Files
pytorch/Dockerfile
SandishKumarHN 9fedf41b60 Dockerfile should set the syntax directive to v1 (#125632)
Fixes #125526 [#1811](https://github.com/pytorch/builder/issues/1811)

Adopt syntax=docker/dockerfile:1 whcih has been stable since 2018, while still best practice to declare in 2024.
- Syntax features dependent upon the [syntax directive version are documented here](https://hub.docker.com/r/docker/dockerfile).
- While you can set a fixed minor version, [Docker officially advises to only pin the major version]

```
(https://docs.docker.com/build/dockerfile/frontend/#stable-channel):
We recommend using docker/dockerfile:1, which always points to the latest stable release of the version 1 syntax, and receives both "minor" and "patch" updates for the version 1 release cycle.
BuildKit automatically checks for updates of the syntax when performing a build, making sure you are using the most current version.
```

**Support for building with Docker prior to v23 (released on Feb 2023)**
NOTE: 18.06 may not be the accurate minimum version for using docker/dockerfile:1, according to the [DockerHub tag history](https://hub.docker.com/layers/docker/dockerfile/1.0/images/sha256-92f5351b2fca8f7e2f452aa9aec1c34213cdd2702ca92414eee6466fab21814a?context=explore) 1.0 of the syntax seems to be from Dec 2018, which is probably why docker/dockerfile:experimental was paired with it in this file.

Personally, I'd favor only supporting builds with Docker v23. This is only relevant for someone building this Dockerfile locally, the user could still extend the already built and published image from a registry on older versions of Docker without any concern for this directive which only applies to building this Dockerfile, not images that extend it.

However if you're reluctant, you may want to refer others to [this Docker docs page](https://docs.docker.com/build/buildkit/#getting-started) where they should only need the ENV DOCKER_BUILDKIT=1, presumably the requirement for experimental was dropped with syntax=docker/dockerfile:1 with releases of Docker since Dec 2018. Affected users can often quite easily install a newer version of Docker on their OS, as per Dockers official guidance (usually via including an additional repo to the package manager).

**Reference links**
Since one of these was already included in the inline note (now a broken link), I've included relevant links mentioned above. You could alternatively rely on git blame with a commit message referencing the links or this PR for more information.

Feel free to remove any of the reference links, they're mostly only relevant to maintainers to be aware of (which this PR itself has detailed adequately above).

Pull Request resolved: https://github.com/pytorch/pytorch/pull/125632
Approved by: https://github.com/malfet
2024-05-08 01:52:56 +00:00

107 lines
4.0 KiB
Docker

# syntax=docker/dockerfile:1
# NOTE: Building this image require's docker version >= 23.0.
#
# For reference:
# - https://docs.docker.com/build/dockerfile/frontend/#stable-channel
ARG BASE_IMAGE=ubuntu:22.04
ARG PYTHON_VERSION=3.11
FROM ${BASE_IMAGE} as dev-base
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
build-essential \
ca-certificates \
ccache \
cmake \
curl \
git \
libjpeg-dev \
libpng-dev && \
rm -rf /var/lib/apt/lists/*
RUN /usr/sbin/update-ccache-symlinks
RUN mkdir /opt/ccache && ccache --set-config=cache_dir=/opt/ccache
ENV PATH /opt/conda/bin:$PATH
FROM dev-base as conda
ARG PYTHON_VERSION=3.11
# Automatically set by buildx
ARG TARGETPLATFORM
# translating Docker's TARGETPLATFORM into miniconda arches
RUN case ${TARGETPLATFORM} in \
"linux/arm64") MINICONDA_ARCH=aarch64 ;; \
*) MINICONDA_ARCH=x86_64 ;; \
esac && \
curl -fsSL -v -o ~/miniconda.sh -O "https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh"
COPY requirements.txt .
# Manually invoke bash on miniconda script per https://github.com/conda/conda/issues/10431
RUN chmod +x ~/miniconda.sh && \
bash ~/miniconda.sh -b -p /opt/conda && \
rm ~/miniconda.sh && \
/opt/conda/bin/conda install -y python=${PYTHON_VERSION} cmake conda-build pyyaml numpy ipython && \
/opt/conda/bin/python -mpip install -r requirements.txt && \
/opt/conda/bin/conda clean -ya
FROM dev-base as submodule-update
WORKDIR /opt/pytorch
COPY . .
RUN git submodule update --init --recursive
FROM conda as build
ARG CMAKE_VARS
WORKDIR /opt/pytorch
COPY --from=conda /opt/conda /opt/conda
COPY --from=submodule-update /opt/pytorch /opt/pytorch
RUN make triton
RUN --mount=type=cache,target=/opt/ccache \
export eval ${CMAKE_VARS} && \
TORCH_CUDA_ARCH_LIST="7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0 9.0a" TORCH_NVCC_FLAGS="-Xfatbin -compress-all" \
CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" \
python setup.py install
FROM conda as conda-installs
ARG PYTHON_VERSION=3.11
ARG CUDA_VERSION=12.1
ARG CUDA_CHANNEL=nvidia
ARG INSTALL_CHANNEL=pytorch-nightly
# Automatically set by buildx
# Note conda needs to be pinned to 23.5.2 see: https://github.com/pytorch/pytorch/issues/106470
RUN /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -y python=${PYTHON_VERSION} conda=23.5.2
ARG TARGETPLATFORM
# On arm64 we can only install wheel packages.
RUN case ${TARGETPLATFORM} in \
"linux/arm64") pip install --extra-index-url https://download.pytorch.org/whl/cpu/ torch torchvision torchaudio ;; \
*) /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -c "${CUDA_CHANNEL}" -y "python=${PYTHON_VERSION}" pytorch torchvision torchaudio "pytorch-cuda=$(echo $CUDA_VERSION | cut -d'.' -f 1-2)" ;; \
esac && \
/opt/conda/bin/conda clean -ya
RUN /opt/conda/bin/pip install torchelastic
FROM ${BASE_IMAGE} as official
ARG PYTORCH_VERSION
ARG TRITON_VERSION
ARG TARGETPLATFORM
ARG CUDA_VERSION
LABEL com.nvidia.volumes.needed="nvidia_driver"
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
ca-certificates \
libjpeg-dev \
libpng-dev \
&& rm -rf /var/lib/apt/lists/*
COPY --from=conda-installs /opt/conda /opt/conda
RUN if test -n "${TRITON_VERSION}" -a "${TARGETPLATFORM}" != "linux/arm64"; then \
DEBIAN_FRONTEND=noninteractive apt install -y --no-install-recommends gcc; \
rm -rf /var/lib/apt/lists/*; \
fi
ENV PATH /opt/conda/bin:$PATH
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
ENV LD_LIBRARY_PATH /usr/local/nvidia/lib:/usr/local/nvidia/lib64
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
ENV PYTORCH_VERSION ${PYTORCH_VERSION}
WORKDIR /workspace
FROM official as dev
# Should override the already installed version from the official-image stage
COPY --from=build /opt/conda /opt/conda