grace-gpu-containers/vllm/Dockerfile

ARG CUDA_VERSION=12.8.1
ARG IMAGE_DISTRO=ubuntu24.04
ARG PYTHON_VERSION=3.12

# ---------- Builder Base ----------
FROM nvcr.io/nvidia/cuda:${CUDA_VERSION}-devel-${IMAGE_DISTRO} AS base

# Set arch lists for all targets
# 'a' suffix is not forward compatible but enables all optimizations
ARG TORCH_CUDA_ARCH_LIST="9.0a"
ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
ENV UV_TORCH_BACKEND=cu128
ARG VLLM_FA_CMAKE_GPU_ARCHES="90a-real"
ENV VLLM_FA_CMAKE_GPU_ARCHES=${VLLM_FA_CMAKE_GPU_ARCHES}

# Update apt packages and install dependencies
ENV DEBIAN_FRONTEND=noninteractive
RUN apt update
RUN apt upgrade -y
RUN apt install -y --no-install-recommends \
        curl \
        git \
        libibverbs-dev \
        zlib1g-dev \
        libnuma-dev

# Clean apt cache
RUN apt clean
RUN rm -rf /var/lib/apt/lists/*
RUN rm -rf /var/cache/apt/archives

# Set compiler paths
ENV CC=/usr/bin/gcc
ENV CXX=/usr/bin/g++
ENV QEMU_CPU=max

# Install uv
RUN curl -LsSf https://astral.sh/uv/install.sh | env UV_INSTALL_DIR=/usr/local/bin sh

# Setup build workspace
WORKDIR /workspace

# Prep build venv
ARG PYTHON_VERSION
RUN uv venv -p ${PYTHON_VERSION} --seed --python-preference only-managed
ENV VIRTUAL_ENV=/workspace/.venv
ENV PATH=${VIRTUAL_ENV}/bin:${PATH}
ENV CUDA_HOME=/usr/local/cuda
ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}
ENV CPLUS_INCLUDE_PATH=${CUDA_HOME}/include/cccl
ENV C_INCLUDE_PATH=${CUDA_HOME}/include/cccl
ENV PATH=${CUDA_HOME}/cuda/bin:${PATH}

RUN apt-get update && apt install -y wget

RUN uv pip install numpy==2.0.0
# Install pytorch nightly
RUN uv pip install torch==2.7.0+cu128 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128

FROM base AS build-base
RUN mkdir /wheels

# Install build deps that aren't in project requirements files
# Pin setuptools to <81 for LMCache compatibility (needs >=77.0.3,<81.0.0)
RUN uv pip install -U build cmake ninja pybind11 "setuptools>=77.0.3,<81.0.0" wheel

# Use PyPI triton wheel instead of building (QEMU segfaults during triton build)
FROM build-base AS build-triton
RUN mkdir -p /wheels && \
    pip download triton==3.6.0 --platform manylinux_2_27_aarch64 --only-binary=:all: --no-deps -d /wheels

# Skip xformers - vLLM has built-in FlashAttention kernels
# xformers requires TORCH_STABLE_ONLY which needs PyTorch headers not in 2.9.0
# FROM build-base AS build-xformers
# RUN git clone https://github.com/facebookresearch/xformers.git
# RUN cd xformers && \
#     git submodule sync && \
#     git submodule update --init --recursive -j 8 && \
#     MAX_JOBS=8 uv build --wheel --no-build-isolation -o /wheels

FROM build-base AS build-flashinfer
ARG FLASHINFER_ENABLE_AOT=1
ARG FLASHINFER_REF=v0.6.6
ARG FLASHINFER_BUILD_SUFFIX=cu130
ENV FLASHINFER_LOCAL_VERSION=${FLASHINFER_BUILD_SUFFIX:-}
RUN git clone https://github.com/flashinfer-ai/flashinfer.git
RUN cd flashinfer && \
    git checkout ${FLASHINFER_REF} && \
    git submodule sync && \
    git submodule update --init --recursive -j 8 && \
    uv build --wheel --no-build-isolation -o /wheels

FROM build-base AS build-lmcache
# Bleeding edge: build from dev branch (v0.4.2+)
RUN git clone https://github.com/LMCache/LMCache.git && \
    cd LMCache && \
    git checkout dev && \
# PyTorch version is dated in LMCache
    sed -i '/torch/d' pyproject.toml && \
    uv pip install setuptools_scm && \
    NO_CUDA_EXT=1 MAX_JOBS=8 python -m build --wheel --no-isolation && \
    cp dist/*.whl /wheels/


FROM build-base AS build-flash-attention
RUN apt-get update && apt-get install -y build-essential cmake gcc && \
    git clone https://github.com/Dao-AILab/flash-attention flash-attention && \
    cd flash-attention/hopper && \
    mkdir wheels && \
    export MAX_JOBS=8 && \
    export NVCC_THREADS=4 && \
    export CMAKE_BUILD_PARALLEL_LEVEL=$MAX_JOBS && \
    MAX_JOBS=$MAX_JOBS \
    CMAKE_BUILD_PARALLEL_LEVEL=$MAX_JOBS \
    FLASH_ATTENTION_FORCE_BUILD="TRUE" \
    FLASH_ATTENTION_FORCE_CXX11_ABI="FALSE" \
    FLASH_ATTENTION_SKIP_CUDA_BUILD="FALSE" \
    pip wheel . -v --no-deps --no-build-isolation -w ./wheels/ && \
    cp wheels/*.whl /wheels/

# ==============================================================================
# NOTE: Temporarily using PyPI vLLM wheel for QEMU testing
# To restore native build on GH200, uncomment the block below and comment out
# the PyPI download section.
# ==============================================================================
FROM build-base AS build-vllm
# Bleeding edge: build from main branch
ARG VLLM_REF=main
# Install ccache for faster compilation
RUN apt-get update && apt-get install -y ccache
RUN git clone https://github.com/vllm-project/vllm.git
RUN cd vllm && \
    git checkout ${VLLM_REF} && \
    git submodule sync && \
    git submodule update --init --recursive -j 8 && \
    sed -i 's/GIT_TAG [a-f0-9]\{40\}/GIT_TAG main/' cmake/external_projects/vllm_flash_attn.cmake && \
    sed -i '/size_t fail_idx = 0;/d' csrc/cache_kernels.cu && \
    sed -i 's/, \&fail_idx,/,/' csrc/cache_kernels.cu && \
    sed -i 's/"cuMemcpyBatchAsync failed at index ",\s*fail_idx, " with error "/"cuMemcpyBatchAsync failed with error "/' csrc/cache_kernels.cu && \
    export MAX_JOBS=8 && \
    export CMAKE_BUILD_PARALLEL_LEVEL=$MAX_JOBS && \
    python use_existing_torch.py && \
    uv pip install -r requirements/build.txt && \
    CCACHE_NOHASHDIR="true" uv build --wheel --no-build-isolation -o /wheels

# Use PyPI vLLM wheel (QEMU cmake fails during try_compile)
# FROM build-base AS build-vllm
# ARG VLLM_VERSION=0.18.1
# RUN mkdir -p /wheels && \
#     pip download vllm==${VLLM_VERSION} --platform manylinux_2_31_aarch64 --only-binary=:all: --no-deps -d /wheels

# Build infinistore after vllm to avoid cache invalidation
FROM build-base AS build-infinistore
# Install additional dependencies needed for building infinistore on aarch64
RUN apt update && apt install -y cmake pybind11-dev python3-dev libuv1-dev libspdlog-dev libboost-dev libboost-all-dev meson

# Build flatbuffers from source with proper CMake version
RUN git clone -b v1.12.0 https://github.com/google/flatbuffers.git && \
  cd flatbuffers && \
  cmake -B build -DFLATBUFFERS_BUILD_TESTS=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5 && \
  cmake --build build -j && \
  cmake --install build

# Build InfiniStore from source as a Python package
RUN git clone https://github.com/bytedance/InfiniStore && \
    cd InfiniStore && \
    uv pip install meson && \
    uv pip install --no-deps --no-build-isolation -e . && \
    uv pip uninstall infinistore && \
    python -m build --wheel --no-isolation && \
    cp dist/*.whl /wheels/

FROM base AS vllm-openai
COPY --from=build-flash-attention /wheels/* wheels/
COPY --from=build-flashinfer /wheels/* wheels/
COPY --from=build-triton /wheels/* wheels/
COPY --from=build-vllm /wheels/* wheels/
COPY --from=build-lmcache /wheels/* wheels/
COPY --from=build-infinistore /wheels/* wheels/

# Install wheels (infinistore is now built as a wheel)
RUN uv pip install wheels/*
RUN rm -r wheels

# Install pynvml
RUN uv pip install pynvml pandas

# Add additional packages for vLLM OpenAI
# Bleeding edge: latest transformers
RUN uv pip install accelerate hf_transfer modelscope bitsandbytes timm boto3 runai-model-streamer runai-model-streamer[s3] tensorizer transformers --upgrade

# Clean uv cache
RUN uv clean

# Install build tools and dependencies
RUN uv pip install -U build cmake ninja pybind11 setuptools==79.0.1 wheel

# Enable hf-transfer
ENV HF_HUB_ENABLE_HF_TRANSFER=1
RUN uv pip install datasets aiohttp

# Install nsys for profiling
ARG NSYS_URL=https://developer.nvidia.com/downloads/assets/tools/secure/nsight-systems/2025_5/
ARG NSYS_PKG=nsight-systems-cli-2025.5.1_2025.5.1.121-1_arm64.deb
RUN apt-get update && apt install -y wget libglib2.0-0
RUN wget ${NSYS_URL}${NSYS_PKG} && dpkg -i $NSYS_PKG && rm $NSYS_PKG
RUN apt install -y --no-install-recommends tmux cmake

# Deprecated cleanup
RUN uv pip uninstall pynvml && uv pip install nvidia-ml-py

# API server entrypoint
# ENTRYPOINT ["vllm", "serve"]
CMD ["/bin/bash"]