002f9879b2
Replaces cherry-picked upstream_ref with complete source trees. xllm/ — Iluvatar official C++ inference engine (15MB, 1470 files) Complete: kernels → layers → models → runtime → scheduler → api Excluded: .git, binary images, third_party submodule checkouts ds_vllm/ — Iluvatar official vllm fork (8MB, 703 files) Included: csrc/ (ALL CUDA kernels), fused_moe/, qwen3_5 model, _custom_ops Excluded: tests, benchmarks, docs, examples (not needed for reference) Critical call chains now fully traceable: MoE: moe_topk_softmax_kernels.cuh → ixformer.h → fused_moe.cpp → layer GDN: qwen3_gated_delta_net_base.cpp → qwen3_5_gated_delta_net.cpp Attention: ixformer.h → xllm_paged_attention → attention.cpp
169 lines
6.5 KiB
Docker
Executable File
169 lines
6.5 KiB
Docker
Executable File
|
|
ARG CUDA_VERSION=12.8.0
|
|
ARG BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu24.04
|
|
|
|
FROM ${BASE_IMAGE}
|
|
|
|
|
|
ARG CUDA_VERSION
|
|
ARG PYTHON_VERSION=3.11
|
|
ARG CMAKE_VERSION=3.27.9
|
|
|
|
ARG TORCH_VERSION=2.7.1
|
|
ARG TORCH_INDEX_URL=https://download.pytorch.org/whl/cu128
|
|
ARG TORCH_CUDA_ARCH_LIST="8.0 8.9 9.0a 10.0a 12.0a"
|
|
|
|
ARG FLASHINFER_VERSION=0.6.2
|
|
ARG FLASHINFER_CUDA_ARCH_LIST="8.0 8.9 9.0a 10.0a 12.0a"
|
|
|
|
ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
|
|
|
|
# Optional: use proxy for network access
|
|
ARG http_proxy=
|
|
ARG https_proxy=
|
|
# Optional: use pypi mirror (e.g. https://pypi.tuna.tsinghua.edu.cn/simple)
|
|
ARG UV_DEFAULT_INDEX=
|
|
# Optional: accelerate github https clone (e.g. https://gh-proxy.com/)
|
|
ARG GIT_HTTPS_MIRROR_PREFIX=
|
|
# Optional: use rustup proxy (example: https://rsproxy.cn)
|
|
ARG RUSTUP_DIST_SERVER=
|
|
ARG RUSTUP_UPDATE_ROOT=
|
|
|
|
|
|
SHELL ["/bin/bash", "-o", "pipefail", "-c"]
|
|
WORKDIR /workspace
|
|
|
|
ENV CUDA_VERSION=${CUDA_VERSION}
|
|
ENV PYTHON_VERSION=${PYTHON_VERSION}
|
|
ENV CMAKE_VERSION=${CMAKE_VERSION}
|
|
|
|
ENV TORCH_VERSION=${TORCH_VERSION}
|
|
ENV TORCH_INDEX_URL=${TORCH_INDEX_URL}
|
|
ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
|
|
|
|
|
|
ENV DEBIAN_FRONTEND=noninteractive
|
|
ENV http_proxy=${http_proxy}
|
|
ENV https_proxy=${https_proxy}
|
|
ENV PIP_DISABLE_PIP_VERSION_CHECK=1
|
|
ENV PYTHONDONTWRITEBYTECODE=1
|
|
|
|
# # Install system dependencies including build tools
|
|
RUN set -eux; \
|
|
echo "tzdata tzdata/Areas select Asia" | debconf-set-selections; \
|
|
echo "tzdata tzdata/Zones/Asia select Shanghai" | debconf-set-selections; \
|
|
apt-get update -y && apt-get install -y --no-install-recommends \
|
|
ca-certificates gdb wget curl vim git zip unzip bison build-essential \
|
|
libx11-dev libxft-dev libxext-dev tar \
|
|
autoconf automake libtool pkg-config \
|
|
libxi-dev libxtst-dev ccache \
|
|
libwayland-dev libxrandr-dev libxinerama-dev libxcursor-dev libepoxy-dev libgtk-3-dev \
|
|
libnuma-dev libibverbs-dev nasm software-properties-common; \
|
|
add-apt-repository ppa:deadsnakes/ppa -y; \
|
|
apt install -y --no-install-recommends python${PYTHON_VERSION}-full python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv; \
|
|
update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; \
|
|
update-alternatives --install /usr/bin/python python /usr/bin/python${PYTHON_VERSION} 1; \
|
|
update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION}; \
|
|
update-alternatives --set python /usr/bin/python${PYTHON_VERSION}; \
|
|
ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config; \
|
|
rm -rf /var/lib/apt/lists/*; \
|
|
wget -q https://bootstrap.pypa.io/get-pip.py && python get-pip.py --break-system-packages && rm get-pip.py; \
|
|
# Allow pip to install packages globally (PEP 668 workaround for Ubuntu 24.04)
|
|
python -m pip config set global.break-system-packages true; \
|
|
python --version && pip --version; \
|
|
apt-get clean
|
|
|
|
|
|
# UV envs
|
|
# Enable bytecode compilation
|
|
ENV UV_COMPILE_BYTECODE=1
|
|
ENV UV_DEFAULT_INDEX=${UV_DEFAULT_INDEX}
|
|
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
|
|
# Reference: https://github.com/astral-sh/uv/pull/1694
|
|
ENV UV_HTTP_TIMEOUT=500
|
|
ENV UV_INDEX_STRATEGY="unsafe-best-match"
|
|
# Use copy mode to avoid hardlink failures with Docker cache mounts
|
|
ENV UV_LINK_MODE=copy
|
|
# Omit development dependencies
|
|
ENV UV_NO_DEV=1
|
|
# Ensure installed tools can be executed out of the box
|
|
ENV UV_TOOL_BIN_DIR=/usr/local/bin
|
|
|
|
ENV PIP_INDEX_URL=${PIP_INDEX_URL}
|
|
# Install python dependencies
|
|
RUN set -eux; \
|
|
pip install "torch==${TORCH_VERSION}" --index-url "${TORCH_INDEX_URL}"; \
|
|
pip install uv ninja "numpy<2" nvshmem4py-cu12 nvidia-nvshmem-cu12 \
|
|
aiohttp requests tqdm transformers
|
|
|
|
ENV NVSHMEM_HOME=/usr/local
|
|
|
|
# Install cmake
|
|
RUN set -eux; \
|
|
url="https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-x86_64.tar.gz"; \
|
|
curl -fL --retry 3 --retry-delay 2 -o /tmp/cmake.tar.gz "${url}"; \
|
|
mkdir -p /opt/cmake; \
|
|
tar -xzf /tmp/cmake.tar.gz -C /opt/cmake --strip-components=1; \
|
|
rm -f /tmp/cmake.tar.gz; \
|
|
ln -sf /opt/cmake/bin/cmake /usr/local/bin/cmake; \
|
|
cmake --version
|
|
|
|
# Install Rust (minimal profile)
|
|
ENV RUSTUP_DIST_SERVER=${RUSTUP_DIST_SERVER}
|
|
ENV RUSTUP_UPDATE_ROOT=${RUSTUP_UPDATE_ROOT}
|
|
ENV CARGO_HOME=/root/.cargo
|
|
ENV RUSTUP_HOME=/root/.rustup
|
|
ENV PATH=${CARGO_HOME}/bin:$PATH
|
|
RUN set -eux; \
|
|
curl -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal --default-toolchain stable; \
|
|
rustup --version; cargo --version; rustc --version
|
|
|
|
# Optional: use https mirror for git clone if provided
|
|
ENV GIT_HTTPS_MIRROR_PREFIX=${GIT_HTTPS_MIRROR_PREFIX}
|
|
RUN if [[ -n "${GIT_HTTPS_MIRROR_PREFIX}" ]]; then \
|
|
git config --global url."${GIT_HTTPS_MIRROR_PREFIX}".insteadOf https://; \
|
|
fi
|
|
|
|
# Install tvm-ffi
|
|
RUN set -eux; \
|
|
git clone --depth 1 --recurse-submodules https://github.com/apache/tvm-ffi.git /workspace/tvm-ffi; \
|
|
git -C /workspace/tvm-ffi submodule update --init --recursive; \
|
|
cmake -S /workspace/tvm-ffi -B /workspace/tvm-ffi/build_cpp -DCMAKE_BUILD_TYPE=RelWithDebInfo; \
|
|
cmake --build /workspace/tvm-ffi/build_cpp --parallel --config RelWithDebInfo --target tvm_ffi_shared; \
|
|
cmake --install /workspace/tvm-ffi/build_cpp --config RelWithDebInfo; \
|
|
rm -rf /workspace/tvm-ffi
|
|
|
|
# Install flashinfer + AOT compile
|
|
ENV FLASHINFER_VERSION=${FLASHINFER_VERSION}
|
|
ENV FLASHINFER_CUDA_ARCH_LIST=${FLASHINFER_CUDA_ARCH_LIST}
|
|
RUN set -eux; \
|
|
git clone --depth 1 --branch "v${FLASHINFER_VERSION}" --recurse-submodules https://github.com/flashinfer-ai/flashinfer.git /workspace/flashinfer; \
|
|
cd /workspace/flashinfer; \
|
|
uv venv --python ${PYTHON_VERSION} && source .venv/bin/activate; \
|
|
uv pip install -v .; \
|
|
uv pip install "torch==${TORCH_VERSION}" --index-url "${TORCH_INDEX_URL}"; \
|
|
uv pip install "cuda-python==${CUDA_VERSION}"; \
|
|
python -m flashinfer.aot; \
|
|
deactivate; \
|
|
ln -sf /root/.cache/flashinfer/${FLASHINFER_VERSION}/*/cached_ops /flashinfer_ops; \
|
|
rm -rf /workspace/flashinfer
|
|
|
|
ENV FLASHINFER_OPS_PATH=/flashinfer_ops
|
|
ENV LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
|
|
ENV LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH
|
|
|
|
# Install VCPKG
|
|
RUN set -eux; \
|
|
git clone https://github.com/microsoft/vcpkg.git /vcpkg
|
|
|
|
ENV VCPKG_ROOT=/vcpkg
|
|
|
|
# Cleanup
|
|
RUN if [[ -n "${GIT_HTTPS_MIRROR_PREFIX}" ]]; then \
|
|
git config --global --unset-all url."${GIT_HTTPS_MIRROR_PREFIX}".insteadOf 2>/dev/null || true; \
|
|
fi; \
|
|
pip cache purge; \
|
|
uv cache clean
|
|
|
|
USER root
|
|
CMD ["/bin/bash"] |