132 lines
4.8 KiB
Docker
132 lines
4.8 KiB
Docker
# ---------- llama.cpp Build Stage
|
|
FROM nvidia/cuda:13.2.1-cudnn-devel-ubuntu24.04 AS builder
|
|
|
|
ENV DEBIAN_FRONTEND=noninteractive
|
|
|
|
# Install build tools, CMake, Git, and documentation tools
|
|
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
|
|
--mount=type=cache,target=/var/lib/apt,sharing=locked \
|
|
apt-get update && apt-get install -y --no-install-recommends \
|
|
build-essential \
|
|
cmake \
|
|
git \
|
|
ninja-build \
|
|
ca-certificates \
|
|
libgomp1 \
|
|
pandoc \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
WORKDIR /workspace
|
|
|
|
# Clone llama.cpp repository
|
|
ARG LLAMA_CPP_TAG=master
|
|
RUN git clone --depth 1 --branch ${LLAMA_CPP_TAG} https://github.com/ggerganov/llama.cpp.git .
|
|
|
|
# Build llama.cpp, the ldconfig is important
|
|
RUN echo "/usr/local/cuda-13.2/compat" > /etc/ld.so.conf.d/988-cuda-compat.conf && ldconfig
|
|
RUN cmake -B build -S . \
|
|
-DCMAKE_BUILD_TYPE=Release \
|
|
-DGGML_CUDA=ON \
|
|
-DCMAKE_INSTALL_PREFIX=/install \
|
|
-DCMAKE_CUDA_ARCHITECTURES=120 \
|
|
-DLLAMA_BUILD_TESTS=OFF \
|
|
-DLLAMA_BUILD_EXAMPLES=ON \
|
|
&& cmake --build build --config Release -j $(nproc) \
|
|
&& cmake --install build \
|
|
&& rm -rf build
|
|
# Generate or stage manpages into /install/share/man
|
|
# If manpages are prebuilt or markdown docs exist, compile/place them in standard sections:
|
|
RUN mkdir -p /install/share/man/man1 && \
|
|
for doc in docs/*.md; do \
|
|
[ -f "$doc" ] || continue; \
|
|
base=$(basename "$doc" .md); \
|
|
pandoc "$doc" -s -t man -o "/install/share/man/man1/${base}.1" 2>/dev/null || true; \
|
|
done
|
|
|
|
# ---------- Python virtual environment builder
|
|
|
|
FROM nvidia/cuda:13.2.1-cudnn-devel-ubuntu24.04 AS venv-builder
|
|
|
|
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
|
|
--mount=type=cache,target=/var/lib/apt,sharing=locked \
|
|
apt update && apt-get install -y --no-install-recommends \
|
|
python3-pip \
|
|
python3-dev \
|
|
python3-venv \
|
|
git
|
|
|
|
WORKDIR /workspace
|
|
RUN python3 -m venv /workspace/py-venv
|
|
ENV PATH="/workspace/py-venv/bin:${PATH}"
|
|
|
|
# llama.cpp requirements
|
|
COPY --from=builder --chmod=0644 /workspace/requirements.txt /workspace/requirements-llama.txt
|
|
COPY --from=builder /workspace/requirements /workspace/requirements
|
|
RUN --mount=type=cache,target=/root/.cache/pip,sharing=locked pip3 install -r /workspace/requirements-llama.txt
|
|
|
|
# Upgrade torch after llama's downgrade
|
|
RUN --mount=type=cache,target=/root/.cache/pip,sharing=locked pip3 install torch torchvision torchaudio --upgrade --index-url \
|
|
https://download.pytorch.org/whl/cu132/
|
|
|
|
# Fetch convert_to_quant, and install it into our python venv.
|
|
RUN git clone --depth=1 https://github.com/silveroxides/convert_to_quant && \
|
|
cd convert_to_quant && pip3 install -r requirements.txt && pip3 install .
|
|
|
|
# Install heretic-llm and compressed-tensors
|
|
RUN --mount=type=cache,target=/root/.cache/pip,sharing=locked pip3 install heretic-llm compressed-tensors
|
|
|
|
# Install zefie's nvfp4 quant requirements
|
|
COPY --chmod=0644 requirements.txt /workspace/requirements-nvfp4.txt
|
|
RUN --mount=type=cache,target=/root/.cache/pip,sharing=locked pip3 install -r /workspace/requirements-nvfp4.txt
|
|
|
|
# ---------- Main image
|
|
|
|
FROM nvidia/cuda:13.2.1-cudnn-devel-ubuntu24.04 AS runner
|
|
ENV DEBIAN_FRONTEND=noninteractive
|
|
|
|
# "Unminify" nvidia container to enable man pages
|
|
RUN rm -f /etc/dpkg/dpkg.cfg.d/excludes && \
|
|
rm -f /etc/apt/apt.conf.d/docker-clean && \
|
|
rm -f /etc/apt/apt.conf.d/01-nodoc && \
|
|
rm -f /usr/bin/man && \
|
|
dpkg-divert --quiet --remove /usr/bin/man
|
|
|
|
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
|
|
--mount=type=cache,target=/var/lib/apt,sharing=locked \
|
|
apt update && apt-get install -y \
|
|
python3 \
|
|
man \
|
|
man-db \
|
|
wget \
|
|
bzip2 \
|
|
curl
|
|
|
|
WORKDIR /opt/nvfp4
|
|
|
|
RUN mkdir nvfp4_gguf_quant
|
|
COPY --exclude=__pycache__ nvfp4gguf ./nvfp4_gguf_quant/nvfp4gguf
|
|
COPY --chmod=0644 --exclude=ollama_nvfp4.py *.py ./nvfp4_gguf_quant/
|
|
|
|
COPY --chmod=0644 docker/README.txt /workspace/README.txt
|
|
COPY --chmod=0755 --exclude=README.txt docker/* /usr/local/bin/
|
|
|
|
# Install manpages for zefie's nvfp4 quant scripts, ctq and heretic
|
|
RUN mkdir -p /usr/local/share/man/man1
|
|
COPY --chmod=0644 --exclude=man/man1/ollama_nvfp4.1 --exclude=man/man1/ollama_nvfp4.py.1 man/man1/* /usr/local/share/man/man1/
|
|
|
|
# Copy llama.cpp compiled binaries, manpages and shared libraries from builder
|
|
COPY --from=builder /install/bin/ /usr/local/bin/
|
|
COPY --from=builder /install/lib/ /usr/local/lib/
|
|
COPY --from=builder /install/share/man/ /usr/share/man/
|
|
|
|
# Set up a minimal llama.cpp source for conversions
|
|
RUN mkdir /opt/nvfp4/llama.cpp
|
|
COPY --from=builder /workspace/conversion /opt/nvfp4/llama.cpp/conversion/
|
|
COPY --from=builder --chmod=0644 /workspace/*.py /opt/nvfp4/llama.cpp/
|
|
|
|
COPY --from=venv-builder /workspace/py-venv /opt/nvfp4/py-venv
|
|
ENV PATH="/opt/nvfp4/py-venv/bin:${PATH}"
|
|
|
|
# Finalize and settle at /workspace
|
|
ENV DEBIAN_FRONTEND=
|
|
WORKDIR /workspace |