Dockerfile.trtllm

# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# NOTE FOR dynamo_base AND wheel_builder STAGES:
#
# All changes to dynamo_base and wheel_builder stages should be replicated across
# Dockerfile and Dockerfile.<framework> images.:
#   - Dockerfile
#   - Dockerfile.vllm
#   - Dockerfile.sglang
#   - Dockerfile.trtllm
# This duplication was introduced purposely to quickly enable Docker layer caching and
# deduplication. Please ensure these stages stay in sync until the duplication can be
# addressed.
#
# Throughout this file, we make certain paths group-writable because this allows
# both the dynamo user (UID 1000) and Dev Container users (UID != 1000) to work
# properly without needing slow chown -R operations (which can add 2-10 extra
# minutes).
#
# DEVELOPMENT PATHS THAT MUST BE GROUP-WRITABLE (for virtualenv containers):
#   /workspace            - Users create/modify project files
#   /home/dynamo          - Users create config/cache files
#   /opt/dynamo/venv      - TensorRT-LLM uses venv, so entire venv must be writable for pip install
#
# HOW TO ACHIEVE GROUP-WRITABLE PERMISSIONS:
# 1. SHELL + /etc/profile.d - Login shell sources umask 002 globally for all RUN commands (775/664)
# 2. COPY --chmod=775       - Sets permissions on copied children (not destination)
# 3. chmod g+w (no -R)      - Fixes destination dirs only (milliseconds vs minutes)

# This section contains build arguments that are common and shared with
# the plain Dockerfile, so they should NOT have a default. The source of truth is from build.sh.
ARG BASE_IMAGE
ARG BASE_IMAGE_TAG

ARG PYTHON_VERSION
ARG ENABLE_KVBM
ARG ENABLE_MEDIA_NIXL
ARG CARGO_BUILD_JOBS

ARG PYTORCH_BASE_IMAGE="nvcr.io/nvidia/pytorch"
ARG PYTORCH_BASE_IMAGE_TAG="25.10-py3"
ARG RUNTIME_IMAGE="nvcr.io/nvidia/cuda-dl-base"
ARG RUNTIME_IMAGE_TAG="25.10-cuda13.0-runtime-ubuntu24.04"

# TensorRT-LLM specific configuration
ARG HAS_TRTLLM_CONTEXT=0
ARG TENSORRTLLM_PIP_WHEEL="tensorrt-llm"
ARG TENSORRTLLM_INDEX_URL="https://pypi.nvidia.com/"
ARG GITHUB_TRTLLM_COMMIT

# SCCACHE configuration
ARG USE_SCCACHE
ARG SCCACHE_BUCKET=""
ARG SCCACHE_REGION=""

# NIXL configuration
ARG NIXL_UCX_REF
ARG NIXL_REF
ARG NIXL_GDRCOPY_REF

# Define general architecture ARGs for supporting both x86 and aarch64 builds.
#   ARCH: Used for package suffixes (e.g., amd64, arm64)
#   ARCH_ALT: Used for Rust targets, manylinux suffix (e.g., x86_64, aarch64)
#
# Default values are for x86/amd64:
#   --build-arg ARCH=amd64 --build-arg ARCH_ALT=x86_64
#
# For arm64/aarch64, build with:
#   --build-arg ARCH=arm64 --build-arg ARCH_ALT=aarch64
#
# NOTE: There isn't an easy way to define one of these values based on the other value
# without adding if statements everywhere, so just define both as ARGs for now.
ARG ARCH=amd64
ARG ARCH_ALT=x86_64

# Copy artifacts from NGC PyTorch image
FROM ${PYTORCH_BASE_IMAGE}:${PYTORCH_BASE_IMAGE_TAG} AS pytorch_base

##################################
########## Base Image ############
##################################

FROM ${BASE_IMAGE}:${BASE_IMAGE_TAG} AS dynamo_base

ARG ARCH
ARG ARCH_ALT

USER root
WORKDIR /opt/dynamo

# Install uv package manager
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/

# Install NATS server
ENV NATS_VERSION="v2.10.28"
RUN --mount=type=cache,target=/var/cache/apt \
    wget --tries=3 --waitretry=5 https://github.com/nats-io/nats-server/releases/download/${NATS_VERSION}/nats-server-${NATS_VERSION}-${ARCH}.deb && \
    dpkg -i nats-server-${NATS_VERSION}-${ARCH}.deb && rm nats-server-${NATS_VERSION}-${ARCH}.deb

# Install etcd
ENV ETCD_VERSION="v3.5.21"
RUN wget --tries=3 --waitretry=5 https://github.com/etcd-io/etcd/releases/download/$ETCD_VERSION/etcd-$ETCD_VERSION-linux-${ARCH}.tar.gz -O /tmp/etcd.tar.gz && \
    mkdir -p /usr/local/bin/etcd && \
    tar -xvf /tmp/etcd.tar.gz -C /usr/local/bin/etcd --strip-components=1 && \
    rm /tmp/etcd.tar.gz
ENV PATH=/usr/local/bin/etcd/:$PATH

# Rust Setup
# Rust environment setup
ENV RUSTUP_HOME=/usr/local/rustup \
    CARGO_HOME=/usr/local/cargo \
    PATH=/usr/local/cargo/bin:$PATH \
    RUST_VERSION=1.90.0

# Define Rust target based on ARCH_ALT ARG
ARG RUSTARCH=${ARCH_ALT}-unknown-linux-gnu

# Install Rust
RUN wget --tries=3 --waitretry=5 "https://static.rust-lang.org/rustup/archive/1.28.1/${RUSTARCH}/rustup-init" && \
    chmod +x rustup-init && \
    ./rustup-init -y --no-modify-path --profile minimal --default-toolchain $RUST_VERSION --default-host ${RUSTARCH} && \
    rm rustup-init && \
    chmod -R a+w $RUSTUP_HOME $CARGO_HOME


##################################
##### Wheel Build Image ##########
##################################

# Redeclare ARCH_ALT ARG so it's available for interpolation in the FROM instruction
ARG ARCH_ALT

FROM quay.io/pypa/manylinux_2_28_${ARCH_ALT} AS wheel_builder

# Redeclare ARGs for this stage
ARG ARCH
ARG ARCH_ALT
ARG CARGO_BUILD_JOBS

WORKDIR /workspace

# Copy CUDA from base stage
COPY --from=dynamo_base /usr/local/cuda /usr/local/cuda
COPY --from=dynamo_base /etc/ld.so.conf.d/hpcx.conf /etc/ld.so.conf.d/hpcx.conf

# Set environment variables first so they can be used in COPY commands
ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16} \
    RUSTUP_HOME=/usr/local/rustup \
    CARGO_HOME=/usr/local/cargo \
    CARGO_TARGET_DIR=/opt/dynamo/target \
    PATH=/usr/local/cargo/bin:$PATH

# Copy artifacts from base stage
COPY --from=dynamo_base $RUSTUP_HOME $RUSTUP_HOME
COPY --from=dynamo_base $CARGO_HOME $CARGO_HOME
# Install system dependencies
RUN yum groupinstall -y 'Development Tools' &&  \
    dnf install -y almalinux-release-synergy && \
    dnf config-manager --set-enabled powertools && \
    dnf install -y \
        # Build tools
        cmake \
        ninja-build \
        clang-devel \
        gcc-c++ \
        flex \
        wget \
        # Kernel module build dependencies
        dkms \
        # Protobuf support
        protobuf-compiler \
        # RDMA/InfiniBand support (required for UCX build with --with-verbs)
        libibverbs \
        libibverbs-devel \
        rdma-core \
        rdma-core-devel \
        libibumad \
        libibumad-devel \
        librdmacm-devel \
        numactl-devel

# Ensure a modern protoc is available (required for --experimental_allow_proto3_optional)
RUN set -eux; \
    PROTOC_VERSION=25.3; \
    case "${ARCH_ALT}" in \
      x86_64) PROTOC_ZIP="protoc-${PROTOC_VERSION}-linux-x86_64.zip" ;; \
      aarch64) PROTOC_ZIP="protoc-${PROTOC_VERSION}-linux-aarch_64.zip" ;; \
      *) echo "Unsupported architecture: ${ARCH_ALT}" >&2; exit 1 ;; \
    esac; \
    wget --tries=3 --waitretry=5 -O /tmp/protoc.zip "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/${PROTOC_ZIP}"; \
    rm -f /usr/local/bin/protoc /usr/bin/protoc; \
    unzip -o /tmp/protoc.zip -d /usr/local bin/protoc include/*; \
    chmod +x /usr/local/bin/protoc; \
    ln -s /usr/local/bin/protoc /usr/bin/protoc; \
    protoc --version

# Point build tools explicitly at the modern protoc
ENV PROTOC=/usr/local/bin/protoc

ENV CUDA_PATH=/usr/local/cuda \
    PATH=/usr/local/cuda/bin:$PATH \
    LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/lib:/usr/local/lib64:$LD_LIBRARY_PATH \
    NVIDIA_DRIVER_CAPABILITIES=video,compute,utility

# Create virtual environment for building wheels
ARG PYTHON_VERSION
ENV VIRTUAL_ENV=/workspace/.venv
RUN uv venv ${VIRTUAL_ENV} --python $PYTHON_VERSION && \
    uv pip install --upgrade meson pybind11 patchelf maturin[patchelf]

ARG NIXL_UCX_REF
ARG NIXL_REF
ARG NIXL_GDRCOPY_REF

# Build and install gdrcopy
RUN git clone --depth 1 --branch ${NIXL_GDRCOPY_REF} https://github.com/NVIDIA/gdrcopy.git && \
    cd gdrcopy/packages && \
    CUDA=/usr/local/cuda ./build-rpm-packages.sh && \
    rpm -Uvh gdrcopy-kmod-*.el8.noarch.rpm && \
    rpm -Uvh gdrcopy-*.el8.${ARCH_ALT}.rpm && \
    rpm -Uvh gdrcopy-devel-*.el8.noarch.rpm

# Install SCCACHE if requested
ARG USE_SCCACHE
ARG SCCACHE_BUCKET
ARG SCCACHE_REGION
COPY container/use-sccache.sh /tmp/use-sccache.sh
RUN if [ "$USE_SCCACHE" = "true" ]; then \
        /tmp/use-sccache.sh install; \
    fi

# Set SCCACHE environment variables
ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET}} \
    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}}

# Build and install UCX
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
    if [ "$USE_SCCACHE" = "true" ]; then \
        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
    fi && \
    cd /usr/local/src && \
     git clone https://github.com/openucx/ucx.git && \
     cd ucx && 			     \
     git checkout $NIXL_UCX_REF &&	 \
     ./autogen.sh &&      \
     ./contrib/configure-release    \
        --prefix=/usr/local/ucx     \
        --enable-shared             \
        --disable-static            \
        --disable-doxygen-doc       \
        --enable-optimizations      \
        --enable-cma                \
        --enable-devel-headers      \
        --with-cuda=/usr/local/cuda \
        --with-verbs                \
        --with-dm                   \
        --with-gdrcopy=/usr/local   \
        --with-efa                  \
        --enable-mt &&              \
     make -j &&                      \
     make -j install-strip &&        \
     /tmp/use-sccache.sh show-stats "UCX" && \
     echo "/usr/local/ucx/lib" > /etc/ld.so.conf.d/ucx.conf && \
     echo "/usr/local/ucx/lib/ucx" >> /etc/ld.so.conf.d/ucx.conf && \
     ldconfig

# build and install nixl
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
    if [ "$USE_SCCACHE" = "true" ]; then \
        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    git clone --depth 1 --branch ${NIXL_REF} "https://github.com/ai-dynamo/nixl.git" && \
    cd nixl && \
    mkdir build && \
    meson setup build/ --prefix=/opt/nvidia/nvda_nixl --buildtype=release \
    -Dcudapath_lib="/usr/local/cuda/lib64" \
    -Dcudapath_inc="/usr/local/cuda/include" \
    -Ducx_path="/usr/local/ucx" && \
    cd build && \
    ninja && \
    ninja install && \
    /tmp/use-sccache.sh show-stats "NIXL"

ENV NIXL_LIB_DIR=/opt/nvidia/nvda_nixl/lib64  \
    NIXL_PLUGIN_DIR=/opt/nvidia/nvda_nixl/lib64/plugins \
    NIXL_PREFIX=/opt/nvidia/nvda_nixl
ENV LD_LIBRARY_PATH=${NIXL_LIB_DIR}:${NIXL_PLUGIN_DIR}:/usr/local/ucx/lib:/usr/local/ucx/lib/ucx:${LD_LIBRARY_PATH}

RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
    echo "$NIXL_PLUGIN_DIR" >> /etc/ld.so.conf.d/nixl.conf && \
    ldconfig

RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
    if [ "$USE_SCCACHE" = "true" ]; then \
        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
    fi && \
    cd /workspace/nixl && \
    uv build . --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION

# Copy source code (order matters for layer caching)
COPY pyproject.toml README.md LICENSE Cargo.toml Cargo.lock rust-toolchain.toml hatch_build.py /opt/dynamo/
COPY launch/ /opt/dynamo/launch/
COPY lib/ /opt/dynamo/lib/
COPY components/ /opt/dynamo/components/

# Build dynamo wheels
ARG ENABLE_KVBM
ARG USE_SCCACHE
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
    if [ "$USE_SCCACHE" = "true" ]; then \
        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export RUSTC_WRAPPER="sccache"; \
    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \
    cd /opt/dynamo/lib/bindings/python && \
    if [ "$ENABLE_MEDIA_NIXL" = "true" ]; then \
        maturin build --release --features dynamo-llm/media-nixl --out /opt/dynamo/dist; \
    else \
        maturin build --release --out /opt/dynamo/dist; \
    fi && \
    if [ "$ENABLE_KVBM" = "true" ]; then \
        cd /opt/dynamo/lib/bindings/kvbm && \
        maturin build --release --out target/wheels && \
        auditwheel repair \
            --exclude libnixl.so \
            --exclude libnixl_build.so \
            --exclude libnixl_common.so \
            --plat manylinux_2_28_${ARCH_ALT} \
            --wheel-dir /opt/dynamo/dist \
            target/wheels/*.whl; \
    fi && \
    /tmp/use-sccache.sh show-stats "Dynamo"

##################################################
########## Framework Builder Stage ##############
##################################################
#
# PURPOSE: Build TensorRT-LLM with root privileges
#
# This stage handles TensorRT-LLM installation which requires:
# - Root access for apt operations (CUDA repos, TensorRT installation)
# - System-level modifications in install_tensorrt.sh
# - Virtual environment population with PyTorch and TensorRT-LLM
#
# The completed venv is then copied to runtime stage with dynamo ownership

FROM ${BASE_IMAGE}:${BASE_IMAGE_TAG} AS framework

ARG ARCH_ALT
COPY --from=dynamo_base /bin/uv /bin/uvx /bin/

# Install minimal dependencies needed for TensorRT-LLM installation
ARG PYTHON_VERSION
RUN apt-get update && \
    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        python${PYTHON_VERSION}-dev \
        python3-pip \
        curl \
        git \
        git-lfs \
        ca-certificates && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Create virtual environment
RUN mkdir -p /opt/dynamo/venv && \
    uv venv /opt/dynamo/venv --python $PYTHON_VERSION

ENV VIRTUAL_ENV=/opt/dynamo/venv \
    PATH="/opt/dynamo/venv/bin:${PATH}"

# Copy pytorch installation from NGC PyTorch
ARG TORCH_VER=2.9.0a0+145a3a7bda.nv25.10
ARG TORCH_TENSORRT_VER=2.9.0a0
ARG TORCHVISION_VER=0.24.0a0+094e7af5
ARG JINJA2_VER=3.1.6
ARG SYMPY_VER=1.14.0
ARG FLASH_ATTN_VER=2.7.4.post1+25.10

COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torch ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torch
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torch-${TORCH_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torch-${TORCH_VER}.dist-info
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torchgen ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torchgen
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torchvision ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torchvision
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torchvision-${TORCHVISION_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torchvision-${TORCHVISION_VER}.dist-info
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torchvision.libs ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torchvision.libs
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/functorch ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/functorch
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/jinja2 ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/jinja2
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/jinja2-${JINJA2_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/jinja2-${JINJA2_VER}.dist-info
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/sympy ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/sympy
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/sympy-${SYMPY_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/sympy-${SYMPY_VER}.dist-info
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/flash_attn ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/flash_attn
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/flash_attn-${FLASH_ATTN_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/flash_attn-${FLASH_ATTN_VER}.dist-info
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/flash_attn_2_cuda.cpython-*-*-linux-gnu.so ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torch_tensorrt ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torch_tensorrt
COPY --from=pytorch_base /usr/local/lib/python${PYTHON_VERSION}/dist-packages/torch_tensorrt-${TORCH_TENSORRT_VER}.dist-info ${VIRTUAL_ENV}/lib/python${PYTHON_VERSION}/site-packages/torch_tensorrt-${TORCH_TENSORRT_VER}.dist-info

# Install TensorRT-LLM and related dependencies
ARG HAS_TRTLLM_CONTEXT
ARG TENSORRTLLM_PIP_WHEEL
ARG TENSORRTLLM_INDEX_URL
ARG GITHUB_TRTLLM_COMMIT

# Copy only wheel files and commit info from trtllm_wheel stage from build_context
COPY --from=trtllm_wheel /*.whl /trtllm_wheel/
COPY --from=trtllm_wheel /*.txt /trtllm_wheel/

RUN uv pip install --no-cache "cuda-python==13.0.2"

# Note: TensorRT needs to be uninstalled before installing the TRTLLM wheel
# because there might be mismatched versions of TensorRT between the NGC PyTorch
# and the TRTLLM wheel.
RUN [ -f /etc/pip/constraint.txt ] && : > /etc/pip/constraint.txt || true && \
    # Clean up any existing conflicting CUDA repository configurations and GPG keys
    rm -f /etc/apt/sources.list.d/cuda*.list && \
    rm -f /usr/share/keyrings/cuda-archive-keyring.gpg && \
    rm -f /etc/apt/trusted.gpg.d/cuda*.gpg

RUN if [ "$HAS_TRTLLM_CONTEXT" = "1" ]; then \
        # Download and run install_tensorrt.sh from TensorRT-LLM GitHub before installing the wheel
        curl -fsSL --retry 5 --retry-delay 10 --max-time 1800 -o /tmp/install_tensorrt.sh "https://github.com/NVIDIA/TensorRT-LLM/raw/${GITHUB_TRTLLM_COMMIT}/docker/common/install_tensorrt.sh" && \
        # Modify the script to use virtual environment pip instead of system pip3
        sed -i 's/pip3 install/uv pip install/g' /tmp/install_tensorrt.sh && \
        bash /tmp/install_tensorrt.sh && \
        # Install from local wheel directory in build context
        WHEEL_FILE="$(find /trtllm_wheel -name "*.whl" | head -n 1)"; \
        if [ -n "$WHEEL_FILE" ]; then \
            uv pip install --no-cache "$WHEEL_FILE" triton==3.5.0; \
        else \
            echo "No wheel file found in /trtllm_wheel directory."; \
            exit 1; \
        fi; \
    else \
        # Download and run install_tensorrt.sh from TensorRT-LLM GitHub before installing the wheel
        TRTLLM_VERSION=$(echo "${TENSORRTLLM_PIP_WHEEL}" | sed -E 's/.*==([0-9a-zA-Z.+-]+).*/\1/') && \
        (curl -fsSL --retry 5 --retry-delay 10 --max-time 1800 -o /tmp/install_tensorrt.sh "https://github.com/NVIDIA/TensorRT-LLM/raw/v${TRTLLM_VERSION}/docker/common/install_tensorrt.sh" || \
         curl -fsSL --retry 5 --retry-delay 10 --max-time 1800 -o /tmp/install_tensorrt.sh "https://github.com/NVIDIA/TensorRT-LLM/raw/${GITHUB_TRTLLM_COMMIT}/docker/common/install_tensorrt.sh") && \
        # Modify the script to use virtual environment pip instead of system pip3
        sed -i 's/pip3 install/uv pip install/g' /tmp/install_tensorrt.sh && \
        bash /tmp/install_tensorrt.sh && \
        # Install TensorRT-LLM wheel from the provided index URL, allow dependencies from PyPI
        # TRTLLM 1.2.0rc5 has issues installing from pypi with uv, installing from direct wheel link works best
        # explicitly installing triton 3.5.0 as trtllm only lists triton as dependency on x64_64 for some reason
        if echo "${TENSORRTLLM_PIP_WHEEL}" | grep -q '^tensorrt-llm=='; then \
            TRTLLM_VERSION=$(echo "${TENSORRTLLM_PIP_WHEEL}" | sed -E 's/tensorrt-llm==([0-9a-zA-Z.+-]+).*/\1/'); \
            PYTHON_TAG="cp$(echo ${PYTHON_VERSION} | tr -d '.')"; \
            DIRECT_URL="https://pypi.nvidia.com/tensorrt-llm/tensorrt_llm-${TRTLLM_VERSION}-${PYTHON_TAG}-${PYTHON_TAG}-linux_${ARCH_ALT}.whl"; \
            uv pip install --no-cache --index-strategy=unsafe-best-match --extra-index-url "${TENSORRTLLM_INDEX_URL}" "${DIRECT_URL}" triton==3.5.0; \
        else \
            uv pip install --no-cache --index-strategy=unsafe-best-match --extra-index-url "${TENSORRTLLM_INDEX_URL}" "${TENSORRTLLM_PIP_WHEEL}" triton==3.5.0; \
        fi; \
    fi

##################################################
########## Runtime Image ########################
##################################################
#
# PURPOSE: Production runtime environment
#
# This stage creates a lightweight production-ready image containing:
# - Pre-compiled TensorRT-LLM and framework dependencies
# - Dynamo runtime libraries and Python packages
# - Essential runtime dependencies and configurations
# - Optimized for inference workloads and deployment
#
# Use this stage when you need:
# - Production deployment of Dynamo with TensorRT-LLM
# - Minimal runtime footprint without build tools
# - Ready-to-run inference server environment
# - Base for custom application containers
#

FROM ${RUNTIME_IMAGE}:${RUNTIME_IMAGE_TAG} AS runtime

ARG ARCH_ALT
WORKDIR /workspace
ENV ENV=${ENV:-/etc/shinit_v2}
ENV VIRTUAL_ENV=/opt/dynamo/venv
ENV PATH="${VIRTUAL_ENV}/bin:${PATH}"
# workaround for pickle lib issue
ENV OMPI_MCA_coll_ucc_enable=0

# Copy CUDA development tools (nvcc, headers, dependencies, etc.) from PyTorch base image
COPY --from=pytorch_base /usr/local/cuda/bin/nvcc /usr/local/cuda/bin/nvcc
COPY --from=pytorch_base /usr/local/cuda/bin/cudafe++ /usr/local/cuda/bin/cudafe++
COPY --from=pytorch_base /usr/local/cuda/bin/ptxas /usr/local/cuda/bin/ptxas
COPY --from=pytorch_base /usr/local/cuda/bin/fatbinary /usr/local/cuda/bin/fatbinary
COPY --from=pytorch_base /usr/local/cuda/include/ /usr/local/cuda/include/
COPY --from=pytorch_base /usr/local/cuda/nvvm /usr/local/cuda/nvvm
COPY --from=pytorch_base /usr/local/cuda/lib64/libcudart.so* /usr/local/cuda/lib64/
COPY --from=pytorch_base /usr/local/cuda/lib64/libcupti* /usr/local/cuda/lib64/
COPY --from=pytorch_base /usr/local/lib/lib* /usr/local/lib/
COPY --from=pytorch_base /usr/local/cuda/bin/cuobjdump /usr/local/cuda/bin/cuobjdump
COPY --from=pytorch_base /usr/local/cuda/bin/nvdisasm /usr/local/cuda/bin/nvdisasm

ENV CUDA_HOME=/usr/local/cuda \
    TRITON_CUPTI_PATH=/usr/local/cuda/include \
    TRITON_CUDACRT_PATH=/usr/local/cuda/include \
    TRITON_CUOBJDUMP_PATH=/usr/local/cuda/bin/cuobjdump \
    TRITON_NVDISASM_PATH=/usr/local/cuda/bin/nvdisasm \
    TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas \
    TRITON_CUDART_PATH=/usr/local/cuda/include

# Copy OpenMPI from PyTorch base image
COPY --from=pytorch_base /opt/hpcx/ompi /opt/hpcx/ompi
# Copy NUMA library from PyTorch base image
COPY --from=pytorch_base /usr/lib/${ARCH_ALT}-linux-gnu/libnuma.so* /usr/lib/${ARCH_ALT}-linux-gnu/

# Copy UCX libraries, libucc.so is needed by pytorch. May not need to copy whole hpcx dir but only /opt/hpcx/ucc/
COPY --from=pytorch_base /opt/hpcx /opt/hpcx
# This is needed to make libucc.so visible so pytorch can use it.
ENV LD_LIBRARY_PATH="/opt/hpcx/ucc/lib:${LD_LIBRARY_PATH}"
# Might not need to copy cusparseLt in the future once it's included in DLFW cuda container
# networkx, packaging, setuptools get overridden by trtllm installation, so not copying them
# pytorch-triton is copied after trtllm installation.
COPY --from=pytorch_base /usr/local/cuda/lib64/libcusparseLt* /usr/local/cuda/lib64/

# Copy nats and etcd from dynamo_base image
COPY --from=dynamo_base /usr/bin/nats-server /usr/bin/nats-server
COPY --from=dynamo_base /usr/local/bin/etcd/ /usr/local/bin/etcd/
# Add ETCD and CUDA binaries to PATH so cicc and other CUDA tools are accessible
ENV PATH=/usr/local/bin/etcd/:/usr/local/cuda/nvvm/bin:$PATH

# Copy uv to system /bin
COPY --from=dynamo_base /bin/uv /bin/uvx /bin/

# Create dynamo user with group 0 for OpenShift compatibility
RUN userdel -r ubuntu > /dev/null 2>&1 || true \
    && useradd -m -s /bin/bash -g 0 dynamo \
    && [ `id -u dynamo` -eq 1000 ] \
    && mkdir -p /home/dynamo/.cache /opt/dynamo \
    # Non-recursive chown - only the directories themselves, not contents
    && chown dynamo:0 /home/dynamo /home/dynamo/.cache /opt/dynamo /workspace \
    # No chmod needed: umask 002 handles new files, COPY --chmod handles copied content
    # Set umask globally for all subsequent RUN commands (must be done as root before USER dynamo)
    # NOTE: Setting ENV UMASK=002 does NOT work - umask is a shell builtin, not an environment variable
    && mkdir -p /etc/profile.d && echo 'umask 002' > /etc/profile.d/00-umask.sh

# Install Python, build-essential and python3-dev as apt dependencies
ARG PYTHON_VERSION
RUN if [ ${ARCH_ALT} = "x86_64" ]; then \
        ARCH_FOR_GPG=${ARCH_ALT}; \
    else \
        ARCH_FOR_GPG="sbsa"; \
    fi && \
    curl -fsSL \
        https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/${ARCH_FOR_GPG}/cuda-archive-keyring.gpg \
        -o /usr/share/keyrings/cuda-archive-keyring.gpg &&\
    echo "deb [signed-by=/usr/share/keyrings/cuda-archive-keyring.gpg] \
        https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/${ARCH_FOR_GPG} /" \
        | tee /etc/apt/sources.list.d/cuda.repo.list > /dev/null &&\
    apt-get update && \
    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        # Build tools
        build-essential \
        g++ \
        ninja-build \
        git \
        git-lfs \
        # Python runtime - CRITICAL for virtual environment to work
        python${PYTHON_VERSION}-dev \
        python3-pip \
        # jq for polling various endpoints and health checks
        jq \
        # CUDA/ML libraries
        libcudnn9-cuda-13 \
        libnvshmem3-cuda-13 \
        # Network and communication libraries
        libzmq3-dev \
        # RDMA/UCX libraries required to find RDMA devices
        ibverbs-providers \
        ibverbs-utils \
        libibumad3 \
        libibverbs1 \
        libnuma1 \
        librdmacm1 \
        rdma-core \
        # OpenMPI dependencies
        openssh-client \
        openssh-server \
        # System utilities and dependencies
        curl && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Switch to dynamo user
USER dynamo
ENV HOME=/home/dynamo
# This picks up the umask 002 from the /etc/profile.d/00-umask.sh file for subsequent RUN commands
SHELL ["/bin/bash", "-l", "-o", "pipefail", "-c"]

ENV DYNAMO_HOME=/workspace
ENV NIXL_PREFIX=/opt/nvidia/nvda_nixl
ENV NIXL_LIB_DIR=$NIXL_PREFIX/lib/${ARCH_ALT}-linux-gnu
ENV NIXL_PLUGIN_DIR=$NIXL_LIB_DIR/plugins

# Copy libgomp.so from framework image
COPY --from=framework /usr/local/tensorrt /usr/local/tensorrt
COPY --from=framework /usr/lib/${ARCH_ALT}-linux-gnu/libgomp.so* /usr/lib/${ARCH_ALT}-linux-gnu/

# Copy pre-built venv with PyTorch and TensorRT-LLM from framework stage
# Pattern: COPY --chmod=775 <path>; chmod g+w <path> done later as root because COPY --chmod only affects <path>/*, not <path>
COPY --chmod=775 --chown=dynamo:0 --from=framework ${VIRTUAL_ENV} ${VIRTUAL_ENV}

# Copy UCX from framework image as plugin for NIXL
# Copy NIXL source from framework image
# Copy dynamo wheels for gitlab artifacts (read-only, no group-write needed)
COPY --chown=dynamo: --from=wheel_builder /usr/local/ucx /usr/local/ucx
COPY --chown=dynamo: --from=wheel_builder $NIXL_PREFIX $NIXL_PREFIX
COPY --chown=dynamo: --from=wheel_builder /opt/nvidia/nvda_nixl/lib64/. ${NIXL_LIB_DIR}/
COPY --chown=dynamo: --from=wheel_builder /opt/dynamo/dist/nixl/ /opt/dynamo/wheelhouse/nixl/
COPY --chown=dynamo: --from=wheel_builder /workspace/nixl/build/src/bindings/python/nixl-meta/nixl-*.whl /opt/dynamo/wheelhouse/nixl/

ENV TENSORRT_LIB_DIR=/usr/local/tensorrt/targets/${ARCH_ALT}-linux-gnu/lib
ENV PATH="/usr/local/ucx/bin:${VIRTUAL_ENV}/bin:/opt/hpcx/ompi/bin:/usr/local/bin/etcd/:/usr/local/cuda/bin:/usr/local/cuda/nvvm/bin:$PATH"
ENV LD_LIBRARY_PATH=\
$NIXL_LIB_DIR:\
$NIXL_PLUGIN_DIR:\
/usr/local/ucx/lib:\
/usr/local/ucx/lib/ucx:\
/opt/hpcx/ompi/lib:\
/usr/lib/${ARCH_ALT}-linux-gnu/nvshmem/13/:\
$TENSORRT_LIB_DIR:\
/opt/dynamo/venv/lib/python${PYTHON_VERSION}/site-packages/torch/lib:\
/opt/dynamo/venv/lib/python${PYTHON_VERSION}/site-packages/torch_tensorrt/lib:\
$LD_LIBRARY_PATH
ENV OPAL_PREFIX=/opt/hpcx/ompi

COPY --chmod=664 --chown=dynamo:0 ATTRIBUTION* LICENSE /workspace/
COPY --chmod=775 --chown=dynamo:0 benchmarks/ /workspace/benchmarks/

# Install dynamo, NIXL, and dynamo-specific dependencies
# Pattern: COPY --chmod=775 <path>; chmod g+w <path> done later as root because COPY --chmod only affects <path>/*, not <path>
ARG ENABLE_KVBM
COPY --chmod=775 --chown=dynamo:0 --from=wheel_builder /opt/dynamo/dist/*.whl /opt/dynamo/wheelhouse/
RUN uv pip install \
      --no-cache \
      /opt/dynamo/wheelhouse/ai_dynamo_runtime*.whl \
      /opt/dynamo/wheelhouse/ai_dynamo*any.whl \
      /opt/dynamo/wheelhouse/nixl/nixl*.whl && \
    if [ "${ENABLE_KVBM}" = "true" ]; then \
        KVBM_WHEEL=$(ls /opt/dynamo/wheelhouse/kvbm*.whl 2>/dev/null | head -1); \
        if [ -z "$KVBM_WHEEL" ]; then \
            echo "ERROR: ENABLE_KVBM is true but no KVBM wheel found in wheelhouse" >&2; \
            exit 1; \
        fi; \
        uv pip install --no-cache "$KVBM_WHEEL"; \
    fi && \
    cd /workspace/benchmarks && \
    UV_GIT_LFS=1 uv pip install --no-cache . && \
    # pip/uv bypasses umask when creating .egg-info files, but chmod -R is fast here (small directory)
    chmod -R g+w /workspace/benchmarks

# Install common and test dependencies
RUN --mount=type=bind,source=./container/deps/requirements.txt,target=/tmp/requirements.txt \
    --mount=type=bind,source=./container/deps/requirements.test.txt,target=/tmp/requirements.test.txt \
    UV_GIT_LFS=1 uv pip install \
        --no-cache \
        --index-strategy unsafe-best-match \
        --extra-index-url https://download.pytorch.org/whl/cu130 \
        --requirement /tmp/requirements.txt \
        --requirement /tmp/requirements.test.txt \
        cupy-cuda13x

# Copy tests, deploy and components for CI with correct ownership
# Pattern: COPY --chmod=775 <path>; chmod g+w <path> done later as root because COPY --chmod only affects <path>/*, not <path>
COPY --chmod=775 --chown=dynamo:0 tests /workspace/tests
COPY --chmod=775 --chown=dynamo:0 examples /workspace/examples
COPY --chmod=775 --chown=dynamo:0 deploy /workspace/deploy
COPY --chmod=775 --chown=dynamo:0 components/ /workspace/components/
COPY --chmod=775 --chown=dynamo:0 recipes/ /workspace/recipes/

# Setup launch banner in common directory accessible to all users
RUN --mount=type=bind,source=./container/launch_message/runtime.txt,target=/opt/dynamo/launch_message.txt \
    sed '/^#\s/d' /opt/dynamo/launch_message.txt > /opt/dynamo/.launch_screen

# Setup environment for all users
USER root
# Fix directory permissions: COPY --chmod only affects contents, not the directory itself
RUN chmod g+w ${VIRTUAL_ENV} /workspace /workspace/* /opt/dynamo /opt/dynamo/* && \
    chown dynamo:0 ${VIRTUAL_ENV} /workspace /opt/dynamo/ && \
    chmod 755 /opt/dynamo/.launch_screen && \
    echo 'source /opt/dynamo/venv/bin/activate' >> /etc/bash.bashrc && \
    echo 'cat /opt/dynamo/.launch_screen' >> /etc/bash.bashrc

USER dynamo
ARG DYNAMO_COMMIT_SHA
ENV DYNAMO_COMMIT_SHA=$DYNAMO_COMMIT_SHA

ENTRYPOINT ["/opt/nvidia/nvidia_entrypoint.sh"]
CMD []

###########################################################
########## Development (run.sh, runs as root user) ########
###########################################################
#
# PURPOSE: Local development environment for use with run.sh (not Dev Container plug-in)
#
# This stage runs as root and provides:
# - Development tools and utilities for local debugging
# - Support for vscode/cursor development outside the Dev Container plug-in
#
# Use this stage if you need a full-featured development environment with extra tools,
# but do not use it with the Dev Container plug-in.

FROM runtime AS dev

# Don't want ubuntu to be editable, just change uid and gid.
ARG WORKSPACE_DIR=/workspace

# Switch to root for system package installation
USER root

# Install utilities as root
RUN apt-get update -y && \
    apt-get install -y --no-install-recommends  \
        # Install utilities
        nvtop \
        wget \
        tmux \
        vim \
        git \
        iproute2 \
        rsync \
        zip \
        unzip \
        htop \
        # Build Dependencies
        autoconf \
        automake \
        cmake \
        libtool \
        meson \
        net-tools \
        pybind11-dev \
        # Rust build dependencies
        clang \
        libclang-dev \
        protobuf-compiler && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Set umask for group-writable files in dev stage (runs as root)
RUN mkdir -p /etc/profile.d && echo 'umask 002' > /etc/profile.d/00-umask.sh
SHELL ["/bin/bash", "-l", "-o", "pipefail", "-c"]

# Set workspace directory variable
ENV WORKSPACE_DIR=${WORKSPACE_DIR} \
    DYNAMO_HOME=${WORKSPACE_DIR} \
    RUSTUP_HOME=/usr/local/rustup \
    CARGO_HOME=/usr/local/cargo \
    CARGO_TARGET_DIR=/workspace/target \
    VIRTUAL_ENV=/opt/dynamo/venv \
    PATH=/usr/local/cargo/bin:$PATH

# Copy rust installation from dynamo_base to avoid duplication efforts
# Pattern: COPY --chmod=775 <path>; chmod g+w <path> because COPY --chmod only affects <path>/*, not <path>
COPY --from=dynamo_base --chmod=775 /usr/local/rustup /usr/local/rustup
COPY --from=dynamo_base --chmod=775 /usr/local/cargo /usr/local/cargo
RUN chmod g+w /usr/local/rustup /usr/local/cargo

# Install maturin, for maturin develop
RUN uv pip install --no-cache maturin[patchelf]

# Editable install of dynamo
COPY pyproject.toml README.md hatch_build.py /workspace/
RUN uv pip install --no-cache --no-deps -e .

CMD []