[Docker] Use uv-managed Python 3.12 and uv for all installs in the CUDA image (#42612)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
This commit is contained in:
Mohammad Miadh Angkad
2026-10-06 14:10:09 -07:00
committed by GitHub
co-authored by Mohammad Angkad
parent 3eec178e48
commit 38bdf25f8d
+83 -67
View File
@@ -1,4 +1,9 @@
ARG CUDA_VERSION=13.0.3
ARG UV_VERSION=latest
ARG PYTHON_VERSION=3.12
FROM ghcr.io/astral-sh/uv:${UV_VERSION} AS uv
FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu24.04 AS base
ARG TARGETARCH
@@ -32,20 +37,26 @@ RUN if [ -n "$UBUNTU_MIRROR" ]; then \
sed -i "s|http://.*security.ubuntu.com|$UBUNTU_MIRROR|g" /etc/apt/sources.list; \
fi
# Python setup (combined with apt update to reduce layers)
# Ubuntu 24.04 ships Python 3.12 in main, so we no longer need the deadsnakes
# PPA. Dropping it avoids transient Launchpad 504s in `add-apt-repository`.
RUN --mount=type=cache,target=/var/cache/apt,id=base-apt \
apt update && apt install -y --no-install-recommends wget software-properties-common \
&& apt install -y --no-install-recommends python3.12-full python3.12-dev \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 2 \
&& update-alternatives --set python3 /usr/bin/python3.12 \
# Fix for apt-add-repository
&& cd /usr/lib/python3/dist-packages/ \
&& ln -s apt_pkg.cpython-312-*-linux-gnu.so apt_pkg.so
# create virtual env for sglang to avoid conflict with system python packages
RUN python3 -m venv /opt/sglang
ARG PYTHON_VERSION
COPY --from=uv /uv /uvx /usr/local/bin/
ENV UV_PYTHON_INSTALL_DIR=/opt/python \
UV_PYTHON_PREFERENCE=only-managed \
UV_LINK_MODE=copy \
VIRTUAL_ENV=/opt/sglang
RUN --mount=type=cache,target=/root/.cache/uv \
if [ -n "${PIP_DEFAULT_INDEX}" ]; then \
mkdir -p /etc/uv \
&& printf '[[index]]\nurl = "%s"\ndefault = true\n' "${PIP_DEFAULT_INDEX}" > /etc/uv/uv.toml; \
fi \
&& if [ "${GITHUB_ARTIFACTORY}" != "github.com" ]; then \
export UV_PYTHON_INSTALL_MIRROR="https://${GITHUB_ARTIFACTORY}/astral-sh/python-build-standalone/releases/download"; \
fi \
&& uv --version \
&& uv python install ${PYTHON_VERSION} \
&& uv venv --seed --python ${PYTHON_VERSION} /opt/sglang \
&& uv pip install --upgrade pip \
&& /opt/sglang/bin/python -VV
ENV PATH="/opt/sglang/bin:${PATH}"
# Install system dependencies (organized by category for better caching)
@@ -53,7 +64,6 @@ RUN --mount=type=cache,target=/var/cache/apt,id=base-apt \
apt-get update && apt-get install -y --no-install-recommends \
# Core system utilities
ca-certificates \
software-properties-common \
netcat-openbsd \
kmod \
unzip \
@@ -115,7 +125,9 @@ RUN --mount=type=cache,target=/var/cache/apt,id=base-apt \
check \
libsubunit0 \
libsubunit-dev \
&& ln -sf /usr/bin/python3.12 /usr/bin/python \
&& rm -f /usr/bin/python \
&& printf '#!/bin/sh\nexec /opt/sglang/bin/python "$@"\n' > /usr/bin/python \
&& chmod +x /usr/bin/python \
&& rm -rf /var/lib/apt/lists/* \
&& apt-get clean
@@ -181,17 +193,17 @@ RUN curl --proto '=https' --tlsv1.2 --retry 3 --retry-delay 2 -sSf https://sh.ru
&& rustc --version && cargo --version
# Install sgl-kernel (from pre-built wheel)
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --upgrade pip setuptools wheel html5lib six \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --upgrade pip setuptools wheel html5lib six \
&& case "$CUDA_VERSION" in \
13.0.3) ;; \
*) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
esac \
# --no-deps prevents pip from pulling torch from default PyPI
&& python3 -m pip install sglang-kernel==${SGL_KERNEL_VERSION} --force-reinstall --no-deps
# --no-deps prevents uv from pulling torch from default PyPI
&& uv pip install sglang-kernel==${SGL_KERNEL_VERSION} --reinstall --no-deps
# Copy the dependency spec, Rust extension sources, and shared protocol crate.
# Split from the pip install so changes to these build inputs invalidate the
# Split from the dependency install so changes to these build inputs invalidate the
# dependency layer, while unrelated Python source changes do not.
COPY python/pyproject.toml /tmp/sglang_deps/python/pyproject.toml
COPY rust/sglang-grpc /tmp/sglang_deps/rust/sglang-grpc
@@ -200,7 +212,7 @@ COPY proto /tmp/sglang_deps/proto
# Install sglang dependencies (torch, transformers, etc.).
# Generate constraints.txt to prevent reinstalling these deps in later stages.
RUN --mount=type=cache,target=/root/.cache/pip \
RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=cache,target=/root/.cargo/registry \
case "$CUDA_VERSION" in \
13.0.3) CUINDEX=130 ;; \
@@ -212,14 +224,11 @@ RUN --mount=type=cache,target=/root/.cache/pip \
&& echo '__version__ = "0.0.0"' > sglang/version.py \
&& touch README.md \
&& touch LICENSE \
&& python3 -m pip install --extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} ".[${BUILD_TYPE}]" \
&& uv pip install --extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} \
--index-strategy unsafe-best-match ".[${BUILD_TYPE}]" \
&& cd /sgl-workspace \
&& rm -rf /tmp/sglang_deps \
&& pip freeze | grep -v "^sglang==" > /sgl-workspace/constraints.txt
# distro resolves to the apt python3-distro under /usr/lib/python3, which the runtime
# stage does not COPY; force a pip copy into /usr/local so it survives the stage split.
RUN python3 -m pip install --ignore-installed --no-deps distro
&& uv pip freeze | grep -vE "^sglang( |==|@)" > /sgl-workspace/constraints.txt
# Cosmos3 guardrails default ON: sglang enables them whenever `cosmos_guardrail`
# is importable (Cosmos3Pipeline.create_pipeline_stages); opt out at runtime with
@@ -229,11 +238,11 @@ RUN python3 -m pip install --ignore-installed --no-deps distro
# by retinaface-py — would shadow our opencv-python-headless cv2). The deps not
# already in the image are installed explicitly; the find_spec check mirrors
# sglang's availability probe plus cosmos_guardrail's own imports.
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --no-deps \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --no-deps \
cosmos-guardrail==0.3.1 \
retinaface-py==0.0.2 \
&& python3 -m pip install \
&& uv pip install \
"better-profanity==0.7.0" \
"nltk==3.9.1" \
"peft==0.18.1" \
@@ -254,8 +263,7 @@ WORKDIR /build
# The kernels target Hopper (sm90a) only, so skip non-x86_64 images.
# setup.py derives the version from `git rev-parse`, so keep the .git dir
# (a source zip archive would not build).
RUN --mount=type=cache,target=/root/.cache/pip \
mkdir -p /wheels && \
RUN mkdir -p /wheels && \
if [ "$(uname -m)" = "x86_64" ]; then \
git clone https://github.com/Tencent/hpc-ops.git && \
cd hpc-ops && \
@@ -273,18 +281,18 @@ ARG INSTALL_FLASHINFER_JIT_CACHE
ARG FLASHINFER_VERSION
# Stage jit-cache/cubin artifacts into /flashinfer_jit_output for clean COPY later
RUN --mount=type=cache,target=/root/.cache/pip \
RUN --mount=type=cache,target=/root/.cache/uv \
case "$CUDA_VERSION" in \
13.0.3) CUINDEX=130 ;; \
*) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
esac \
&& mkdir -p /flashinfer_jit_output \
# flashinfer-cubin is CUDA-version-agnostic, unlike jit-cache, so its index-url has no cu${CUINDEX} suffix
&& python3 -m pip install flashinfer-cubin==${FLASHINFER_VERSION} --index-url https://flashinfer.ai/whl \
&& uv pip install flashinfer-cubin==${FLASHINFER_VERSION} --index-url https://flashinfer.ai/whl \
&& cp -r /opt/sglang/lib/python3.12/site-packages/flashinfer_cubin /flashinfer_jit_output/ \
&& cp -r /opt/sglang/lib/python3.12/site-packages/flashinfer_cubin-*.dist-info /flashinfer_jit_output/ \
&& if [ "$INSTALL_FLASHINFER_JIT_CACHE" = "1" ]; then \
python3 -m pip install flashinfer-jit-cache==${FLASHINFER_VERSION} --index-url https://flashinfer.ai/whl/cu${CUINDEX} \
uv pip install flashinfer-jit-cache==${FLASHINFER_VERSION} --index-url https://flashinfer.ai/whl/cu${CUINDEX} \
&& cp -r /opt/sglang/lib/python3.12/site-packages/flashinfer_jit_cache /flashinfer_jit_output/ \
&& cp -r /opt/sglang/lib/python3.12/site-packages/flashinfer_jit_cache*.dist-info /flashinfer_jit_output/ ; \
fi
@@ -357,10 +365,10 @@ WORKDIR /build
COPY sgl-model-gateway /build/sgl-model-gateway
# Install Rust, build gateway binary and Python bindings, then clean up Rust toolchain
RUN --mount=type=cache,target=/root/.cache/pip \
RUN --mount=type=cache,target=/root/.cache/uv \
curl --proto '=https' --tlsv1.2 --retry 3 --retry-delay 2 -sSf https://sh.rustup.rs | sh -s -- -y \
&& export PATH="/root/.cargo/bin:${PATH}" \
&& python3 -m pip install maturin \
&& uv pip install maturin \
&& cd /build/sgl-model-gateway/bindings/python \
&& ulimit -n 65536 && maturin build --release --features vendored-openssl --out /build/gateway_wheels \
&& cd /build/sgl-model-gateway \
@@ -396,9 +404,9 @@ WORKDIR /sgl-workspace
# Copy HPC-Ops wheel and install (empty on non-x86_64; kernels are sm90a-only)
COPY --from=hpc_ops_builder /wheels /tmp/wheels/hpc-ops
RUN --mount=type=cache,target=/root/.cache/pip \
RUN --mount=type=cache,target=/root/.cache/uv \
if ls /tmp/wheels/hpc-ops/*.whl >/dev/null 2>&1; then \
pip install --no-deps /tmp/wheels/hpc-ops/*.whl; \
uv pip install --no-deps /tmp/wheels/hpc-ops/*.whl; \
fi && rm -rf /tmp/wheels/hpc-ops
# Copy flashinfer cubin (always) and jit-cache (if installed) packages
@@ -460,16 +468,16 @@ RUN --mount=type=cache,target=/var/cache/apt,id=framework-apt \
# =============================================================================
# Install Mooncake
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}
# Install MSCCL++ Python dependencies and package (builds extension via CMake through pip)
RUN --mount=type=cache,target=/root/.cache/pip \
# Install MSCCL++ Python dependencies and package (builds extension via CMake through uv)
RUN --mount=type=cache,target=/root/.cache/uv \
git clone --depth=1 --branch ${MSCCLPP_VERSION} https://${GITHUB_ARTIFACTORY}/microsoft/mscclpp.git /tmp/mscclpp \
&& case "${CUDA_VERSION}" in \
13.*) \
CMAKE_ARGS="-DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON -DMSCCLPP_GPU_ARCHS=80,90,100,100a,103,103a" \
python3 -m pip install "/tmp/mscclpp[cuda13]"; \
uv pip install "/tmp/mscclpp[cuda13]"; \
;; \
*) \
echo "Unsupported CUDA version for MSCCL++: ${CUDA_VERSION}" && exit 1; \
@@ -478,15 +486,14 @@ RUN --mount=type=cache,target=/root/.cache/pip \
&& rm -rf /tmp/mscclpp
# Install essential Python packages (use constraints to prevent conflicts)
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install -c /sgl-workspace/constraints.txt \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install -c /sgl-workspace/constraints.txt \
datamodel_code_generator \
pre-commit \
pytest \
black \
isort \
icdiff \
uv \
wheel \
scikit-build-core \
py-spy \
@@ -500,20 +507,21 @@ RUN --mount=type=cache,target=/root/.cache/pip \
# DeepEP needs SGL_NCCL_VERSION at runtime even if torch pins another version;
# install it after the constrained dependency solve so the solve cannot revert it.
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --force-reinstall --no-deps \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --reinstall --no-deps \
"nvidia-nccl-cu13==${SGL_NCCL_VERSION}"
# The `nixl` stub package is needed (it owns the `nixl` import path) but
# unconditionally requires nixl-cu12, so install it with --no-deps and pair it
# with the nixl-cu13 binary to avoid shipping wrong-CUDA libs.
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install nixl nixl-cu13 --no-deps \
&& python3 -m pip install "cuda-python>=13,<14"
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install nixl nixl-cu13 --no-deps \
&& uv pip install "cuda-python>=13,<14"
# Install the latest available AI Dynamo prerelease from NVIDIA's package index.
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --pre --extra-index-url https://pypi.nvidia.com/ ai-dynamo
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --prerelease=allow --index-strategy unsafe-best-match \
--extra-index-url https://pypi.nvidia.com/ ai-dynamo
# Add yank script
COPY --chown=root:root --chmod=755 docker/configs/yank /usr/local/bin/yank
@@ -534,7 +542,8 @@ COPY docker/configs/.zshrc /root/.zshrc
# libsqlite3-0: CVE-2025-{6965,7709}
# libtasn1-6: CVE-2025-13151
# dpkg: CVE-2025-6297
RUN python3 -m pip install --upgrade "urllib3>=2.6.3" "pillow>=12.1.1"
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --upgrade "urllib3>=2.6.3" "pillow>=12.1.1"
RUN --mount=type=cache,target=/var/cache/apt,id=framework-apt \
apt-get update && apt-get install -y --only-upgrade \
binutils binutils-common binutils-x86-64-linux-gnu libbinutils \
@@ -577,9 +586,9 @@ RUN if [ "$BRANCH_TYPE" = "local" ]; then \
# Editable install (fast - dependencies already installed via constraints)
# Clean up __pycache__/tests/pyc in same RUN to avoid writing ~28k files to layer
RUN --mount=type=cache,target=/root/.cache/pip \
RUN --mount=type=cache,target=/root/.cache/uv \
cd /sgl-workspace/sglang \
&& python3 -m pip install --no-deps -e "python[${BUILD_TYPE}]" \
&& uv pip install --no-deps -e "python[${BUILD_TYPE}]" \
&& kernels lock python \
&& ( success=0; \
# aarch64: kernels-community/sgl-flash-attn3 ships no arm variants; JIT-compile at runtime.
@@ -613,8 +622,8 @@ ENV SGLANG_RUST_BUILD_MODE=never
# Install pre-built gateway artifacts from parallel builder
COPY --from=gateway_builder /build/sgl-model-gateway-bin /opt/sglang/bin/sgl-model-gateway
COPY --from=gateway_builder /build/gateway_wheels /tmp/gateway_wheels
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --force-reinstall /tmp/gateway_wheels/*.whl \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --reinstall /tmp/gateway_wheels/*.whl \
&& rm -rf /tmp/gateway_wheels
# Set workspace directory
@@ -658,6 +667,7 @@ FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu24.04 AS runtime
ARG CUDA_VERSION
ARG TARGETARCH
ARG GDRCOPY_VERSION=2.5.1
ARG PYTHON_VERSION
ENV DEBIAN_FRONTEND=noninteractive \
CUDA_HOME=/usr/local/cuda \
@@ -669,12 +679,8 @@ ENV PATH="${PATH}:/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/cuda/nvvm
LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
# Install runtime dependencies (devel base provides gcc/g++/build tools)
# Python 3.12 ships in Ubuntu 24.04 main, so no deadsnakes PPA needed.
RUN --mount=type=cache,target=/var/cache/apt,id=runtime-apt \
apt-get update && apt-get install -y --no-install-recommends --allow-change-held-packages \
# Python runtime
python3.12-full \
python3.12-dev \
wget \
# Core system utilities
ca-certificates \
@@ -715,14 +721,23 @@ RUN --mount=type=cache,target=/var/cache/apt,id=runtime-apt \
# GPG key verification
gnupg2 \
linux-libc-dev \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 2 \
&& update-alternatives --set python3 /usr/bin/python3.12 \
&& ln -sf /usr/bin/python3.12 /usr/bin/python \
&& for py in python python3; do \
rm -f /usr/bin/$py \
&& printf '#!/bin/sh\nexec /opt/sglang/bin/python "$@"\n' > /usr/bin/$py \
&& chmod +x /usr/bin/$py; \
done \
&& rm -rf /var/lib/apt/lists/* \
&& apt-get clean
# create virtual env for sglang to avoid conflict with system python packages
RUN python3 -m venv /opt/sglang
COPY --from=uv /uv /uvx /usr/local/bin/
COPY --from=base /opt/python /opt/python
ENV UV_PYTHON_INSTALL_DIR=/opt/python \
UV_PYTHON_PREFERENCE=only-managed \
UV_LINK_MODE=copy \
VIRTUAL_ENV=/opt/sglang
RUN UV_PYTHON_DOWNLOADS=never uv venv --python ${PYTHON_VERSION} /opt/sglang \
&& /opt/sglang/bin/python -VV
ENV PATH="/opt/sglang/bin:${PATH}"
# Set up locale
@@ -747,6 +762,7 @@ RUN --mount=type=cache,target=/var/cache/apt,id=runtime-apt \
# Copy Python site-packages from framework (already cleaned of __pycache__/tests/pyc files)
COPY --from=framework_final /opt/sglang/lib/python3.12/site-packages /opt/sglang/lib/python3.12/site-packages
COPY --from=framework_final /opt/sglang/bin/pip* /opt/sglang/bin/
# Copy SGLang workspace
COPY --from=framework_final /sgl-workspace /sgl-workspace