build(docker): switch from pip to uv for package management (#3229)

- Upgrade base image from Python 3.10 to Python 3.11 slim on Debian Bullseye.
- Replace manual PyTorch and dependencies installation with uv environment management and installation.
- Add installation of large dependencies (torch, triton, tensorflow, onnxruntime-gpu) explicitly to optimize image layer caching.
- Switch from pip to uv for package management and dependency synchronization.
- Replace pillow with pillow-simd on amd64 architecture using uv-managed pip for performance improvements.
- Adjust environment variables to use the virtual environment under /venv instead of local user installs.
- Create /venv directory with appropriate permissions and copy the virtual environment from build stage.
- Modify CMD to add --noverify flag when launching kohya_gui.py.
- Remove requirements_linux_docker.txt file, consolidating dependency management into uv and pyproject.toml.

Signed-off-by: CHEN, CHUN <[email protected]>

Author: jim60105

Dockerfile

@@ -3,81 +3,116 @@ ARG UID=1000
 ARG VERSION=EDGE
 ARG RELEASE=0
 
-FROM python:3.10-slim as build
+########################################
+# Base stage
+########################################
+FROM docker.io/library/python:3.11-slim-bullseye AS base
+
+# RUN mount cache for multi-arch: https://github.com/docker/buildx/issues/549#issuecomment-1788297892
+ARG TARGETARCH
+ARG TARGETVARIANT
+
+WORKDIR /tmp
+
+ENV NVIDIA_VISIBLE_DEVICES=all
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+
+# Install CUDA partially
+# https://docs.nvidia.com/cuda/cuda-installation-guide-linux/#debian
+# Installing the complete CUDA Toolkit system-wide usually adds around 8GB to the image size.
+# Since most CUDA packages already installed through pip, there's no need to download the entire toolkit.
+# Therefore, we opt to install only the essential libraries.
+# Here is the package list for your reference: https://developer.download.nvidia.com/compute/cuda/repos/debian11/x86_64
+
+ADD https://developer.download.nvidia.com/compute/cuda/repos/debian11/x86_64/cuda-keyring_1.1-1_all.deb /tmp/cuda-keyring_x86_64.deb
+RUN --mount=type=cache,id=apt-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/cache/apt \
+    --mount=type=cache,id=aptlists-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/lib/apt/lists \
+    dpkg -i cuda-keyring_x86_64.deb && \
+    rm -f cuda-keyring_x86_64.deb && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends \
+    # !If you experience any related issues, replace the following line with `cuda-12-4` to obtain the complete CUDA package.
+    cuda-nvcc-12-4
+
+ENV PATH="/usr/local/cuda/bin${PATH:+:${PATH}}"
+ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64
+ENV CUDA_VERSION=12.4
+ENV NVIDIA_REQUIRE_CUDA=cuda>=12.4
+ENV CUDA_HOME=/usr/local/cuda
+
+########################################
+# Build stage
+########################################
+FROM base AS build
 
 # RUN mount cache for multi-arch: https://github.com/docker/buildx/issues/549#issuecomment-1788297892
 ARG TARGETARCH
 ARG TARGETVARIANT
 
 WORKDIR /app
 
-# Install under /root/.local
-ENV PIP_USER="true"
-ARG PIP_NO_WARN_SCRIPT_LOCATION=0
-ARG PIP_ROOT_USER_ACTION="ignore"
+# Install uv
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
+
+ENV UV_PROJECT_ENVIRONMENT=/venv
+ENV VIRTUAL_ENV=/venv
+ENV UV_LINK_MODE=copy
+ENV UV_PYTHON_DOWNLOADS=0
+ENV UV_INDEX=https://download.pytorch.org/whl/cu124
 
 # Install build dependencies
 RUN --mount=type=cache,id=apt-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/cache/apt \
     --mount=type=cache,id=aptlists-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/lib/apt/lists \
     apt-get update && apt-get upgrade -y && \
     apt-get install -y --no-install-recommends python3-launchpadlib git curl
 
-# Install PyTorch
-# The versions must align and be in sync with the requirements_linux_docker.txt
-# hadolint ignore=SC2102
-RUN --mount=type=cache,id=pip-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/root/.cache/pip \
-    pip install --upgrade pip setuptools wheel ninja &&\
-    pip install -U \
-    --index-url https://download.pytorch.org/whl/cu124 \
-    --extra-index-url https://pypi.nvidia.com \
-    torch==2.5.0+cu124 \
-    torchvision==0.20.0+cu124 &&\
-    pip install -U xformers --index-url https://download.pytorch.org/whl/cu124
-
-# Install requirements
-RUN --mount=type=cache,id=pip-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/root/.cache/pip \
-    --mount=source=requirements_linux_docker.txt,target=requirements_linux_docker.txt \
-    --mount=source=requirements.txt,target=requirements.txt \
-    --mount=source=setup/docker_setup.py,target=setup.py \
-    --mount=source=sd-scripts,target=sd-scripts,rw \
-    pip install -r requirements_linux_docker.txt -r requirements.txt
+# Install big dependencies separately for layer caching
+# !Please note that the version restrictions should be the same as pyproject.toml
+# No packages listed should be removed in the next `uv sync` command
+# If this happens, please update the version restrictions or update the uv.lock file
+RUN --mount=type=cache,id=uv-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/root/.cache/uv \
+    uv venv --system-site-packages /venv && \
+    uv pip install --no-deps \
+    # torch (866.2MiB)
+    torch==2.5.1+cu124 \ 
+    # triton (199.8MiB)
+    triton==3.1.0 \
+    # tensorflow (615.0MiB)
+    tensorflow>=2.16.1 \
+    # onnxruntime-gpu (215.7MiB)
+    onnxruntime-gpu==1.19.2
+
+# Install dependencies
+RUN --mount=type=cache,id=uv-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/root/.cache/uv \
+    --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
+    --mount=type=bind,source=uv.lock,target=uv.lock \
+    --mount=type=bind,source=sd-scripts,target=sd-scripts,rw \
+    uv sync --frozen --no-dev --no-install-project --no-editable
 
 # Replace pillow with pillow-simd (Only for x86)
 ARG TARGETPLATFORM
 RUN --mount=type=cache,id=apt-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/cache/apt \
     --mount=type=cache,id=aptlists-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/lib/apt/lists \
     if [ "$TARGETPLATFORM" = "linux/amd64" ]; then \
     apt-get update && apt-get install -y --no-install-recommends zlib1g-dev libjpeg62-turbo-dev build-essential && \
-    pip uninstall -y pillow && \
-    CC="cc -mavx2" pip install -U --force-reinstall pillow-simd; \
+    uv pip uninstall pillow && \
+    CC="cc -mavx2" uv pip install pillow-simd; \
     fi
 
-FROM python:3.10-slim as final
+########################################
+# Final stage
+########################################
+FROM base AS final
 
 ARG TARGETARCH
 ARG TARGETVARIANT
 
-ENV NVIDIA_VISIBLE_DEVICES all
-ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
-
 WORKDIR /tmp
 
-
-# Install CUDA partially
-ADD https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb .
-RUN --mount=type=cache,id=apt-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/cache/apt \
-    --mount=type=cache,id=aptlists-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/lib/apt/lists \
-    dpkg -i cuda-keyring_1.1-1_all.deb && \
-    rm cuda-keyring_1.1-1_all.deb && \
-    sed -i 's/^Components: main$/& contrib/' /etc/apt/sources.list.d/debian.sources && \
-    apt-get update && \
-    apt-get install -y --no-install-recommends \
-    cuda-toolkit-12-4
-
 # Install runtime dependencies
 RUN --mount=type=cache,id=apt-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/cache/apt \
     --mount=type=cache,id=aptlists-$TARGETARCH$TARGETVARIANT,sharing=locked,target=/var/lib/apt/lists \
-    apt-get update && \
+    apt-get update && apt-get upgrade -y && \
     apt-get install -y --no-install-recommends libgl1 libglib2.0-0 libjpeg62 libtcl8.6 libtk8.6 libgoogle-perftools-dev dumb-init
 
 # Fix missing libnvinfer7
@@ -92,20 +127,23 @@ RUN groupadd -g $UID $UID && \
 # Create directories with correct permissions
 RUN install -d -m 775 -o $UID -g 0 /dataset && \
     install -d -m 775 -o $UID -g 0 /licenses && \
-    install -d -m 775 -o $UID -g 0 /app
+    install -d -m 775 -o $UID -g 0 /app && \
+    install -d -m 775 -o $UID -g 0 /venv
 
 # Copy licenses (OpenShift Policy)
 COPY --link --chmod=775 LICENSE.md /licenses/LICENSE.md
 
 # Copy dependencies and code (and support arbitrary uid for OpenShift best practice)
-COPY --link --chown=$UID:0 --chmod=775 --from=build /root/.local /home/$UID/.local
+COPY --link --chown=$UID:0 --chmod=775 --from=build /venv /venv
 COPY --link --chown=$UID:0 --chmod=775 . /app
 
-ENV PATH="/usr/local/cuda/lib:/usr/local/cuda/lib64:/home/$UID/.local/bin:$PATH"
-ENV PYTHONPATH="${PYTHONPATH}:/home/$UID/.local/lib/python3.10/site-packages" 
-ENV LD_LIBRARY_PATH="/usr/local/cuda/lib:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"
+ENV PATH="/venv/bin${PATH:+:${PATH}}"
+ENV PYTHONPATH="/venv/lib/python3.11/site-packages"
+
+ENV LD_LIBRARY_PATH="/venv/lib/python3.11/site-packages/nvidia/cudnn/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
 ENV LD_PRELOAD=libtcmalloc.so
 ENV PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
 # Rich logging
 # https://rich.readthedocs.io/en/stable/console.html#interactive-mode
 ENV FORCE_COLOR="true"
@@ -124,7 +162,7 @@ STOPSIGNAL SIGINT
 
 # Use dumb-init as PID 1 to handle signals properly
 ENTRYPOINT ["dumb-init", "--"]
-CMD ["python3", "kohya_gui.py", "--listen", "0.0.0.0", "--server_port", "7860", "--headless"]
+CMD ["python3", "kohya_gui.py", "--listen", "0.0.0.0", "--server_port", "7860", "--headless", "--noverify"]
 
 ARG VERSION
 ARG RELEASE
@@ -138,4 +176,4 @@ LABEL name="bmaltais/kohya_ss" \
     release=${RELEASE} \
     io.k8s.display-name="kohya_ss" \
     summary="Kohya's GUI: This repository provides a Gradio GUI for Kohya's Stable Diffusion trainers(https://github.com/kohya-ss/sd-scripts)." \
-    description="The GUI allows you to set the training parameters and generate and run the required CLI commands to train the model. This is the docker image for Kohya's GUI. For more information about this tool, please visit the following website: https://github.com/bmaltais/kohya_ss."
+    description="The GUI allows you to set the training parameters and generate and run the required CLI commands to train the model. This is the docker image for Kohya's GUI. For more information about this tool, please visit the following website: https://github.com/bmaltais/kohya_ss."