|><|code_end|>
#pretty<|t_0.32|><|code_start|><|1818|><|1747|><|692|><|733|><|1010|><|534|><|406|><|1697|><|1053|><|1521|><|1355|><|1274|><|816|><|1398|><|211|><|1218|><|817|><|1472|><|1703|><|686|><|13|><|822|><|445|><|1068|><|code_end|>
#remarkable<|t_0.68|><|code_start|><|230|><|1048|><|1705|><|355|><|706|><|1149|><|1535|><|1787|><|1356|><|1396|><|835|><|1583|><|486|><|1249|><|286|><|937|><|1076|><|1150|><|614|><|42|><|1058|><|705|><|681|><|798|><|934|><|490|><|514|><|1399|><|572|><|1446|><|1703|><|1346|><|1040|><|1426|><|1304|><|664|><|171|><|1530|><|625|><|64|><|1708|><|1830|><|1030|><|443|><|1509|><|1063|><|1605|><|1785|><|721|><|1440|><|923|><|code_end|>
#sure<|t_0.36|><|code_start|><|792|><|1780|><|923|><|1640|><|265|><|261|><|1525|><|567|><|1491|><|1250|><|1730|><|362|><|919|><|1766|><|543|><|1|><|333|><|113|><|970|><|252|><|1606|><|133|><|302|><|1810|><|1046|><|1190|><|1675|><|code_end|>
#i<|t_0.08|><|code_start|><|123|><|439|><|1074|><|705|><|1799|><|637|><|code_end|>
#have<|t_0.16|><|code_start|><|1509|><|599|><|518|><|1170|><|552|><|1029|><|1267|><|864|><|419|><|143|><|1061|><|0|><|code_end|>
#some<|t_0.16|><|code_start|><|619|><|400|><|1270|><|62|><|1370|><|1832|><|917|><|1661|><|167|><|269|><|1366|><|1508|><|code_end|>
#critiques<|t_0.60|><|code_start|><|559|><|584|><|1163|><|1129|><|1313|><|1728|><|721|><|1146|><|1093|><|577|><|928|><|27|><|630|><|1080|><|1346|><|1337|><|320|><|1382|><|1175|><|1682|><|1556|><|990|><|1683|><|860|><|1721|><|110|><|786|><|376|><|1085|><|756|><|1523|><|234|><|1334|><|1506|><|1578|><|659|><|612|><|1108|><|1466|><|1647|><|308|><|1470|><|746|><|556|><|1061|><|code_end|>
#about<|t_0.29|><|code_start|><|26|><|1649|><|545|><|1367|><|1263|><|1728|><|450|><|859|><|1434|><|497|><|1220|><|1285|><|179|><|755|><|1154|><|779|><|179|><|1229|><|1213|><|922|><|1774|><|1408|><|code_end|>
#some<|t_0.23|><|code_start|><|986|><|28|><|1649|><|778|><|858|><|1519|><|1|><|18|><|26|><|1042|><|1174|><|1309|><|1499|><|1712|><|1692|><|1516|><|1574|><|code_end|>
#of<|t_0.07|><|code_start|><|197|><|716|><|1039|><|1662|><|64|><|code_end|>
#the<|t_0.08|><|code_start|><|1811|><|1568|><|569|><|886|><|1025|><|1374|><|code_end|>
#gameplay<|t_0.48|><|code_start|><|1269|><|1092|><|933|><|1362|><|1762|><|1700|><|1675|><|215|><|781|><|1086|><|461|><|838|><|1022|><|759|><|649|><|1416|><|1004|><|551|><|909|><|787|><|343|><|830|><|1391|><|1040|><|1622|><|1779|><|1360|><|1231|><|1187|><|1317|><|76|><|997|><|989|><|978|><|737|><|189|><|code_end|>
#aspects<|t_0.56|><|code_start|><|1423|><|797|><|1316|><|1222|><|147|><|719|><|1347|><|386|><|1390|><|1558|><|154|><|440|><|634|><|592|><|1097|><|1718|><|712|><|763|><|1118|><|1721|><|1311|><|868|><|580|><|362|><|1435|><|868|><|247|><|221|><|886|><|1145|><|1274|><|1284|><|457|><|1043|><|1459|><|1818|><|62|><|599|><|1035|><|62|><|1649|><|778|><|code_end|>
#but<|t_0.20|><|code_start|><|780|><|1825|><|1681|><|1007|><|861|><|710|><|702|><|939|><|1669|><|1491|><|613|><|1739|><|823|><|1469|><|648|><|code_end|>
#its<|t_0.09|><|code_start|><|92|><|688|><|1623|><|962|><|1670|><|527|><|599|><|code_end|>
#still<|t_0.27|><|code_start|><|636|><|10|><|1217|><|344|><|713|><|957|><|823|><|154|><|1649|><|1286|><|508|><|214|><|1760|><|1250|><|456|><|1352|><|1368|><|921|><|615|><|5|><|code_end|>
#really<|t_0.36|><|code_start|><|55|><|420|><|1008|><|1659|><|27|><|644|><|1266|><|617|><|761|><|1712|><|109|><|1465|><|1587|><|503|><|1541|><|619|><|197|><|1019|><|817|><|269|><|377|><|362|><|1381|><|507|><|1488|><|4|><|1695|><|code_end|>
#enjoyable<|t_0.49|><|code_start|><|678|><|501|><|864|><|319|><|288|><|1472|><|1341|><|686|><|562|><|1463|><|619|><|1563|><|471|><|911|><|730|><|1811|><|1006|><|520|><|861|><|1274|><|125|><|1431|><|638|><|621|><|153|><|876|><|1770|><|437|><|987|><|1653|><|1109|><|898|><|1285|><|80|><|593|><|1709|><|843|><|code_end|>
#and<|t_0.15|><|code_start|><|1285|><|987|><|303|><|1037|><|730|><|1164|><|502|><|120|><|1737|><|1655|><|1318|><|code_end|>
#it<|t_0.09|><|code_start|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|><|code_end|>
#looks<|t_0.27|><|code_start|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|code_end|>
#lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|code_end|>"""
# TODO: tokenization is slow for some reason - here is pre-tokenized input
suffix = [ 151667, 198, 1782, 155780, 151669, 151929, 152412, 152308, 152585, 152460, 153375, 151670, 198, 74455,
155808, 151669, 151799, 151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470, 151970, 153413,
152419, 153334, 153289, 153374, 153199, 152040, 153260, 152721, 152680, 153297, 152419, 153248, 152400,
152691, 153368, 153437, 151670, 198, 1722, 155828, 151669, 152607, 152256, 152991, 152299, 152688, 153163,
153016, 152789, 153198, 152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207, 152461, 153321,
153309, 151750, 152137, 153340, 152573, 152267, 153347, 151789, 152681, 153339, 151992, 152512, 151751,
152179, 153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904, 152311, 151670, 198, 1499, 155791,
151669, 152276, 152454, 153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226, 153043, 152325,
153267, 152622, 151670, 198, 4250, 155797, 151669, 153454, 153342, 151989, 152458, 153420, 152303, 152271,
152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213, 152112, 153204, 151722, 152542, 151670, 198,
19789, 155796, 151669, 153353, 153182, 152345, 152471, 152477, 153014, 152002, 152191, 151734, 152312, 152810,
152237, 153224, 153169, 153224, 152244, 153387, 153404, 151670, 198, 16069, 155811, 151669, 152265, 151946,
151808, 152412, 152363, 152305, 153156, 152733, 152810, 153157, 152016, 152100, 152069, 153234, 152317,
152589, 152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504, 153376, 152272, 152433, 152325,
151941, 151670, 198, 285, 155788, 151669, 152238, 152255, 153427, 152318, 153009, 152381, 152474, 152680,
152157, 153255, 152324, 151682, 151670, 198, 32955, 155804, 151669, 153490, 153419, 152364, 152405, 152682,
152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488, 153070, 151883, 152890, 152489, 153144,
153375, 152358, 151685, 152494, 152117, 152740, 151670, 198, 37448, 480, 155840, 151669, 151902, 152720,
153377, 152027, 152378, 152821, 153207, 153459, 153028, 153068, 152507, 153255, 152158, 152921, 151958,
152609, 152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470, 152606, 152162, 152186, 153071,
152244, 153118, 153375, 153018, 152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736, 153380,
153502, 152702, 152115, 153181, 152735, 153277, 153457, 152393, 153112, 152595, 151670, 198, 19098, 155808,
151669, 152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239, 153163, 152922, 153402, 152034,
152591, 153438, 152215, 151673, 152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482, 152718,
152862, 153347, 151670, 198, 72, 155780, 151669, 151795, 152111, 152746, 152377, 153471, 152309, 151670, 198,
19016, 155788, 151669, 153181, 152271, 152190, 152842, 152224, 152701, 152939, 152536, 152091, 151815, 152733,
151672, 151670, 198, 14689, 155788, 151669, 152291, 152072, 152942, 151734, 153042, 153504, 152589, 153333,
151839, 151941, 153038, 153180, 151670, 198, 36996, 8303, 155832, 151669, 152231, 152256, 152835, 152801,
152985, 153400, 152393, 152818, 152765, 152249, 152600, 151699, 152302, 152752, 153018, 153009, 151992,
153054, 152847, 153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458, 152048, 152757, 152428,
153195, 151906, 153006, 153178, 153250, 152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418,
152228, 152733, 151670, 198, 9096, 155801, 151669, 151698, 153321, 152217, 153039, 152935, 153400, 152122,
152531, 153106, 152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851, 152901, 152885, 152594,
153446, 153080, 151670, 198, 14689, 155795, 151669, 152658, 151700, 153321, 152450, 152530, 153191, 151673,
151690, 151698, 152714, 152846, 152981, 153171, 153384, 153364, 153188, 153246, 151670, 198, 1055, 155779,
151669, 151869, 152388, 152711, 153334, 151736, 151670, 198, 1782, 155780, 151669, 153483, 153240, 152241,
152558, 152697, 153046, 151670, 198, 5804, 1363, 155820, 151669, 152941, 152764, 152605, 153034, 153434,
153372, 153347, 151887, 152453, 152758, 152133, 152510, 152694, 152431, 152321, 153088, 152676, 152223,
152581, 152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032, 152903, 152859, 152989, 151748,
152669, 152661, 152650, 152409, 151861, 151670, 198, 300, 7973, 155828, 151669, 153095, 152469, 152988,
152894, 151819, 152391, 153019, 152058, 153062, 153230, 151826, 152112, 152306, 152264, 152769, 153390,
152384, 152435, 152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540, 151919, 151893, 152558,
152817, 152946, 152956, 152129, 152715, 153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450,
151670, 198, 8088, 155792, 151669, 152452, 153497, 153353, 152679, 152533, 152382, 152374, 152611, 153341,
153163, 152285, 153411, 152495, 153141, 152320, 151670, 198, 1199, 155781, 151669, 151764, 152360, 153295,
152634, 153342, 152199, 152271, 151670, 198, 43366, 155799, 151669, 152308, 151682, 152889, 152016, 152385,
152629, 152495, 151826, 153321, 152958, 152180, 151886, 153432, 152922, 152128, 153024, 153040, 152593,
152287, 151677, 151670, 198, 53660, 155808, 151669, 151727, 152092, 152680, 153331, 151699, 152316, 152938,
152289, 152433, 153384, 151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691, 152489, 151941,
152049, 152034, 153053, 152179, 153160, 151676, 153367, 151670, 198, 268, 4123, 480, 155821, 151669, 152350,
152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234, 153135, 152291, 153235, 152143, 152583,
152402, 153483, 152678, 152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825, 152548, 153442,
152109, 152659, 153325, 152781, 152570, 152957, 151752, 152265, 153381, 152515, 151670, 198, 437, 155787,
151669, 152957, 152659, 151975, 152709, 152402, 152836, 152174, 151792, 153409, 153327, 152990, 151670, 198,
275, 155781, 151669, 152520, 153038, 152067, 153273, 153185, 152265, 152974, 151670, 198, 94273, 155799,
151669, 152953, 152938, 153427, 152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331, 152257,
152987, 152777, 153448, 152408, 151696, 152408, 152326, 152699, 151670, 198, 385, 16239, 155828, 151669,
152306, 152268, 153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110, 152918, 152923, 152467,
152331, 153053, 153330, 151889, 153444, 152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751,
152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499, 152109, 152255, 151739, 152267, 152759,
153318, 153165, 153349, 151670, ]
response = requests.post(
host_llm + "/completion",
json={
"prompt": [prefix + words, *suffix],
"n_predict": 1024,
"cache_prompt": True,
"return_tokens": True,
"samplers": ["top_k"],
"top_k": 16,
"seed": 1003,
}
)
response_json = response.json()
#print(json.dumps(response_json, indent=4))
#print(json.dumps(response_json["prompt"], indent=4).replace("\\n", "\n"))
#print(json.dumps(response_json["timings"], indent=4))
#print(json.dumps(response_json["tokens"], indent=4))
codes = response_json["tokens"]
codes = [t - 151672 for t in codes if t >= 151672 and t <= 155772]
response = requests.post(
host_dec + "/embeddings",
json={
"input": [*codes],
}
)
response_json = response.json()
#print(json.dumps(response_json, indent=4))
# spectrogram
embd = response_json[0]["embedding"]
n_codes = len(embd)
n_embd = len(embd[0])
print('spectrogram generated: n_codes: %d, n_embd: %d' % (n_codes, n_embd))
# post-process the spectrogram to convert to audio
# TODO: see the tts.cpp:embd_to_audio() and implement it in Python
print('converting to audio ...')
print('TODO: see the tts.cpp:embd_to_audio() and implement it in Python')

932
examples/tts/tts.cpp Normal file
View file

@ -0,0 +1,932 @@
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#define _USE_MATH_DEFINES // For M_PI on MSVC
#include <algorithm>
#include <cmath>
#include <cstdio>
#include <fstream>
#include <map>
#include <regex>
#include <string>
#include <thread>
#include <vector>
//
// Terminal utils
//
#define SQR(X) ((X) * (X))
#define UNCUBE(x) x < 48 ? 0 : x < 115 ? 1 : (x - 35) / 40
/**
* Quantizes 24-bit RGB to xterm256 code range [16,256).
*/
static int rgb2xterm256(int r, int g, int b) {
unsigned char cube[] = {0, 0137, 0207, 0257, 0327, 0377};
int av, ir, ig, ib, il, qr, qg, qb, ql;
av = r * .299 + g * .587 + b * .114 + .5;
ql = (il = av > 238 ? 23 : (av - 3) / 10) * 10 + 8;
qr = cube[(ir = UNCUBE(r))];
qg = cube[(ig = UNCUBE(g))];
qb = cube[(ib = UNCUBE(b))];
if (SQR(qr - r) + SQR(qg - g) + SQR(qb - b) <=
SQR(ql - r) + SQR(ql - g) + SQR(ql - b))
return ir * 36 + ig * 6 + ib + 020;
return il + 0350;
}
static std::string set_xterm256_foreground(int r, int g, int b) {
int x = rgb2xterm256(r, g, b);
std::ostringstream oss;
oss << "\033[38;5;" << x << "m";
return oss.str();
}
const std::vector<std::string> k_colors = {
set_xterm256_foreground(220, 5, 12),
set_xterm256_foreground(232, 96, 28),
set_xterm256_foreground(241, 147, 45),
set_xterm256_foreground(246, 193, 65),
set_xterm256_foreground(247, 240, 86),
set_xterm256_foreground(144, 201, 135),
set_xterm256_foreground( 78, 178, 101),
};
static void print_usage(int, char ** argv) {
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -p \"Hello!\"\n", argv[0]);
LOG("\n");
}
struct wav_header {
char riff[4] = {'R', 'I', 'F', 'F'};
uint32_t chunk_size;
char wave[4] = {'W', 'A', 'V', 'E'};
char fmt[4] = {'f', 'm', 't', ' '};
uint32_t fmt_chunk_size = 16;
uint16_t audio_format = 1; // PCM
uint16_t num_channels = 1; // Mono
uint32_t sample_rate;
uint32_t byte_rate;
uint16_t block_align;
uint16_t bits_per_sample = 16;
char data[4] = {'d', 'a', 't', 'a'};
uint32_t data_size;
};
static void save_wav16(const std::string & fname, const std::vector<float> & data, int sample_rate) {
std::ofstream file(fname, std::ios::binary);
if (!file) {
LOG_ERR("%s: Failed to open file '%s' for writing", __func__, fname.c_str());
return;
}
wav_header header;
header.sample_rate = sample_rate;
header.byte_rate = header.sample_rate * header.num_channels * (header.bits_per_sample / 8);
header.block_align = header.num_channels * (header.bits_per_sample / 8);
header.data_size = data.size() * (header.bits_per_sample / 8);
header.chunk_size = 36 + header.data_size;
file.write(reinterpret_cast<const char*>(&header), sizeof(header));
for (const auto & sample : data) {
int16_t pcm_sample = static_cast<int16_t>(std::clamp(sample * 32767.0, -32768.0, 32767.0));
file.write(reinterpret_cast<const char*>(&pcm_sample), sizeof(pcm_sample));
}
file.close();
}
static void fill_hann_window(int length, bool periodic, float * output) {
int offset = -1;
if (periodic) {
offset = 0;
}
for (int i = 0; i < length; i++) {
output[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
}
}
// very poor-man fft
static void twiddle(float * real, float * imag, int k, int N) {
float angle = 2 * M_PI * k / N;
*real = cos(angle);
*imag = sin(angle);
}
static void irfft(int n, const float * inp_cplx, float * out_real) {
int N = n / 2 + 1;
std::vector<float> real_input(N);
std::vector<float> imag_input(N);
for (int i = 0; i < N; ++i) {
real_input[i] = inp_cplx[2 * i];
imag_input[i] = inp_cplx[2 * i + 1];
}
std::vector<float> real_output(n);
std::vector<float> imag_output(n);
for (int k = 0; k < n; ++k) {
real_output[k] = 0.0f;
imag_output[k] = 0.0f;
for (int m = 0; m < N; ++m) {
float twiddle_real;
float twiddle_imag;
twiddle(&twiddle_real, &twiddle_imag, k * m, n);
real_output[k] += real_input[m] * twiddle_real - imag_input[m] * twiddle_imag;
imag_output[k] += real_input[m] * twiddle_imag + imag_input[m] * twiddle_real;
}
}
for (int i = 0; i < n; ++i) {
out_real[i] = real_output[i] / N;
}
}
//
// y = torch.nn.functional.fold(
// data, output_size=(1, output_size), kernel_size=(1, self.win_length), stride=(1, self.hop_length),
// )[:, 0, 0, pad:-pad]
//
// data.shape = torch.Size([1, 1280, 261])
// output_size = 84480
// win_length = 1280
// hop_length = 320
// pad = 480
//
static void fold(const std::vector<float> & data, int64_t n_out, int64_t n_win, int64_t n_hop, int64_t n_pad, std::vector<float> & output) {
int64_t output_height = n_out;
int64_t kernel_w = n_win;
int64_t stride_w = n_hop;
int64_t width = n_out;
output.resize(width, 0.0f);
int64_t col_idx = 0;
for (int64_t w_col = 0; w_col < width; ++w_col) {
int64_t start = w_col * stride_w - n_pad;
int64_t end = start + kernel_w;
for (int64_t w_im = start; w_im < end; ++w_im) {
if (w_im >= 0 && w_im < output_height && col_idx < (int64_t) data.size()) {
output[w_im] += data[col_idx];
}
col_idx++;
}
}
output.resize(n_out - 2 * n_pad);
}
// TODO: not optimized at all
static std::vector<float> embd_to_audio(
const float * embd,
const int n_codes,
const int n_embd,
const int n_thread) {
const int n_fft = 1280;
const int n_hop = 320;
const int n_win = 1280;
const int n_pad = (n_win - n_hop)/2;
const int n_out = (n_codes - 1)*n_hop + n_win;
std::vector<float> hann(n_fft);
fill_hann_window(hann.size(), true, hann.data());
int n_spec = n_embd*n_codes;
std::vector<float> E (n_spec);
std::vector<float> S (n_spec);
std::vector<float> ST(n_spec);
for (int l = 0; l < n_codes; ++l) {
for (int k = 0; k < n_embd; ++k) {
E[k*n_codes + l] = embd[l*n_embd + k];
}
}
for (int k = 0; k < n_embd/2; ++k) {
for (int l = 0; l < n_codes; ++l) {
float mag = E[(k )*n_codes + l];
float phi = E[(k + n_embd/2)*n_codes + l];
mag = exp(mag);
if (mag > 1e2) {
mag = 1e2;
}
S[2*(k*n_codes + l) + 0] = mag*cosf(phi);
S[2*(k*n_codes + l) + 1] = mag*sinf(phi);
}
}
for (int l = 0; l < n_codes; ++l) {
for (int k = 0; k < n_embd/2; ++k) {
ST[l*n_embd + 2*k + 0] = S[2*(k*n_codes + l) + 0];
ST[l*n_embd + 2*k + 1] = S[2*(k*n_codes + l) + 1];
}
}
std::vector<float> res (n_codes*n_fft);
std::vector<float> hann2(n_codes*n_fft);
std::vector<std::thread> workers(n_thread);
for (int i = 0; i < n_thread; ++i) {
workers[i] = std::thread([&, i]() {
for (int l = i; l < n_codes; l += n_thread) {
irfft(n_fft, ST.data() + l*n_embd, res.data() + l*n_fft);
for (int j = 0; j < n_fft; ++j) {
res [l*n_fft + j] *= hann[j];
hann2[l*n_fft + j] = hann[j] * hann[j];
}
}
});
}
for (int i = 0; i < n_thread; ++i) {
workers[i].join();
}
std::vector<float> audio;
std::vector<float> env;
fold(res, n_out, n_win, n_hop, n_pad, audio);
fold(hann2, n_out, n_win, n_hop, n_pad, env); // TODO: can be done once
for (size_t i = 0; i < audio.size(); ++i) {
audio[i] /= env[i];
}
return audio;
}
static const std::map<int, std::string> ones = {
{0, "zero"}, {1, "one"}, {2, "two"}, {3, "three"}, {4, "four"},
{5, "five"}, {6, "six"}, {7, "seven"}, {8, "eight"}, {9, "nine"},
{10, "ten"}, {11, "eleven"}, {12, "twelve"}, {13, "thirteen"}, {14, "fourteen"},
{15, "fifteen"}, {16, "sixteen"}, {17, "seventeen"}, {18, "eighteen"}, {19, "nineteen"}
};
static const std::map<int, std::string> tens = {
{2, "twenty"}, {3, "thirty"}, {4, "forty"}, {5, "fifty"},
{6, "sixty"}, {7, "seventy"}, {8, "eighty"}, {9, "ninety"}
};
// Convert a number less than 1000 to words
static std::string convert_less_than_thousand(int num) {
std::string result;
if (num >= 100) {
result += ones.at(num / 100) + " hundred ";
num %= 100;
}
if (num >= 20) {
result += tens.at(num / 10);
if (num % 10 > 0) {
result += "-" + ones.at(num % 10);
}
} else if (num > 0) {
result += ones.at(num);
}
return result;
}
static std::string number_to_words(const std::string & number_str) {
try {
size_t decimal_pos = number_str.find('.');
std::string integer_part = number_str.substr(0, decimal_pos);
int int_number = std::stoi(integer_part);
std::string result;
if (int_number == 0) {
result = "zero";
} else {
if (int_number >= 1000000000) {
int billions = int_number / 1000000000;
result += convert_less_than_thousand(billions) + " billion ";
int_number %= 1000000000;
}
if (int_number >= 1000000) {
int millions = int_number / 1000000;
result += convert_less_than_thousand(millions) + " million ";
int_number %= 1000000;
}
if (int_number >= 1000) {
int thousands = int_number / 1000;
result += convert_less_than_thousand(thousands) + " thousand ";
int_number %= 1000;
}
if (int_number > 0) {
result += convert_less_than_thousand(int_number);
}
}
// Handle decimal part
if (decimal_pos != std::string::npos) {
result += " point";
std::string decimal_part = number_str.substr(decimal_pos + 1);
for (char digit : decimal_part) {
result += " " + ones.at(digit - '0');
}
}
return result;
} catch (const std::exception& e) {
// Skip if fails
return " ";
}
}
static std::string replace_numbers_with_words(const std::string & input_text) {
std::regex number_pattern(R"(\d+(\.\d+)?)");
std::string result;
auto it = std::sregex_iterator(input_text.begin(), input_text.end(), number_pattern);
auto end = std::sregex_iterator();
size_t last_pos = 0;
for (std::sregex_iterator i = it; i != end; ++i) {
const std::smatch& match = *i;
result.append(input_text, last_pos, match.position() - last_pos);
result.append(number_to_words(match.str()));
last_pos = match.position() + match.length();
}
result.append(input_text, last_pos);
return result;
}
// Based on: https://github.com/edwko/OuteTTS/blob/a613e79c489d8256dd657ea9168d78de75895d82/outetts/version/v1/prompt_processor.py#L39
static std::string process_text(const std::string & text) {
// For now I skipped text romanization as I am unsure how to handle
// uroman and MeCab implementations in C++
// maybe something like https://github.com/anyascii/anyascii/ could work.
// currently only English would be supported in this function
std::string processed_text = replace_numbers_with_words(text);
std::transform(processed_text.begin(), processed_text.end(),
processed_text.begin(), ::tolower);
std::regex special_chars(R"([-_/,\.\\])");
processed_text = std::regex_replace(processed_text, special_chars, " ");
std::regex non_alpha(R"([^a-z\s])");
processed_text = std::regex_replace(processed_text, non_alpha, "");
std::regex multiple_spaces(R"(\s+)");
processed_text = std::regex_replace(processed_text, multiple_spaces, " ");
processed_text = std::regex_replace(processed_text, std::regex(R"(^\s+|\s+$)"), "");
/*
Replace spaces with the separator token same as in line 365
for (auto & c : prompt_user) {
if (c == ' ') {
prompt_clean += "<|text_sep|>";
*/
processed_text = std::regex_replace(processed_text, std::regex(R"(\s)"), "<|text_sep|>");
return processed_text;
}
static void prompt_add(llama_tokens & prompt, llama_token token) {
prompt.push_back(token);
}
static void prompt_add(llama_tokens & prompt, const llama_tokens & tokens) {
prompt.insert(prompt.end(), tokens.begin(), tokens.end());
}
static void prompt_add(llama_tokens & prompt, const llama_model * model, const std::string & txt, bool add_special, bool parse_special) {
auto tmp = common_tokenize(model, txt, add_special, parse_special);
prompt_add(prompt, tmp);
}
static void prompt_init(llama_tokens & prompt, const llama_model * model) {
prompt.clear();
prompt_add(prompt, model, "<|im_start|>\n", true, true);
}
int main(int argc, char ** argv) {
common_params params;
params.prompt = "";
params.n_predict = 4096;
params.n_batch = 8192;
params.n_ctx = 8192;
params.sampling.top_k = 4;
params.sampling.samplers = { COMMON_SAMPLER_TYPE_TOP_K, };
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_TTS, print_usage)) {
return 1;
}
const int n_parallel = params.n_parallel;
const int n_predict = params.n_predict;
common_init();
// init LLM
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model_ttc = NULL; // text-to-codes
llama_model * model_cts = NULL; // codes-to-speech
llama_context * ctx_ttc = NULL;
llama_context * ctx_cts = NULL;
common_init_result llama_init_ttc = common_init_from_params(params);
model_ttc = llama_init_ttc.model;
ctx_ttc = llama_init_ttc.context;
// TODO: refactor in a common struct
params.model = params.vocoder.model;
params.model_url = params.vocoder.model_url;
params.hf_repo = params.vocoder.hf_repo;
params.hf_file = params.vocoder.hf_file;
params.embedding = true;
common_init_result llama_init_cts = common_init_from_params(params);
model_cts = llama_init_cts.model;
ctx_cts = llama_init_cts.context;
std::vector<common_sampler *> smpl(n_parallel);
for (int i = 0; i < n_parallel; ++i) {
params.sampling.no_perf = (i != 0);
params.sampling.seed = params.sampling.seed + 1;
smpl[i] = common_sampler_init(model_ttc, params.sampling);
}
LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl[0]));
LOG_INF("sampler params: \n%s\n", params.sampling.print().c_str());
LOG_INF("sampler chain: %s\n", common_sampler_print(smpl[0]).c_str());
LOG_INF("%s: loading done\n", __func__);
const auto t_main_start = ggml_time_us();
std::vector<llama_token> codes;
// process prompt and generate voice codes
{
LOG_INF("%s: constructing prompt ..\n", __func__);
std::vector<llama_token> prompt_inp;
prompt_init(prompt_inp, model_ttc);
prompt_add(prompt_inp, model_ttc, "<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>", false, true);
// convert the input text into the necessary format expected by OuteTTS
{
std::string prompt_clean = process_text(params.prompt);
LOG_INF("%s: prompt: '%s'\n", __func__, prompt_clean.c_str());
prompt_add(prompt_inp, model_ttc, prompt_clean, false, true);
}
prompt_add(prompt_inp, model_ttc, "<|text_end|>\n", false, true);
// disabled to save time on tokenizing each time
// TODO: load voices from the json files
#if 0
const std::string voice_data = R"(<|audio_start|>
the<|t_0.08|><|code_start|><|257|><|740|><|636|><|913|><|788|><|1703|><|code_end|>
overall<|t_0.36|><|code_start|><|127|><|201|><|191|><|774|><|700|><|532|><|1056|><|557|><|798|><|298|><|1741|><|747|><|1662|><|1617|><|1702|><|1527|><|368|><|1588|><|1049|><|1008|><|1625|><|747|><|1576|><|728|><|1019|><|1696|><|1765|><|code_end|>
package<|t_0.56|><|code_start|><|935|><|584|><|1319|><|627|><|1016|><|1491|><|1344|><|1117|><|1526|><|1040|><|239|><|1435|><|951|><|498|><|723|><|1180|><|535|><|789|><|1649|><|1637|><|78|><|465|><|1668|><|901|><|595|><|1675|><|117|><|1009|><|1667|><|320|><|840|><|79|><|507|><|1762|><|1508|><|1228|><|1768|><|802|><|1450|><|1457|><|232|><|639|><|code_end|>
from<|t_0.19|><|code_start|><|604|><|782|><|1682|><|872|><|1532|><|1600|><|1036|><|1761|><|647|><|1554|><|1371|><|653|><|1595|><|950|><|code_end|>
just<|t_0.25|><|code_start|><|1782|><|1670|><|317|><|786|><|1748|><|631|><|599|><|1155|><|1364|><|1524|><|36|><|1591|><|889|><|1535|><|541|><|440|><|1532|><|50|><|870|><|code_end|>
two<|t_0.24|><|code_start|><|1681|><|1510|><|673|><|799|><|805|><|1342|><|330|><|519|><|62|><|640|><|1138|><|565|><|1552|><|1497|><|1552|><|572|><|1715|><|1732|><|code_end|>
people<|t_0.39|><|code_start|><|593|><|274|><|136|><|740|><|691|><|633|><|1484|><|1061|><|1138|><|1485|><|344|><|428|><|397|><|1562|><|645|><|917|><|1035|><|1449|><|1669|><|487|><|442|><|1484|><|1329|><|1832|><|1704|><|600|><|761|><|653|><|269|><|code_end|>
is<|t_0.16|><|code_start|><|566|><|583|><|1755|><|646|><|1337|><|709|><|802|><|1008|><|485|><|1583|><|652|><|10|><|code_end|>
pretty<|t_0.32|><|code_start|><|1818|><|1747|><|692|><|733|><|1010|><|534|><|406|><|1697|><|1053|><|1521|><|1355|><|1274|><|816|><|1398|><|211|><|1218|><|817|><|1472|><|1703|><|686|><|13|><|822|><|445|><|1068|><|code_end|>
remarkable<|t_0.68|><|code_start|><|230|><|1048|><|1705|><|355|><|706|><|1149|><|1535|><|1787|><|1356|><|1396|><|835|><|1583|><|486|><|1249|><|286|><|937|><|1076|><|1150|><|614|><|42|><|1058|><|705|><|681|><|798|><|934|><|490|><|514|><|1399|><|572|><|1446|><|1703|><|1346|><|1040|><|1426|><|1304|><|664|><|171|><|1530|><|625|><|64|><|1708|><|1830|><|1030|><|443|><|1509|><|1063|><|1605|><|1785|><|721|><|1440|><|923|><|code_end|>
sure<|t_0.36|><|code_start|><|792|><|1780|><|923|><|1640|><|265|><|261|><|1525|><|567|><|1491|><|1250|><|1730|><|362|><|919|><|1766|><|543|><|1|><|333|><|113|><|970|><|252|><|1606|><|133|><|302|><|1810|><|1046|><|1190|><|1675|><|code_end|>
i<|t_0.08|><|code_start|><|123|><|439|><|1074|><|705|><|1799|><|637|><|code_end|>
have<|t_0.16|><|code_start|><|1509|><|599|><|518|><|1170|><|552|><|1029|><|1267|><|864|><|419|><|143|><|1061|><|0|><|code_end|>
some<|t_0.16|><|code_start|><|619|><|400|><|1270|><|62|><|1370|><|1832|><|917|><|1661|><|167|><|269|><|1366|><|1508|><|code_end|>
critiques<|t_0.60|><|code_start|><|559|><|584|><|1163|><|1129|><|1313|><|1728|><|721|><|1146|><|1093|><|577|><|928|><|27|><|630|><|1080|><|1346|><|1337|><|320|><|1382|><|1175|><|1682|><|1556|><|990|><|1683|><|860|><|1721|><|110|><|786|><|376|><|1085|><|756|><|1523|><|234|><|1334|><|1506|><|1578|><|659|><|612|><|1108|><|1466|><|1647|><|308|><|1470|><|746|><|556|><|1061|><|code_end|>
about<|t_0.29|><|code_start|><|26|><|1649|><|545|><|1367|><|1263|><|1728|><|450|><|859|><|1434|><|497|><|1220|><|1285|><|179|><|755|><|1154|><|779|><|179|><|1229|><|1213|><|922|><|1774|><|1408|><|code_end|>
some<|t_0.23|><|code_start|><|986|><|28|><|1649|><|778|><|858|><|1519|><|1|><|18|><|26|><|1042|><|1174|><|1309|><|1499|><|1712|><|1692|><|1516|><|1574|><|code_end|>
of<|t_0.07|><|code_start|><|197|><|716|><|1039|><|1662|><|64|><|code_end|>
the<|t_0.08|><|code_start|><|1811|><|1568|><|569|><|886|><|1025|><|1374|><|code_end|>
gameplay<|t_0.48|><|code_start|><|1269|><|1092|><|933|><|1362|><|1762|><|1700|><|1675|><|215|><|781|><|1086|><|461|><|838|><|1022|><|759|><|649|><|1416|><|1004|><|551|><|909|><|787|><|343|><|830|><|1391|><|1040|><|1622|><|1779|><|1360|><|1231|><|1187|><|1317|><|76|><|997|><|989|><|978|><|737|><|189|><|code_end|>
aspects<|t_0.56|><|code_start|><|1423|><|797|><|1316|><|1222|><|147|><|719|><|1347|><|386|><|1390|><|1558|><|154|><|440|><|634|><|592|><|1097|><|1718|><|712|><|763|><|1118|><|1721|><|1311|><|868|><|580|><|362|><|1435|><|868|><|247|><|221|><|886|><|1145|><|1274|><|1284|><|457|><|1043|><|1459|><|1818|><|62|><|599|><|1035|><|62|><|1649|><|778|><|code_end|>
but<|t_0.20|><|code_start|><|780|><|1825|><|1681|><|1007|><|861|><|710|><|702|><|939|><|1669|><|1491|><|613|><|1739|><|823|><|1469|><|648|><|code_end|>
its<|t_0.09|><|code_start|><|92|><|688|><|1623|><|962|><|1670|><|527|><|599|><|code_end|>
still<|t_0.27|><|code_start|><|636|><|10|><|1217|><|344|><|713|><|957|><|823|><|154|><|1649|><|1286|><|508|><|214|><|1760|><|1250|><|456|><|1352|><|1368|><|921|><|615|><|5|><|code_end|>
really<|t_0.36|><|code_start|><|55|><|420|><|1008|><|1659|><|27|><|644|><|1266|><|617|><|761|><|1712|><|109|><|1465|><|1587|><|503|><|1541|><|619|><|197|><|1019|><|817|><|269|><|377|><|362|><|1381|><|507|><|1488|><|4|><|1695|><|code_end|>
enjoyable<|t_0.49|><|code_start|><|678|><|501|><|864|><|319|><|288|><|1472|><|1341|><|686|><|562|><|1463|><|619|><|1563|><|471|><|911|><|730|><|1811|><|1006|><|520|><|861|><|1274|><|125|><|1431|><|638|><|621|><|153|><|876|><|1770|><|437|><|987|><|1653|><|1109|><|898|><|1285|><|80|><|593|><|1709|><|843|><|code_end|>
and<|t_0.15|><|code_start|><|1285|><|987|><|303|><|1037|><|730|><|1164|><|502|><|120|><|1737|><|1655|><|1318|><|code_end|>
it<|t_0.09|><|code_start|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|><|code_end|>
looks<|t_0.27|><|code_start|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|code_end|>
lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|code_end|>)";
auto tmp = common_tokenize(model_ttc, voice_data, false, true);
printf("\n\n");
for (int i = 0; i < tmp.size(); ++i) {
printf("%d, ", tmp[i]);
}
printf("\n\n");
#else
prompt_add(prompt_inp, llama_tokens {
151667, 198, 1782, 155780, 151669, 151929, 152412, 152308, 152585,
152460, 153375, 151670, 198, 74455, 155808, 151669, 151799,
151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470,
151970, 153413, 152419, 153334, 153289, 153374, 153199, 152040,
153260, 152721, 152680, 153297, 152419, 153248, 152400, 152691,
153368, 153437, 151670, 198, 1722, 155828, 151669, 152607,
152256, 152991, 152299, 152688, 153163, 153016, 152789, 153198,
152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207,
152461, 153321, 153309, 151750, 152137, 153340, 152573, 152267,
153347, 151789, 152681, 153339, 151992, 152512, 151751, 152179,
153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904,
152311, 151670, 198, 1499, 155791, 151669, 152276, 152454,
153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226,
153043, 152325, 153267, 152622, 151670, 198, 4250, 155797,
151669, 153454, 153342, 151989, 152458, 153420, 152303, 152271,
152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213,
152112, 153204, 151722, 152542, 151670, 198, 19789, 155796,
151669, 153353, 153182, 152345, 152471, 152477, 153014, 152002,
152191, 151734, 152312, 152810, 152237, 153224, 153169, 153224,
152244, 153387, 153404, 151670, 198, 16069, 155811, 151669,
152265, 151946, 151808, 152412, 152363, 152305, 153156, 152733,
152810, 153157, 152016, 152100, 152069, 153234, 152317, 152589,
152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504,
153376, 152272, 152433, 152325, 151941, 151670, 198, 285,
155788, 151669, 152238, 152255, 153427, 152318, 153009, 152381,
152474, 152680, 152157, 153255, 152324, 151682, 151670, 198,
32955, 155804, 151669, 153490, 153419, 152364, 152405, 152682,
152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488,
153070, 151883, 152890, 152489, 153144, 153375, 152358, 151685,
152494, 152117, 152740, 151670, 198, 37448, 480, 155840, 151669,
151902, 152720, 153377, 152027, 152378, 152821, 153207, 153459,
153028, 153068, 152507, 153255, 152158, 152921, 151958, 152609,
152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470,
152606, 152162, 152186, 153071, 152244, 153118, 153375, 153018,
152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736,
153380, 153502, 152702, 152115, 153181, 152735, 153277, 153457,
152393, 153112, 152595, 151670, 198, 19098, 155808, 151669,
152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239,
153163, 152922, 153402, 152034, 152591, 153438, 152215, 151673,
152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482,
152718, 152862, 153347, 151670, 198, 72, 155780, 151669, 151795,
152111, 152746, 152377, 153471, 152309, 151670, 198, 19016,
155788, 151669, 153181, 152271, 152190, 152842, 152224, 152701,
152939, 152536, 152091, 151815, 152733, 151672, 151670, 198,
14689, 155788, 151669, 152291, 152072, 152942, 151734, 153042,
153504, 152589, 153333, 151839, 151941, 153038, 153180, 151670,
198, 36996, 8303, 155832, 151669, 152231, 152256, 152835,
152801, 152985, 153400, 152393, 152818, 152765, 152249, 152600,
151699, 152302, 152752, 153018, 153009, 151992, 153054, 152847,
153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458,
152048, 152757, 152428, 153195, 151906, 153006, 153178, 153250,
152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418,
152228, 152733, 151670, 198, 9096, 155801, 151669, 151698,
153321, 152217, 153039, 152935, 153400, 152122, 152531, 153106,
152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851,
152901, 152885, 152594, 153446, 153080, 151670, 198, 14689,
155795, 151669, 152658, 151700, 153321, 152450, 152530, 153191,
151673, 151690, 151698, 152714, 152846, 152981, 153171, 153384,
153364, 153188, 153246, 151670, 198, 1055, 155779, 151669,
151869, 152388, 152711, 153334, 151736, 151670, 198, 1782,
155780, 151669, 153483, 153240, 152241, 152558, 152697, 153046,
151670, 198, 5804, 1363, 155820, 151669, 152941, 152764, 152605,
153034, 153434, 153372, 153347, 151887, 152453, 152758, 152133,
152510, 152694, 152431, 152321, 153088, 152676, 152223, 152581,
152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032,
152903, 152859, 152989, 151748, 152669, 152661, 152650, 152409,
151861, 151670, 198, 300, 7973, 155828, 151669, 153095, 152469,
152988, 152894, 151819, 152391, 153019, 152058, 153062, 153230,
151826, 152112, 152306, 152264, 152769, 153390, 152384, 152435,
152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540,
151919, 151893, 152558, 152817, 152946, 152956, 152129, 152715,
153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450,
151670, 198, 8088, 155792, 151669, 152452, 153497, 153353,
152679, 152533, 152382, 152374, 152611, 153341, 153163, 152285,
153411, 152495, 153141, 152320, 151670, 198, 1199, 155781,
151669, 151764, 152360, 153295, 152634, 153342, 152199, 152271,
151670, 198, 43366, 155799, 151669, 152308, 151682, 152889,
152016, 152385, 152629, 152495, 151826, 153321, 152958, 152180,
151886, 153432, 152922, 152128, 153024, 153040, 152593, 152287,
151677, 151670, 198, 53660, 155808, 151669, 151727, 152092,
152680, 153331, 151699, 152316, 152938, 152289, 152433, 153384,
151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691,
152489, 151941, 152049, 152034, 153053, 152179, 153160, 151676,
153367, 151670, 198, 268, 4123, 480, 155821, 151669, 152350,
152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234,
153135, 152291, 153235, 152143, 152583, 152402, 153483, 152678,
152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825,
152548, 153442, 152109, 152659, 153325, 152781, 152570, 152957,
151752, 152265, 153381, 152515, 151670, 198, 437, 155787,
151669, 152957, 152659, 151975, 152709, 152402, 152836, 152174,
151792, 153409, 153327, 152990, 151670, 198, 275, 155781,
151669, 152520, 153038, 152067, 153273, 153185, 152265, 152974,
151670, 198, 94273, 155799, 151669, 152953, 152938, 153427,
152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331,
152257, 152987, 152777, 153448, 152408, 151696, 152408, 152326,
152699, 151670, 198, 385, 16239, 155828, 151669, 152306, 152268,
153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110,
152918, 152923, 152467, 152331, 153053, 153330, 151889, 153444,
152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751,
152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499,
152109, 152255, 151739, 152267, 152759, 153318, 153165, 153349,
151670,});
#endif
// print the prompt token-by-token
LOG("\n");
for (auto id : prompt_inp) {
LOG("%s", common_token_to_piece(ctx_ttc, id).c_str());
}
LOG_INF("%s: prompt size: %d\n", __func__, (int) prompt_inp.size());
LOG("\n");
// create a llama_batch
// we use this object to submit token data for decoding
llama_batch batch = llama_batch_init(std::max(prompt_inp.size(), (size_t) n_parallel), 0, n_parallel);
std::vector<llama_seq_id> seq_ids(n_parallel, 0);
for (int32_t i = 0; i < n_parallel; ++i) {
seq_ids[i] = i;
}
// evaluate the initial prompt
for (size_t i = 0; i < prompt_inp.size(); ++i) {
common_batch_add(batch, prompt_inp[i], i, seq_ids, false);
}
GGML_ASSERT(batch.n_tokens == (int) prompt_inp.size());
// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;
if (llama_decode(ctx_ttc, batch) != 0) {
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
if (n_parallel > 1) {
LOG_INF("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
}
llama_synchronize(ctx_ttc);
LOG_INF("%s: time for prompt: %.3f ms\n\n", __func__, (ggml_time_us() - t_main_start) / 1000.0f);
const auto t_dec_start = ggml_time_us();
// main loop
// remember the batch index of the last token for each parallel sequence
// we need this to determine which logits to sample from
std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
int n_past = batch.n_tokens;
int n_decode = 0;
while (n_decode <= n_predict) {
// prepare the next batch
common_batch_clear(batch);
// sample the next token for each parallel sequence / stream
for (int32_t i = 0; i < n_parallel; ++i) {
if (i_batch[i] < 0) {
// the stream has already finished
continue;
}
const llama_token new_token_id = common_sampler_sample(smpl[i], ctx_ttc, i_batch[i]);
common_sampler_accept(smpl[i], new_token_id, true);
codes.push_back(new_token_id);
const auto * cands = common_sampler_get_candidates(smpl[i]);
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model_ttc, new_token_id) || n_decode == n_predict) {
std::string reason;
if (llama_token_is_eog(model_ttc, new_token_id)) {
reason = "eos";
} else {
reason = "n_predict";
}
i_batch[i] = -1;
LOG("\n");
if (n_parallel > 1) {
LOG_CNT("\n");
LOG_INF("%s: stream %d finished at n_past = %d, reason = '%s'\n", __func__, i, n_past, reason.c_str());
}
continue;
}
{
const float p = cands->data[cands->selected].p;
const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) ((3*p)*float(k_colors.size()))));
LOG_CNT("%s%d%s", k_colors[col].c_str(), i, "\033[0m");
//LOG_CNT("%d", i);
}
i_batch[i] = batch.n_tokens;
// push this new token for next evaluation
common_batch_add(batch, new_token_id, n_past, { i }, true);
}
// all streams are finished
if (batch.n_tokens == 0) {
break;
}
n_decode += 1;
n_past += 1;
// evaluate the current batch with the transformer model
if (llama_decode(ctx_ttc, batch)) {
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
llama_batch_free(batch);
LOG("\n");
LOG_INF("%s: time for decoder: %.3f ms\n", __func__, (ggml_time_us() - t_dec_start) / 1000.0f);
}
common_perf_print(ctx_ttc, smpl[0]);
//std::vector<llama_token> codes = {198, 88225, 155856, 151669, 152205,
// 153064, 152537, 153421, 153209, 152524, 151689, 152993, 152438, 152695,
// 153091, 152945, 152829, 152534, 152934, 153020, 151997, 152263, 153010,
// 153146, 152399, 153208, 152496, 151793, 152848, 152263, 152571, 153286,
// 152227, 153300, 152934, 152263, 153208, 152263, 152965, 152430, 152296,
// 153146, 152920, 152376, 152556, 153363, 151775, 152044, 152972, 152690,
// 153379, 152368, 152233, 153422, 152490, 151996, 152022, 151694, 152061,
// 153238, 152539, 153356, 152640, 153021, 153123, 151962, 153094, 151670,
// 198, 20339, 13189, 155824, 151669, 152070, 152007, 152910, 151683,
// 152000, 152373, 152760, 152046, 151735, 152334, 152394, 153073, 152908,
// 151856, 151953, 153247, 153293, 151903, 153480, 153168, 152478, 153359,
// 153429, 151905, 151678, 152567, 152411, 152165, 152556, 153075, 153424,
// 151993, 152999, 153078, 152151, 152088, 153389, 152484, 151874, 151670,
// 198, 285, 155784, 151669, 152226, 152126, 152638, 153215, 151729,
// 152959, 153479, 153059, 151838, 151670, 198, 1782, 155783, 151669,
// 153288, 153055, 153314, 152497, 152962, 152741, 152076, 153253, 151670,
// 198, 471, 16488, 155825, 151669, 152060, 152916, 151893, 153469, 152501,
// 152080, 152743, 151932, 153161, 152096, 152761, 152698, 153401, 153242,
// 153336, 152441, 152838, 153467, 152706, 153496, 153310, 152422, 153360,
// 153115, 152763, 151998, 152373, 153450, 152554, 151968, 153323, 152055,
// 152468, 153111, 153358, 152813, 152010, 151770, 152823, 152960, 151670,
// 198, 22627, 155823, 151669, 152814, 152366, 153484, 152931, 153441,
// 152164, 152877, 152915, 153463, 151692, 152911, 152747, 152776, 151831,
// 153449, 151882, 152975, 152031, 152513, 153150, 152448, 152667, 153133,
// 153189, 152619, 153466, 152054, 152106, 153119, 152277, 152439, 153109,
// 152997, 152141, 153154, 153256, 153311, 151922, 151670, 198, 1055,
// 155781, 151669, 152633, 151850, 153060, 153270, 152560, 153348, 152729,
// 151670, 198, 25312, 155803, 151669, 152521, 153403, 152561, 153337,
// 153383, 152199, 153493, 153326, 151830, 152254, 152248, 152349, 152153,
// 153007, 151823, 153037, 152575, 152457, 152406, 152592, 153116, 153365,
// 153456, 151670, 198, 88225, 155817, 151669, 153271, 151925, 152218,
// 152418, 152253, 153140, 151903, 153151, 152626, 152338, 152647, 153464,
// 152785, 152768, 151711, 152037, 152033, 151804, 152216, 151701, 151855,
// 152348, 152995, 152955, 152905, 152342, 152340, 153391, 153453, 152418,
// 153415, 151990, 153083, 152884, 151670, 198, 151668, 198, 151645};
{
const std::string inp_txt = common_detokenize(ctx_ttc, codes, true);
LOG("\n");
LOG_INF("codes: '%s'\n", inp_txt.c_str());
LOG_INF("%s: codes size: %d\n", __func__, (int) codes.size());
}
// remove all non-audio tokens (i.e. < 151672 || > 155772)
codes.erase(std::remove_if(codes.begin(), codes.end(), [](llama_token t) { return t < 151672 || t > 155772; }), codes.end());
{
const std::string inp_txt = common_detokenize(ctx_ttc, codes, true);
LOG_INF("codes audio: '%s'\n", inp_txt.c_str());
LOG_INF("%s: codes audio size: %d\n", __func__, (int) codes.size());
}
for (auto & token : codes) {
token -= 151672;
}
const auto t_voc_start = ggml_time_us();
const int n_codes = codes.size();
llama_batch batch = llama_batch_init(n_codes, 0, 1);
for (size_t i = 0; i < codes.size(); ++i) {
common_batch_add(batch, codes[i], i, { 0 }, true); // TODO: all logits?
}
GGML_ASSERT(batch.n_tokens == n_codes);
if (llama_decode(ctx_cts, batch) != 0) {
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
llama_synchronize(ctx_cts);
LOG_INF("%s: time for vocoder: %.3f ms\n", __func__, (ggml_time_us() - t_voc_start) / 1000.0f);
const auto t_spec_start = ggml_time_us();
#if 1
// spectral operations
const int n_embd = llama_n_embd(model_cts);
const float * embd = llama_get_embeddings(ctx_cts);
auto audio = embd_to_audio(embd, n_codes, n_embd, params.cpuparams.n_threads);
#else
// read the spectrogram from a file for debugging purposes
std::vector<float> audio;
{
std::ifstream fin("out.bin", std::ios::binary);
if (!fin) {
LOG_ERR("%s: failed to open file '%s'\n", __func__, "out.bin");
return 1;
}
std::vector<float> embd;
int n_codes;
int n_embd;
fin.read(reinterpret_cast<char *>(&n_codes), sizeof(int));
fin.read(reinterpret_cast<char *>(&n_embd), sizeof(int));
embd.resize(n_codes * n_embd);
fin.read(reinterpret_cast<char *>(embd.data()), n_codes * n_embd * sizeof(float));
fin.close();
LOG_INF("%s: n_codes: %d, n_embd: %d\n", __func__, n_codes, n_embd);
audio = embd_to_audio(embd.data(), n_codes, n_embd, params.cpuparams.n_threads);
}
#endif
const std::string fname = "output.wav";
const int n_sr = 24000; // sampling rate
// zero out first 0.25 seconds
for (int i = 0; i < 24000/4; ++i) {
audio[i] = 0.0f;
}
LOG_INF("%s: time for spectral ops: %.3f ms\n", __func__, (ggml_time_us() - t_spec_start) / 1000.0f);
LOG_INF("%s: total time: %.3f ms\n", __func__, (ggml_time_us() - t_main_start) / 1000.0f);
save_wav16(fname, audio, n_sr);
LOG_INF("%s: audio written to file '%s'\n", __func__, fname.c_str());
llama_free(ctx_ttc);
llama_free_model(model_ttc);
llama_free(ctx_cts);
llama_free_model(model_cts);
llama_backend_free();
return 0;
}

View file

@ -74,10 +74,10 @@ if (NOT GGML_CUDA_GRAPHS_DEFAULT)
endif()
# general
option(GGML_STATIC "ggml: static link libraries" OFF)
option(GGML_NATIVE "ggml: enable -march=native flag" ${GGML_NATIVE_DEFAULT})
option(GGML_LTO "ggml: enable link time optimization" OFF)
option(GGML_CCACHE "ggml: use ccache if available" ON)
option(GGML_STATIC "ggml: static link libraries" OFF)
option(GGML_NATIVE "ggml: optimize the build for the current system" ${GGML_NATIVE_DEFAULT})
option(GGML_LTO "ggml: enable link time optimization" OFF)
option(GGML_CCACHE "ggml: use ccache if available" ON)
# debug
option(GGML_ALL_WARNINGS "ggml: enable all compiler warnings" ON)
@ -120,8 +120,9 @@ endif()
option(GGML_LASX "ggml: enable lasx" ON)
option(GGML_LSX "ggml: enable lsx" ON)
option(GGML_RVV "ggml: enable rvv" ON)
option(GGML_SVE "ggml: enable SVE" OFF)
option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
if (WIN32)

View file

@ -1564,17 +1564,6 @@ extern "C" {
int d1, // dilation dimension 1
bool is_2D);
GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1
GGML_API struct ggml_tensor * ggml_conv_1d(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
@ -1592,6 +1581,23 @@ extern "C" {
int s, // stride
int d); // dilation
// depthwise
// TODO: this is very likely wrong for some cases! - needs more testing
GGML_API struct ggml_tensor * ggml_conv_1d_dw(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride
int p0, // padding
int d0); // dilation
GGML_API struct ggml_tensor * ggml_conv_1d_dw_ph(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride
int d0); // dilation
GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
@ -1611,7 +1617,6 @@ extern "C" {
int d0, // dilation dimension 0
int d1); // dilation dimension 1
// kernel size is a->ne[0] x a->ne[1]
// stride is equal to kernel size
// padding is zero
@ -1638,6 +1643,18 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// depthwise
GGML_API struct ggml_tensor * ggml_conv_2d_dw(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1
GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,

View file

@ -234,6 +234,7 @@ function(ggml_add_backend_library backend)
# write the shared library to the output directory
set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
add_dependencies(ggml ${backend})
else()
add_library(${backend} ${ARGN})
target_link_libraries(ggml PUBLIC ${backend})

View file

@ -534,7 +534,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
hn->buffer_id = buffer_id;
hn->offset = offset;
return;
}
}

View file

@ -66,6 +66,26 @@
#include "ggml-kompute.h"
#endif
// disable C++17 deprecation warning for std::codecvt_utf8
#if defined(__clang__)
# pragma clang diagnostic push
# pragma clang diagnostic ignored "-Wdeprecated-declarations"
#endif
static std::wstring utf8_to_utf16(const std::string & str) {
std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
return converter.from_bytes(str);
}
static std::string utf16_to_utf8(const std::wstring & str) {
std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
return converter.to_bytes(str);
}
#if defined(__clang__)
# pragma clang diagnostic pop
#endif
#ifdef _WIN32
using dl_handle = std::remove_pointer_t<HMODULE>;
@ -88,11 +108,6 @@ static dl_handle * dl_load_library(const std::wstring & path) {
return handle;
}
static dl_handle * dl_load_library(const std::string & path) {
std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
return dl_load_library(converter.from_bytes(path));
}
static void * dl_get_sym(dl_handle * handle, const char * name) {
DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
@ -114,8 +129,8 @@ struct dl_handle_deleter {
}
};
static void * dl_load_library(const std::string & path) {
dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
static void * dl_load_library(const std::wstring & path) {
dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);
return handle;
}
@ -202,11 +217,11 @@ struct ggml_backend_registry {
devices.push_back(device);
}
ggml_backend_reg_t load_backend(const char * path, bool silent) {
ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
dl_handle_ptr handle { dl_load_library(path) };
if (!handle) {
if (!silent) {
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
}
return nullptr;
}
@ -214,7 +229,7 @@ struct ggml_backend_registry {
auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
if (score_fn && score_fn() == 0) {
if (!silent) {
GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
}
return nullptr;
}
@ -222,7 +237,7 @@ struct ggml_backend_registry {
auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
if (!backend_init_fn) {
if (!silent) {
GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
}
return nullptr;
}
@ -231,16 +246,16 @@ struct ggml_backend_registry {
if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
if (!silent) {
if (!reg) {
GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
} else {
GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
__func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
__func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
}
}
return nullptr;
}
GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
register_backend(reg, std::move(handle));
@ -376,14 +391,14 @@ ggml_backend_t ggml_backend_init_best(void) {
// Dynamic loading
ggml_backend_reg_t ggml_backend_load(const char * path) {
return get_reg().load_backend(path, false);
return get_reg().load_backend(utf8_to_utf16(path), false);
}
void ggml_backend_unload(ggml_backend_reg_t reg) {
get_reg().unload_backend(reg, true);
}
static std::string get_executable_path() {
static std::wstring get_executable_path() {
#if defined(__APPLE__)
// get executable path
std::vector<char> path;
@ -401,13 +416,17 @@ static std::string get_executable_path() {
if (last_slash != std::string::npos) {
base_path = base_path.substr(0, last_slash);
}
return base_path + "/";
#elif defined(__linux__)
return utf8_to_utf16(base_path + "/");
#elif defined(__linux__) || defined(__FreeBSD__)
std::string base_path = ".";
std::vector<char> path(1024);
while (true) {
// get executable path
# if defined(__linux__)
ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
# elif defined(__FreeBSD__)
ssize_t len = readlink("/proc/curproc/file", path.data(), path.size());
# endif
if (len == -1) {
break;
}
@ -423,57 +442,63 @@ static std::string get_executable_path() {
path.resize(path.size() * 2);
}
return base_path + "/";
return utf8_to_utf16(base_path + "/");
#elif defined(_WIN32)
std::vector<char> path(MAX_PATH);
DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
std::vector<wchar_t> path(MAX_PATH);
DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
if (len == 0) {
return "";
return {};
}
std::string base_path(path.data(), len);
std::wstring base_path(path.data(), len);
// remove executable name
auto last_slash = base_path.find_last_of('\\');
if (last_slash != std::string::npos) {
base_path = base_path.substr(0, last_slash);
}
return base_path + "\\";
return base_path + L"\\";
#else
return {};
#endif
}
static std::string backend_filename_prefix() {
static std::wstring backend_filename_prefix() {
#ifdef _WIN32
return "ggml-";
return L"ggml-";
#else
return "libggml-";
return L"libggml-";
#endif
}
static std::string backend_filename_suffix() {
static std::wstring backend_filename_suffix() {
#ifdef _WIN32
return ".dll";
return L".dll";
#else
return ".so";
return L".so";
#endif
}
static std::wstring path_separator() {
#ifdef _WIN32
return L"\\";
#else
return L"/";
#endif
}
static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
// enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
// TODO: search system paths
std::string file_prefix = backend_filename_prefix() + name + "-";
std::vector<std::string> search_paths;
std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
std::vector<std::wstring> search_paths;
if (user_search_path == nullptr) {
search_paths.push_back("./");
search_paths.push_back(L"." + path_separator());
search_paths.push_back(get_executable_path());
} else {
#if defined(_WIN32)
search_paths.push_back(std::string(user_search_path) + "\\");
#else
search_paths.push_back(std::string(user_search_path) + "/");
#endif
search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
}
int best_score = 0;
std::string best_path;
std::wstring best_path;
namespace fs = std::filesystem;
for (const auto & search_path : search_paths) {
@ -483,27 +508,27 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
for (const auto & entry : dir_it) {
if (entry.is_regular_file()) {
std::string filename = entry.path().filename().string();
std::string ext = entry.path().extension().string();
std::wstring filename = entry.path().filename().wstring();
std::wstring ext = entry.path().extension().wstring();
if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
if (!handle && !silent) {
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
}
if (handle) {
auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
if (score_fn) {
int s = score_fn();
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
#endif
if (s > best_score) {
best_score = s;
best_path = entry.path().string();
best_path = entry.path().wstring();
}
} else {
if (!silent) {
GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
}
}
}
@ -515,15 +540,15 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
if (best_score == 0) {
// try to load the base backend
for (const auto & search_path : search_paths) {
std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
if (fs::exists(path)) {
return get_reg().load_backend(path.c_str(), silent);
return get_reg().load_backend(path, silent);
}
}
return nullptr;
}
return get_reg().load_backend(best_path.c_str(), silent);
return get_reg().load_backend(best_path, silent);
}
void ggml_backend_load_all() {

View file

@ -74,112 +74,96 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR
CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
(NOT CMAKE_OSX_ARCHITECTURES AND
NOT CMAKE_GENERATOR_PLATFORM_LWR AND
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
message(STATUS "ARM detected")
if (MSVC)
list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
list(APPEND ARCH_DEFINITIONS __ARM_NEON)
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
if (GGML_COMPILER_SUPPORT_DOTPROD)
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
message(STATUS "ARM feature DOTPROD enabled")
endif ()
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
message(STATUS "ARM feature MATMUL_INT8 enabled")
endif ()
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
endif ()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
elseif (APPLE)
if (GGML_NATIVE)
set(USER_PROVIDED_MARCH FALSE)
foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
set(USER_PROVIDED_MARCH TRUE)
break()
endif()
endforeach()
if (NOT USER_PROVIDED_MARCH)
set(MARCH_FLAGS "-march=armv8.2a")
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
if (GGML_COMPILER_SUPPORT_DOTPROD)
set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
message(STATUS "ARM feature DOTPROD enabled")
endif ()
set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
message(STATUS "ARM feature MATMUL_INT8 enabled")
endif ()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
endif ()
endif ()
if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
else()
check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
# Raspberry Pi 1, Zero
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
# Android armeabi-v7a
list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
else()
# Raspberry Pi 2
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
if (GGML_NATIVE)
# -mcpu=native does not always enable all the features in some compilers,
# so we check for them manually and enable them if available
execute_process(
COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v -
INPUT_FILE "/dev/null"
OUTPUT_QUIET
ERROR_VARIABLE ARM_MCPU
RESULT_VARIABLE ARM_MCPU_RESULT
)
if (NOT ARM_MCPU_RESULT)
string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
endif()
if ("${ARM_MCPU_FLAG}" STREQUAL "")
set(ARM_MCPU_FLAG -mcpu=native)
message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
endif()
include(CheckCXXSourceRuns)
function(check_arm_feature tag code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
check_cxx_source_runs(
"${code}"
GGML_MACHINE_SUPPORTS_${tag}
)
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
else()
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endfunction()
check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
else()
if (GGML_CPU_ARM_ARCH)
list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
endif()
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
# Android arm64-v8a
# Raspberry Pi 3, 4, Zero 2 (32-bit)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
# show enabled features
if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
set(FEAT_INPUT_FILE "NUL")
else()
set(FEAT_INPUT_FILE "/dev/null")
endif()
if (GGML_SVE)
list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
execute_process(
COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
INPUT_FILE ${FEAT_INPUT_FILE}
OUTPUT_VARIABLE ARM_FEATURE
RESULT_VARIABLE ARM_FEATURE_RESULT
)
if (ARM_FEATURE_RESULT)
message(WARNING "Failed to get ARM features")
else()
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
if (NOT ${feature_pos} EQUAL -1)
message(STATUS "ARM feature ${feature} enabled")
endif()
endforeach()
endif()
endif()
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
message(STATUS "x86 detected")
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
@ -339,6 +323,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
if (GGML_BACKEND_DL)
if (GGML_NATIVE)
# the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
endif()
# The feature detection code is compiled as a separate target so that
# it can be built without the architecture flags
# Since multiple variants of the CPU backend may be included in the same

View file

@ -564,21 +564,21 @@ static void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *)vx;
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
for (int c = 0; c < nc; c += ncols_interleaved) {
const block_q8_0 * a_ptr = (const block_q8_0 *)vy;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
float32x4_t acc = vdupq_n_f32(0);
for (int b = 0; b < nb; b++) {
int8x16_t b0 = vld1q_s8((const int8_t *)b_ptr->qs);
int8x16_t b1 = vld1q_s8((const int8_t *)b_ptr->qs + 16);
int8x16_t b2 = vld1q_s8((const int8_t *)b_ptr->qs + 32);
int8x16_t b3 = vld1q_s8((const int8_t *)b_ptr->qs + 48);
float16x4_t bd = vld1_f16((const __fp16 *)b_ptr->d);
int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
int8x16_t a0 = vld1q_s8(a_ptr->qs);
int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
float16x4_t ad = vld1_dup_f16((const __fp16 *)&a_ptr->d);
float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
int32x4_t ret = vdupq_n_s32(0);
@ -647,72 +647,52 @@ static void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
__asm__ __volatile__(
"movi v2.16b, #0x4\n"
"movi v1.16b, #0xf0\n"
"add %x[b_ptr], %x[b_ptr], #0x8\n"
"1:" // Column loop
"add x23, %x[a_ptr], #0x2\n"
"movi v0.16b, #0x0\n"
"mov x22, %x[nb]\n"
"2:" // Block loop
"ldr q31, [%x[b_ptr], #0x0]\n"
"ldr q30, [%x[b_ptr], #0x10]\n"
"mov x21, x23\n"
"movi v29.4s, #0x0\n"
"ldr q28, [%x[b_ptr], #0x20]\n"
"ldr q27, [%x[b_ptr], #0x30]\n"
"movi v26.4s, #0x0\n"
"sub x20, x23, #0x2\n"
"ld1r { v25.8h }, [x20]\n"
"ldr q24, [%x[b_ptr], #-0x8]\n"
"sub x22, x22, #0x1\n"
"add x23, x23, #0x22\n"
"ld1r { v23.2d }, [x21], #0x8\n"
"sshl v22.16b, v31.16b, v2.16b\n"
"sshl v16.16b, v30.16b, v2.16b\n"
"add %x[b_ptr], %x[b_ptr], #0x48\n"
"ld1r { v21.2d }, [x21], #0x8\n"
"sshl v20.16b, v28.16b, v2.16b\n"
"sshl v19.16b, v27.16b, v2.16b\n"
"ld1r { v18.2d }, [x21], #0x8\n"
"ld1r { v17.2d }, [x21], #0x8\n"
"and v31.16b, v31.16b, v1.16b\n"
"and v30.16b, v30.16b, v1.16b\n"
".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n"
".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n"
"and v28.16b, v28.16b, v1.16b\n"
"and v27.16b, v27.16b, v1.16b\n"
"fcvtl v25.4s, v25.4h\n"
"fcvtl v16.4s, v24.4h\n"
".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n"
".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n"
"fmul v16.4s, v16.4s, v25.4s\n"
".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n"
".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n"
".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n"
".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n"
"addp v29.4s, v29.4s, v26.4s\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"fmla v0.4s, v29.4s, v16.4s\n"
"cbnz x22, 2b\n"
"sub %x[nc], %x[nc], #0x4\n"
"str q0, [%x[res_ptr], #0x0]\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"cbnz %x[nc], 1b\n"
: [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
: [a_ptr] "r" (a_ptr), [nb] "r" (nb)
: "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
);
for (int c = 0; c < nc; c += ncols_interleaved) {
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
float32x4_t acc = vdupq_n_f32(0);
for (int b = 0; b < nb; b++) {
int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs);
int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1);
int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2);
int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3);
float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
int32x4_t ret0 = vdupq_n_s32(0);
int32x4_t ret1 = vdupq_n_s32(0);
ret0 = vdotq_s32(ret0, b0 << 4, a0);
ret1 = vdotq_s32(ret1, b1 << 4, a0);
ret0 = vdotq_s32(ret0, b2 << 4, a1);
ret1 = vdotq_s32(ret1, b3 << 4, a1);
ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2);
ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2);
ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3);
ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3);
int32x4_t ret = vpaddq_s32(ret0, ret1);
acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
a_ptr++;
b_ptr++;
}
vst1q_f32(s, acc);
s += ncols_interleaved;
}
return;
}
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
float sumf[4];
int sumi;

View file

@ -986,7 +986,7 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
#define GGML_F16_STEP 32
#define GGML_F16_EPR 4
static inline __m128 __sse_f16x4_load(ggml_fp16_t *x) {
static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]);
@ -997,7 +997,7 @@ static inline __m128 __sse_f16x4_load(ggml_fp16_t *x) {
return _mm_loadu_ps(tmp);
}
static inline void __sse_f16x4_store(ggml_fp16_t *x, __m128 y) {
static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
float arr[4];
_mm_storeu_ps(arr, y);
@ -7419,14 +7419,14 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(params,
ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
src1->type,
dst->type))
@ -7471,14 +7471,14 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(params,
ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
vec_dot_type,
dst->type))

View file

@ -394,8 +394,11 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ3_XXS &&
op->type != GGML_TYPE_IQ3_S &&
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ2_S &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
@ -519,6 +522,12 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
if (ggml_cpu_has_sve()) {
features.push_back({ "SVE", "1" });
}
if (ggml_cpu_has_dotprod()) {
features.push_back({ "DOTPROD", "1" });
}
if (ggml_cpu_has_matmul_int8()) {
features.push_back({ "MATMUL_INT8", "1" });
}
if (ggml_cpu_get_sve_cnt() > 0) {
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });

View file

@ -53,6 +53,8 @@
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#include <atomic>
#ifdef _MSC_VER
#define NOINLINE __declspec(noinline)
#else
@ -134,6 +136,16 @@ inline __m512 madd(__m512 a, __m512 b, __m512 c) {
return _mm512_fmadd_ps(a, b, c);
}
#endif
#if defined(__AVX512BF16__)
template <>
inline __m512 madd(__m512bh a, __m512bh b, __m512 c) {
return _mm512_dpbf16_ps(c, a, b);
}
template <>
inline __m256 madd(__m256bh a, __m256bh b, __m256 c) {
return _mm256_dpbf16_ps(c, a, b);
}
#endif
#endif
#if defined(__ARM_FEATURE_FMA)
@ -204,6 +216,7 @@ template <> inline float32x4_t load(const float *p) {
return vld1q_f32(p);
}
#if !defined(_MSC_VER)
// FIXME: this should check for __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
template <> inline float16x8_t load(const ggml_fp16_t *p) {
return vld1q_f16((const float16_t *)p);
}
@ -225,6 +238,13 @@ template <> inline __m256 load(const float *p) {
}
#endif // __AVX__
#if defined(__AVX2__) || defined(__AVX512F__)
template <> inline __m256 load(const ggml_bf16_t *p) {
return _mm256_castsi256_ps(
_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)p)), 16));
}
#endif // __AVX2__
#if defined(__F16C__)
template <> inline __m256 load(const ggml_fp16_t *p) {
return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
@ -238,8 +258,27 @@ template <> inline __m512 load(const float *p) {
template <> inline __m512 load(const ggml_fp16_t *p) {
return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
}
template <> inline __m512 load(const ggml_bf16_t *p) {
return _mm512_castsi512_ps(
_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)p)), 16));
}
#endif // __AVX512F__
#if defined(__AVX512BF16__)
template <> inline __m512bh load(const ggml_bf16_t *p) {
return (__m512bh)_mm512_loadu_ps((const float *)p);
}
template <> inline __m256bh load(const ggml_bf16_t *p) {
return (__m256bh)_mm256_loadu_ps((const float *)p);
}
template <> inline __m512bh load(const float *p) {
return _mm512_cvtne2ps_pbh(_mm512_loadu_ps(p + 16), _mm512_loadu_ps(p));
}
template <> inline __m256bh load(const float *p) {
return _mm512_cvtneps_pbh(_mm512_loadu_ps(p));
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// CONSTANTS
@ -251,199 +290,170 @@ static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
////////////////////////////////////////////////////////////////////////////////////////////////////
// FLOATING POINT MATRIX MULTIPLICATION
template <int M>
static inline int64_t BLOCK_SIZE(size_t m) {
const int64_t NB_BLOC_M = (m + M - 1) / M;
return (m % NB_BLOC_M == 0) ? m / NB_BLOC_M : (m / NB_BLOC_M) + 1;
}
static constexpr inline int64_t BLOC_POS(int64_t ib, int64_t ibN, int64_t bloc_size) {
return ib < ibN ? ib * bloc_size : ibN * bloc_size + (ib - ibN) * (bloc_size - 1);
}
template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
class tinyBLAS {
public:
tinyBLAS(int64_t k,
tinyBLAS(const ggml_compute_params * params, int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
TC *C, int64_t ldc)
: params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
bool matmul(int64_t m, int64_t n) {
if (k % KN != 0)
return false;
// compute RM for only need tile with size RM&RM-1
#if VECTOR_REGISTERS == 32
if (m % 16 == 0 && (m/16 >= params->nth)) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 4>(m, n, SIZE_N, 12);
return true;
}
if (m % 8 == 0 ) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 2>(m, n, SIZE_N, 12);
return true;
}
if (m % 4 == 0) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 1>(m, n, SIZE_N, 12);
return true;
}
#else // VECTOR_REGISTERS == 16
if (m % 16 == 0 && (m/16 >= params->nth)) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 4>(m, n, SIZE_N, 24);
return true;
}
if (m % 8 == 0 ) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 2>(m, n, SIZE_N, 24);
return true;
}
if (m % 4 == 0) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 1>(m, n, SIZE_N, 24);
return true;
}
#endif
return false;
}
private:
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
#if VECTOR_REGISTERS == 32
case 0x55:
mc = 5;
nc = 5;
gemm<5, 5>(m0, m, n0, n);
break;
case 0x45:
mc = 4;
nc = 5;
gemm<4, 5>(m0, m, n0, n);
break;
case 0x54:
mc = 5;
nc = 4;
gemm<5, 4>(m0, m, n0, n);
break;
case 0x44:
mc = 4;
nc = 4;
gemm<4, 4>(m0, m, n0, n);
break;
case 0x53:
mc = 5;
nc = 3;
gemm<5, 3>(m0, m, n0, n);
break;
case 0x35:
mc = 3;
nc = 5;
gemm<3, 5>(m0, m, n0, n);
break;
case 0x43:
mc = 4;
nc = 3;
gemm<4, 3>(m0, m, n0, n);
break;
#else
case 0x55:
case 0x54:
case 0x53:
case 0x45:
case 0x44:
case 0x43:
mc = 4;
nc = 3;
gemm<4, 3>(m0, m, n0, n);
break;
case 0x35:
#endif
case 0x34:
mc = 3;
nc = 4;
gemm<3, 4>(m0, m, n0, n);
break;
case 0x52:
mc = 5;
nc = 2;
gemm<5, 2>(m0, m, n0, n);
break;
case 0x33:
mc = 3;
nc = 3;
gemm<3, 3>(m0, m, n0, n);
break;
case 0x25:
mc = 2;
nc = 5;
gemm<2, 5>(m0, m, n0, n);
break;
case 0x42:
mc = 4;
nc = 2;
gemm<4, 2>(m0, m, n0, n);
break;
case 0x24:
mc = 2;
nc = 4;
gemm<2, 4>(m0, m, n0, n);
break;
case 0x32:
mc = 3;
nc = 2;
gemm<3, 2>(m0, m, n0, n);
break;
case 0x23:
mc = 2;
nc = 3;
gemm<2, 3>(m0, m, n0, n);
break;
case 0x51:
mc = 5;
nc = 1;
gemm<5, 1>(m0, m, n0, n);
break;
case 0x41:
mc = 4;
nc = 1;
gemm<4, 1>(m0, m, n0, n);
break;
case 0x22:
mc = 2;
nc = 2;
gemm<2, 2>(m0, m, n0, n);
break;
case 0x15:
mc = 1;
nc = 5;
gemm<1, 5>(m0, m, n0, n);
break;
case 0x14:
mc = 1;
nc = 4;
gemm<1, 4>(m0, m, n0, n);
break;
case 0x31:
mc = 3;
nc = 1;
gemm<3, 1>(m0, m, n0, n);
break;
case 0x13:
mc = 1;
nc = 3;
gemm<1, 3>(m0, m, n0, n);
break;
case 0x21:
mc = 2;
nc = 1;
gemm<2, 1>(m0, m, n0, n);
break;
case 0x12:
mc = 1;
nc = 2;
gemm<1, 2>(m0, m, n0, n);
break;
case 0x11:
mc = 1;
nc = 1;
gemm<1, 1>(m0, m, n0, n);
break;
default:
return;
template <int RM, int RN, int BM>
inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
if (SIZE_N == RN) {
return gemm<RM, RN, BM>(m, n, BN);
}
if constexpr (RN > 1) {
return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
} else {
GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
GGML_ASSERT(false); // we have miss something.
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
D Cv[RN][RM] = {};
for (int64_t l = 0; l < k; l += KN)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
load<V>(B + ldb * (jj + j) + l),
Cv[j][i]);
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
inline void gemm_bloc(int64_t ii, int64_t jj) {
D Cv[RN][RM] = {};
for (int64_t l = 0; l < k; l += KN) {
// help compiler for op order.
if constexpr (RM <= RN) {
V Av[RM];
for (int64_t i = 0; i < RM; ++i) {
Av[i] = load<V>(A + lda * (ii + i) + l);
}
for (int64_t j = 0; j < RN; ++j) {
V Bv = load<V>(B + ldb * (jj + j) + l);
for (int64_t i = 0; i < RM; ++i) {
Cv[j][i] = madd(Av[i], Bv, Cv[j][i]);
}
}
} else {
V Bv[RN];
for (int64_t j = 0; j < RN; ++j) {
Bv[j] = load<V>(B + ldb * (jj + j) + l);
}
for (int64_t i = 0; i < RM; ++i) {
V Av = load<V>(A + lda * (ii + i) + l);
for (int64_t j = 0; j < RN; ++j) {
Cv[j][i] = madd(Av, Bv[j], Cv[j][i]);
}
}
}
}
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
template <int RM, int RN, int BM>
NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
static std::atomic<int64_t> current_chunk;
GGML_ASSERT(m % (RM * BM) == 0);
const int64_t ytiles = m / (RM * BM);
const int64_t xtiles = (n + RN -1) / RN;
const int64_t jj_RN = (xtiles - (xtiles * RN - n));
// "round" bloc_size to "nearest" BN
const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
const int64_t nb_job = ytiles * NB_BN;
if (params->ith == 0) {
GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
// Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start.
std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
}
ggml_barrier(params->threadpool);
int64_t job = params->ith;
while (job < nb_job) {
const int64_t ii = (job % ytiles) * RM * BM;
const int64_t jb = job / ytiles;
const int64_t jr0 = BLOC_POS(jb , jj_BN, SIZE_BN);
const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
for (int64_t bi = 0; bi < BM * RM; bi += RM) {
int64_t jj = jj0;
for (; jj < jj1; jj += RN) {
gemm_bloc<RM, RN>(ii + bi, jj);
}
if constexpr (RN > 1) {
for (; jj < jj2; jj += RN - 1) {
gemm_bloc<RM, RN-1>(ii + bi, jj);
}
}
GGML_ASSERT(jj == jj2);
}
// next step.
job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
}
ggml_barrier(params->threadpool);
return;
}
const ggml_compute_params * params;
const TA *const A;
const TB *const B;
TC *const C;
@ -451,8 +461,6 @@ class tinyBLAS {
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
//////////////////////////////////////////////////////////////////////////////////////////
@ -1656,8 +1664,9 @@ class tinyBLAS_PPC {
* @param Ctype is GGML data type of `C`
* @return true if this function was able to service the matmul request
*/
bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
int64_t ldc, int Atype, int Btype, int Ctype) {
assert(m >= 0);
assert(n >= 0);
@ -1665,8 +1674,8 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
assert(lda >= k);
assert(ldb >= k);
assert(ldc >= m);
assert(nth > 0);
assert(ith < nth);
assert(params->nth > 0);
assert(params->ith < params->nth);
// only enable sgemm for prompt processing
if (n < 2)
@ -1681,37 +1690,25 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
if (Btype != GGML_TYPE_F32)
return false;
#if defined(__AVX512F__)
if (k % 16)
return false;
tinyBLAS<16, __m512, __m512, float, float, float> tb{
tinyBLAS<16, __m512, __m512, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__AVX__) || defined(__AVX2__)
if (k % 8)
return false;
tinyBLAS<8, __m256, __m256, float, float, float> tb{
tinyBLAS<8, __m256, __m256, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__ARM_NEON)
if (n < 4)
return false;
if (k % 4)
return false;
tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__MMA__)
if (k % 8)
return false;
@ -1719,7 +1716,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
@ -1727,60 +1724,71 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
#endif
}
case GGML_TYPE_BF16: {
#if defined(__AVX512BF16__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__AVX512F__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__AVX2__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<8, __m256, __m256, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#endif
return false;
}
case GGML_TYPE_F16: {
#if defined(__AVX512F__)
if (k % 16)
return false;
if (Btype != GGML_TYPE_F32)
return false;
tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
k, (const ggml_fp16_t *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
if (Btype == GGML_TYPE_F16) {
tinyBLAS<16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
(const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
if (k % 8)
return false;
if (Btype != GGML_TYPE_F32)
return false;
tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
k, (const ggml_fp16_t *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
if (Btype == GGML_TYPE_F16) {
tinyBLAS<8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
(const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
if (n < 8)
return false;
if (k % 8)
return false;
if (Btype != GGML_TYPE_F16)
return false;
tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
k, (const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
if (Btype == GGML_TYPE_F16) {
tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
k, (const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__ARM_NEON) && !defined(_MSC_VER)
if (k % 4)
return false;
if (Btype != GGML_TYPE_F32)
return false;
tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
k, (const ggml_fp16_t *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
if (Btype == GGML_TYPE_F32) {
tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{ params,
k, (const ggml_fp16_t *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#endif
return false;
}
case GGML_TYPE_Q8_0: {
@ -1791,7 +1799,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
@ -1799,7 +1807,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
@ -1815,7 +1823,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_q4_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
@ -1823,7 +1831,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_q4_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
@ -1839,7 +1847,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_q5_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
@ -1855,7 +1863,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
k, (const block_iq4_nl *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
@ -1867,6 +1875,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
return false;
}
(void)params;
(void)m;
(void)n;
(void)k;
@ -1876,8 +1885,6 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(void)ldb;
(void)C;
(void)ldc;
(void)ith;
(void)nth;
(void)Atype;
(void)Btype;
(void)Ctype;

View file

@ -5,8 +5,8 @@
extern "C" {
#endif
bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
const void *, int64_t, void *, int64_t, int, int,
bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
const void *, int64_t, const void *, int64_t, void *, int64_t,
int, int, int);
#ifdef __cplusplus

View file

@ -551,6 +551,22 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
// expose GGUF internals for test code
GGML_API size_t gguf_type_size(enum gguf_type type);
GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);
struct gguf_buf {
void * data;
size_t size;
size_t offset;
};
GGML_API struct gguf_buf gguf_buf_init(size_t size);
GGML_API void gguf_buf_free(struct gguf_buf buf);
GGML_API void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta);
#ifdef __cplusplus
}
#endif

View file

@ -11,6 +11,8 @@
//
#include "common.hpp"
#include "ggml-backend-impl.h"
#include "ggml-impl.h"
int get_current_device_id() {
@ -65,9 +67,9 @@ void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
const ggml_sycl_op_flatten_t op) try {
const bool use_src1 = src1 != nullptr;
GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT( dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
if(use_src1)
GGML_ASSERT(strcmp(src1->buffer->buft->iface.get_name(src1->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
// dd = data device
float * src0_ddf = (float *) src0->data;

View file

@ -26,7 +26,11 @@
#define GGML_COMMON_DECL_SYCL
#define GGML_COMMON_IMPL_SYCL
/* suppress warning spam */
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wnested-anon-types"
#include "ggml-common.h"
#pragma clang diagnostic pop
void* ggml_sycl_host_malloc(size_t size);
void ggml_sycl_host_free(void* ptr);

View file

@ -288,10 +288,8 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor) try {
ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
if (tensor->view_src != NULL && tensor->view_offs == 0) {
if (tensor->view_src != NULL) {
assert(tensor->view_src->buffer->buft == buffer->buft);
tensor->backend = tensor->view_src->backend;
tensor->extra = tensor->view_src->extra;
return;
}
@ -539,7 +537,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
auto dev_count = ggml_backend_sycl_get_device_count();
if (device>=dev_count or device<0) {
printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
device, dev_count-1);
GGML_ASSERT(device<dev_count);
}
@ -567,7 +565,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_conte
int device = ctx->device;
if (device>=ggml_sycl_info().device_count or device<0) {
printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
device, ggml_sycl_info().device_count-1);
GGML_ASSERT(device<ggml_sycl_info().device_count);
}
@ -746,7 +744,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
// FIXME: do not crash if cudaMalloc fails
// FIXME: do not crash if SYCL Buffer alloc fails
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
ggml_sycl_set_device(i);
const queue_ptr stream = ctx->streams[i];
@ -788,7 +786,6 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
CHECK_TRY_ERROR(extra->events[i][is] = new sycl::event()));
}
}
tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT;
tensor->extra = extra;
}
catch (sycl::exception const &exc) {
@ -2349,12 +2346,22 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
dpct::memcpy_direction kind;
char * src_ptr;
if (src->backend == GGML_BACKEND_TYPE_CPU) {
if (ggml_backend_buffer_is_host(src->buffer)) {
kind = dpct::host_to_device;
//GGML_SYCL_DEBUG("%s: Host buffer type src tensor\n", __func__);
src_ptr = (char *) src->data;
// GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
} else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
} else if (ggml_backend_buffer_is_sycl(src->buffer)) {
// If buffer is a SYCL buffer
//GGML_SYCL_DEBUG("%s: SYCL buffer type src tensor\n", __func__);
kind = dpct::device_to_device;
src_ptr = (char *) src->data;
} else if (ggml_backend_buffer_is_sycl_split(src->buffer)) {
/*
If buffer is a SYCL split buffer
*/
//GGML_SYCL_DEBUG("%s: Split buffer type src tensor\n", __func__);
GGML_ASSERT(i1_low == 0 && i1_high == src->ne[1]);
kind = dpct::device_to_device;
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
int id;
@ -2857,8 +2864,8 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src1->buffer));
GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
@ -2878,7 +2885,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
GGML_ASSERT(!(split && ne02 > 1));
GGML_ASSERT(!(split && ne03 > 1));
GGML_ASSERT(!(split && ne02 < ne12));
@ -3198,7 +3205,7 @@ static void ggml_sycl_mul_mat_vec_p021(ggml_backend_sycl_context & ctx, const gg
const ggml_tensor *src1,
ggml_tensor *dst) try {
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
GGML_ASSERT(src0->type == GGML_TYPE_F16);
@ -3231,7 +3238,7 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_is_permuted(src0));
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@ -3293,7 +3300,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
ggml_tensor *dst) try {
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_TENSOR_BINARY_OP_LOCALS
@ -4638,10 +4645,9 @@ static ggml_backend_dev_t ggml_backend_sycl_reg_get_device(ggml_backend_reg_t re
static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name) {
GGML_UNUSED(reg);
// TODO: update to the current function signature
//if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
// return (void *)ggml_backend_sycl_split_buffer_type;
//}
if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
return (void *)ggml_backend_sycl_split_buffer_type;
}
// SYCL doesn't support registering host memory, left here for reference
// "ggml_backend_register_host_buffer"

View file

@ -245,6 +245,7 @@ struct vk_device_struct {
vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
vk_pipeline pipeline_timestep_embedding_f32;
vk_pipeline pipeline_pool2d_f32;
vk_pipeline pipeline_rwkv_wkv6_f32;
// [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
@ -528,6 +529,13 @@ struct vk_op_pool2d_push_constants {
int32_t p0; int32_t p1;
};
struct vk_op_rwkv_wkv6_push_constants {
uint32_t B;
uint32_t T;
uint32_t C;
uint32_t H;
};
// Allow pre-recording command buffers
struct vk_staging_memcpy {
vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@ -1363,7 +1371,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
// Needs to be kept up to date on shader changes
const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1;
const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
const uint32_t warps = warptile[0] / device->subgroup_size;
const uint32_t warps = warptile[0] / warptile[10];
const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
@ -1377,8 +1385,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
std::cerr << "ggml_vulkan: Compiling shaders";
// some shaders require the subgroup size to be 16 or larger
// some shaders have a minimum subgroup size
const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);
// mulmat
std::vector<uint32_t> l_warptile, m_warptile, s_warptile,
@ -1445,7 +1454,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 };
@ -1846,53 +1855,58 @@ static void ggml_vk_load_shaders(vk_device& device) {
// mul mat vec
// AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
uint32_t rm = 1;
if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
rm = 2;
// the number of rows computed per shader depends on GPU model and quant
uint32_t rm_stdq = 1;
uint32_t rm_kq = 2;
if (device->vendor_id == VK_VENDOR_ID_AMD) {
if (device->subgroup_min_size == 64 && device->subgroup_max_size == 64) { // GCN
rm_stdq = 2;
rm_kq = 4;
}
} else if (device->vendor_id == VK_VENDOR_ID_INTEL)
rm_stdq = 2;
// computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32", mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32", mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32", mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32", mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
// dequant shaders
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@ -2014,6 +2028,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
for (auto &c : compiles) {
c.wait();
}
@ -3194,8 +3210,8 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
GGML_ABORT("fatal error");
}
// Check if src is pinned memory
vk_buffer buf;
size_t buf_offset;
vk_buffer buf = nullptr;
size_t buf_offset = 0;
ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset);
const uint64_t ne0 = tensor->ne[0];
@ -3258,7 +3274,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
VkBufferCopy buf_copy{ 0, offset, copy_size };
ggml_vk_sync_buffers(subctx);
vkCmdCopyBuffer(subctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
for (uint64_t i3 = 0; i3 < ne3; i3++) {
for (uint64_t i2 = 0; i2 < ne2; i2++) {
@ -3291,7 +3307,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
}
// Check if src is pinned memory
vk_buffer buf = nullptr;
size_t buf_offset;
size_t buf_offset = 0;
ggml_vk_host_get(dst->device, src, buf, buf_offset);
if (buf != nullptr) {
@ -3333,7 +3349,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
copy_size};
ggml_vk_sync_buffers(subctx);
vkCmdCopyBuffer(subctx->s->buffer, staging_buffer->buffer, dst->buffer, 1, &buf_copy);
vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
if (width == spitch) {
deferred_memcpy((uint8_t *)staging_buffer->ptr, src, width * height, &subctx->in_memcpys);
@ -3389,7 +3405,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
// Check if dst is pinned memory
vk_buffer buf = nullptr;
size_t buf_offset;
size_t buf_offset = 0;
ggml_vk_host_get(src->device, dst, buf, buf_offset);
std::vector<vk::BufferCopy> slices(1);
@ -3469,7 +3485,7 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds
VkBufferCopy bc{ src_offset, dst_offset, size };
vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
vkCmdCopyBuffer(ctx->s->buffer, (VkBuffer)src->buffer, (VkBuffer)dst->buffer, 1, &bc);
}
static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
@ -3721,9 +3737,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
vk_buffer d_Qx;
vk_buffer d_Qx = nullptr;
size_t qx_buf_offset = 0;
vk_buffer d_Qy;
vk_buffer d_Qy = nullptr;
size_t qy_buf_offset = 0;
bool src0_uma = false;
@ -3923,9 +3939,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
vk_buffer d_Qx;
vk_buffer d_Qx = nullptr;
size_t qx_buf_offset = 0;
vk_buffer d_Qy;
vk_buffer d_Qy = nullptr;
size_t qy_buf_offset = 0;
bool src0_uma = false;
@ -4101,7 +4117,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
vk_buffer d_Qy;
vk_buffer d_Qy = nullptr;
size_t qy_buf_offset = 0;
bool src1_uma = false;
@ -4289,11 +4305,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
vk_buffer d_Qx;
vk_buffer d_Qx = nullptr;
size_t qx_buf_offset = 0;
vk_buffer d_Qy;
vk_buffer d_Qy = nullptr;
size_t qy_buf_offset = 0;
vk_buffer d_ids;
vk_buffer d_ids = nullptr;
size_t ids_buf_offset = 0;
bool src0_uma = false;
@ -4494,11 +4510,11 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
vk_buffer d_Qx;
vk_buffer d_Qx = nullptr;
size_t qx_buf_offset = 0;
vk_buffer d_Qy;
vk_buffer d_Qy = nullptr;
size_t qy_buf_offset = 0;
vk_buffer d_ids;
vk_buffer d_ids = nullptr;
size_t ids_buf_offset = 0;
bool src0_uma = false;
@ -4757,8 +4773,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
ggml_vk_sync_buffers(subctx);
vk_buffer d_Q, d_K, d_V, d_D, d_M;
uint64_t q_buf_offset, k_buf_offset, v_buf_offset, d_buf_offset, m_buf_offset;
vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr;
size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0;
bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false;
@ -5022,6 +5038,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_pool2d_f32;
}
return nullptr;
case GGML_OP_RWKV_WKV6:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_rwkv_wkv6_f32;
}
return nullptr;
case GGML_OP_LEAKY_RELU:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_leaky_relu_f32;
@ -5424,6 +5445,134 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
}, dryrun);
}
static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, bool dryrun = false) {
const ggml_tensor * k = dst->src[0];
const ggml_tensor * v = dst->src[1];
const ggml_tensor * r = dst->src[2];
const ggml_tensor * tf = dst->src[3];
const ggml_tensor * td = dst->src[4];
const ggml_tensor * state = dst->src[5];
GGML_ASSERT(!ggml_is_quantized(k->type));
GGML_ASSERT(!ggml_is_quantized(v->type));
GGML_ASSERT(!ggml_is_quantized(r->type));
GGML_ASSERT(!ggml_is_quantized(tf->type));
GGML_ASSERT(!ggml_is_quantized(td->type));
GGML_ASSERT(!ggml_is_quantized(state->type));
GGML_ASSERT(dst->buffer != nullptr);
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, k, v, r, dst, GGML_OP_RWKV_WKV6);
GGML_ASSERT(pipeline != nullptr);
if (dryrun) {
ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
return;
}
ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
ggml_backend_vk_buffer_context * r_buf_ctx = (ggml_backend_vk_buffer_context *)r->buffer->context;
ggml_backend_vk_buffer_context * tf_buf_ctx = (ggml_backend_vk_buffer_context *)tf->buffer->context;
ggml_backend_vk_buffer_context * td_buf_ctx = (ggml_backend_vk_buffer_context *)td->buffer->context;
ggml_backend_vk_buffer_context * state_buf_ctx = (ggml_backend_vk_buffer_context *)state->buffer->context;
ggml_vk_sync_buffers(subctx);
vk_buffer d_D = nullptr, d_K = nullptr, d_V = nullptr, d_R = nullptr, d_TF = nullptr, d_TD = nullptr, d_State = nullptr;
size_t k_offset = 0, v_offset = 0, r_offset = 0, tf_offset = 0, td_offset = 0, state_offset = 0, dst_offset = 0;
bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false;
if (ctx->device->uma) {
ggml_vk_host_get(ctx->device, k->data, d_K, k_offset);
ggml_vk_host_get(ctx->device, v->data, d_V, v_offset);
ggml_vk_host_get(ctx->device, r->data, d_R, r_offset);
ggml_vk_host_get(ctx->device, tf->data, d_TF, tf_offset);
ggml_vk_host_get(ctx->device, td->data, d_TD, td_offset);
ggml_vk_host_get(ctx->device, state->data, d_State, state_offset);
ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
K_uma = d_K != nullptr;
V_uma = d_V != nullptr;
R_uma = d_R != nullptr;
TF_uma = d_TF != nullptr;
TD_uma = d_TD != nullptr;
STATE_uma = d_State != nullptr;
DST_uma = d_D != nullptr;
}
if (!K_uma) {
d_K = k_buf_ctx->dev_buffer;
k_offset = vk_tensor_offset(k) + k->view_offs;
}
if (!V_uma) {
d_V = v_buf_ctx->dev_buffer;
v_offset = vk_tensor_offset(v) + v->view_offs;
}
if (!R_uma) {
d_R = r_buf_ctx->dev_buffer;
r_offset = vk_tensor_offset(r) + r->view_offs;
}
if (!TF_uma) {
d_TF = tf_buf_ctx->dev_buffer;
tf_offset = vk_tensor_offset(tf) + tf->view_offs;
}
if (!TD_uma) {
d_TD = td_buf_ctx->dev_buffer;
td_offset = vk_tensor_offset(td) + td->view_offs;
}
if (!STATE_uma) {
d_State = state_buf_ctx->dev_buffer;
state_offset = vk_tensor_offset(state) + state->view_offs;
}
if (!DST_uma) {
d_D = dst_buf_ctx->dev_buffer;
dst_offset = vk_tensor_offset(dst) + dst->view_offs;
}
const uint64_t k_size = ggml_nbytes(k);
const uint64_t v_size = ggml_nbytes(v);
const uint64_t r_size = ggml_nbytes(r);
const uint64_t tf_size = ggml_nbytes(tf);
const uint64_t td_size = ggml_nbytes(td);
const uint64_t state_size = ggml_nbytes(state);
const uint64_t dst_size = ggml_nbytes(dst);
std::array<uint32_t, 3> elements = {
(uint32_t)(pc.B * pc.H),
1,
1
};
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
vk_subbuffer{ d_K, k_offset, k_size },
vk_subbuffer{ d_V, v_offset, v_size },
vk_subbuffer{ d_R, r_offset, r_size },
vk_subbuffer{ d_TF, tf_offset, tf_size },
vk_subbuffer{ d_TD, td_offset, td_size },
vk_subbuffer{ d_State, state_offset, state_size },
vk_subbuffer{ d_D, dst_offset, dst_size }
}, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
}
static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
const size_t seq_length = dst->src[0]->ne[3];
const size_t n_embed = dst->ne[0];
const size_t n_heads = dst->src[0]->ne[2];
const size_t n_seqs = dst->src[5]->ne[1];
ggml_vk_op_f32_rwkv6(
ctx, subctx, dst,
{
(uint32_t)n_seqs,
(uint32_t)seq_length,
(uint32_t)n_embed,
(uint32_t)n_heads,
},
dryrun
);
}
static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
int * op_params = (int *)dst->op_params;
@ -6569,6 +6718,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_RWKV_WKV6:
case GGML_OP_LEAKY_RELU:
case GGML_OP_FLASH_ATTN_EXT:
break;
@ -6768,6 +6918,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_FLASH_ATTN_EXT:
ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node, dryrun);
break;
case GGML_OP_RWKV_WKV6:
ggml_vk_rwkv_wkv6(ctx, compute_ctx, node, dryrun);
break;
default:
return false;
@ -6848,6 +7003,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_RWKV_WKV6:
case GGML_OP_LEAKY_RELU:
case GGML_OP_REPEAT:
buf = tensor->buffer;
@ -7724,6 +7880,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_POOL_2D:
case GGML_OP_RWKV_WKV6:
case GGML_OP_LEAKY_RELU:
return true;
default:
@ -8300,7 +8457,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
} else if (tensor->op == GGML_OP_LEAKY_RELU) {
const float * op_params = (const float *)tensor->op_params;
tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);
} else {
} else if (tensor->op == GGML_OP_RWKV_WKV6) {
tensor_clone = ggml_rwkv_wkv6(ggml_ctx, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3],
tensor->src[4], tensor->src[5]);
}
else {
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
GGML_ABORT("fatal error");
}

View file

@ -10,9 +10,10 @@ float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = unpack8(uint32_t(bl.block.qs[(idx & 0xE) >> 1]))[idx & 1];
uint32_t qs = uint32_t(bl.block.qs[(idx & 0xE) >> 1]);
qs >>= shift;
qs &= 0xF;
qs &= 0x0F0F;
qs = unpack8(qs)[idx & 1];
float16_t ret = (float16_t(qs) - float16_t(8)) * d;
return ret;
}
@ -152,15 +153,17 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4
block_q4_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed16 {
block_q4_K_packed16 block;
};
float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl);
const uint idx = coordInBlock[1];
const uint iqs = idx;
const uint n = iqs / 64; // 0,1,2,3
const uint b = (iqs % 64) / 32; // 0,1
const uint b = (idx & 0x20) >> 5; // 0,1
const uint is = (idx & 0xE0) >> 5; // 0..7
const uint qsi = n * 32 + (iqs % 32); // 0..127
const f16vec2 loadd = bl.block.d;
@ -184,9 +187,11 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2
const float16_t d = loadd.x * float16_t(sc);
const float16_t m = loadd.y * float16_t(mbyte);
uint32_t dmask = 0xF << (b * 4);
uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
qs = (qs >> (b * 4)) & 0x0F0F;
qs = unpack8(qs)[idx & 1];
float16_t ret = d * float16_t((bl.block.qs[qsi ] & dmask) >> (b * 4)) - m;
float16_t ret = d * float16_t(qs) - m;
return ret;
}
@ -195,18 +200,19 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5
block_q5_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed16 {
block_q5_K_packed16 block;
};
float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl);
const uint idx = coordInBlock[1];
const uint iqs = idx;
const uint n = iqs / 64; // 0,1,2,3
const uint b = (iqs % 64) / 32; // 0,1
const uint b = (idx & 0x20) >> 5; // 0,1
const uint is = (idx & 0xE0) >> 5; // 0..7
const uint qsi = n * 32 + (iqs % 32); // 0..127
const uint qhi = (iqs % 32); // 0..31
const uint8_t hm = uint8_t(1 << (iqs / 32));
const uint32_t hm = 0x0101 << is;
const f16vec2 loadd = bl.block.d;
@ -230,9 +236,15 @@ float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2
const float16_t d = loadd.x * float16_t(sc);
const float16_t m = loadd.y * float16_t(mbyte);
uint32_t dmask = 0xF << (b * 4);
uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]);
qh = qh & hm;
qh = unpack8(qh)[idx & 1];
float16_t ret = d * (float16_t((bl.block.qs[qsi ] & dmask) >> (b * 4)) + float16_t((bl.block.qh[qhi ] & hm) != 0 ? 16 : 0)) - m;
uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
qs = (qs >> (b * 4)) & 0x0F0F;
qs = unpack8(qs)[idx & 1];
float16_t ret = d * (float16_t(qs) + (qh != 0 ? float16_t(16) : float16_t(0))) - m;
return ret;
}
@ -241,22 +253,30 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_
block_q6_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ6_K_packed16 {
block_q6_K_packed16 block;
};
float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ6_K_packed16 bl16 = decodeBufQ6_K_packed16(bl);
const uint idx = coordInBlock[1];
const uint iqs = idx;
const uint n = iqs / 128; // 0,1
const uint b = (iqs % 128) / 64; // 0,1
const uint is_b = (iqs % 32) / 16; // 0,1
const uint qhshift = ((iqs % 128) / 32) * 2;// 0,2,4,6
const uint is = 8 * n + qhshift + is_b; // 0..15
const uint qsi = n * 64 + (iqs % 64); // 0..127
const uint qhi = n * 32 + (iqs % 32); // 0..63
const uint b = (idx & 0x40) >> 6; // 0,1
const uint qhshift = (idx & 0x60) >> 4; // 0,2,4,6
const uint is = (idx & 0xF0) >> 4; // 0..15
const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]);
float16_t ret = dscale * float16_t(int8_t(((bl.block.ql[qsi ] >> (b * 4)) & 0xF) | (((bl.block.qh[qhi ] >> qhshift) & 3) << 4)) - 32);
uint ql = uint32_t(bl16.block.ql[((idx & 0x80) >> 2) + ((idx & 0x3E) >> 1)]);
ql = (ql >> (b * 4)) & 0x0F0F;
uint qh = uint32_t(bl16.block.qh[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
qh = ((qh >> qhshift) & 0x0303) << 4;
int q = unpack8(ql | qh)[idx & 1];
float16_t ret = dscale * float16_t(q - 32);
return ret;
}

View file

@ -6,21 +6,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -38,15 +32,15 @@ void main() {
const uint s_offset = 8*v_im;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
@ -56,58 +50,84 @@ void main() {
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;
uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));
uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
}
temp[n] = fma(dall, sum1, fma(-dmin, sum2, temp[n]));
}
temp = fma(dall, sum1, fma(-dmin, sum2, temp));
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -6,21 +6,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -35,19 +29,21 @@ void main() {
const uint8_t m = uint8_t(1 << (4 * v_im));
const uint l0 = 2*v_in; // 0...15
const uint l0 = 2*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
const uint s_shift = 4 * v_im;
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
@ -57,44 +53,68 @@ void main() {
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
u8vec2 s0 = unpack8(s0_16);
u8vec2 s2 = unpack8(s2_16);
u8vec2 s4 = unpack8(s4_16);
u8vec2 s6 = unpack8(s6_16);
u8vec2 s8 = unpack8(s8_16);
u8vec2 s10 = unpack8(s10_16);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
u8vec2 s0 = unpack8(s0_16);
u8vec2 s2 = unpack8(s2_16);
u8vec2 s4 = unpack8(s4_16);
u8vec2 s6 = unpack8(s6_16);
u8vec2 s8 = unpack8(s8_16);
u8vec2 s10 = unpack8(s10_16);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
}
temp[n] = fma(d, sum, temp[n]);
}
temp = fma(d, sum, temp);
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -7,21 +7,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -31,8 +25,8 @@ void main() {
const uint step = 4;
const uint il = itid/step; // 0...3
const uint ir = itid - step*il; // 0...7 or 0...3
const uint il = itid/step; // 0...3
const uint ir = itid - step*il; // 0...7 or 0...3
const uint n = 4;
const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
@ -42,90 +36,116 @@ void main() {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec4 scale0 = uvec4(unpack8(scale0_u32));
uvec4 scale4 = uvec4(unpack8(scale4_u32));
uvec4 scale8 = uvec4(unpack8(scale8_u32));
const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f);
const uint32_t sc2 = ( scale4.x & 0x3f);
const uint32_t sc3 = ( scale4.y & 0x3f);
const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2));
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
const uint32_t q4_0 = qs0_lo4.x;
const uint32_t q4_1 = qs0_lo4.y;
const uint32_t q4_2 = qs0_lo4.z;
const uint32_t q4_3 = qs0_lo4.w;
const uint32_t q4_4 = qs0_hi4.x;
const uint32_t q4_5 = qs0_hi4.y;
const uint32_t q4_6 = qs0_hi4.z;
const uint32_t q4_7 = qs0_hi4.w;
const uint32_t q4_8 = qs64_lo4.x;
const uint32_t q4_9 = qs64_lo4.y;
const uint32_t q4_10 = qs64_lo4.z;
const uint32_t q4_11 = qs64_lo4.w;
const uint32_t q4_12 = qs64_hi4.x;
const uint32_t q4_13 = qs64_hi4.y;
const uint32_t q4_14 = qs64_hi4.z;
const uint32_t q4_15 = qs64_hi4.w;
B_TYPE_VEC4 by10 = data_b_v4[(b_offset + y1_idx) / 4];
B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
B_TYPE_VEC4 by20 = data_b_v4[(b_offset + y2_idx) / 4];
B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec4 scale0 = uvec4(unpack8(scale0_u32));
uvec4 scale4 = uvec4(unpack8(scale4_u32));
uvec4 scale8 = uvec4(unpack8(scale8_u32));
const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f);
const uint32_t sc2 = ( scale4.x & 0x3f);
const uint32_t sc3 = ( scale4.y & 0x3f);
const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2));
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
const uint32_t q4_0 = qs0_lo4.x;
const uint32_t q4_1 = qs0_lo4.y;
const uint32_t q4_2 = qs0_lo4.z;
const uint32_t q4_3 = qs0_lo4.w;
const uint32_t q4_4 = qs0_hi4.x;
const uint32_t q4_5 = qs0_hi4.y;
const uint32_t q4_6 = qs0_hi4.z;
const uint32_t q4_7 = qs0_hi4.w;
const uint32_t q4_8 = qs64_lo4.x;
const uint32_t q4_9 = qs64_lo4.y;
const uint32_t q4_10 = qs64_lo4.z;
const uint32_t q4_11 = qs64_lo4.w;
const uint32_t q4_12 = qs64_hi4.x;
const uint32_t q4_13 = qs64_hi4.y;
const uint32_t q4_14 = qs64_hi4.z;
const uint32_t q4_15 = qs64_hi4.w;
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
}
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -7,21 +7,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -39,74 +33,16 @@ void main() {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec4 scale0 = uvec4(unpack8(scale0_u32));
uvec4 scale4 = uvec4(unpack8(scale4_u32));
uvec4 scale8 = uvec4(unpack8(scale8_u32));
const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f);
const uint32_t sc2 = ( scale4.x & 0x3f);
const uint32_t sc3 = ( scale4.y & 0x3f);
const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2));
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
qs0_16_u32_lo4 += qs0_16_lo4_offset16;
qs0_16_u32_hi4 += qs0_16_hi4_offset16;
qs64_80_u32_lo4 += qs64_80_lo4_offset16;
qs64_80_u32_hi4 += qs64_80_hi4_offset16;
uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
const uint32_t q4_0 = qs0_16_lo4.x;
const uint32_t q4_1 = qs0_16_lo4.y;
const uint32_t q4_2 = qs0_16_lo4.z;
const uint32_t q4_3 = qs0_16_lo4.w;
const uint32_t q4_4 = qs0_16_hi4.x;
const uint32_t q4_5 = qs0_16_hi4.y;
const uint32_t q4_6 = qs0_16_hi4.z;
const uint32_t q4_7 = qs0_16_hi4.w;
const uint32_t q4_8 = qs64_80_lo4.x;
const uint32_t q4_9 = qs64_80_lo4.y;
const uint32_t q4_10 = qs64_80_lo4.z;
const uint32_t q4_11 = qs64_80_lo4.w;
const uint32_t q4_12 = qs64_80_hi4.x;
const uint32_t q4_13 = qs64_80_hi4.y;
const uint32_t q4_14 = qs64_80_hi4.z;
const uint32_t q4_15 = qs64_80_hi4.w;
B_TYPE_VEC2 by10 = data_b_v2[(b_offset + y1_idx) / 2];
B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
@ -116,45 +52,129 @@ void main() {
B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
const FLOAT_TYPE sx =
fma(FLOAT_TYPE(by10.x), q4_0,
fma(FLOAT_TYPE(by10.y), q4_1,
fma(FLOAT_TYPE(by116.x), q4_2,
FLOAT_TYPE(by116.y) * q4_3)));
const FLOAT_TYPE sy =
fma(FLOAT_TYPE(by132.x), q4_4,
fma(FLOAT_TYPE(by132.y), q4_5,
fma(FLOAT_TYPE(by148.x), q4_6,
FLOAT_TYPE(by148.y) * q4_7)));
const FLOAT_TYPE sz =
fma(FLOAT_TYPE(by20.x), q4_8,
fma(FLOAT_TYPE(by20.y), q4_9,
fma(FLOAT_TYPE(by216.x), q4_10,
FLOAT_TYPE(by216.y) * q4_11)));
const FLOAT_TYPE sw =
fma(FLOAT_TYPE(by232.x), q4_12,
fma(FLOAT_TYPE(by232.y), q4_13,
fma(FLOAT_TYPE(by248.x), q4_14,
FLOAT_TYPE(by248.y) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
(FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec4 scale0 = uvec4(unpack8(scale0_u32));
uvec4 scale4 = uvec4(unpack8(scale4_u32));
uvec4 scale8 = uvec4(unpack8(scale8_u32));
const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f);
const uint32_t sc2 = ( scale4.x & 0x3f);
const uint32_t sc3 = ( scale4.y & 0x3f);
const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2));
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
qs0_16_u32_lo4 += qs0_16_lo4_offset16;
qs0_16_u32_hi4 += qs0_16_hi4_offset16;
qs64_80_u32_lo4 += qs64_80_lo4_offset16;
qs64_80_u32_hi4 += qs64_80_hi4_offset16;
uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
const uint32_t q4_0 = qs0_16_lo4.x;
const uint32_t q4_1 = qs0_16_lo4.y;
const uint32_t q4_2 = qs0_16_lo4.z;
const uint32_t q4_3 = qs0_16_lo4.w;
const uint32_t q4_4 = qs0_16_hi4.x;
const uint32_t q4_5 = qs0_16_hi4.y;
const uint32_t q4_6 = qs0_16_hi4.z;
const uint32_t q4_7 = qs0_16_hi4.w;
const uint32_t q4_8 = qs64_80_lo4.x;
const uint32_t q4_9 = qs64_80_lo4.y;
const uint32_t q4_10 = qs64_80_lo4.z;
const uint32_t q4_11 = qs64_80_lo4.w;
const uint32_t q4_12 = qs64_80_hi4.x;
const uint32_t q4_13 = qs64_80_hi4.y;
const uint32_t q4_14 = qs64_80_hi4.z;
const uint32_t q4_15 = qs64_80_hi4.w;
const FLOAT_TYPE sx =
fma(FLOAT_TYPE(by10.x), q4_0,
fma(FLOAT_TYPE(by10.y), q4_1,
fma(FLOAT_TYPE(by116.x), q4_2,
FLOAT_TYPE(by116.y) * q4_3)));
const FLOAT_TYPE sy =
fma(FLOAT_TYPE(by132.x), q4_4,
fma(FLOAT_TYPE(by132.y), q4_5,
fma(FLOAT_TYPE(by148.x), q4_6,
FLOAT_TYPE(by148.y) * q4_7)));
const FLOAT_TYPE sz =
fma(FLOAT_TYPE(by20.x), q4_8,
fma(FLOAT_TYPE(by20.y), q4_9,
fma(FLOAT_TYPE(by216.x), q4_10,
FLOAT_TYPE(by216.y) * q4_11)));
const FLOAT_TYPE sw =
fma(FLOAT_TYPE(by232.x), q4_12,
fma(FLOAT_TYPE(by232.y), q4_13,
fma(FLOAT_TYPE(by248.x), q4_14,
FLOAT_TYPE(by248.y) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
(FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
}
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -7,21 +7,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -42,69 +36,95 @@ void main() {
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
FLOAT_TYPE scales[4];
scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32;
uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32;
uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
uvec4 q0 = uvec4(unpack8(q0_u32));
uvec4 q1 = uvec4(unpack8(q1_u32));
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
const uint y_idx = i * QUANT_K + y_offset;
B_TYPE_VEC4 by0 = data_b_v4[(b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
FLOAT_TYPE scales[4];
scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32;
uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32;
uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
uvec4 q0 = uvec4(unpack8(q0_u32));
uvec4 q1 = uvec4(unpack8(q1_u32));
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
}
temp[n] += sum * d;
}
temp += sum * d;
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -32,7 +32,7 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
void soft_max(uint num_iters) {
const uint tid = gl_LocalInvocationID.x;
const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint rowy = rowx % p.KY;
const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0;
if (rowx >= p.nrows_x) {
return;

View file

@ -78,7 +78,8 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
}
PROCESS_INFORMATION pi;
STARTUPINFOA si = { sizeof(STARTUPINFOA) };
STARTUPINFOA si = {};
si.cb = sizeof(STARTUPINFOA);
si.dwFlags = STARTF_USESTDHANDLES;
si.hStdOutput = stdout_write;
si.hStdError = stderr_write;
@ -479,6 +480,8 @@ void process_shaders() {
string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
for (auto &c : compiles) {
c.wait();
}

Some files were not shown because too many files have changed in this diff Show more